-
Notifications
You must be signed in to change notification settings - Fork 29
/
adv_encode.py
294 lines (241 loc) · 12.8 KB
/
adv_encode.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
import torch
import numpy as np
import itertools
from math import gcd
from comfy import model_management
from comfy.sdxl_clip import SDXLClipModel, SDXLRefinerClipModel, SDXLClipG
def _grouper(n, iterable):
it = iter(iterable)
while True:
chunk = list(itertools.islice(it, n))
if not chunk:
return
yield chunk
def _norm_mag(w, n):
d = w - 1
return 1 + np.sign(d) * np.sqrt(np.abs(d)**2 / n)
#return np.sign(w) * np.sqrt(np.abs(w)**2 / n)
def divide_length(word_ids, weights):
sums = dict(zip(*np.unique(word_ids, return_counts=True)))
sums[0] = 1
weights = [[_norm_mag(w, sums[id]) if id != 0 else 1.0
for w, id in zip(x, y)] for x, y in zip(weights, word_ids)]
return weights
def shift_mean_weight(word_ids, weights):
delta = 1 - np.mean([w for x, y in zip(weights, word_ids) for w, id in zip(x,y) if id != 0])
weights = [[w if id == 0 else w+delta
for w, id in zip(x, y)] for x, y in zip(weights, word_ids)]
return weights
def scale_to_norm(weights, word_ids, w_max):
top = np.max(weights)
w_max = min(top, w_max)
weights = [[w_max if id == 0 else (w/top) * w_max
for w, id in zip(x, y)] for x, y in zip(weights, word_ids)]
return weights
def from_zero(weights, base_emb):
weight_tensor = torch.tensor(weights, dtype=base_emb.dtype, device=base_emb.device)
weight_tensor = weight_tensor.reshape(1,-1,1).expand(base_emb.shape)
return base_emb * weight_tensor
def mask_word_id(tokens, word_ids, target_id, mask_token):
new_tokens = [[mask_token if wid == target_id else t
for t, wid in zip(x,y)] for x,y in zip(tokens, word_ids)]
mask = np.array(word_ids) == target_id
return (new_tokens, mask)
def batched_clip_encode(tokens, length, encode_func, num_chunks):
embs = []
for e in _grouper(32, tokens):
enc, pooled = encode_func(e)
enc = enc.reshape((len(e), length, -1))
embs.append(enc)
embs = torch.cat(embs)
embs = embs.reshape((len(tokens) // num_chunks, length * num_chunks, -1))
return embs
def from_masked(tokens, weights, word_ids, base_emb, length, encode_func, m_token=266):
pooled_base = base_emb[0,length-1:length,:]
wids, inds = np.unique(np.array(word_ids).reshape(-1), return_index=True)
weight_dict = dict((id,w)
for id,w in zip(wids ,np.array(weights).reshape(-1)[inds])
if w != 1.0)
if len(weight_dict) == 0:
return torch.zeros_like(base_emb), base_emb[0,length-1:length,:]
weight_tensor = torch.tensor(weights, dtype=base_emb.dtype, device=base_emb.device)
weight_tensor = weight_tensor.reshape(1,-1,1).expand(base_emb.shape)
#m_token = (clip.tokenizer.end_token, 1.0) if clip.tokenizer.pad_with_end else (0,1.0)
#TODO: find most suitable masking token here
m_token = (m_token, 1.0)
ws = []
masked_tokens = []
masks = []
#create prompts
for id, w in weight_dict.items():
masked, m = mask_word_id(tokens, word_ids, id, m_token)
masked_tokens.extend(masked)
m = torch.tensor(m, dtype=base_emb.dtype, device=base_emb.device)
m = m.reshape(1,-1,1).expand(base_emb.shape)
masks.append(m)
ws.append(w)
#batch process prompts
embs = batched_clip_encode(masked_tokens, length, encode_func, len(tokens))
masks = torch.cat(masks)
embs = (base_emb.expand(embs.shape) - embs)
pooled = embs[0,length-1:length,:]
embs *= masks
embs = embs.sum(axis=0, keepdim=True)
pooled_start = pooled_base.expand(len(ws), -1)
ws = torch.tensor(ws).reshape(-1,1).expand(pooled_start.shape)
pooled = (pooled - pooled_start) * (ws - 1)
pooled = pooled.mean(axis=0, keepdim=True)
return ((weight_tensor - 1) * embs), pooled_base + pooled
def mask_inds(tokens, inds, mask_token):
clip_len = len(tokens[0])
inds_set = set(inds)
new_tokens = [[mask_token if i*clip_len + j in inds_set else t
