Refactor node cache code to more easily add other types of cache.

* add dependency aware cache that removed a cached node as soon as all of its decendents have executed. This allows users with lower RAM to run workflows they would otherwise not be able to run. The downside is that every workflow will fully run each time even if no nodes have changed.

* remove test code

* tidy code

CODEOWNERS

@@ -19,5 +19,6 @@
 /app/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata
 /utils/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata
 
-# Extra nodes
-/comfy_extras/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink
+# Node developers
+/comfy_extras/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink @webfiltered
+/comfy/comfy_types/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink @webfiltered

README.md

@@ -69,6 +69,8 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
    - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/)
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
+- 3D Models
+   - [Hunyuan3D 2.0](https://docs.comfy.org/tutorials/3d/hunyuan3D-2)
 - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
 - Asynchronous Queue system
 - Many optimizations: Only re-executes the parts of the workflow that changes between executions.

app/app_settings.py

@@ -9,8 +9,14 @@ class AppSettings():
         self.user_manager = user_manager
 
     def get_settings(self, request):
-        file = self.user_manager.get_request_user_filepath(
-            request, "comfy.settings.json")
+        try:
+            file = self.user_manager.get_request_user_filepath(
+                request,
+                "comfy.settings.json"
+            )
+        except KeyError as e:
+            logging.error("User settings not found.")
+            raise web.HTTPUnauthorized() from e
         if os.path.isfile(file):
             try:
                 with open(file) as f:

app/frontend_management.py

@@ -11,33 +11,62 @@ from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
 from typing import TypedDict, Optional
+from importlib.metadata import version
 
 import requests
 from typing_extensions import NotRequired
 
 from comfy.cli_args import DEFAULT_VERSION_STRING
+import app.logger
+
+# The path to the requirements.txt file
+req_path = Path(__file__).parents[1] / "requirements.txt"
 
 
 def frontend_install_warning_message():
-    req_path = os.path.abspath(os.path.join(os.path.dirname(__file__), '..', 'requirements.txt'))
+    """The warning message to display when the frontend version is not up to date."""
+
     extra = ""
     if sys.flags.no_user_site:
         extra = "-s "
-    return f"Please install the updated requirements.txt file by running:\n{sys.executable} {extra}-m pip install -r {req_path}\n\nThis error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.\n\nIf you are on the portable package you can run: update\\update_comfyui.bat to solve this problem"
+    return f"""
+Please install the updated requirements.txt file by running:
+{sys.executable} {extra}-m pip install -r {req_path}
 
-try:
-    import comfyui_frontend_package
-except ImportError:
-    # TODO: Remove the check after roll out of 0.3.16
-    logging.error(f"\n\n********** ERROR ***********\n\ncomfyui-frontend-package is not installed. {frontend_install_warning_message()}\n********** ERROR **********\n")
-    exit(-1)
+This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.
+
+If you are on the portable package you can run: update\\update_comfyui.bat to solve this problem
+""".strip()
 
 
-try:
-    frontend_version = tuple(map(int, comfyui_frontend_package.__version__.split(".")))
-except:
-    frontend_version = (0,)
-    pass
+def check_frontend_version():
+    """Check if the frontend version is up to date."""
+
+    def parse_version(version: str) -> tuple[int, int, int]:
+        return tuple(map(int, version.split(".")))
+
+    try:
+        frontend_version_str = version("comfyui-frontend-package")
+        frontend_version = parse_version(frontend_version_str)
+        with open(req_path, "r", encoding="utf-8") as f:
+            required_frontend = parse_version(f.readline().split("=")[-1])
+        if frontend_version < required_frontend:
+            app.logger.log_startup_warning(
+                f"""
+________________________________________________________________________
+WARNING WARNING WARNING WARNING WARNING
+
+Installed frontend version {".".join(map(str, frontend_version))} is lower than the recommended version {".".join(map(str, required_frontend))}.
+
+{frontend_install_warning_message()}
+________________________________________________________________________
+""".strip()
+            )
+        else:
+            logging.info("ComfyUI frontend version: {}".format(frontend_version_str))
+    except Exception as e:
+        logging.error(f"Failed to check frontend version: {e}")
+
 
 REQUEST_TIMEOUT = 10  # seconds
 
@@ -133,9 +162,28 @@ def download_release_asset_zip(release: Release, destination_path: str) -> None:
 
 
 class FrontendManager:
-    DEFAULT_FRONTEND_PATH = str(importlib.resources.files(comfyui_frontend_package) / "static")
     CUSTOM_FRONTENDS_ROOT = str(Path(__file__).parents[1] / "web_custom_versions")
 
+    @classmethod
+    def default_frontend_path(cls) -> str:
+        try:
+            import comfyui_frontend_package
+
+            return str(importlib.resources.files(comfyui_frontend_package) / "static")
+        except ImportError:
+            logging.error(
+                f"""
+********** ERROR ***********
+
+comfyui-frontend-package is not installed.
+
+{frontend_install_warning_message()}
+
+********** ERROR ***********
+""".strip()
+            )
+            sys.exit(-1)
+
     @classmethod
     def parse_version_string(cls, value: str) -> tuple[str, str, str]:
         """
@@ -156,7 +204,9 @@ class FrontendManager:
         return match_result.group(1), match_result.group(2), match_result.group(3)
 
     @classmethod
-    def init_frontend_unsafe(cls, version_string: str, provider: Optional[FrontEndProvider] = None) -> str:
+    def init_frontend_unsafe(
+        cls, version_string: str, provider: Optional[FrontEndProvider] = None
+    ) -> str:
         """
         Initializes the frontend for the specified version.
 
@@ -172,17 +222,26 @@ class FrontendManager:
             main error source might be request timeout or invalid URL.
         """
         if version_string == DEFAULT_VERSION_STRING:
-            return cls.DEFAULT_FRONTEND_PATH
+            check_frontend_version()
+            return cls.default_frontend_path()
 
         repo_owner, repo_name, version = cls.parse_version_string(version_string)
 
         if version.startswith("v"):
-            expected_path = str(Path(cls.CUSTOM_FRONTENDS_ROOT) / f"{repo_owner}_{repo_name}" / version.lstrip("v"))
+            expected_path = str(
+                Path(cls.CUSTOM_FRONTENDS_ROOT)
+                / f"{repo_owner}_{repo_name}"
+                / version.lstrip("v")
+            )
             if os.path.exists(expected_path):
-                logging.info(f"Using existing copy of specific frontend version tag: {repo_owner}/{repo_name}@{version}")
+                logging.info(
+                    f"Using existing copy of specific frontend version tag: {repo_owner}/{repo_name}@{version}"
+                )
                 return expected_path
 
-        logging.info(f"Initializing frontend: {repo_owner}/{repo_name}@{version}, requesting version details from GitHub...")
+        logging.info(
+            f"Initializing frontend: {repo_owner}/{repo_name}@{version}, requesting version details from GitHub..."
+        )
 
         provider = provider or FrontEndProvider(repo_owner, repo_name)
         release = provider.get_release(version)
@@ -225,4 +284,5 @@ class FrontendManager:
         except Exception as e:
             logging.error("Failed to initialize frontend: %s", e)
             logging.info("Falling back to the default frontend.")
-            return cls.DEFAULT_FRONTEND_PATH
+            check_frontend_version()
+            return cls.default_frontend_path()

app/logger.py

@@ -82,3 +82,17 @@ def setup_logger(log_level: str = 'INFO', capacity: int = 300, use_stdout: bool
         logger.addHandler(stdout_handler)
 
     logger.addHandler(stream_handler)
+
+
+STARTUP_WARNINGS = []
+
+
+def log_startup_warning(msg):
+    logging.warning(msg)
+    STARTUP_WARNINGS.append(msg)
+
+
+def print_startup_warnings():
+    for s in STARTUP_WARNINGS:
+        logging.warning(s)
+    STARTUP_WARNINGS.clear()

comfy/cli_args.py

@@ -79,6 +79,7 @@ fpte_group.add_argument("--fp8_e4m3fn-text-enc", action="store_true", help="Stor
 fpte_group.add_argument("--fp8_e5m2-text-enc", action="store_true", help="Store text encoder weights in fp8 (e5m2 variant).")
 fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text encoder weights in fp16.")
 fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
+fpte_group.add_argument("--bf16-text-enc", action="store_true", help="Store text encoder weights in bf16.")
 
 parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
 
@@ -100,12 +101,14 @@ parser.add_argument("--preview-size", type=int, default=512, help="Sets the maxi
 cache_group = parser.add_mutually_exclusive_group()
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
+cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
 attn_group.add_argument("--use-quad-cross-attention", action="store_true", help="Use the sub-quadratic cross attention optimization . Ignored when xformers is used.")
 attn_group.add_argument("--use-pytorch-cross-attention", action="store_true", help="Use the new pytorch 2.0 cross attention function.")
 attn_group.add_argument("--use-sage-attention", action="store_true", help="Use sage attention.")
+attn_group.add_argument("--use-flash-attention", action="store_true", help="Use FlashAttention.")
 
 parser.add_argument("--disable-xformers", action="store_true", help="Disable xformers.")
 

comfy/clip_vision.py

@@ -9,6 +9,7 @@ import comfy.model_patcher
 import comfy.model_management
 import comfy.utils
 import comfy.clip_model
+import comfy.image_encoders.dino2
 
 class Output:
     def __getitem__(self, key):
@@ -34,6 +35,12 @@ def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], s
     image = torch.clip((255. * image), 0, 255).round() / 255.0
     return (image - mean.view([3,1,1])) / std.view([3,1,1])
 
+IMAGE_ENCODERS = {
+    "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "dinov2": comfy.image_encoders.dino2.Dinov2Model,
+}
+
 class ClipVisionModel():
     def __init__(self, json_config):
         with open(json_config) as f:
@@ -42,10 +49,11 @@ class ClipVisionModel():
         self.image_size = config.get("image_size", 224)
         self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073])
         self.image_std = config.get("image_std", [0.26862954, 0.26130258, 0.27577711])
+        model_class = IMAGE_ENCODERS.get(config.get("model_type", "clip_vision_model"))
         self.load_device = comfy.model_management.text_encoder_device()
         offload_device = comfy.model_management.text_encoder_offload_device()
         self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
-        self.model = comfy.clip_model.CLIPVisionModelProjection(config, self.dtype, offload_device, comfy.ops.manual_cast)
+        self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
         self.model.eval()
 
         self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
@@ -102,15 +110,21 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
     elif "vision_model.encoder.layers.30.layer_norm1.weight" in sd:
         json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_h.json")
     elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
+        embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0]
         if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152:
-            json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
-        elif sd["vision_model.embeddings.position_embedding.weight"].shape[0] == 577:
+            if embed_shape == 729:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
+            elif embed_shape == 1024:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
+        elif embed_shape == 577:
             if "multi_modal_projector.linear_1.bias" in sd:
                 json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json")
             else:
                 json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336.json")
         else:
             json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl.json")
+    elif "embeddings.patch_embeddings.projection.weight" in sd:
+        json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino2_giant.json")
     else:
         return None
 

comfy/clip_vision_siglip_512.json

@@ -0,0 +1,13 @@
+{
+  "num_channels": 3,
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "image_size": 512,
+  "intermediate_size": 4304,
+  "model_type": "siglip_vision_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 27,
+  "patch_size": 16,
+  "image_mean": [0.5, 0.5, 0.5],
+  "image_std": [0.5, 0.5, 0.5]
+}

comfy/comfy_types/node_typing.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 from typing import Literal, TypedDict
+from typing_extensions import NotRequired
 from abc import ABC, abstractmethod
 from enum import Enum
 
@@ -26,6 +27,7 @@ class IO(StrEnum):
     BOOLEAN = "BOOLEAN"
     INT = "INT"
     FLOAT = "FLOAT"
+    COMBO = "COMBO"
     CONDITIONING = "CONDITIONING"
     SAMPLER = "SAMPLER"
     SIGMAS = "SIGMAS"
@@ -66,6 +68,7 @@ class IO(StrEnum):
         b = frozenset(value.split(","))
         return not (b.issubset(a) or a.issubset(b))
 
+
 class RemoteInputOptions(TypedDict):
     route: str
     """The route to the remote source."""
@@ -80,6 +83,14 @@ class RemoteInputOptions(TypedDict):
     refresh: int
     """The TTL of the remote input's value in milliseconds. Specifies the interval at which the remote input's value is refreshed."""
 
+
+class MultiSelectOptions(TypedDict):
+    placeholder: NotRequired[str]
+    """The placeholder text to display in the multi-select widget when no items are selected."""
+    chip: NotRequired[bool]
+    """Specifies whether to use chips instead of comma separated values for the multi-select widget."""
+
+
 class InputTypeOptions(TypedDict):
     """Provides type hinting for the return type of the INPUT_TYPES node function.
 
@@ -91,9 +102,13 @@ class InputTypeOptions(TypedDict):
     default: bool | str | float | int | list | tuple
     """The default value of the widget"""
     defaultInput: bool
-    """Defaults to an input slot rather than a widget"""
+    """@deprecated in v1.16 frontend. v1.16 frontend allows input socket and widget to co-exist.
+    - defaultInput on required inputs should be dropped.
+    - defaultInput on optional inputs should be replaced with forceInput.
+    Ref: https://github.com/Comfy-Org/ComfyUI_frontend/pull/3364
+    """
     forceInput: bool
-    """`defaultInput` and also don't allow converting to a widget"""
+    """Forces the input to be an input slot rather than a widget even a widget is available for the input type."""
     lazy: bool
     """Declares that this input uses lazy evaluation"""
     rawLink: bool
@@ -133,9 +148,22 @@ class InputTypeOptions(TypedDict):
     """Specifies which folder to get preview images from if the input has the ``image_upload`` flag.
     """
     remote: RemoteInputOptions
-    """Specifies the configuration for a remote input."""
+    """Specifies the configuration for a remote input.
+    Available after ComfyUI frontend v1.9.7
+    https://github.com/Comfy-Org/ComfyUI_frontend/pull/2422"""
     control_after_generate: bool
     """Specifies whether a control widget should be added to the input, adding options to automatically change the value after each prompt is queued. Currently only used for INT and COMBO types."""
+    options: NotRequired[list[str | int | float]]
+    """COMBO type only. Specifies the selectable options for the combo widget.
+    Prefer:
+    ["COMBO", {"options": ["Option 1", "Option 2", "Option 3"]}]
+    Over:
+    [["Option 1", "Option 2", "Option 3"]]
+    """
+    multi_select: NotRequired[MultiSelectOptions]
+    """COMBO type only. Specifies the configuration for a multi-select widget.
+    Available after ComfyUI frontend v1.13.4
+    https://github.com/Comfy-Org/ComfyUI_frontend/pull/2987"""
 
 
 class HiddenInputTypeDict(TypedDict):

