mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2025-01-11 10:25:16 +00:00
bddb02660c
* PixArt initial version * PixArt Diffusers convert logic * pos_emb and interpolation logic * Reduce duplicate code * Formatting * Use optimized attention * Edit empty token logic * Basic PixArt LoRA support * Fix aspect ratio logic * PixArtAlpha text encode with conds * Use same detection key logic for PixArt diffusers
202 lines
8.2 KiB
Python
202 lines
8.2 KiB
Python
# Based on:
|
|
# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
|
|
# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from .blocks import (
|
|
t2i_modulate,
|
|
CaptionEmbedder,
|
|
AttentionKVCompress,
|
|
MultiHeadCrossAttention,
|
|
T2IFinalLayer,
|
|
)
|
|
from comfy.ldm.modules.diffusionmodules.mmdit import PatchEmbed, TimestepEmbedder, Mlp, get_1d_sincos_pos_embed_from_grid_torch
|
|
|
|
|
|
class PixArtBlock(nn.Module):
|
|
"""
|
|
A PixArt block with adaptive layer norm (adaLN-single) conditioning.
|
|
"""
|
|
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0, input_size=None, sampling=None, sr_ratio=1, qk_norm=False, **block_kwargs):
|
|
super().__init__()
|
|
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
self.attn = AttentionKVCompress(
|
|
hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
|
|
qk_norm=qk_norm, **block_kwargs
|
|
)
|
|
self.cross_attn = MultiHeadCrossAttention(hidden_size, num_heads, **block_kwargs)
|
|
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
# to be compatible with lower version pytorch
|
|
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
|
self.mlp = Mlp(in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0)
|
|
self.drop_path = nn.Identity() #DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
|
self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
|
|
self.sampling = sampling
|
|
self.sr_ratio = sr_ratio
|
|
|
|
def forward(self, x, y, t, mask=None, **kwargs):
|
|
B, N, C = x.shape
|
|
|
|
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None] + t.reshape(B, 6, -1)).chunk(6, dim=1)
|
|
x = x + self.drop_path(gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)).reshape(B, N, C))
|
|
x = x + self.cross_attn(x, y, mask)
|
|
x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
|
|
|
|
return x
|
|
|
|
|
|
### Core PixArt Model ###
|
|
class PixArt(nn.Module):
|
|
"""
|
|
Diffusion model with a Transformer backbone.
|
|
"""
|
|
def __init__(
|
|
self,
|
|
input_size=32,
|
|
patch_size=2,
|
|
in_channels=4,
|
|
hidden_size=1152,
|
|
depth=28,
|
|
num_heads=16,
|
|
mlp_ratio=4.0,
|
|
class_dropout_prob=0.1,
|
|
pred_sigma=True,
|
|
drop_path: float = 0.,
|
|
caption_channels=4096,
|
|
pe_interpolation=1.0,
|
|
pe_precision=None,
|
|
config=None,
|
|
model_max_length=120,
|
|
qk_norm=False,
|
|
kv_compress_config=None,
|
|
**kwargs,
|
|
):
|
|
super().__init__()
|
|
self.pred_sigma = pred_sigma
|
|
self.in_channels = in_channels
|
|
self.out_channels = in_channels * 2 if pred_sigma else in_channels
|
|
self.patch_size = patch_size
|
|
self.num_heads = num_heads
|
|
self.pe_interpolation = pe_interpolation
|
|
self.pe_precision = pe_precision
|
|
self.depth = depth
|
|
|
|
self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
|
|
self.t_embedder = TimestepEmbedder(hidden_size)
|
|
num_patches = self.x_embedder.num_patches
|
|
self.base_size = input_size // self.patch_size
|
|
# Will use fixed sin-cos embedding:
|
|
self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))
|
|
|
|
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
|
self.t_block = nn.Sequential(
|
|
nn.SiLU(),
|
|
nn.Linear(hidden_size, 6 * hidden_size, bias=True)
|
|
)
|
|
self.y_embedder = CaptionEmbedder(
|
|
in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
|
|
act_layer=approx_gelu, token_num=model_max_length
|
|
)
|
|
drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)] # stochastic depth decay rule
|
|
self.kv_compress_config = kv_compress_config
|
|
if kv_compress_config is None:
|
|
self.kv_compress_config = {
|
|
'sampling': None,
|
|
'scale_factor': 1,
|
|
'kv_compress_layer': [],
|
|
}
|
|
self.blocks = nn.ModuleList([
|
|
PixArtBlock(
|
|
hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
|
|
input_size=(input_size // patch_size, input_size // patch_size),
|
|
sampling=self.kv_compress_config['sampling'],
|
|
sr_ratio=int(
|
|
self.kv_compress_config['scale_factor']
|
|
) if i in self.kv_compress_config['kv_compress_layer'] else 1,
|
|
qk_norm=qk_norm,
|
|
)
|
|
for i in range(depth)
|
|
])
|
|
self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)
|
|
|
|
def forward_raw(self, x, t, y, mask=None, data_info=None):
|
|
"""
|
|
Original forward pass of PixArt.
