mirrors/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-09-15 05:57:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix issues with new hunyuan img2vid model and bumb version to v0.3.26

Browse Source
This commit is contained in:
comfyanonymous
2025-03-09 04:59:15 -04:00
parent 528d1b3563
commit 9aac21f894
4 changed files with 16 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
comfy/ldm/flux/layers.py
View File

@@ -159,20 +159,20 @@ class DoubleStreamBlock(nn.Module):
)
self.flipped_img_txt = flipped_img_txt
def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None):
def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None):
img_mod1, img_mod2 = self.img_mod(vec)
txt_mod1, txt_mod2 = self.txt_mod(vec)
# prepare image for attention
img_modulated = self.img_norm1(img)
img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims)
img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
img_qkv = self.img_attn.qkv(img_modulated)
img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
# prepare txt for attention
txt_modulated = self.txt_norm1(txt)
txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims)
txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt)
txt_qkv = self.txt_attn.qkv(txt_modulated)
txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
@@ -195,12 +195,12 @@ class DoubleStreamBlock(nn.Module):
txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
# calculate the img bloks
img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims)
img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims)), img_mod2.gate, None, modulation_dims)
img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
# calculate the txt bloks
txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims)
txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims)), txt_mod2.gate, None, modulation_dims)
txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt)
txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt)
if txt.dtype == torch.float16:
txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)