comfy/image_encoders/dino2.py

@@ -0,0 +1,141 @@
+import torch
+from comfy.text_encoders.bert import BertAttention
+import comfy.model_management
+from comfy.ldm.modules.attention import optimized_attention_for_device
+
+
+class Dino2AttentionOutput(torch.nn.Module):
+    def __init__(self, input_dim, output_dim, layer_norm_eps, dtype, device, operations):
+        super().__init__()
+        self.dense = operations.Linear(input_dim, output_dim, dtype=dtype, device=device)
+
+    def forward(self, x):
+        return self.dense(x)
+
+
+class Dino2AttentionBlock(torch.nn.Module):
+    def __init__(self, embed_dim, heads, layer_norm_eps, dtype, device, operations):
+        super().__init__()
+        self.attention = BertAttention(embed_dim, heads, dtype, device, operations)
+        self.output = Dino2AttentionOutput(embed_dim, embed_dim, layer_norm_eps, dtype, device, operations)
+
+    def forward(self, x, mask, optimized_attention):
+        return self.output(self.attention(x, mask, optimized_attention))
+
+
+class LayerScale(torch.nn.Module):
+    def __init__(self, dim, dtype, device, operations):
+        super().__init__()
+        self.lambda1 = torch.nn.Parameter(torch.empty(dim, device=device, dtype=dtype))
+
+    def forward(self, x):
+        return x * comfy.model_management.cast_to_device(self.lambda1, x.device, x.dtype)
+
+
+class SwiGLUFFN(torch.nn.Module):
+    def __init__(self, dim, dtype, device, operations):
+        super().__init__()
+        in_features = out_features = dim
+        hidden_features = int(dim * 4)
+        hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
+
+        self.weights_in = operations.Linear(in_features, 2 * hidden_features, bias=True, device=device, dtype=dtype)
+        self.weights_out = operations.Linear(hidden_features, out_features, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.weights_in(x)
+        x1, x2 = x.chunk(2, dim=-1)
+        x = torch.nn.functional.silu(x1) * x2
+        return self.weights_out(x)
+
+
+class Dino2Block(torch.nn.Module):
+    def __init__(self, dim, num_heads, layer_norm_eps, dtype, device, operations):
+        super().__init__()
+        self.attention = Dino2AttentionBlock(dim, num_heads, layer_norm_eps, dtype, device, operations)
+        self.layer_scale1 = LayerScale(dim, dtype, device, operations)
+        self.layer_scale2 = LayerScale(dim, dtype, device, operations)
+        self.mlp = SwiGLUFFN(dim, dtype, device, operations)
+        self.norm1 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
+        self.norm2 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
+
+    def forward(self, x, optimized_attention):
+        x = x + self.layer_scale1(self.attention(self.norm1(x), None, optimized_attention))
+        x = x + self.layer_scale2(self.mlp(self.norm2(x)))
+        return x
+
+
+class Dino2Encoder(torch.nn.Module):
+    def __init__(self, dim, num_heads, layer_norm_eps, num_layers, dtype, device, operations):
+        super().__init__()
+        self.layer = torch.nn.ModuleList([Dino2Block(dim, num_heads, layer_norm_eps, dtype, device, operations) for _ in range(num_layers)])
+
+    def forward(self, x, intermediate_output=None):
+        optimized_attention = optimized_attention_for_device(x.device, False, small_input=True)
+
+        if intermediate_output is not None:
+            if intermediate_output < 0:
+                intermediate_output = len(self.layer) + intermediate_output
+
+        intermediate = None
+        for i, l in enumerate(self.layer):
+            x = l(x, optimized_attention)
+            if i == intermediate_output:
+                intermediate = x.clone()
+        return x, intermediate
+
+
+class Dino2PatchEmbeddings(torch.nn.Module):
+    def __init__(self, dim, num_channels=3, patch_size=14, image_size=518, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.projection = operations.Conv2d(
+            in_channels=num_channels,
+            out_channels=dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=True,
+            dtype=dtype,
+            device=device
+        )
+
+    def forward(self, pixel_values):
+        return self.projection(pixel_values).flatten(2).transpose(1, 2)
+
+
+class Dino2Embeddings(torch.nn.Module):
+    def __init__(self, dim, dtype, device, operations):
+        super().__init__()
+        patch_size = 14
+        image_size = 518
+
+        self.patch_embeddings = Dino2PatchEmbeddings(dim, patch_size=patch_size, image_size=image_size, dtype=dtype, device=device, operations=operations)
+        self.position_embeddings = torch.nn.Parameter(torch.empty(1, (image_size // patch_size) ** 2 + 1, dim, dtype=dtype, device=device))
+        self.cls_token = torch.nn.Parameter(torch.empty(1, 1, dim, dtype=dtype, device=device))
+        self.mask_token = torch.nn.Parameter(torch.empty(1, dim, dtype=dtype, device=device))
+
+    def forward(self, pixel_values):
+        x = self.patch_embeddings(pixel_values)
+        # TODO: mask_token?
+        x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
+        x = x + comfy.model_management.cast_to_device(self.position_embeddings, x.device, x.dtype)
+        return x
+
+
+class Dinov2Model(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        num_layers = config_dict["num_hidden_layers"]
+        dim = config_dict["hidden_size"]
+        heads = config_dict["num_attention_heads"]
+        layer_norm_eps = config_dict["layer_norm_eps"]
+
+        self.embeddings = Dino2Embeddings(dim, dtype, device, operations)
+        self.encoder = Dino2Encoder(dim, heads, layer_norm_eps, num_layers, dtype, device, operations)
+        self.layernorm = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
+
+    def forward(self, pixel_values, attention_mask=None, intermediate_output=None):
+        x = self.embeddings(pixel_values)
+        x, i = self.encoder(x, intermediate_output=intermediate_output)
+        x = self.layernorm(x)
+        pooled_output = x[:, 0, :]
+        return x, i, pooled_output, None

comfy/image_encoders/dino2_giant.json

@@ -0,0 +1,21 @@
+{
+  "attention_probs_dropout_prob": 0.0,
+  "drop_path_rate": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.0,
+  "hidden_size": 1536,
+  "image_size": 518,
+  "initializer_range": 0.02,
+  "layer_norm_eps": 1e-06,
+  "layerscale_value": 1.0,
+  "mlp_ratio": 4,
+  "model_type": "dinov2",
+  "num_attention_heads": 24,
+  "num_channels": 3,
+  "num_hidden_layers": 40,
+  "patch_size": 14,
+  "qkv_bias": true,
+  "use_swiglu_ffn": true,
+  "image_mean": [0.485, 0.456, 0.406],
+  "image_std": [0.229, 0.224, 0.225]
+}

comfy/k_diffusion/sampling.py

@@ -688,10 +688,10 @@ def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=N
     if len(sigmas) <= 1:
         return x
 
+    extra_args = {} if extra_args is None else extra_args
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     seed = extra_args.get("seed", None)
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
-    extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
     sigma_fn = lambda t: t.neg().exp()
     t_fn = lambda sigma: sigma.log().neg()
@@ -762,10 +762,10 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     if solver_type not in {'heun', 'midpoint'}:
         raise ValueError('solver_type must be \'heun\' or \'midpoint\'')
 
+    extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
-    extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
 
     old_denoised = None
@@ -808,10 +808,10 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     if len(sigmas) <= 1:
         return x
 
+    extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
-    extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
 
     denoised_1, denoised_2 = None, None
@@ -858,7 +858,7 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
 def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     if len(sigmas) <= 1:
         return x
-
+    extra_args = {} if extra_args is None else extra_args
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
@@ -867,7 +867,7 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
 def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
     if len(sigmas) <= 1:
         return x
-
+    extra_args = {} if extra_args is None else extra_args
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
@@ -876,7 +876,7 @@ def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
 def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     if len(sigmas) <= 1:
         return x
-
+    extra_args = {} if extra_args is None else extra_args
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=r)
@@ -1366,3 +1366,59 @@ def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None,
             x = x + d_bar * dt
         old_d = d
     return x
+
+@torch.no_grad()
+def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, noise_scaler=None, max_stage=3):
+    """
+    Extended Reverse-Time SDE solver (VE ER-SDE-Solver-3). Arxiv: https://arxiv.org/abs/2309.06169.
+    Code reference: https://github.com/QinpengCui/ER-SDE-Solver/blob/main/er_sde_solver.py.
+    """
+    extra_args = {} if extra_args is None else extra_args
+    seed = extra_args.get("seed", None)
+    noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+
+    def default_noise_scaler(sigma):
+        return sigma * ((sigma ** 0.3).exp() + 10.0)
+    noise_scaler = default_noise_scaler if noise_scaler is None else noise_scaler
+    num_integration_points = 200.0
+    point_indice = torch.arange(0, num_integration_points, dtype=torch.float32, device=x.device)
+
+    old_denoised = None
+    old_denoised_d = None
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        stage_used = min(max_stage, i + 1)
+        if sigmas[i + 1] == 0:
+            x = denoised
+        elif stage_used == 1:
+            r = noise_scaler(sigmas[i + 1]) / noise_scaler(sigmas[i])
+            x = r * x + (1 - r) * denoised
+        else:
+            r = noise_scaler(sigmas[i + 1]) / noise_scaler(sigmas[i])
+            x = r * x + (1 - r) * denoised
+
+            dt = sigmas[i + 1] - sigmas[i]
+            sigma_step_size = -dt / num_integration_points
+            sigma_pos = sigmas[i + 1] + point_indice * sigma_step_size
+            scaled_pos = noise_scaler(sigma_pos)
+
+            # Stage 2
+            s = torch.sum(1 / scaled_pos) * sigma_step_size
+            denoised_d = (denoised - old_denoised) / (sigmas[i] - sigmas[i - 1])
+            x = x + (dt + s * noise_scaler(sigmas[i + 1])) * denoised_d
+
+            if stage_used >= 3:
+                # Stage 3
+                s_u = torch.sum((sigma_pos - sigmas[i]) / scaled_pos) * sigma_step_size
+                denoised_u = (denoised_d - old_denoised_d) / ((sigmas[i] - sigmas[i - 2]) / 2)
+                x = x + ((dt ** 2) / 2 + s_u * noise_scaler(sigmas[i + 1])) * denoised_u
+            old_denoised_d = denoised_d
+
+        if s_noise != 0 and sigmas[i + 1] > 0:
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * (sigmas[i + 1] ** 2 - sigmas[i] ** 2 * r ** 2).sqrt().nan_to_num(nan=0.0)
+        old_denoised = denoised
+    return x

comfy/latent_formats.py

@@ -456,3 +456,13 @@ class Wan21(LatentFormat):
         latents_mean = self.latents_mean.to(latent.device, latent.dtype)
         latents_std = self.latents_std.to(latent.device, latent.dtype)
         return latent * latents_std / self.scale_factor + latents_mean
+
+class Hunyuan3Dv2(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+    scale_factor = 0.9990943042622529
+
+class Hunyuan3Dv2mini(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+    scale_factor = 1.0188137142395404

comfy/ldm/flux/math.py

@@ -10,10 +10,11 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor:
     q_shape = q.shape
     k_shape = k.shape
 
-    q = q.float().reshape(*q.shape[:-1], -1, 1, 2)
-    k = k.float().reshape(*k.shape[:-1], -1, 1, 2)
-    q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v)
-    k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
+    if pe is not None:
+        q = q.to(dtype=pe.dtype).reshape(*q.shape[:-1], -1, 1, 2)
+        k = k.to(dtype=pe.dtype).reshape(*k.shape[:-1], -1, 1, 2)
+        q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v)
+        k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
 
     heads = q.shape[1]
     x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask)
@@ -36,8 +37,8 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
 
 
 def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-    xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
-    xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
+    xq_ = xq.to(dtype=freqs_cis.dtype).reshape(*xq.shape[:-1], -1, 1, 2)
+    xk_ = xk.to(dtype=freqs_cis.dtype).reshape(*xk.shape[:-1], -1, 1, 2)
     xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
     xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
     return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)

comfy/ldm/flux/model.py

@@ -115,8 +115,11 @@ class Flux(nn.Module):
         vec = vec + self.vector_in(y[:,:self.params.vec_in_dim])
         txt = self.txt_in(txt)
 
-        ids = torch.cat((txt_ids, img_ids), dim=1)
-        pe = self.pe_embedder(ids)
+        if img_ids is not None:
+            ids = torch.cat((txt_ids, img_ids), dim=1)
+            pe = self.pe_embedder(ids)
+        else:
+            pe = None
 
         blocks_replace = patches_replace.get("dit", {})
         for i, block in enumerate(self.double_blocks):

comfy/ldm/hunyuan3d/model.py

@@ -0,0 +1,135 @@
+import torch
+from torch import nn
+from comfy.ldm.flux.layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding,
+)
+
+
+class Hunyuan3Dv2(nn.Module):
+    def __init__(
+        self,
+        in_channels=64,
+        context_in_dim=1536,
+        hidden_size=1024,
+        mlp_ratio=4.0,
+        num_heads=16,
+        depth=16,
+        depth_single_blocks=32,
+        qkv_bias=True,
+        guidance_embed=False,
+        image_model=None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.dtype = dtype
+
+        if hidden_size % num_heads != 0:
+            raise ValueError(
+                f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}"
+            )
+
+        self.max_period = 1000  # While reimplementing the model I noticed that they messed up. This 1000 value was meant to be the time_factor but they set the max_period instead
+        self.latent_in = operations.Linear(in_channels, hidden_size, bias=True, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations) if guidance_embed else None
+        )
+        self.cond_in = operations.Linear(context_in_dim, hidden_size, dtype=dtype, device=device)
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth_single_blocks)
+            ]
+        )
+        self.final_layer = LastLayer(hidden_size, 1, in_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
+        x = x.movedim(-1, -2)
+        timestep = 1.0 - timestep
+        txt = context
+        img = self.latent_in(x)
+
+        vec = self.time_in(timestep_embedding(timestep, 256, self.max_period).to(dtype=img.dtype))
+        if self.guidance_in is not None:
+            if guidance is not None:
+                vec = vec + self.guidance_in(timestep_embedding(guidance, 256, self.max_period).to(img.dtype))
+
+        txt = self.cond_in(txt)
+        pe = None
+        attn_mask = None
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"], out["txt"] = block(img=args["img"],
+                                                   txt=args["txt"],
+                                                   vec=args["vec"],
+                                                   pe=args["pe"],
+                                                   attn_mask=args.get("attn_mask"))
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img,
+                                                           "txt": txt,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask},
+                                                          {"original_block": block_wrap})
+                txt = out["txt"]
+                img = out["img"]
+            else:
+                img, txt = block(img=img,
+                                 txt=txt,
+                                 vec=vec,
+                                 pe=pe,
+                                 attn_mask=attn_mask)
+
+        img = torch.cat((txt, img), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if ("single_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"],
+                                       vec=args["vec"],
+                                       pe=args["pe"],
+                                       attn_mask=args.get("attn_mask"))
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask},
+                                                          {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+
+        img = img[:, txt.shape[1]:, ...]
+        img = self.final_layer(img, vec)
+        return img.movedim(-2, -1) * (-1.0)

