comfyanonymous/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-05-01 03:41:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement NAG on all the models based on the Flux code. (#12500)

Browse Source 
Use the Normalized Attention Guidance node.

Flux, Flux2, Klein, Chroma, Chroma radiance, Hunyuan Video, etc..
This commit is contained in:
comfyanonymous
2026-02-16 20:30:34 -08:00
committed by GitHub
parent 8a6fbc2dc2
commit 18927538a1
7 changed files with 128 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
comfy/ldm/flux/layers.py
View File

@@ -196,6 +196,9 @@ class DoubleStreamBlock(nn.Module):
else:
(img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
transformer_patches = transformer_options.get("patches", {})
extra_options = transformer_options.copy()
# 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)
@@ -224,6 +227,12 @@ class DoubleStreamBlock(nn.Module):
attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
del q, k, v
if "attn1_output_patch" in transformer_patches:
extra_options["img_slice"] = [txt.shape[1], attn.shape[1]]
patch = transformer_patches["attn1_output_patch"]
for p in patch:
attn = p(attn, extra_options)
txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
# calculate the img bloks
@@ -303,6 +312,9 @@ class SingleStreamBlock(nn.Module):
else:
mod = vec
transformer_patches = transformer_options.get("patches", {})
extra_options = transformer_options.copy()
qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1)
q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
@@ -312,6 +324,12 @@ class SingleStreamBlock(nn.Module):
# compute attention
attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
del q, k, v
if "attn1_output_patch" in transformer_patches:
patch = transformer_patches["attn1_output_patch"]
for p in patch:
attn = p(attn, extra_options)
# compute activation in mlp stream, cat again and run second linear layer
if self.yak_mlp:
mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2]

2
comfy/ldm/flux/model.py
View File

@@ -142,6 +142,7 @@ class Flux(nn.Module):
attn_mask: Tensor = None,
) -> Tensor:
transformer_options = transformer_options.copy()
patches = transformer_options.get("patches", {})
patches_replace = transformer_options.get("patches_replace", {})
if img.ndim != 3 or txt.ndim != 3:
@@ -231,6 +232,7 @@ class Flux(nn.Module):
transformer_options["total_blocks"] = len(self.single_blocks)
transformer_options["block_type"] = "single"
transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
for i, block in enumerate(self.single_blocks):
transformer_options["block_index"] = i
if ("single_block", i) in blocks_replace: