nodes/bvhari-ComfyUI_PerpWeight
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix error with multi-chunk prompts

Browse Source
This commit is contained in:
BVH
2023-09-12 11:29:33 +05:30
committed by GitHub
parent c8d49ca6bc
commit 6c9df5dad6
1 changed files with 10 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
clipperpweight.py
+10 -12
View File
@@ -15,9 +15,7 @@ class CLIPTextEncodePerpWeight:
def encode(self, clip, text):
empty_tokens = clip.tokenize("")
sdxl_flag = False
if isinstance(empty_tokens, dict):
sdxl_flag = True
sdxl_flag = isinstance(empty_tokens, dict)
if sdxl_flag:
empty_cond, empty_cond_pooled = clip.encode_from_tokens(empty_tokens, return_pooled=True)
@@ -30,10 +28,10 @@ class CLIPTextEncodePerpWeight:
cond = torch.clone(unweighted_cond)
for i in range(unweighted_cond.shape[0]):
for j in range(unweighted_cond.shape[1]):
weight_l = tokens["l"][i][j][1]
weight_l = tokens["l"][(j//77)][(j%77)][1]
if weight_l != 1.0:
token_vector_l = unweighted_cond[i][j][:768]
zero_vector_l = empty_cond[0][j][:768]
zero_vector_l = empty_cond[0][(j%77)][:768]
perp_l = ((torch.mul(zero_vector_l, token_vector_l).sum())/(torch.norm(token_vector_l)**2)) * token_vector_l
if weight_l > 1.0:
cond[i][j][:768] = token_vector_l + (weight_l * perp_l)
@@ -42,12 +40,12 @@ class CLIPTextEncodePerpWeight:
elif weight_l < 0.0:
cond[i][j][:768] = token_vector_l + (weight_l * perp_l)
elif weight_l == 0.0:
cond[i][j][:768] = empty_cond[0][j][:768]
cond[i][j][:768] = empty_cond[0][(j%77)][:768]
weight_g = tokens["g"][i][j][1]
weight_g = tokens["g"][(j//77)][(j%77)][1]
if weight_g != 1.0:
token_vector_g = unweighted_cond[i][j][768:]
zero_vector_g = empty_cond[0][j][768:]
zero_vector_g = empty_cond[0][(j%77)][768:]
perp_g = ((torch.mul(zero_vector_g, token_vector_g).sum())/(torch.norm(token_vector_g)**2)) * token_vector_g
if weight_g > 1.0:
cond[i][j][768:] = token_vector_g + (weight_g * perp_g)
@@ -56,7 +54,7 @@ class CLIPTextEncodePerpWeight:
elif (weight_g < 0.0):
cond[i][j][768:] = token_vector_g + (weight_g * perp_g)
elif weight_g == 0.0:
cond[i][j][768:] = empty_cond[0][j][768:]
cond[i][j][768:] = empty_cond[0][(j%77)][768:]
else:
empty_cond, empty_cond_pooled = clip.encode_from_tokens(empty_tokens, return_pooled=True)
tokens = clip.tokenize(text)
@@ -66,10 +64,10 @@ class CLIPTextEncodePerpWeight:
cond = torch.clone(unweighted_cond)
for i in range(unweighted_cond.shape[0]):
for j in range(unweighted_cond.shape[1]):
weight = tokens[i][j][1]
weight = tokens[(j//77)][(j%77)][1]
if weight != 1.0:
token_vector = unweighted_cond[i][j]
zero_vector = empty_cond[0][j]
zero_vector = empty_cond[0][(j%77)]
perp = ((torch.mul(zero_vector, token_vector).sum())/(torch.norm(token_vector)**2)) * token_vector
if weight > 1.0:
cond[i][j] = token_vector + (weight * perp)
@@ -78,7 +76,7 @@ class CLIPTextEncodePerpWeight:
elif (weight < 0.0):
cond[i][j] = token_vector + (weight * perp)
elif weight == 0.0:
cond[i][j] = empty_cond[0][j]
cond[i][j] = empty_cond[0][(j%77)]
return ([[cond, {"pooled_output": unweighted_pooled}]], )