comfy/ldm/hunyuan3d/vae.py

@@ -0,0 +1,587 @@
+# Original: https://github.com/Tencent/Hunyuan3D-2/blob/main/hy3dgen/shapegen/models/autoencoders/model.py
+# Since the header on their VAE source file was a bit confusing we asked for permission to use this code from tencent under the GPL license used in ComfyUI.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+from typing import Union, Tuple, List, Callable, Optional
+
+import numpy as np
+from einops import repeat, rearrange
+from tqdm import tqdm
+import logging
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def generate_dense_grid_points(
+    bbox_min: np.ndarray,
+    bbox_max: np.ndarray,
+    octree_resolution: int,
+    indexing: str = "ij",
+):
+    length = bbox_max - bbox_min
+    num_cells = octree_resolution
+
+    x = np.linspace(bbox_min[0], bbox_max[0], int(num_cells) + 1, dtype=np.float32)
+    y = np.linspace(bbox_min[1], bbox_max[1], int(num_cells) + 1, dtype=np.float32)
+    z = np.linspace(bbox_min[2], bbox_max[2], int(num_cells) + 1, dtype=np.float32)
+    [xs, ys, zs] = np.meshgrid(x, y, z, indexing=indexing)
+    xyz = np.stack((xs, ys, zs), axis=-1)
+    grid_size = [int(num_cells) + 1, int(num_cells) + 1, int(num_cells) + 1]
+
+    return xyz, grid_size, length
+
+
+class VanillaVolumeDecoder:
+    @torch.no_grad()
+    def __call__(
+        self,
+        latents: torch.FloatTensor,
+        geo_decoder: Callable,
+        bounds: Union[Tuple[float], List[float], float] = 1.01,
+        num_chunks: int = 10000,
+        octree_resolution: int = None,
+        enable_pbar: bool = True,
+        **kwargs,
+    ):
+        device = latents.device
+        dtype = latents.dtype
+        batch_size = latents.shape[0]
+
+        # 1. generate query points
+        if isinstance(bounds, float):
+            bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds]
+
+        bbox_min, bbox_max = np.array(bounds[0:3]), np.array(bounds[3:6])
+        xyz_samples, grid_size, length = generate_dense_grid_points(
+            bbox_min=bbox_min,
+            bbox_max=bbox_max,
+            octree_resolution=octree_resolution,
+            indexing="ij"
+        )
+        xyz_samples = torch.from_numpy(xyz_samples).to(device, dtype=dtype).contiguous().reshape(-1, 3)
+
+        # 2. latents to 3d volume
+        batch_logits = []
+        for start in tqdm(range(0, xyz_samples.shape[0], num_chunks), desc="Volume Decoding",
+                          disable=not enable_pbar):
+            chunk_queries = xyz_samples[start: start + num_chunks, :]
+            chunk_queries = repeat(chunk_queries, "p c -> b p c", b=batch_size)
+            logits = geo_decoder(queries=chunk_queries, latents=latents)
+            batch_logits.append(logits)
+
+        grid_logits = torch.cat(batch_logits, dim=1)
+        grid_logits = grid_logits.view((batch_size, *grid_size)).float()
+
+        return grid_logits
+
+
+class FourierEmbedder(nn.Module):
+    """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
+    each feature dimension of `x[..., i]` into:
+        [
+            sin(x[..., i]),
+            sin(f_1*x[..., i]),
+            sin(f_2*x[..., i]),
+            ...
+            sin(f_N * x[..., i]),
+            cos(x[..., i]),
+            cos(f_1*x[..., i]),
+            cos(f_2*x[..., i]),
+            ...
+            cos(f_N * x[..., i]),
+            x[..., i]     # only present if include_input is True.
+        ], here f_i is the frequency.
+
+    Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs].
+    If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...];
+    Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)].
+
+    Args:
+        num_freqs (int): the number of frequencies, default is 6;
+        logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+            otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)];
+        input_dim (int): the input dimension, default is 3;
+        include_input (bool): include the input tensor or not, default is True.
+
+    Attributes:
+        frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+                otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1);
+
+        out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1),
+            otherwise, it is input_dim * num_freqs * 2.
+
+    """
+
+    def __init__(self,
+                 num_freqs: int = 6,
+                 logspace: bool = True,
+                 input_dim: int = 3,
+                 include_input: bool = True,
+                 include_pi: bool = True) -> None:
+
+        """The initialization"""
+
+        super().__init__()
+
+        if logspace:
+            frequencies = 2.0 ** torch.arange(
+                num_freqs,
+                dtype=torch.float32
+            )
+        else:
+            frequencies = torch.linspace(
+                1.0,
+                2.0 ** (num_freqs - 1),
+                num_freqs,
+                dtype=torch.float32
+            )
+
+        if include_pi:
+            frequencies *= torch.pi
+
+        self.register_buffer("frequencies", frequencies, persistent=False)
+        self.include_input = include_input
+        self.num_freqs = num_freqs
+
+        self.out_dim = self.get_dims(input_dim)
+
+    def get_dims(self, input_dim):
+        temp = 1 if self.include_input or self.num_freqs == 0 else 0
+        out_dim = input_dim * (self.num_freqs * 2 + temp)
+
+        return out_dim
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """ Forward process.
+
+        Args:
+            x: tensor of shape [..., dim]
+
+        Returns:
+            embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)]
+                where temp is 1 if include_input is True and 0 otherwise.
+        """
+
+        if self.num_freqs > 0:
+            embed = (x[..., None].contiguous() * self.frequencies.to(device=x.device, dtype=x.dtype)).view(*x.shape[:-1], -1)
+            if self.include_input:
+                return torch.cat((x, embed.sin(), embed.cos()), dim=-1)
+            else:
+                return torch.cat((embed.sin(), embed.cos()), dim=-1)
+        else:
+            return x
+
+
+class CrossAttentionProcessor:
+    def __call__(self, attn, q, k, v):
+        out = F.scaled_dot_product_attention(q, k, v)
+        return out
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+
+    def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+        This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+        the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+        See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+        changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+        'survival rate' as the argument.
+
+        """
+        if self.drop_prob == 0. or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+    def extra_repr(self):
+        return f'drop_prob={round(self.drop_prob, 3):0.3f}'
+
+
+class MLP(nn.Module):
+    def __init__(
+        self, *,
+        width: int,
+        expand_ratio: int = 4,
+        output_width: int = None,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.c_fc = ops.Linear(width, width * expand_ratio)
+        self.c_proj = ops.Linear(width * expand_ratio, output_width if output_width is not None else width)
+        self.gelu = nn.GELU()
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        return self.drop_path(self.c_proj(self.gelu(self.c_fc(x))))
+
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+        self.attn_processor = CrossAttentionProcessor()
+
+    def forward(self, q, kv):
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+        attn_ch = width // self.heads // 2
+        q = q.view(bs, n_ctx, self.heads, -1)
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        out = self.attn_processor(self, q, k, v)
+        out = out.transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        data_width: Optional[int] = None,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        kv_cache: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = ops.Linear(width, width, bias=qkv_bias)
+        self.c_kv = ops.Linear(self.data_width, width * 2, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadCrossAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.kv_cache = kv_cache
+        self.data = None
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        if self.kv_cache:
+            if self.data is None:
+                self.data = self.c_kv(data)
+                logging.info('Save kv cache,this should be called only once for one mesh')
+            data = self.data
+        else:
+            data = self.c_kv(data)
+        x = self.attention(x, data)
+        x = self.c_proj(x)
+        return x
+
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        data_width: Optional[int] = None,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.ln_2 = norm_layer(data_width, elementwise_affine=True, eps=1e-6)
+        self.ln_3 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, expand_ratio=mlp_expand_ratio)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        out = F.scaled_dot_product_attention(q, k, v).transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.c_qkv = ops.Linear(width, width * 3, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = self.attention(x)
+        x = self.drop_path(self.c_proj(x))
+        return x
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0,
+    ):
+        super().__init__()
+        self.attn = MultiheadAttention(
+            width=width,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, drop_path_rate=drop_path_rate)
+        self.ln_2 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        layers: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    width=width,
+                    heads=heads,
+                    qkv_bias=qkv_bias,
+                    norm_layer=norm_layer,
+                    qk_norm=qk_norm,
+                    drop_path_rate=drop_path_rate
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x
+
+
+class CrossAttentionDecoder(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        out_channels: int,
+        fourier_embedder: FourierEmbedder,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        downsample_ratio: int = 1,
+        enable_ln_post: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary"
+    ):
+        super().__init__()
+
+        self.enable_ln_post = enable_ln_post
+        self.fourier_embedder = fourier_embedder
+        self.downsample_ratio = downsample_ratio
+        self.query_proj = ops.Linear(self.fourier_embedder.out_dim, width)
+        if self.downsample_ratio != 1:
+            self.latents_proj = ops.Linear(width * downsample_ratio, width)
+        if self.enable_ln_post == False:
+            qk_norm = False
+        self.cross_attn_decoder = ResidualCrossAttentionBlock(
+            width=width,
+            mlp_expand_ratio=mlp_expand_ratio,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm
+        )
+
+        if self.enable_ln_post:
+            self.ln_post = ops.LayerNorm(width)
+        self.output_proj = ops.Linear(width, out_channels)
+        self.label_type = label_type
+        self.count = 0
+
+    def forward(self, queries=None, query_embeddings=None, latents=None):
+        if query_embeddings is None:
+            query_embeddings = self.query_proj(self.fourier_embedder(queries).to(latents.dtype))
+        self.count += query_embeddings.shape[1]
+        if self.downsample_ratio != 1:
+            latents = self.latents_proj(latents)
+        x = self.cross_attn_decoder(query_embeddings, latents)
+        if self.enable_ln_post:
+            x = self.ln_post(x)
+        occ = self.output_proj(x)
+        return occ
+
+
+class ShapeVAE(nn.Module):
+    def __init__(
+        self,
+        *,
+        embed_dim: int,
+        width: int,
+        heads: int,
+        num_decoder_layers: int,
+        geo_decoder_downsample_ratio: int = 1,
+        geo_decoder_mlp_expand_ratio: int = 4,
+        geo_decoder_ln_post: bool = True,
+        num_freqs: int = 8,
+        include_pi: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary",
+        drop_path_rate: float = 0.0,
+        scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        self.geo_decoder_ln_post = geo_decoder_ln_post
+
+        self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi)
+
+        self.post_kl = ops.Linear(embed_dim, width)
+
+        self.transformer = Transformer(
+            width=width,
+            layers=num_decoder_layers,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+
+        self.geo_decoder = CrossAttentionDecoder(
+            fourier_embedder=self.fourier_embedder,
+            out_channels=1,
+            mlp_expand_ratio=geo_decoder_mlp_expand_ratio,
+            downsample_ratio=geo_decoder_downsample_ratio,
+            enable_ln_post=self.geo_decoder_ln_post,
+            width=width // geo_decoder_downsample_ratio,
+            heads=heads // geo_decoder_downsample_ratio,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            label_type=label_type,
+        )
+
+        self.volume_decoder = VanillaVolumeDecoder()
+        self.scale_factor = scale_factor
+
+    def decode(self, latents, **kwargs):
+        latents = self.post_kl(latents.movedim(-2, -1))
+        latents = self.transformer(latents)
+
+        bounds = kwargs.get("bounds", 1.01)
+        num_chunks = kwargs.get("num_chunks", 8000)
+        octree_resolution = kwargs.get("octree_resolution", 256)
+        enable_pbar = kwargs.get("enable_pbar", True)
+
+        grid_logits = self.volume_decoder(latents, self.geo_decoder, bounds=bounds, num_chunks=num_chunks, octree_resolution=octree_resolution, enable_pbar=enable_pbar)
+        return grid_logits.movedim(-2, -1)
+
+    def encode(self, x):
+        return None

comfy/ldm/modules/attention.py

@@ -24,6 +24,13 @@ if model_management.sage_attention_enabled():
         logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention")
         exit(-1)
 
+if model_management.flash_attention_enabled():
+    try:
+        from flash_attn import flash_attn_func
+    except ModuleNotFoundError:
+        logging.error(f"\n\nTo use the `--use-flash-attention` feature, the `flash-attn` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install flash-attn")
+        exit(-1)
+
 from comfy.cli_args import args
 import comfy.ops
 ops = comfy.ops.disable_weight_init
@@ -464,7 +471,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
 def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
     if skip_reshape:
         b, _, _, dim_head = q.shape
-        tensor_layout="HND"
+        tensor_layout = "HND"
     else:
         b, _, dim_head = q.shape
         dim_head //= heads
@@ -472,7 +479,7 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
             lambda t: t.view(b, -1, heads, dim_head),
             (q, k, v),
         )
-        tensor_layout="NHD"
+        tensor_layout = "NHD"
 
     if mask is not None:
         # add a batch dimension if there isn't already one
@@ -482,7 +489,17 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
         if mask.ndim == 3:
             mask = mask.unsqueeze(1)
 
-    out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+    try:
+        out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+    except Exception as e:
+        logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e))
+        if tensor_layout == "NHD":
+            q, k, v = map(
+                lambda t: t.transpose(1, 2),
+                (q, k, v),
+            )
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape)
+
     if tensor_layout == "HND":
         if not skip_output_reshape:
             out = (
@@ -496,6 +513,63 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
     return out
 
 
+try:
+    @torch.library.custom_op("flash_attention::flash_attn", mutates_args=())
+    def flash_attn_wrapper(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+                    dropout_p: float = 0.0, causal: bool = False) -> torch.Tensor:
+        return flash_attn_func(q, k, v, dropout_p=dropout_p, causal=causal)
+
+
+    @flash_attn_wrapper.register_fake
+    def flash_attn_fake(q, k, v, dropout_p=0.0, causal=False):
+        # Output shape is the same as q
+        return q.new_empty(q.shape)
+except AttributeError as error:
+    FLASH_ATTN_ERROR = error
+
+    def flash_attn_wrapper(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+                    dropout_p: float = 0.0, causal: bool = False) -> torch.Tensor:
+        assert False, f"Could not define flash_attn_wrapper: {FLASH_ATTN_ERROR}"
+
+
+def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    if skip_reshape:
+        b, _, _, dim_head = q.shape
+    else:
+        b, _, dim_head = q.shape
+        dim_head //= heads
+        q, k, v = map(
+            lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
+            (q, k, v),
+        )
+
+    if mask is not None:
+        # add a batch dimension if there isn't already one
+        if mask.ndim == 2:
+            mask = mask.unsqueeze(0)
+        # add a heads dimension if there isn't already one
+        if mask.ndim == 3:
+            mask = mask.unsqueeze(1)
+
+    try:
+        assert mask is None
+        out = flash_attn_wrapper(
+            q.transpose(1, 2),
+            k.transpose(1, 2),
+            v.transpose(1, 2),
+            dropout_p=0.0,
+            causal=False,
+        ).transpose(1, 2)
+    except Exception as e:
+        logging.warning(f"Flash Attention failed, using default SDPA: {e}")
+        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+    if not skip_output_reshape:
+        out = (
+            out.transpose(1, 2).reshape(b, -1, heads * dim_head)
+        )
+    return out
+
+
 optimized_attention = attention_basic
 
 if model_management.sage_attention_enabled():
@@ -504,6 +578,9 @@ if model_management.sage_attention_enabled():
 elif model_management.xformers_enabled():
     logging.info("Using xformers attention")
     optimized_attention = attention_xformers
+elif model_management.flash_attention_enabled():
+    logging.info("Using Flash Attention")
+    optimized_attention = attention_flash
 elif model_management.pytorch_attention_enabled():
     logging.info("Using pytorch attention")
     optimized_attention = attention_pytorch
@@ -770,6 +847,7 @@ class SpatialTransformer(nn.Module):
         if not isinstance(context, list):
             context = [context] * len(self.transformer_blocks)
         b, c, h, w = x.shape
+        transformer_options["activations_shape"] = list(x.shape)
         x_in = x
         x = self.norm(x)
         if not self.use_linear:
@@ -885,6 +963,7 @@ class SpatialVideoTransformer(SpatialTransformer):
         transformer_options={}
     ) -> torch.Tensor:
         _, _, h, w = x.shape
+        transformer_options["activations_shape"] = list(x.shape)
         x_in = x
         spatial_context = None
         if exists(context):

comfy/ldm/wan/model.py

@@ -384,6 +384,7 @@ class WanModel(torch.nn.Module):
         context,
         clip_fea=None,
         freqs=None,
+        transformer_options={},
     ):
         r"""
         Forward pass through the diffusion model
@@ -423,14 +424,18 @@ class WanModel(torch.nn.Module):
             context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
             context = torch.concat([context_clip, context], dim=1)
 
-        # arguments
-        kwargs = dict(
-            e=e0,
-            freqs=freqs,
-            context=context)
-
-        for block in self.blocks:
-            x = block(x, **kwargs)
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context)
 
         # head
         x = self.head(x, e)
@@ -439,7 +444,7 @@ class WanModel(torch.nn.Module):
         x = self.unpatchify(x, grid_sizes)
         return x
 
-    def forward(self, x, timestep, context, clip_fea=None, **kwargs):
+    def forward(self, x, timestep, context, clip_fea=None, transformer_options={},**kwargs):
         bs, c, t, h, w = x.shape
         x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
         patch_size = self.patch_size
@@ -453,7 +458,7 @@ class WanModel(torch.nn.Module):
         img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
 
         freqs = self.rope_embedder(img_ids).movedim(1, 2)
-        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs)[:, :, :t, :h, :w]
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options)[:, :, :t, :h, :w]
 
     def unpatchify(self, x, grid_sizes):
         r"""

comfy/lora_convert.py

@@ -1,4 +1,5 @@
 import torch
+import comfy.utils
 
 
 def convert_lora_bfl_control(sd): #BFL loras for Flux
@@ -11,7 +12,13 @@ def convert_lora_bfl_control(sd): #BFL loras for Flux
     return sd_out
 