|
|
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
|
|
t: (N,) tensor of diffusion timesteps
|
|
y: (N, 1, 120, C) tensor of class labels
|
|
"""
|
|
x = x.to(self.dtype)
|
|
timestep = t.to(self.dtype)
|
|
y = y.to(self.dtype)
|
|
pos_embed = self.pos_embed.to(self.dtype)
|
|
x = self.x_embedder(x) + pos_embed # (N, T, D), where T = H * W / patch_size ** 2
|
|
t = self.t_embedder(timestep.to(x.dtype)) # (N, D)
|
|
t0 = self.t_block(t)
|
|
y = self.y_embedder(y, self.training) # (N, 1, L, D)
|
|
if mask is not None:
|
|
if mask.shape[0] != y.shape[0]:
|
|
mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
|
|
mask = mask.squeeze(1).squeeze(1)
|
|
y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
|
|
y_lens = mask.sum(dim=1).tolist()
|
|
else:
|
|
y_lens = [y.shape[2]] * y.shape[0]
|
|
y = y.squeeze(1).view(1, -1, x.shape[-1])
|
|
for block in self.blocks:
|
|
x = block(x, y, t0, y_lens) # (N, T, D)
|
|
x = self.final_layer(x, t) # (N, T, patch_size ** 2 * out_channels)
|
|
x = self.unpatchify(x) # (N, out_channels, H, W)
|
|
return x
|
|
|
|
def forward(self, x, timesteps, context, y=None, **kwargs):
|
|
"""
|
|
Forward pass that adapts comfy input to original forward function
|
|
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
|
|
timesteps: (N,) tensor of diffusion timesteps
|
|
context: (N, 1, 120, C) conditioning
|
|
y: extra conditioning.
|
|
"""
|
|
## Still accepts the input w/o that dim but returns garbage
|
|
if len(context.shape) == 3:
|
|
context = context.unsqueeze(1)
|
|
|
|
## run original forward pass
|
|
out = self.forward_raw(
|
|
x = x.to(self.dtype),
|
|
t = timesteps.to(self.dtype),
|
|
y = context.to(self.dtype),
|
|
)
|
|
|
|
## only return EPS
|
|
out = out.to(torch.float)
|
|
eps, _ = out[:, :self.in_channels], out[:, self.in_channels:]
|
|
return eps
|
|
|
|
def unpatchify(self, x):
|
|
"""
|
|
x: (N, T, patch_size**2 * C)
|
|
imgs: (N, H, W, C)
|
|
"""
|
|
c = self.out_channels
|
|
p = self.x_embedder.patch_size[0]
|
|
h = w = int(x.shape[1] ** 0.5)
|
|
assert h * w == x.shape[1]
|
|
|
|
x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
|
|
x = torch.einsum('nhwpqc->nchpwq', x)
|
|
imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))
|
|
return imgs
|
|
|
|
def get_2d_sincos_pos_embed_torch(embed_dim, w, h, pe_interpolation=1.0, base_size=16, device=None, dtype=torch.float32):
|
|
grid_h, grid_w = torch.meshgrid(
|
|
torch.arange(h, device=device, dtype=dtype) / (h/base_size) / pe_interpolation,
|
|
torch.arange(w, device=device, dtype=dtype) / (w/base_size) / pe_interpolation,
|
|
indexing='ij'
|
|
)
|
|
emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
|
|
emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
|
|
emb = torch.cat([emb_w, emb_h], dim=1) # (H*W, D)
|
|
return emb
|