1
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-08-02 23:14:49 +08:00

Compare commits

...

2 Commits

Author SHA1 Message Date
comfyanonymous
8e69e2ddfd Bump ComfyUI version to v0.3.17 2025-02-26 17:59:10 -05:00
comfyanonymous
0270a0b41c Reduce artifacts on Wan by doing the patch embedding in fp32. 2025-02-26 16:59:26 -05:00
3 changed files with 5 additions and 5 deletions

View File

@@ -18,7 +18,7 @@ def sinusoidal_embedding_1d(dim, position):
# preprocess
assert dim % 2 == 0
half = dim // 2
position = position.type(torch.float64)
position = position.type(torch.float32)
# calculation
sinusoid = torch.outer(
@@ -353,7 +353,7 @@ class WanModel(torch.nn.Module):
# embeddings
self.patch_embedding = operations.Conv3d(
in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=torch.float32)
self.text_embedding = nn.Sequential(
operations.Linear(text_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
@@ -411,7 +411,7 @@ class WanModel(torch.nn.Module):
List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
"""
# embeddings
x = self.patch_embedding(x)
x = self.patch_embedding(x.float()).to(x.dtype)
grid_sizes = x.shape[2:]
x = x.flatten(2).transpose(1, 2)

View File

@@ -1,3 +1,3 @@
# This file is automatically generated by the build process when version is
# updated in pyproject.toml.
__version__ = "0.3.16"
__version__ = "0.3.17"

View File

@@ -1,6 +1,6 @@
[project]
name = "ComfyUI"
version = "0.3.16"
version = "0.3.17"
readme = "README.md"
license = { file = "LICENSE" }
requires-python = ">=3.9"