 
+def convert_lora_wan_fun(sd): #Wan Fun loras
+    return comfy.utils.state_dict_prefix_replace(sd, {"lora_unet__": "lora_unet_"})
+
+
 def convert_lora(sd):
     if "img_in.lora_A.weight" in sd and "single_blocks.0.norm.key_norm.scale" in sd:
         return convert_lora_bfl_control(sd)
+    if "lora_unet__blocks_0_cross_attn_k.lora_down.weight" in sd:
+        return convert_lora_wan_fun(sd)
     return sd

comfy/model_base.py

@@ -36,6 +36,7 @@ import comfy.ldm.hunyuan_video.model
 import comfy.ldm.cosmos.model
 import comfy.ldm.lumina.model
 import comfy.ldm.wan.model
+import comfy.ldm.hunyuan3d.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -58,6 +59,7 @@ class ModelType(Enum):
     FLOW = 6
     V_PREDICTION_CONTINUOUS = 7
     FLUX = 8
+    IMG_TO_IMG = 9
 
 
 from comfy.model_sampling import EPS, V_PREDICTION, EDM, ModelSamplingDiscrete, ModelSamplingContinuousEDM, StableCascadeSampling, ModelSamplingContinuousV
@@ -88,6 +90,8 @@ def model_sampling(model_config, model_type):
     elif model_type == ModelType.FLUX:
         c = comfy.model_sampling.CONST
         s = comfy.model_sampling.ModelSamplingFlux
+    elif model_type == ModelType.IMG_TO_IMG:
+        c = comfy.model_sampling.IMG_TO_IMG
 
     class ModelSampling(s, c):
         pass
@@ -139,6 +143,7 @@ class BaseModel(torch.nn.Module):
     def _apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         sigma = t
         xc = self.model_sampling.calculate_input(sigma, x)
+
         if c_concat is not None:
             xc = torch.cat([xc] + [c_concat], dim=1)
 
@@ -600,6 +605,19 @@ class SDXL_instructpix2pix(IP2P, SDXL):
         else:
             self.process_ip2p_image_in = lambda image: image #diffusers ip2p
 
+class Lotus(BaseModel):
+    def extra_conds(self, **kwargs):
+        out = {}
+        cross_attn = kwargs.get("cross_attn", None)
+        out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+        device = kwargs["device"]
+        task_emb = torch.tensor([1, 0]).float().to(device)
+        task_emb = torch.cat([torch.sin(task_emb), torch.cos(task_emb)]).unsqueeze(0)
+        out['y'] = comfy.conds.CONDRegular(task_emb)
+        return out
+
+    def __init__(self, model_config, model_type=ModelType.IMG_TO_IMG, device=None):
+        super().__init__(model_config, model_type, device=device)
 
 class StableCascade_C(BaseModel):
     def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None):
@@ -973,29 +991,42 @@ class WAN21(BaseModel):
         self.image_to_video = image_to_video
 
     def concat_cond(self, **kwargs):
-        if not self.image_to_video:
+        noise = kwargs.get("noise", None)
+        extra_channels = self.diffusion_model.patch_embedding.weight.shape[1] - noise.shape[1]
+        if extra_channels == 0:
             return None
 
         image = kwargs.get("concat_latent_image", None)
-        noise = kwargs.get("noise", None)
         device = kwargs["device"]
 
         if image is None:
-            image = torch.zeros_like(noise)
+            shape_image = list(noise.shape)
+            shape_image[1] = extra_channels
+            image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
+        else:
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            for i in range(0, image.shape[1], 16):
+                image[:, i: i + 16] = self.process_latent_in(image[:, i: i + 16])
+            image = utils.resize_to_batch_size(image, noise.shape[0])
 
-        image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
-        image = self.process_latent_in(image)
-        image = utils.resize_to_batch_size(image, noise.shape[0])
+        if not self.image_to_video or extra_channels == image.shape[1]:
+            return image
+
+        if image.shape[1] > (extra_channels - 4):
+            image = image[:, :(extra_channels - 4)]
 
         mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
         if mask is None:
             mask = torch.zeros_like(noise)[:, :4]
         else:
-            mask = 1.0 - torch.mean(mask, dim=1, keepdim=True)
+            if mask.shape[1] != 4:
+                mask = torch.mean(mask, dim=1, keepdim=True)
+            mask = 1.0 - mask
             mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
             if mask.shape[-3] < noise.shape[-3]:
                 mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
-            mask = mask.repeat(1, 4, 1, 1, 1)
+            if mask.shape[1] == 1:
+                mask = mask.repeat(1, 4, 1, 1, 1)
             mask = utils.resize_to_batch_size(mask, noise.shape[0])
 
         return torch.cat((mask, image), dim=1)
@@ -1010,3 +1041,18 @@ class WAN21(BaseModel):
         if clip_vision_output is not None:
             out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
         return out
+
+class Hunyuan3Dv2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        guidance = kwargs.get("guidance", 5.0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+        return out

comfy/model_detection.py

@@ -154,7 +154,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["guidance_embed"] = len(guidance_keys) > 0
         return dit_config
 
-    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys: #Flux
+    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and '{}img_in.weight'.format(key_prefix) in state_dict_keys: #Flux
         dit_config = {}
         dit_config["image_model"] = "flux"
         dit_config["in_channels"] = 16
@@ -323,6 +323,21 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["model_type"] = "t2v"
         return dit_config
 
+    if '{}latent_in.weight'.format(key_prefix) in state_dict_keys:  # Hunyuan 3D
+        in_shape = state_dict['{}latent_in.weight'.format(key_prefix)].shape
+        dit_config = {}
+        dit_config["image_model"] = "hunyuan3d2"
+        dit_config["in_channels"] = in_shape[1]
+        dit_config["context_in_dim"] = state_dict['{}cond_in.weight'.format(key_prefix)].shape[1]
+        dit_config["hidden_size"] = in_shape[0]
+        dit_config["mlp_ratio"] = 4.0
+        dit_config["num_heads"] = 16
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
+        dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
+        dit_config["qkv_bias"] = True
+        dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -667,8 +682,13 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None):
             'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
             'use_temporal_attention': False, 'use_temporal_resblock': False}
 
+    LotusD = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': 4,
+            'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
+    supported_models = [LotusD, SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
 
     for unet_config in supported_models:
         matches = True

comfy/model_management.py

@@ -46,6 +46,32 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
+def get_supported_float8_types():
+    float8_types = []
+    try:
+        float8_types.append(torch.float8_e4m3fn)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e4m3fnuz)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e5m2)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e5m2fnuz)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e8m0fnu)
+    except:
+        pass
+    return float8_types
+
+FLOAT8_TYPES = get_supported_float8_types()
+
 xpu_available = False
 torch_version = ""
 try:
@@ -186,12 +212,21 @@ def get_total_memory(dev=None, torch_total_too=False):
     else:
         return mem_total
 
+def mac_version():
+    try:
+        return tuple(int(n) for n in platform.mac_ver()[0].split("."))
+    except:
+        return None
+
 total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
 total_ram = psutil.virtual_memory().total / (1024 * 1024)
 logging.info("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram))
 
 try:
     logging.info("pytorch version: {}".format(torch_version))
+    mac_ver = mac_version()
+    if mac_ver is not None:
+        logging.info("Mac Version {}".format(mac_ver))
 except:
     pass
 
@@ -692,11 +727,8 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e5m2
 
     fp8_dtype = None
-    try:
-        if weight_dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
-            fp8_dtype = weight_dtype
-    except:
-        pass
+    if weight_dtype in FLOAT8_TYPES:
+        fp8_dtype = weight_dtype
 
     if fp8_dtype is not None:
         if supports_fp8_compute(device): #if fp8 compute is supported the casting is most likely not expensive
@@ -791,6 +823,8 @@ def text_encoder_dtype(device=None):
         return torch.float8_e5m2
     elif args.fp16_text_enc:
         return torch.float16
+    elif args.bf16_text_enc:
+        return torch.bfloat16
     elif args.fp32_text_enc:
         return torch.float32
 
@@ -921,6 +955,9 @@ def cast_to_device(tensor, device, dtype, copy=False):
 def sage_attention_enabled():
     return args.use_sage_attention
 
+def flash_attention_enabled():
+    return args.use_flash_attention
+
 def xformers_enabled():
     global directml_enabled
     global cpu_state
@@ -969,12 +1006,6 @@ def pytorch_attention_flash_attention():
             return True #if you have pytorch attention enabled on AMD it probably supports at least mem efficient attention
     return False
 
-def mac_version():
-    try:
-        return tuple(int(n) for n in platform.mac_ver()[0].split("."))
-    except:
-        return None
-
 def force_upcast_attention_dtype():
     upcast = args.force_upcast_attention
 
@@ -1206,6 +1237,8 @@ def soft_empty_cache(force=False):
         torch.xpu.empty_cache()
     elif is_ascend_npu():
         torch.npu.empty_cache()
+    elif is_mlu():
+        torch.mlu.empty_cache()
     elif torch.cuda.is_available():
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()

comfy/model_patcher.py

@@ -747,6 +747,7 @@ class ModelPatcher:
 
     def partially_unload(self, device_to, memory_to_free=0):
         with self.use_ejected():
+            hooks_unpatched = False
             memory_freed = 0
             patch_counter = 0
             unload_list = self._load_list()
@@ -770,6 +771,10 @@ class ModelPatcher:
                                 move_weight = False
                                 break
 
+                            if not hooks_unpatched:
+                                self.unpatch_hooks()
+                                hooks_unpatched = True
+
                             if bk.inplace_update:
                                 comfy.utils.copy_to_param(self.model, key, bk.weight)
                             else:

comfy/model_sampling.py

@@ -69,6 +69,15 @@ class CONST:
         sigma = sigma.view(sigma.shape[:1] + (1,) * (latent.ndim - 1))
         return latent / (1.0 - sigma)
 
+class X0(EPS):
+    def calculate_denoised(self, sigma, model_output, model_input):
+        return model_output
+
+class IMG_TO_IMG(X0):
+    def calculate_input(self, sigma, noise):
+        return noise
+
+
 class ModelSamplingDiscrete(torch.nn.Module):
     def __init__(self, model_config=None, zsnr=None):
         super().__init__()

comfy/patcher_extension.py

@@ -48,6 +48,7 @@ def get_all_callbacks(call_type: str, transformer_options: dict, is_model_option
 
 class WrappersMP:
     OUTER_SAMPLE = "outer_sample"
+    PREPARE_SAMPLING = "prepare_sampling"
     SAMPLER_SAMPLE = "sampler_sample"
     CALC_COND_BATCH = "calc_cond_batch"
     APPLY_MODEL = "apply_model"

comfy/sampler_helpers.py

@@ -106,6 +106,13 @@ def cleanup_additional_models(models):
 
 
 def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
+    executor = comfy.patcher_extension.WrapperExecutor.new_executor(
+        _prepare_sampling,
+        comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True)
+    )
+    return executor.execute(model, noise_shape, conds, model_options=model_options)
+
+def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
     real_model: BaseModel = None
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)

comfy/samplers.py

@@ -710,7 +710,7 @@ KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_c
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
-                  "gradient_estimation"]
+                  "gradient_estimation", "er_sde"]
 
 class KSAMPLER(Sampler):
     def __init__(self, sampler_function, extra_options={}, inpaint_options={}):

comfy/sd.py

@@ -14,6 +14,7 @@ import comfy.ldm.genmo.vae.model
 import comfy.ldm.lightricks.vae.causal_video_autoencoder
 import comfy.ldm.cosmos.vae
 import comfy.ldm.wan.vae
+import comfy.ldm.hunyuan3d.vae
 import yaml
 import math
 
@@ -264,6 +265,7 @@ class VAE:
         self.process_input = lambda image: image * 2.0 - 1.0
         self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0)
         self.working_dtypes = [torch.bfloat16, torch.float32]
+        self.disable_offload = False
 
         self.downscale_index_formula = None
         self.upscale_index_formula = None
@@ -336,6 +338,7 @@ class VAE:
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
                 self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+                self.disable_offload = True
             elif "blocks.2.blocks.3.stack.5.weight" in sd or "decoder.blocks.2.blocks.3.stack.5.weight" in sd or "layers.4.layers.1.attn_block.attn.qkv.weight" in sd or "encoder.layers.4.layers.1.attn_block.attn.qkv.weight" in sd: #genmo mochi vae
                 if "blocks.2.blocks.3.stack.5.weight" in sd:
                     sd = comfy.utils.state_dict_prefix_replace(sd, {"": "decoder."})
@@ -412,6 +415,17 @@ class VAE:
                 self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
                 self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
+            elif "geo_decoder.cross_attn_decoder.ln_1.bias" in sd:
+                self.latent_dim = 1
+                ln_post = "geo_decoder.ln_post.weight" in sd
+                inner_size = sd["geo_decoder.output_proj.weight"].shape[1]
+                downsample_ratio = sd["post_kl.weight"].shape[0] // inner_size
+                mlp_expand = sd["geo_decoder.cross_attn_decoder.mlp.c_fc.weight"].shape[0] // inner_size
+                self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)  # TODO
+                self.memory_used_decode = lambda shape, dtype: (1024 * 1024 * 1024 * 2.0) * model_management.dtype_size(dtype)  # TODO
+                ddconfig = {"embed_dim": 64, "num_freqs": 8, "include_pi": False, "heads": 16, "width": 1024, "num_decoder_layers": 16, "qkv_bias": False, "qk_norm": True, "geo_decoder_mlp_expand_ratio": mlp_expand, "geo_decoder_downsample_ratio": downsample_ratio, "geo_decoder_ln_post": ln_post}
+                self.first_stage_model = comfy.ldm.hunyuan3d.vae.ShapeVAE(**ddconfig)
+                self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -440,6 +454,10 @@ class VAE:
         self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device)
         logging.info("VAE load device: {}, offload device: {}, dtype: {}".format(self.device, offload_device, self.vae_dtype))
 
+    def throw_exception_if_invalid(self):
+        if self.first_stage_model is None:
+            raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
+
     def vae_encode_crop_pixels(self, pixels):
         downscale_ratio = self.spacial_compression_encode()
 
@@ -494,18 +512,19 @@ class VAE:
         encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
         return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.downscale_ratio, out_channels=self.latent_channels, downscale=True, index_formulas=self.downscale_index_formula, output_device=self.output_device)
 
-    def decode(self, samples_in):
+    def decode(self, samples_in, vae_options={}):
+        self.throw_exception_if_invalid()
         pixel_samples = None
         try:
             memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
-            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+            model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
 
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
-                out = self.process_output(self.first_stage_model.decode(samples).to(self.output_device).float())
+                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).float())
                 if pixel_samples is None:
                     pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
                 pixel_samples[x:x+batch_number] = out
@@ -525,8 +544,9 @@ class VAE:
         return pixel_samples
 
     def decode_tiled(self, samples, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None):
+        self.throw_exception_if_invalid()
         memory_used = self.memory_used_decode(samples.shape, self.vae_dtype) #TODO: calculate mem required for tile
-        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
         dims = samples.ndim - 2
         args = {}
         if tile_x is not None:
@@ -553,13 +573,14 @@ class VAE:
         return output.movedim(1, -1)
 
     def encode(self, pixel_samples):
+        self.throw_exception_if_invalid()
         pixel_samples = self.vae_encode_crop_pixels(pixel_samples)
         pixel_samples = pixel_samples.movedim(-1, 1)
         if self.latent_dim == 3 and pixel_samples.ndim < 5:
             pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
         try:
             memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
-            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+            model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / max(1, memory_used))
             batch_number = max(1, batch_number)
@@ -585,6 +606,7 @@ class VAE:
         return samples
 
     def encode_tiled(self, pixel_samples, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None):
+        self.throw_exception_if_invalid()
         pixel_samples = self.vae_encode_crop_pixels(pixel_samples)
         dims = self.latent_dim
         pixel_samples = pixel_samples.movedim(-1, 1)
@@ -592,7 +614,7 @@ class VAE:
             pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
 
         memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)  # TODO: calculate mem required for tile
-        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
 
         args = {}
         if tile_x is not None:
@@ -899,7 +921,12 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
 
     model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix, metadata=metadata)
     if model_config is None:
-        return None
+        logging.warning("Warning, This is not a checkpoint file, trying to load it as a diffusion model only.")
+        diffusion_model = load_diffusion_model_state_dict(sd, model_options={})
+        if diffusion_model is None:
+            return None
+        return (diffusion_model, None, VAE(sd={}), None)  # The VAE object is there to throw an exception if it's actually used'
+
 
     unet_weight_dtype = list(model_config.supported_inference_dtypes)
     if model_config.scaled_fp8 is not None:

comfy/supported_models.py

@@ -506,6 +506,22 @@ class SDXL_instructpix2pix(SDXL):
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.SDXL_instructpix2pix(self, model_type=self.model_type(state_dict, prefix), device=device)
 
+class LotusD(SD20):
+    unet_config = {
+        "model_channels": 320,
+        "use_linear_in_transformer": True,
+        "use_temporal_attention": False,
+        "adm_in_channels": 4,
+        "in_channels": 4,
+    }
+
+    unet_extra_config = {
+        "num_classes": 'sequential'
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.Lotus(self, device=device)
+
 class SD3(supported_models_base.BASE):
     unet_config = {
         "in_channels": 16,
@@ -953,12 +969,62 @@ class WAN21_I2V(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
         "model_type": "i2v",
+        "in_dim": 36,
     }
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN21(self, image_to_video=True, device=device)
         return out
 
-models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V]
+class WAN21_FunControl2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "i2v",
+        "in_dim": 48,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21(self, image_to_video=False, device=device)
+        return out
+
+class Hunyuan3Dv2(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "hunyuan3d2",
+    }
+
+    unet_extra_config = {}
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 1.0,
+    }
+
+    memory_usage_factor = 3.5
+
+    clip_vision_prefix = "conditioner.main_image_encoder.model."
+    vae_key_prefix = ["vae."]
+
+    latent_format = latent_formats.Hunyuan3Dv2
+
+    def process_unet_state_dict_for_saving(self, state_dict):
+        replace_prefix = {"": "model."}
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Hunyuan3Dv2(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return None
+
+class Hunyuan3Dv2mini(Hunyuan3Dv2):
+    unet_config = {
+        "image_model": "hunyuan3d2",
+        "depth": 8,
+    }
+
+    latent_format = latent_formats.Hunyuan3Dv2mini
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, Hunyuan3Dv2mini, Hunyuan3Dv2]
 
 models += [SVD_img2vid]

comfy_execution/caching.py

@@ -316,3 +316,156 @@ class LRUCache(BasicCache):
             self.children[cache_key].append(self.cache_key_set.get_data_key(child_id))
         return self
 
+
+class DependencyAwareCache(BasicCache):
+    """
+    A cache implementation that tracks dependencies between nodes and manages
+    their execution and caching accordingly. It extends the BasicCache class.
+    Nodes are removed from this cache once all of their descendants have been
+    executed.
+    """
+
+    def __init__(self, key_class):
+        """
+        Initialize the DependencyAwareCache.
+
+        Args:
+            key_class: The class used for generating cache keys.
+        """
+        super().__init__(key_class)
+        self.descendants = {}  # Maps node_id -> set of descendant node_ids
+        self.ancestors = {}    # Maps node_id -> set of ancestor node_ids
+        self.executed_nodes = set()  # Tracks nodes that have been executed
+
+    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+        """
+        Clear the entire cache and rebuild the dependency graph.
+
+        Args:
+            dynprompt: The dynamic prompt object containing node information.
+            node_ids: List of node IDs to initialize the cache for.
+            is_changed_cache: Flag indicating if the cache has changed.
+        """
+        # Clear all existing cache data
+        self.cache.clear()
+        self.subcaches.clear()
+        self.descendants.clear()
+        self.ancestors.clear()
+        self.executed_nodes.clear()
+
+        # Call the parent method to initialize the cache with the new prompt
+        super().set_prompt(dynprompt, node_ids, is_changed_cache)
+
+        # Rebuild the dependency graph
+        self._build_dependency_graph(dynprompt, node_ids)
+
+    def _build_dependency_graph(self, dynprompt, node_ids):
+        """
+        Build the dependency graph for all nodes.
+
+        Args:
+            dynprompt: The dynamic prompt object containing node information.
+            node_ids: List of node IDs to build the graph for.
+        """
+        self.descendants.clear()
+        self.ancestors.clear()
+        for node_id in node_ids:
+            self.descendants[node_id] = set()
+            self.ancestors[node_id] = set()
+
+        for node_id in node_ids:
+            inputs = dynprompt.get_node(node_id)["inputs"]
+            for input_data in inputs.values():
+                if is_link(input_data):  # Check if the input is a link to another node
+                    ancestor_id = input_data[0]
+                    self.descendants[ancestor_id].add(node_id)
+                    self.ancestors[node_id].add(ancestor_id)
+
+    def set(self, node_id, value):
+        """
+        Mark a node as executed and store its value in the cache.
+
+        Args:
+            node_id: The ID of the node to store.
+            value: The value to store for the node.
+        """
+        self._set_immediate(node_id, value)
+        self.executed_nodes.add(node_id)
+        self._cleanup_ancestors(node_id)
+
+    def get(self, node_id):
+        """
+        Retrieve the cached value for a node.
+
+        Args:
+            node_id: The ID of the node to retrieve.
+
+        Returns:
+            The cached value for the node.
+        """
+        return self._get_immediate(node_id)
+
+    def ensure_subcache_for(self, node_id, children_ids):
+        """
+        Ensure a subcache exists for a node and update dependencies.
+
+        Args:
+            node_id: The ID of the parent node.
+            children_ids: List of child node IDs to associate with the parent node.
+
+        Returns:
+            The subcache object for the node.
+        """
+        subcache = super()._ensure_subcache(node_id, children_ids)
+        for child_id in children_ids:
+            self.descendants[node_id].add(child_id)
+            self.ancestors[child_id].add(node_id)
+        return subcache
+
+    def _cleanup_ancestors(self, node_id):
+        """
+        Check if ancestors of a node can be removed from the cache.
+
+        Args:
+            node_id: The ID of the node whose ancestors are to be checked.
+        """
+        for ancestor_id in self.ancestors.get(node_id, []):
+            if ancestor_id in self.executed_nodes:
+                # Remove ancestor if all its descendants have been executed
+                if all(descendant in self.executed_nodes for descendant in self.descendants[ancestor_id]):
+                    self._remove_node(ancestor_id)
+
+    def _remove_node(self, node_id):
+        """
+        Remove a node from the cache.
+
+        Args:
+            node_id: The ID of the node to remove.
+        """
+        cache_key = self.cache_key_set.get_data_key(node_id)
+        if cache_key in self.cache:
+            del self.cache[cache_key]
+        subcache_key = self.cache_key_set.get_subcache_key(node_id)
+        if subcache_key in self.subcaches:
+            del self.subcaches[subcache_key]
+
+    def clean_unused(self):
+        """
+        Clean up unused nodes. This is a no-op for this cache implementation.
+        """
+        pass
+
+    def recursive_debug_dump(self):
+        """
+        Dump the cache and dependency graph for debugging.
+
+        Returns:
+            A list containing the cache state and dependency graph.
+        """
+        result = super().recursive_debug_dump()
+        result.append({
+            "descendants": self.descendants,
+            "ancestors": self.ancestors,
+            "executed_nodes": list(self.executed_nodes),
+        })
+        return result

comfy_extras/nodes_cfg.py

@@ -0,0 +1,45 @@
+import torch
+
+# https://github.com/WeichenFan/CFG-Zero-star
+def optimized_scale(positive, negative):
+    positive_flat = positive.reshape(positive.shape[0], -1)
+    negative_flat = negative.reshape(negative.shape[0], -1)
+
+    # Calculate dot production
+    dot_product = torch.sum(positive_flat * negative_flat, dim=1, keepdim=True)
+
+    # Squared norm of uncondition
+    squared_norm = torch.sum(negative_flat ** 2, dim=1, keepdim=True) + 1e-8
+
+    # st_star = v_cond^T * v_uncond / ||v_uncond||^2
+    st_star = dot_product / squared_norm
+
+    return st_star.reshape([positive.shape[0]] + [1] * (positive.ndim - 1))
+
+class CFGZeroStar:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"model": ("MODEL",),
+                            }}
+    RETURN_TYPES = ("MODEL",)
+    RETURN_NAMES = ("patched_model",)
+    FUNCTION = "patch"
+    CATEGORY = "advanced/guidance"
+
+    def patch(self, model):
+        m = model.clone()
+        def cfg_zero_star(args):
+            guidance_scale = args['cond_scale']
+            x = args['input']
+            cond_p = args['cond_denoised']
+            uncond_p = args['uncond_denoised']
+            out = args["denoised"]
+            alpha = optimized_scale(x - cond_p, x - uncond_p)
+
+            return out + uncond_p * (alpha - 1.0)  + guidance_scale * uncond_p * (1.0 - alpha)
+        m.set_model_sampler_post_cfg_function(cfg_zero_star)
+        return (m, )
+
+NODE_CLASS_MAPPINGS = {
+    "CFGZeroStar": CFGZeroStar
+}

comfy_extras/nodes_hunyuan3d.py

@@ -0,0 +1,634 @@
+import torch
+import os
+import json
+import struct
+import numpy as np
+from comfy.ldm.modules.diffusionmodules.mmdit import get_1d_sincos_pos_embed_from_grid_torch
+import folder_paths
+import comfy.model_management
+from comfy.cli_args import args
+
+
+class EmptyLatentHunyuan3Dv2:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"resolution": ("INT", {"default": 3072, "min": 1, "max": 8192}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."}),
+                             }}
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "generate"
+
+    CATEGORY = "latent/3d"
+
+    def generate(self, resolution, batch_size):
+        latent = torch.zeros([batch_size, 64, resolution], device=comfy.model_management.intermediate_device())
+        return ({"samples": latent, "type": "hunyuan3dv2"}, )
+
+
+class Hunyuan3Dv2Conditioning:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"clip_vision_output": ("CLIP_VISION_OUTPUT",),
+                             }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING")
+    RETURN_NAMES = ("positive", "negative")
+
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, clip_vision_output):
+        embeds = clip_vision_output.last_hidden_state
+        positive = [[embeds, {}]]
+        negative = [[torch.zeros_like(embeds), {}]]
+        return (positive, negative)
+
+
+class Hunyuan3Dv2ConditioningMultiView:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {},
+                "optional": {"front": ("CLIP_VISION_OUTPUT",),
+                             "left": ("CLIP_VISION_OUTPUT",),
+                             "back": ("CLIP_VISION_OUTPUT",),
+                             "right": ("CLIP_VISION_OUTPUT",), }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING")
+    RETURN_NAMES = ("positive", "negative")
+
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, front=None, left=None, back=None, right=None):
+        all_embeds = [front, left, back, right]
+        out = []
+        pos_embeds = None
+        for i, e in enumerate(all_embeds):
+            if e is not None:
+                if pos_embeds is None:
+                    pos_embeds = get_1d_sincos_pos_embed_from_grid_torch(e.last_hidden_state.shape[-1], torch.arange(4))
+                out.append(e.last_hidden_state + pos_embeds[i].reshape(1, 1, -1))
+
+        embeds = torch.cat(out, dim=1)
+        positive = [[embeds, {}]]
+        negative = [[torch.zeros_like(embeds), {}]]
+        return (positive, negative)
+
+
+class VOXEL:
+    def __init__(self, data):
+        self.data = data
+
+
+class VAEDecodeHunyuan3D:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"samples": ("LATENT", ),
+                             "vae": ("VAE", ),
+                             "num_chunks": ("INT", {"default": 8000, "min": 1000, "max": 500000}),
+                             "octree_resolution": ("INT", {"default": 256, "min": 16, "max": 512}),
+                             }}
+    RETURN_TYPES = ("VOXEL",)
+    FUNCTION = "decode"
+
+    CATEGORY = "latent/3d"
+
+    def decode(self, vae, samples, num_chunks, octree_resolution):
+        voxels = VOXEL(vae.decode(samples["samples"], vae_options={"num_chunks": num_chunks, "octree_resolution": octree_resolution}))
+        return (voxels, )
+
+
+def voxel_to_mesh(voxels, threshold=0.5, device=None):
+    if device is None:
+        device = torch.device("cpu")
+    voxels = voxels.to(device)
+
+    binary = (voxels > threshold).float()
+    padded = torch.nn.functional.pad(binary, (1, 1, 1, 1, 1, 1), 'constant', 0)
+
+    D, H, W = binary.shape
+
+    neighbors = torch.tensor([
+        [0, 0, 1],
+        [0, 0, -1],
+        [0, 1, 0],
+        [0, -1, 0],
+        [1, 0, 0],
+        [-1, 0, 0]
+    ], device=device)
+
+    z, y, x = torch.meshgrid(
+        torch.arange(D, device=device),
+        torch.arange(H, device=device),
+        torch.arange(W, device=device),
+        indexing='ij'
+    )
+    voxel_indices = torch.stack([z.flatten(), y.flatten(), x.flatten()], dim=1)
+
+    solid_mask = binary.flatten() > 0
+    solid_indices = voxel_indices[solid_mask]
+
+    corner_offsets = [
+        torch.tensor([
+            [0, 0, 1], [0, 1, 1], [1, 1, 1], [1, 0, 1]
+        ], device=device),
+        torch.tensor([
+            [0, 0, 0], [1, 0, 0], [1, 1, 0], [0, 1, 0]
+        ], device=device),
+        torch.tensor([
+            [0, 1, 0], [1, 1, 0], [1, 1, 1], [0, 1, 1]
+        ], device=device),
+        torch.tensor([
+            [0, 0, 0], [0, 0, 1], [1, 0, 1], [1, 0, 0]
+        ], device=device),
+        torch.tensor([
+            [1, 0, 1], [1, 1, 1], [1, 1, 0], [1, 0, 0]
+        ], device=device),
+        torch.tensor([
+            [0, 1, 0], [0, 1, 1], [0, 0, 1], [0, 0, 0]
+        ], device=device)
+    ]
+
+    all_vertices = []
+    all_indices = []
+
+    vertex_count = 0
+
+    for face_idx, offset in enumerate(neighbors):
+        neighbor_indices = solid_indices + offset
+
+        padded_indices = neighbor_indices + 1
+
+        is_exposed = padded[
+            padded_indices[:, 0],
+            padded_indices[:, 1],
+            padded_indices[:, 2]
+        ] == 0
+
+        if not is_exposed.any():
+            continue
+
+        exposed_indices = solid_indices[is_exposed]
+
+        corners = corner_offsets[face_idx].unsqueeze(0)
+
+        face_vertices = exposed_indices.unsqueeze(1) + corners
+
+        all_vertices.append(face_vertices.reshape(-1, 3))
+
+        num_faces = exposed_indices.shape[0]
+        face_indices = torch.arange(
+            vertex_count,
+            vertex_count + 4 * num_faces,
+            device=device
+        ).reshape(-1, 4)
+
+        all_indices.append(torch.stack([face_indices[:, 0], face_indices[:, 1], face_indices[:, 2]], dim=1))
+        all_indices.append(torch.stack([face_indices[:, 0], face_indices[:, 2], face_indices[:, 3]], dim=1))
+
+        vertex_count += 4 * num_faces
+
+    if len(all_vertices) > 0:
+        vertices = torch.cat(all_vertices, dim=0)
+        faces = torch.cat(all_indices, dim=0)
+    else:
+        vertices = torch.zeros((1, 3))
+        faces = torch.zeros((1, 3))
+
+    v_min = 0
+    v_max = max(voxels.shape)
+
+    vertices = vertices - (v_min + v_max) / 2
+
+    scale = (v_max - v_min) / 2
+    if scale > 0:
+        vertices = vertices / scale
+
+    vertices = torch.fliplr(vertices)
+    return vertices, faces
+
+def voxel_to_mesh_surfnet(voxels, threshold=0.5, device=None):
+    if device is None:
+        device = torch.device("cpu")
+    voxels = voxels.to(device)
+
+    D, H, W = voxels.shape
+
+    padded = torch.nn.functional.pad(voxels, (1, 1, 1, 1, 1, 1), 'constant', 0)
+    z, y, x = torch.meshgrid(
+        torch.arange(D, device=device),
+        torch.arange(H, device=device),
+        torch.arange(W, device=device),
+        indexing='ij'
+    )
+    cell_positions = torch.stack([z.flatten(), y.flatten(), x.flatten()], dim=1)
+
+    corner_offsets = torch.tensor([
+        [0, 0, 0], [1, 0, 0], [0, 1, 0], [1, 1, 0],
+        [0, 0, 1], [1, 0, 1], [0, 1, 1], [1, 1, 1]
+    ], device=device)
+
+    corner_values = torch.zeros((cell_positions.shape[0], 8), device=device)
+    for c, (dz, dy, dx) in enumerate(corner_offsets):
+        corner_values[:, c] = padded[
+            cell_positions[:, 0] + dz,
+            cell_positions[:, 1] + dy,
+            cell_positions[:, 2] + dx
+        ]
+
+    corner_signs = corner_values > threshold
+    has_inside = torch.any(corner_signs, dim=1)
+    has_outside = torch.any(~corner_signs, dim=1)
+    contains_surface = has_inside & has_outside
+
+    active_cells = cell_positions[contains_surface]
+    active_signs = corner_signs[contains_surface]
+    active_values = corner_values[contains_surface]
+
+    if active_cells.shape[0] == 0:
+        return torch.zeros((0, 3), device=device), torch.zeros((0, 3), dtype=torch.long, device=device)
+
+    edges = torch.tensor([
+        [0, 1], [0, 2], [0, 4], [1, 3],
+        [1, 5], [2, 3], [2, 6], [3, 7],
+        [4, 5], [4, 6], [5, 7], [6, 7]
+    ], device=device)
+
+    cell_vertices = {}
+    progress = comfy.utils.ProgressBar(100)
+
+    for edge_idx, (e1, e2) in enumerate(edges):
+        progress.update(1)
+        crossing = active_signs[:, e1] != active_signs[:, e2]
+        if not crossing.any():
+            continue
+
+        cell_indices = torch.nonzero(crossing, as_tuple=True)[0]
+
+        v1 = active_values[cell_indices, e1]
+        v2 = active_values[cell_indices, e2]
+
+        t = torch.zeros_like(v1, device=device)
+        denom = v2 - v1
+        valid = denom != 0
+        t[valid] = (threshold - v1[valid]) / denom[valid]
+        t[~valid] = 0.5
+
+        p1 = corner_offsets[e1].float()
+        p2 = corner_offsets[e2].float()
+
+        intersection = p1.unsqueeze(0) + t.unsqueeze(1) * (p2.unsqueeze(0) - p1.unsqueeze(0))
+
+        for i, point in zip(cell_indices.tolist(), intersection):
+            if i not in cell_vertices:
+                cell_vertices[i] = []
+            cell_vertices[i].append(point)
+
+    # Calculate the final vertices as the average of intersection points for each cell
+    vertices = []
+    vertex_lookup = {}
+
+    vert_progress_mod = round(len(cell_vertices)/50)
+
+    for i, points in cell_vertices.items():
+        if not i % vert_progress_mod:
+            progress.update(1)
+
+        if points:
+            vertex = torch.stack(points).mean(dim=0)
+            vertex = vertex + active_cells[i].float()
+            vertex_lookup[tuple(active_cells[i].tolist())] = len(vertices)
+            vertices.append(vertex)
+
+    if not vertices:
+        return torch.zeros((0, 3), device=device), torch.zeros((0, 3), dtype=torch.long, device=device)
+
+    final_vertices = torch.stack(vertices)
+
+    inside_corners_mask = active_signs
+    outside_corners_mask = ~active_signs
+
+    inside_counts = inside_corners_mask.sum(dim=1, keepdim=True).float()
+    outside_counts = outside_corners_mask.sum(dim=1, keepdim=True).float()
+
+    inside_pos = torch.zeros((active_cells.shape[0], 3), device=device)
+    outside_pos = torch.zeros((active_cells.shape[0], 3), device=device)
+
+    for i in range(8):
+        mask_inside = inside_corners_mask[:, i].unsqueeze(1)
+        mask_outside = outside_corners_mask[:, i].unsqueeze(1)
+        inside_pos += corner_offsets[i].float().unsqueeze(0) * mask_inside
+        outside_pos += corner_offsets[i].float().unsqueeze(0) * mask_outside
+
+    inside_pos /= inside_counts
+    outside_pos /= outside_counts
+    gradients = inside_pos - outside_pos
+
+    pos_dirs = torch.tensor([
+        [1, 0, 0],
+        [0, 1, 0],
+        [0, 0, 1]
+    ], device=device)
+
+    cross_products = [
+        torch.linalg.cross(pos_dirs[i].float(), pos_dirs[j].float())
+        for i in range(3) for j in range(i+1, 3)
+    ]
+
+    faces = []
+    all_keys = set(vertex_lookup.keys())
+
+    face_progress_mod = round(len(active_cells)/38*3)
+
+    for pair_idx, (i, j) in enumerate([(0,1), (0,2), (1,2)]):
+        dir_i = pos_dirs[i]
+        dir_j = pos_dirs[j]
+        cross_product = cross_products[pair_idx]
+
+        ni_positions = active_cells + dir_i
+        nj_positions = active_cells + dir_j
+        diag_positions = active_cells + dir_i + dir_j
+
+        alignments = torch.matmul(gradients, cross_product)
+
+        valid_quads = []
+        quad_indices = []
+
+        for idx, active_cell in enumerate(active_cells):
+            if not idx % face_progress_mod:
+                progress.update(1)
+            cell_key = tuple(active_cell.tolist())
+            ni_key = tuple(ni_positions[idx].tolist())
+            nj_key = tuple(nj_positions[idx].tolist())
+            diag_key = tuple(diag_positions[idx].tolist())
+
+            if cell_key in all_keys and ni_key in all_keys and nj_key in all_keys and diag_key in all_keys:
+                v0 = vertex_lookup[cell_key]
+                v1 = vertex_lookup[ni_key]
+                v2 = vertex_lookup[nj_key]
+                v3 = vertex_lookup[diag_key]
+
+                valid_quads.append((v0, v1, v2, v3))
+                quad_indices.append(idx)
+
+        for q_idx, (v0, v1, v2, v3) in enumerate(valid_quads):
+            cell_idx = quad_indices[q_idx]
+            if alignments[cell_idx] > 0:
+                faces.append(torch.tensor([v0, v1, v3], device=device, dtype=torch.long))
+                faces.append(torch.tensor([v0, v3, v2], device=device, dtype=torch.long))
+            else:
+                faces.append(torch.tensor([v0, v3, v1], device=device, dtype=torch.long))
+                faces.append(torch.tensor([v0, v2, v3], device=device, dtype=torch.long))
+
+    if faces:
+        faces = torch.stack(faces)
+    else:
+        faces = torch.zeros((0, 3), dtype=torch.long, device=device)
+
+    v_min = 0
+    v_max = max(D, H, W)
+
+    final_vertices = final_vertices - (v_min + v_max) / 2
+
+    scale = (v_max - v_min) / 2
+    if scale > 0:
+        final_vertices = final_vertices / scale
+
+    final_vertices = torch.fliplr(final_vertices)
+
+    return final_vertices, faces
+
+class MESH:
+    def __init__(self, vertices, faces):
+        self.vertices = vertices
+        self.faces = faces
+
+
+class VoxelToMeshBasic:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"voxel": ("VOXEL", ),
+                             "threshold": ("FLOAT", {"default": 0.6, "min": -1.0, "max": 1.0, "step": 0.01}),
+                             }}
+    RETURN_TYPES = ("MESH",)
+    FUNCTION = "decode"
+
+    CATEGORY = "3d"
+
+    def decode(self, voxel, threshold):
+        vertices = []
+        faces = []
+        for x in voxel.data:
+            v, f = voxel_to_mesh(x, threshold=threshold, device=None)
+            vertices.append(v)
+            faces.append(f)
+
+        return (MESH(torch.stack(vertices), torch.stack(faces)), )
+
+class VoxelToMesh:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"voxel": ("VOXEL", ),
+                             "algorithm": (["surface net", "basic"], ),
+                             "threshold": ("FLOAT", {"default": 0.6, "min": -1.0, "max": 1.0, "step": 0.01}),
+                             }}
+    RETURN_TYPES = ("MESH",)
+    FUNCTION = "decode"
+
+    CATEGORY = "3d"
+
+    def decode(self, voxel, algorithm, threshold):
+        vertices = []
+        faces = []
+
+        if algorithm == "basic":
+            mesh_function = voxel_to_mesh
+        elif algorithm == "surface net":
+            mesh_function = voxel_to_mesh_surfnet
+
+        for x in voxel.data:
+            v, f = mesh_function(x, threshold=threshold, device=None)
+            vertices.append(v)
+            faces.append(f)
+
+        return (MESH(torch.stack(vertices), torch.stack(faces)), )
+
+
+def save_glb(vertices, faces, filepath, metadata=None):
+    """
+    Save PyTorch tensor vertices and faces as a GLB file without external dependencies.
+
+    Parameters:
+    vertices: torch.Tensor of shape (N, 3) - The vertex coordinates
+    faces: torch.Tensor of shape (M, 3) - The face indices (triangle faces)
+    filepath: str - Output filepath (should end with .glb)
+    """
+
+    # Convert tensors to numpy arrays
+    vertices_np = vertices.cpu().numpy().astype(np.float32)
+    faces_np = faces.cpu().numpy().astype(np.uint32)
+
+    vertices_buffer = vertices_np.tobytes()
+    indices_buffer = faces_np.tobytes()
+
+    def pad_to_4_bytes(buffer):
+        padding_length = (4 - (len(buffer) % 4)) % 4
+        return buffer + b'\x00' * padding_length
+
+    vertices_buffer_padded = pad_to_4_bytes(vertices_buffer)
+    indices_buffer_padded = pad_to_4_bytes(indices_buffer)
+
+    buffer_data = vertices_buffer_padded + indices_buffer_padded
+
+    vertices_byte_length = len(vertices_buffer)
+    vertices_byte_offset = 0
+    indices_byte_length = len(indices_buffer)
+    indices_byte_offset = len(vertices_buffer_padded)
+
+    gltf = {
+        "asset": {"version": "2.0", "generator": "ComfyUI"},
+        "buffers": [
+            {
+                "byteLength": len(buffer_data)
+            }
+        ],
+        "bufferViews": [
+            {
+                "buffer": 0,
+                "byteOffset": vertices_byte_offset,
+                "byteLength": vertices_byte_length,
+                "target": 34962  # ARRAY_BUFFER
+            },
+            {
+                "buffer": 0,
+                "byteOffset": indices_byte_offset,
+                "byteLength": indices_byte_length,
+                "target": 34963  # ELEMENT_ARRAY_BUFFER
+            }
+        ],
+        "accessors": [
+            {
+                "bufferView": 0,
+                "byteOffset": 0,
+                "componentType": 5126,  # FLOAT
+                "count": len(vertices_np),
+                "type": "VEC3",
+                "max": vertices_np.max(axis=0).tolist(),
+                "min": vertices_np.min(axis=0).tolist()
+            },
+            {
+                "bufferView": 1,
+                "byteOffset": 0,
+                "componentType": 5125,  # UNSIGNED_INT
+                "count": faces_np.size,
+                "type": "SCALAR"
+            }
+        ],
+        "meshes": [
+            {
+                "primitives": [
+                    {
+                        "attributes": {
+                            "POSITION": 0
+                        },
+                        "indices": 1,
+                        "mode": 4  # TRIANGLES
+                    }
+                ]
+            }
+        ],
+        "nodes": [
+            {
+                "mesh": 0
+            }
+        ],
+        "scenes": [
+            {
+                "nodes": [0]
+            }
+        ],
+        "scene": 0
+    }
+
+    if metadata is not None:
+        gltf["asset"]["extras"] = metadata
+
+    # Convert the JSON to bytes
+    gltf_json = json.dumps(gltf).encode('utf8')
+
+    def pad_json_to_4_bytes(buffer):
+        padding_length = (4 - (len(buffer) % 4)) % 4
+        return buffer + b' ' * padding_length
+
+    gltf_json_padded = pad_json_to_4_bytes(gltf_json)
+
+    # Create the GLB header
+    # Magic glTF
+    glb_header = struct.pack('<4sII', b'glTF', 2, 12 + 8 + len(gltf_json_padded) + 8 + len(buffer_data))
+
+    # Create JSON chunk header (chunk type 0)
+    json_chunk_header = struct.pack('<II', len(gltf_json_padded), 0x4E4F534A)  # "JSON" in little endian
+
+    # Create BIN chunk header (chunk type 1)
+    bin_chunk_header = struct.pack('<II', len(buffer_data), 0x004E4942)  # "BIN\0" in little endian
+
+    # Write the GLB file
+    with open(filepath, 'wb') as f:
+        f.write(glb_header)
+        f.write(json_chunk_header)
+        f.write(gltf_json_padded)
+        f.write(bin_chunk_header)
+        f.write(buffer_data)
+
+    return filepath
+
+
+class SaveGLB:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"mesh": ("MESH", ),
+                             "filename_prefix": ("STRING", {"default": "mesh/ComfyUI"}), },
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, }
+
+    RETURN_TYPES = ()
+    FUNCTION = "save"
+
+    OUTPUT_NODE = True
+
+    CATEGORY = "3d"
+
+    def save(self, mesh, filename_prefix, prompt=None, extra_pnginfo=None):
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory())
+        results = []
+
+        metadata = {}
+        if not args.disable_metadata:
+            if prompt is not None:
+                metadata["prompt"] = json.dumps(prompt)
+            if extra_pnginfo is not None:
+                for x in extra_pnginfo:
+                    metadata[x] = json.dumps(extra_pnginfo[x])
+
+        for i in range(mesh.vertices.shape[0]):
+            f = f"{filename}_{counter:05}_.glb"
+            save_glb(mesh.vertices[i], mesh.faces[i], os.path.join(full_output_folder, f), metadata)
+            results.append({
+                "filename": f,
+                "subfolder": subfolder,
+                "type": "output"
+            })
+            counter += 1
+        return {"ui": {"3d": results}}
+
+
+NODE_CLASS_MAPPINGS = {
+    "EmptyLatentHunyuan3Dv2": EmptyLatentHunyuan3Dv2,
+    "Hunyuan3Dv2Conditioning": Hunyuan3Dv2Conditioning,
+    "Hunyuan3Dv2ConditioningMultiView": Hunyuan3Dv2ConditioningMultiView,
+    "VAEDecodeHunyuan3D": VAEDecodeHunyuan3D,
+    "VoxelToMeshBasic": VoxelToMeshBasic,
+    "VoxelToMesh": VoxelToMesh,
+    "SaveGLB": SaveGLB,
+}

comfy_extras/nodes_load_3d.py

@@ -19,12 +19,10 @@ class Load3D():
             "image": ("LOAD_3D", {}),
             "width": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
             "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
-            "material": (["original", "normal", "wireframe", "depth"],),
-            "up_direction": (["original", "-x", "+x", "-y", "+y", "-z", "+z"],),
         }}
 
-    RETURN_TYPES = ("IMAGE", "MASK", "STRING")
-    RETURN_NAMES = ("image", "mask", "mesh_path")
+    RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "IMAGE")
+    RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "lineart")
 
     FUNCTION = "process"
     EXPERIMENTAL = True
@@ -34,12 +32,16 @@ class Load3D():
     def process(self, model_file, image, **kwargs):
         image_path = folder_paths.get_annotated_filepath(image['image'])
         mask_path = folder_paths.get_annotated_filepath(image['mask'])
+        normal_path = folder_paths.get_annotated_filepath(image['normal'])
+        lineart_path = folder_paths.get_annotated_filepath(image['lineart'])
 
         load_image_node = nodes.LoadImage()
         output_image, ignore_mask = load_image_node.load_image(image=image_path)
         ignore_image, output_mask = load_image_node.load_image(image=mask_path)
+        normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path)
+        lineart_image, ignore_mask3 = load_image_node.load_image(image=lineart_path)
 
-        return output_image, output_mask, model_file,
+        return output_image, output_mask, model_file, normal_image, lineart_image
 
 class Load3DAnimation():
     @classmethod
@@ -55,12 +57,10 @@ class Load3DAnimation():
             "image": ("LOAD_3D_ANIMATION", {}),
             "width": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
             "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
-            "material": (["original", "normal", "wireframe", "depth"],),
-            "up_direction": (["original", "-x", "+x", "-y", "+y", "-z", "+z"],),
         }}
 
-    RETURN_TYPES = ("IMAGE", "MASK", "STRING")
-    RETURN_NAMES = ("image", "mask", "mesh_path")
+    RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE")
+    RETURN_NAMES = ("image", "mask", "mesh_path", "normal")
 
     FUNCTION = "process"
     EXPERIMENTAL = True
@@ -70,20 +70,20 @@ class Load3DAnimation():
     def process(self, model_file, image, **kwargs):
         image_path = folder_paths.get_annotated_filepath(image['image'])
         mask_path = folder_paths.get_annotated_filepath(image['mask'])
+        normal_path = folder_paths.get_annotated_filepath(image['normal'])
 
         load_image_node = nodes.LoadImage()
         output_image, ignore_mask = load_image_node.load_image(image=image_path)
         ignore_image, output_mask = load_image_node.load_image(image=mask_path)
+        normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path)
 
-        return output_image, output_mask, model_file,
+        return output_image, output_mask, model_file, normal_image
 
 class Preview3D():
     @classmethod
     def INPUT_TYPES(s):
         return {"required": {
             "model_file": ("STRING", {"default": "", "multiline": False}),
-            "material": (["original", "normal", "wireframe", "depth"],),
-            "up_direction": (["original", "-x", "+x", "-y", "+y", "-z", "+z"],),
         }}
 
     OUTPUT_NODE = True
@@ -102,8 +102,6 @@ class Preview3DAnimation():
     def INPUT_TYPES(s):
         return {"required": {
             "model_file": ("STRING", {"default": "", "multiline": False}),
-            "material": (["original", "normal", "wireframe", "depth"],),
-            "up_direction": (["original", "-x", "+x", "-y", "+y", "-z", "+z"],),
         }}
 
     OUTPUT_NODE = True

comfy_extras/nodes_lt.py

@@ -99,12 +99,13 @@ class LTXVAddGuide:
                              "negative": ("CONDITIONING", ),
                              "vae": ("VAE",),
                              "latent": ("LATENT",),
-                             "image": ("IMAGE", {"tooltip": "Image or video to condition the latent video on. Must be 8*n + 1 frames." \
+                             "image": ("IMAGE", {"tooltip": "Image or video to condition the latent video on. Must be 8*n + 1 frames."
                                                  "If the video is not 8*n + 1 frames, it will be cropped to the nearest 8*n + 1 frames."}),
                              "frame_idx": ("INT", {"default": 0, "min": -9999, "max": 9999,
-                                                   "tooltip": "Frame index to start the conditioning at. Must be divisible by 8. " \
-                                                   "If a frame is not divisible by 8, it will be rounded down to the nearest multiple of 8. " \
-                                                   "Negative values are counted from the end of the video."}),
+                                                   "tooltip": "Frame index to start the conditioning at. For single-frame images or "
+                                                   "videos with 1-8 frames, any frame_idx value is acceptable. For videos with 9+ "
+                                                   "frames, frame_idx must be divisible by 8, otherwise it will be rounded down to "
+                                                   "the nearest multiple of 8. Negative values are counted from the end of the video."}),
                              "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
                              }
             }
@@ -127,12 +128,13 @@ class LTXVAddGuide:
         t = vae.encode(encode_pixels)
         return encode_pixels, t
 
-    def get_latent_index(self, cond, latent_length, frame_idx, scale_factors):
+    def get_latent_index(self, cond, latent_length, guide_length, frame_idx, scale_factors):
         time_scale_factor, _, _ = scale_factors
         _, num_keyframes = get_keyframe_idxs(cond)
         latent_count = latent_length - num_keyframes
-        frame_idx = frame_idx if frame_idx >= 0 else max((latent_count - 1) * 8 + 1 + frame_idx, 0)
-        frame_idx = frame_idx // time_scale_factor * time_scale_factor # frame index must be divisible by 8
+        frame_idx = frame_idx if frame_idx >= 0 else max((latent_count - 1) * time_scale_factor + 1 + frame_idx, 0)
+        if guide_length > 1:
+            frame_idx = frame_idx // time_scale_factor * time_scale_factor # frame index must be divisible by 8
 
         latent_idx = (frame_idx + time_scale_factor - 1) // time_scale_factor
 
@@ -191,7 +193,7 @@ class LTXVAddGuide:
         _, _, latent_length, latent_height, latent_width = latent_image.shape
         image, t = self.encode(vae, latent_width, latent_height, image, scale_factors)
 
-        frame_idx, latent_idx = self.get_latent_index(positive, latent_length, frame_idx, scale_factors)
+        frame_idx, latent_idx = self.get_latent_index(positive, latent_length, len(image), frame_idx, scale_factors)
         assert latent_idx + t.shape[2] <= latent_length, "Conditioning frames exceed the length of the latent sequence."
 
         num_prefix_frames = min(self._num_prefix_frames, t.shape[2])
@@ -444,10 +446,9 @@ class LTXVPreprocess:
     CATEGORY = "image"
 
     def preprocess(self, image, img_compression):
-        if img_compression > 0:
-            output_images = []
-            for i in range(image.shape[0]):
-                output_images.append(preprocess(image[i], img_compression))
+        output_images = []
+        for i in range(image.shape[0]):
+            output_images.append(preprocess(image[i], img_compression))
         return (torch.stack(output_images),)
 
 

comfy_extras/nodes_mask.py

@@ -2,6 +2,7 @@ import numpy as np
 import scipy.ndimage
 import torch
 import comfy.utils
+import node_helpers
 
 from nodes import MAX_RESOLUTION
 
@@ -87,6 +88,7 @@ class ImageCompositeMasked:
     CATEGORY = "image"
 
     def composite(self, destination, source, x, y, resize_source, mask = None):
+        destination, source = node_helpers.image_alpha_fix(destination, source)
         destination = destination.clone().movedim(-1, 1)
         output = composite(destination, source.movedim(-1, 1), x, y, mask, 1, resize_source).movedim(1, -1)
         return (output,)

comfy_extras/nodes_model_advanced.py

@@ -20,10 +20,6 @@ class LCM(comfy.model_sampling.EPS):
 
         return c_out * x0 + c_skip * model_input
 
-class X0(comfy.model_sampling.EPS):
-    def calculate_denoised(self, sigma, model_output, model_input):
-        return model_output
-
 class ModelSamplingDiscreteDistilled(comfy.model_sampling.ModelSamplingDiscrete):
     original_timesteps = 50
 
@@ -56,7 +52,7 @@ class ModelSamplingDiscrete:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "model": ("MODEL",),
-                              "sampling": (["eps", "v_prediction", "lcm", "x0"],),
+                              "sampling": (["eps", "v_prediction", "lcm", "x0", "img_to_img"],),
                               "zsnr": ("BOOLEAN", {"default": False}),
                               }}
 
@@ -77,7 +73,9 @@ class ModelSamplingDiscrete:
             sampling_type = LCM
             sampling_base = ModelSamplingDiscreteDistilled
         elif sampling == "x0":
-            sampling_type = X0
+            sampling_type = comfy.model_sampling.X0
+        elif sampling == "img_to_img":
+            sampling_type = comfy.model_sampling.IMG_TO_IMG
 
         class ModelSamplingAdvanced(sampling_base, sampling_type):
             pass

comfy_extras/nodes_model_merging_model_specific.py

@@ -244,6 +244,30 @@ class ModelMergeCosmos14B(comfy_extras.nodes_model_merging.ModelMergeBlocks):
 
         return {"required": arg_dict}
 
+class ModelMergeWAN2_1(comfy_extras.nodes_model_merging.ModelMergeBlocks):
+    CATEGORY = "advanced/model_merging/model_specific"
+    DESCRIPTION = "1.3B model has 30 blocks, 14B model has 40 blocks. Image to video model has the extra img_emb."
+
+    @classmethod
+    def INPUT_TYPES(s):
+        arg_dict = { "model1": ("MODEL",),
+                              "model2": ("MODEL",)}
+
+        argument = ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01})
+
+        arg_dict["patch_embedding."] = argument
+        arg_dict["time_embedding."] = argument
+        arg_dict["time_projection."] = argument
+        arg_dict["text_embedding."] = argument
+        arg_dict["img_emb."] = argument
+
+        for i in range(40):
+            arg_dict["blocks.{}.".format(i)] = argument
+
+        arg_dict["head."] = argument
+
+        return {"required": arg_dict}
+
 NODE_CLASS_MAPPINGS = {
     "ModelMergeSD1": ModelMergeSD1,
     "ModelMergeSD2": ModelMergeSD1, #SD1 and SD2 have the same blocks
@@ -256,4 +280,5 @@ NODE_CLASS_MAPPINGS = {
     "ModelMergeLTXV": ModelMergeLTXV,
     "ModelMergeCosmos7B": ModelMergeCosmos7B,
     "ModelMergeCosmos14B": ModelMergeCosmos14B,
+    "ModelMergeWAN2_1": ModelMergeWAN2_1,
 }

comfy_extras/nodes_morphology.py

@@ -2,6 +2,7 @@ import torch
 import comfy.model_management
 
 from kornia.morphology import dilation, erosion, opening, closing, gradient, top_hat, bottom_hat
+import kornia.color
 
 
 class Morphology:
@@ -40,8 +41,45 @@ class Morphology:
         img_out = output.to(comfy.model_management.intermediate_device()).movedim(1, -1)
         return (img_out,)
 
+
+class ImageRGBToYUV:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": ("IMAGE",),
+                              }}
+
+    RETURN_TYPES = ("IMAGE", "IMAGE", "IMAGE")
+    RETURN_NAMES = ("Y", "U", "V")
+    FUNCTION = "execute"
+
+    CATEGORY = "image/batch"
+
+    def execute(self, image):
+        out = kornia.color.rgb_to_ycbcr(image.movedim(-1, 1)).movedim(1, -1)
+        return (out[..., 0:1].expand_as(image), out[..., 1:2].expand_as(image), out[..., 2:3].expand_as(image))
+
+class ImageYUVToRGB:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"Y": ("IMAGE",),
+                             "U": ("IMAGE",),
+                             "V": ("IMAGE",),
+                              }}
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "execute"
+
+    CATEGORY = "image/batch"
+
+    def execute(self, Y, U, V):
+        image = torch.cat([torch.mean(Y, dim=-1, keepdim=True), torch.mean(U, dim=-1, keepdim=True), torch.mean(V, dim=-1, keepdim=True)], dim=-1)
+        out = kornia.color.ycbcr_to_rgb(image.movedim(-1, 1)).movedim(1, -1)
+        return (out,)
+
 NODE_CLASS_MAPPINGS = {
     "Morphology": Morphology,
+    "ImageRGBToYUV": ImageRGBToYUV,
+    "ImageYUVToRGB": ImageYUVToRGB,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {

comfy_extras/nodes_post_processing.py

@@ -6,7 +6,7 @@ import math
 
 import comfy.utils
 import comfy.model_management
-
+import node_helpers
 
 class Blend:
     def __init__(self):
@@ -34,6 +34,7 @@ class Blend:
     CATEGORY = "image/postprocessing"
 
     def blend_images(self, image1: torch.Tensor, image2: torch.Tensor, blend_factor: float, blend_mode: str):
+        image1, image2 = node_helpers.image_alpha_fix(image1, image2)
         image2 = image2.to(image1.device)
         if image1.shape != image2.shape:
             image2 = image2.permute(0, 3, 1, 2)

comfy_extras/nodes_primitive.py

@@ -0,0 +1,79 @@
+# Primitive nodes that are evaluated at backend.
+from __future__ import annotations
+
+from comfy.comfy_types.node_typing import ComfyNodeABC, InputTypeDict, IO
+
+
+class String(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {"value": (IO.STRING, {})},
+        }
+
+    RETURN_TYPES = (IO.STRING,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/primitive"
+
+    def execute(self, value: str) -> tuple[str]:
+        return (value,)
+
+
+class Int(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {"value": (IO.INT, {"control_after_generate": True})},
+        }
+
+    RETURN_TYPES = (IO.INT,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/primitive"
+
+    def execute(self, value: int) -> tuple[int]:
+        return (value,)
+
+
+class Float(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {"value": (IO.FLOAT, {})},
+        }
+
+    RETURN_TYPES = (IO.FLOAT,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/primitive"
+
+    def execute(self, value: float) -> tuple[float]:
+        return (value,)
+
+
+class Boolean(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {"value": (IO.BOOLEAN, {})},
+        }
+
+    RETURN_TYPES = (IO.BOOLEAN,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/primitive"
+
+    def execute(self, value: bool) -> tuple[bool]:
+        return (value,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "PrimitiveString": String,
+    "PrimitiveInt": Int,
+    "PrimitiveFloat": Float,
+    "PrimitiveBoolean": Boolean,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PrimitiveString": "String",
+    "PrimitiveInt": "Int",
+    "PrimitiveFloat": "Float",
+    "PrimitiveBoolean": "Boolean",
+}

comfy_extras/nodes_wan.py

@@ -3,6 +3,7 @@ import node_helpers
 import torch
 import comfy.model_management
 import comfy.utils
+import comfy.latent_formats
 
 
 class WanImageToVideo:
@@ -49,6 +50,110 @@ class WanImageToVideo:
         return (positive, negative, out_latent)
 
 
+class WanFunControlToVideo:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"positive": ("CONDITIONING", ),
+                             "negative": ("CONDITIONING", ),
+                             "vae": ("VAE", ),
+                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
+                },
+                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
+                             "start_image": ("IMAGE", ),
+                             "control_video": ("IMAGE", ),
+                }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
+    RETURN_NAMES = ("positive", "negative", "latent")
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, control_video=None):
+        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        concat_latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
+        concat_latent = concat_latent.repeat(1, 2, 1, 1, 1)
+
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            concat_latent_image = vae.encode(start_image[:, :, :, :3])
+            concat_latent[:,16:,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
+
+        if control_video is not None:
+            control_video = comfy.utils.common_upscale(control_video[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            concat_latent_image = vae.encode(control_video[:, :, :, :3])
+            concat_latent[:,:16,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
+
+        positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent})
+        negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent})
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return (positive, negative, out_latent)
+
+class WanFunInpaintToVideo:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"positive": ("CONDITIONING", ),
+                             "negative": ("CONDITIONING", ),
+                             "vae": ("VAE", ),
+                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
+                },
+                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
+                             "start_image": ("IMAGE", ),
+                             "end_image": ("IMAGE", ),
+                }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
+    RETURN_NAMES = ("positive", "negative", "latent")
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_output=None):
+        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+        if end_image is not None:
+            end_image = comfy.utils.common_upscale(end_image[-length:].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+
+        image = torch.ones((length, height, width, 3)) * 0.5
+        mask = torch.ones((1, 1, latent.shape[2] * 4, latent.shape[-2], latent.shape[-1]))
+
+        if start_image is not None:
+            image[:start_image.shape[0]] = start_image
+            mask[:, :, :start_image.shape[0] + 3] = 0.0
+
+        if end_image is not None:
+            image[-end_image.shape[0]:] = end_image
+            mask[:, :, -end_image.shape[0]:] = 0.0
+
+        concat_latent_image = vae.encode(image[:, :, :, :3])
+        mask = mask.view(1, mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]).transpose(1, 2)
+        positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+        negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return (positive, negative, out_latent)
+
 NODE_CLASS_MAPPINGS = {
     "WanImageToVideo": WanImageToVideo,
+    "WanFunControlToVideo": WanFunControlToVideo,
+    "WanFunInpaintToVideo": WanFunInpaintToVideo,
 }

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.26"
+__version__ = "0.3.27"

execution.py

@@ -15,7 +15,7 @@ import nodes
 import comfy.model_management
 from comfy_execution.graph import get_input_info, ExecutionList, DynamicPrompt, ExecutionBlocker
 from comfy_execution.graph_utils import is_link, GraphBuilder
-from comfy_execution.caching import HierarchicalCache, LRUCache, CacheKeySetInputSignature, CacheKeySetID
+from comfy_execution.caching import HierarchicalCache, LRUCache, DependencyAwareCache, CacheKeySetInputSignature, CacheKeySetID
 from comfy_execution.validation import validate_node_input
 
 class ExecutionResult(Enum):
@@ -59,20 +59,27 @@ class IsChangedCache:
             self.is_changed[node_id] = node["is_changed"]
         return self.is_changed[node_id]
 
-class CacheSet:
-    def __init__(self, lru_size=None):
-        if lru_size is None or lru_size == 0:
-            self.init_classic_cache()
-        else:
-            self.init_lru_cache(lru_size)
-        self.all = [self.outputs, self.ui, self.objects]
 
-    # Useful for those with ample RAM/VRAM -- allows experimenting without
-    # blowing away the cache every time
-    def init_lru_cache(self, cache_size):
-        self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
-        self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
-        self.objects = HierarchicalCache(CacheKeySetID)
+class CacheType(Enum):
+    CLASSIC = 0
+    LRU = 1
+    DEPENDENCY_AWARE = 2
+
+
+class CacheSet:
+    def __init__(self, cache_type=None, cache_size=None):
+        if cache_type == CacheType.DEPENDENCY_AWARE:
+            self.init_dependency_aware_cache()
+            logging.info("Disabling intermediate node cache.")
+        elif cache_type == CacheType.LRU:
+            if cache_size is None:
+                cache_size = 0
+            self.init_lru_cache(cache_size)
+            logging.info("Using LRU cache")
+        else:
+            self.init_classic_cache()
+
+        self.all = [self.outputs, self.ui, self.objects]
 
     # Performs like the old cache -- dump data ASAP
     def init_classic_cache(self):
@@ -80,6 +87,17 @@ class CacheSet:
         self.ui = HierarchicalCache(CacheKeySetInputSignature)
         self.objects = HierarchicalCache(CacheKeySetID)
 
+    def init_lru_cache(self, cache_size):
+        self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.objects = HierarchicalCache(CacheKeySetID)
+
+    # only hold cached items while the decendents have not executed
+    def init_dependency_aware_cache(self):
+        self.outputs = DependencyAwareCache(CacheKeySetInputSignature)
+        self.ui = DependencyAwareCache(CacheKeySetInputSignature)
+        self.objects = DependencyAwareCache(CacheKeySetID)
+
     def recursive_debug_dump(self):
         result = {
             "outputs": self.outputs.recursive_debug_dump(),
@@ -414,13 +432,14 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
     return (ExecutionResult.SUCCESS, None, None)
 
 class PromptExecutor:
-    def __init__(self, server, lru_size=None):
-        self.lru_size = lru_size
+    def __init__(self, server, cache_type=False, cache_size=None):
+        self.cache_size = cache_size
+        self.cache_type = cache_type
         self.server = server
         self.reset()
 
     def reset(self):
-        self.caches = CacheSet(self.lru_size)
+        self.caches = CacheSet(cache_type=self.cache_type, cache_size=self.cache_size)
         self.status_messages = []
         self.success = True
 
@@ -634,6 +653,13 @@ def validate_inputs(prompt, item, validated):
                 continue
         else:
             try:
+                # Unwraps values wrapped in __value__ key. This is used to pass
+                # list widget value to execution, as by default list value is
+                # reserved to represent the connection between nodes.
+                if isinstance(val, dict) and "__value__" in val:
+                    val = val["__value__"]
+                    inputs[x] = val
+
                 if type_input == "INT":
                     val = int(val)
                     inputs[x] = val
@@ -768,7 +794,7 @@ def validate_prompt(prompt):
                 "details": f"Node ID '#{x}'",
                 "extra_info": {}
             }
-            return (False, error, [], [])
+            return (False, error, [], {})
 
         class_type = prompt[x]['class_type']
         class_ = nodes.NODE_CLASS_MAPPINGS.get(class_type, None)
@@ -779,7 +805,7 @@ def validate_prompt(prompt):
                 "details": f"Node ID '#{x}'",
                 "extra_info": {}
             }
-            return (False, error, [], [])
+            return (False, error, [], {})
 
         if hasattr(class_, 'OUTPUT_NODE') and class_.OUTPUT_NODE is True:
             outputs.add(x)
@@ -791,7 +817,7 @@ def validate_prompt(prompt):
             "details": "",
             "extra_info": {}
         }
-        return (False, error, [], [])
+        return (False, error, [], {})
 
     good_outputs = set()
     errors = []

main.py

@@ -139,7 +139,7 @@ from server import BinaryEventTypes
 import nodes
 import comfy.model_management
 import comfyui_version
-import app.frontend_management
+import app.logger
 
 
 def cuda_malloc_warning():
@@ -156,7 +156,13 @@ def cuda_malloc_warning():
 
 def prompt_worker(q, server_instance):
     current_time: float = 0.0
-    e = execution.PromptExecutor(server_instance, lru_size=args.cache_lru)
+    cache_type = execution.CacheType.CLASSIC
+    if args.cache_lru > 0:
+        cache_type = execution.CacheType.LRU
+    elif args.cache_none:
+        cache_type = execution.CacheType.DEPENDENCY_AWARE
+
+    e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_size=args.cache_lru)
     last_gc_collect = 0
     need_gc = False
     gc_collect_interval = 10.0
@@ -293,28 +299,14 @@ def start_comfyui(asyncio_loop=None):
     return asyncio_loop, prompt_server, start_all
 
 
-def warn_frontend_version(frontend_version):
-    try:
-        required_frontend = (0,)
-        req_path = os.path.join(os.path.dirname(__file__), 'requirements.txt')
-        with open(req_path, 'r') as f:
-            required_frontend = tuple(map(int, f.readline().split('=')[-1].split('.')))
-        if frontend_version < required_frontend:
-            logging.warning("________________________________________________________________________\nWARNING WARNING WARNING WARNING WARNING\n\nInstalled frontend version {} is lower than the recommended version {}.\n\n{}\n________________________________________________________________________".format('.'.join(map(str, frontend_version)), '.'.join(map(str, required_frontend)), app.frontend_management.frontend_install_warning_message()))
-    except:
-        pass
-
-
 if __name__ == "__main__":
     # Running directly, just start ComfyUI.
     logging.info("ComfyUI version: {}".format(comfyui_version.__version__))
-    frontend_version = app.frontend_management.frontend_version
-    logging.info("ComfyUI frontend version: {}".format('.'.join(map(str, frontend_version))))
 
     event_loop, _, start_all_func = start_comfyui()
     try:
         x = start_all_func()
-        warn_frontend_version(frontend_version)
+        app.logger.print_startup_warnings()
         event_loop.run_until_complete(x)
     except KeyboardInterrupt:
         logging.info("\nStopped server")

node_helpers.py

@@ -44,3 +44,11 @@ def string_to_torch_dtype(string):
         return torch.float16
     if string == "bf16":
         return torch.bfloat16
+
+def image_alpha_fix(destination, source):
+    if destination.shape[-1] < source.shape[-1]:
+        source = source[...,:destination.shape[-1]]
+    elif destination.shape[-1] > source.shape[-1]:
+        destination = torch.nn.functional.pad(destination, (0, 1))
+        destination[..., -1] = 1.0
+    return destination, source

nodes.py

@@ -489,7 +489,7 @@ class SaveLatent:
         file = os.path.join(full_output_folder, file)
 
         output = {}
-        output["latent_tensor"] = samples["samples"]
+        output["latent_tensor"] = samples["samples"].contiguous()
         output["latent_format_version_0"] = torch.tensor([])
 
         comfy.utils.save_torch_file(output, file, metadata=metadata)
@@ -770,6 +770,7 @@ class VAELoader:
             vae_path = folder_paths.get_full_path_or_raise("vae", vae_name)
             sd = comfy.utils.load_torch_file(vae_path)
         vae = comfy.sd.VAE(sd=sd)
+        vae.throw_exception_if_invalid()
         return (vae,)
 
 class ControlNetLoader:
@@ -785,6 +786,8 @@ class ControlNetLoader:
     def load_controlnet(self, control_net_name):
         controlnet_path = folder_paths.get_full_path_or_raise("controlnet", control_net_name)
         controlnet = comfy.controlnet.load_controlnet(controlnet_path)
+        if controlnet is None:
+            raise RuntimeError("ERROR: controlnet file is invalid and does not contain a valid controlnet model.")
         return (controlnet,)
 
 class DiffControlNetLoader:
@@ -1005,6 +1008,8 @@ class CLIPVisionLoader:
     def load_clip(self, clip_name):
         clip_path = folder_paths.get_full_path_or_raise("clip_vision", clip_name)
         clip_vision = comfy.clip_vision.load(clip_path)
+        if clip_vision is None:
+            raise RuntimeError("ERROR: clip vision file is invalid and does not contain a valid vision model.")
         return (clip_vision,)
 
 class CLIPVisionEncode:
@@ -1687,6 +1692,9 @@ class LoadImage:
             if 'A' in i.getbands():
                 mask = np.array(i.getchannel('A')).astype(np.float32) / 255.0
                 mask = 1. - torch.from_numpy(mask)
+            elif i.mode == 'P' and 'transparency' in i.info:
+                mask = np.array(i.convert('RGBA').getchannel('A')).astype(np.float32) / 255.0
+                mask = 1. - torch.from_numpy(mask)
             else:
                 mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
             output_images.append(image)
@@ -1785,14 +1793,7 @@ class LoadImageOutput(LoadImage):
 
     DESCRIPTION = "Load an image from the output folder. When the refresh button is clicked, the node will update the image list and automatically select the first image, allowing for easy iteration."
     EXPERIMENTAL = True
-    FUNCTION = "load_image_output"
-
-    def load_image_output(self, image):
-        return self.load_image(f"{image} [output]")
-
-    @classmethod
-    def VALIDATE_INPUTS(s, image):
-        return True
+    FUNCTION = "load_image"
 
 
 class ImageScale:
@@ -2129,21 +2130,25 @@ def get_module_name(module_path: str) -> str:
 
 
 def load_custom_node(module_path: str, ignore=set(), module_parent="custom_nodes") -> bool:
-    module_name = os.path.basename(module_path)
+    module_name = get_module_name(module_path)
     if os.path.isfile(module_path):
         sp = os.path.splitext(module_path)
         module_name = sp[0]
+        sys_module_name = module_name
+    elif os.path.isdir(module_path):
+        sys_module_name = module_path.replace(".", "_x_")
+
     try:
         logging.debug("Trying to load custom node {}".format(module_path))
         if os.path.isfile(module_path):
-            module_spec = importlib.util.spec_from_file_location(module_name, module_path)
+            module_spec = importlib.util.spec_from_file_location(sys_module_name, module_path)
             module_dir = os.path.split(module_path)[0]
         else:
-            module_spec = importlib.util.spec_from_file_location(module_name, os.path.join(module_path, "__init__.py"))
+            module_spec = importlib.util.spec_from_file_location(sys_module_name, os.path.join(module_path, "__init__.py"))
             module_dir = module_path
 
         module = importlib.util.module_from_spec(module_spec)
-        sys.modules[module_name] = module
+        sys.modules[sys_module_name] = module
         module_spec.loader.exec_module(module)
 
         LOADED_MODULE_DIRS[module_name] = os.path.abspath(module_dir)
@@ -2270,6 +2275,10 @@ def init_builtin_extra_nodes():
         "nodes_video.py",
         "nodes_lumina2.py",
         "nodes_wan.py",
+        "nodes_lotus.py",
+        "nodes_hunyuan3d.py",
+        "nodes_primitive.py",
+        "nodes_cfg.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.26"
+version = "0.3.27"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.11.8
+comfyui-frontend-package==1.15.13
 torch
 torchsde
 torchvision

server.py

@@ -48,7 +48,7 @@ async def send_socket_catch_exception(function, message):
 @web.middleware
 async def cache_control(request: web.Request, handler):
     response: web.Response = await handler(request)
-    if request.path.endswith('.js') or request.path.endswith('.css'):
+    if request.path.endswith('.js') or request.path.endswith('.css') or request.path.endswith('index.json'):
         response.headers.setdefault('Cache-Control', 'no-cache')
     return response
 
@@ -657,7 +657,13 @@ class PromptServer():
                     logging.warning("invalid prompt: {}".format(valid[1]))
                     return web.json_response({"error": valid[1], "node_errors": valid[3]}, status=400)
             else:
-                return web.json_response({"error": "no prompt", "node_errors": []}, status=400)
+                error = {
+                    "type": "no_prompt",
+                    "message": "No prompt provided",
+                    "details": "No prompt provided",
+                    "extra_info": {}
+                }
+                return web.json_response({"error": error, "node_errors": {}}, status=400)
 
         @routes.post("/queue")
         async def post_queue(request):

tests-unit/app_test/frontend_manager_test.py

@@ -70,7 +70,7 @@ def test_get_release_invalid_version(mock_provider):
 def test_init_frontend_default():
     version_string = DEFAULT_VERSION_STRING
     frontend_path = FrontendManager.init_frontend(version_string)
-    assert frontend_path == FrontendManager.DEFAULT_FRONTEND_PATH
+    assert frontend_path == FrontendManager.default_frontend_path()
 
 
 def test_init_frontend_invalid_version():
@@ -84,24 +84,29 @@ def test_init_frontend_invalid_provider():
     with pytest.raises(HTTPError):
         FrontendManager.init_frontend_unsafe(version_string)
 
+
 @pytest.fixture
 def mock_os_functions():
-    with patch('app.frontend_management.os.makedirs') as mock_makedirs, \
-         patch('app.frontend_management.os.listdir') as mock_listdir, \
-         patch('app.frontend_management.os.rmdir') as mock_rmdir:
+    with (
+        patch("app.frontend_management.os.makedirs") as mock_makedirs,
+        patch("app.frontend_management.os.listdir") as mock_listdir,
+        patch("app.frontend_management.os.rmdir") as mock_rmdir,
+    ):
         mock_listdir.return_value = []  # Simulate empty directory
         yield mock_makedirs, mock_listdir, mock_rmdir
 
+
 @pytest.fixture
 def mock_download():
-    with patch('app.frontend_management.download_release_asset_zip') as mock:
+    with patch("app.frontend_management.download_release_asset_zip") as mock:
         mock.side_effect = Exception("Download failed")  # Simulate download failure
         yield mock
 
+
 def test_finally_block(mock_os_functions, mock_download, mock_provider):
     # Arrange
     mock_makedirs, mock_listdir, mock_rmdir = mock_os_functions
-    version_string = 'test-owner/test-repo@1.0.0'
+    version_string = "test-owner/test-repo@1.0.0"
 
     # Act & Assert
     with pytest.raises(Exception):
@@ -128,3 +133,42 @@ def test_parse_version_string_invalid():
     version_string = "invalid"
     with pytest.raises(argparse.ArgumentTypeError):
         FrontendManager.parse_version_string(version_string)
+
+
+def test_init_frontend_default_with_mocks():
+    # Arrange
+    version_string = DEFAULT_VERSION_STRING
+
+    # Act
+    with (
+        patch("app.frontend_management.check_frontend_version") as mock_check,
+        patch.object(
+            FrontendManager, "default_frontend_path", return_value="/mocked/path"
+        ),
+    ):
+        frontend_path = FrontendManager.init_frontend(version_string)
+
+    # Assert
+    assert frontend_path == "/mocked/path"
+    mock_check.assert_called_once()
+
+
+def test_init_frontend_fallback_on_error():
+    # Arrange
+    version_string = "test-owner/test-repo@1.0.0"
+
+    # Act
+    with (
+        patch.object(
+            FrontendManager, "init_frontend_unsafe", side_effect=Exception("Test error")
+        ),
+        patch("app.frontend_management.check_frontend_version") as mock_check,
+        patch.object(
+            FrontendManager, "default_frontend_path", return_value="/default/path"
+        ),
+    ):
+        frontend_path = FrontendManager.init_frontend(version_string)
+
+    # Assert
+    assert frontend_path == "/default/path"
+    mock_check.assert_called_once()