for j, t in enumerate(x)] for i, x in enumerate(tokens)]
return new_tokens
def down_weight(tokens, weights, word_ids, base_emb, length, encode_func, m_token=266):
w, w_inv = np.unique(weights,return_inverse=True)
if np.sum(w < 1) == 0:
return base_emb, tokens, base_emb[0,length-1:length,:]
#m_token = (clip.tokenizer.end_token, 1.0) if clip.tokenizer.pad_with_end else (0,1.0)
#using the comma token as a masking token seems to work better than aos tokens for SD 1.x
m_token = (m_token, 1.0)
masked_tokens = []
masked_current = tokens
for i in range(len(w)):
if w[i] >= 1:
continue
masked_current = mask_inds(masked_current, np.where(w_inv == i)[0], m_token)
masked_tokens.extend(masked_current)
embs = batched_clip_encode(masked_tokens, length, encode_func, len(tokens))
embs = torch.cat([base_emb, embs])
w = w[w<=1.0]
w_mix = np.diff([0] + w.tolist())
w_mix = torch.tensor(w_mix, dtype=embs.dtype, device=embs.device).reshape((-1,1,1))
weighted_emb = (w_mix * embs).sum(axis=0, keepdim=True)
return weighted_emb, masked_current, weighted_emb[0,length-1:length,:]
def scale_emb_to_mag(base_emb, weighted_emb):
norm_base = torch.linalg.norm(base_emb)
norm_weighted = torch.linalg.norm(weighted_emb)
embeddings_final = (norm_base / norm_weighted) * weighted_emb
return embeddings_final
def recover_dist(base_emb, weighted_emb):
fixed_std = (base_emb.std() / weighted_emb.std()) * (weighted_emb - weighted_emb.mean())
embeddings_final = fixed_std + (base_emb.mean() - fixed_std.mean())
return embeddings_final
def A1111_renorm(base_emb, weighted_emb):
embeddings_final = (base_emb.mean() / weighted_emb.mean()) * weighted_emb
return embeddings_final
def advanced_encode_from_tokens(tokenized, token_normalization, weight_interpretation, encode_func, m_token=266, length=77, w_max=1.0, return_pooled=False, apply_to_pooled=False):
tokens = [[t for t,_,_ in x] for x in tokenized]
weights = [[w for _,w,_ in x] for x in tokenized]
word_ids = [[wid for _,_,wid in x] for x in tokenized]
#weight normalization
#====================
#distribute down/up weights over word lengths
if token_normalization.startswith("length"):
weights = divide_length(word_ids, weights)
#make mean of word tokens 1
if token_normalization.endswith("mean"):
weights = shift_mean_weight(word_ids, weights)
#weight interpretation
#=====================
pooled = None
if weight_interpretation == "comfy":
weighted_tokens = [[(t,w) for t, w in zip(x, y)] for x, y in zip(tokens, weights)]
weighted_emb, pooled_base = encode_func(weighted_tokens)
pooled = pooled_base
else:
unweighted_tokens = [[(t,1.0) for t, _,_ in x] for x in tokenized]
base_emb, pooled_base = encode_func(unweighted_tokens)
if weight_interpretation == "A1111":
weighted_emb = from_zero(weights, base_emb)
weighted_emb = A1111_renorm(base_emb, weighted_emb)
pooled = pooled_base
if weight_interpretation == "compel":
pos_tokens = [[(t,w) if w >= 1.0 else (t,1.0) for t, w in zip(x, y)] for x, y in zip(tokens, weights)]
weighted_emb, _ = encode_func(pos_tokens)
weighted_emb, _, pooled = down_weight(pos_tokens, weights, word_ids, weighted_emb, length, encode_func)
if weight_interpretation == "comfy++":
weighted_emb, tokens_down, _ = down_weight(unweighted_tokens, weights, word_ids, base_emb, length, encode_func)
weights = [[w if w > 1.0 else 1.0 for w in x] for x in weights]
#unweighted_tokens = [[(t,1.0) for t, _,_ in x] for x in tokens_down]
embs, pooled = from_masked(unweighted_tokens, weights, word_ids, base_emb, length, encode_func)
weighted_emb += embs
if weight_interpretation == "down_weight":
weights = scale_to_norm(weights, word_ids, w_max)
weighted_emb, _, pooled = down_weight(unweighted_tokens, weights, word_ids, base_emb, length, encode_func)
if return_pooled:
if apply_to_pooled:
return weighted_emb, pooled
else:
return weighted_emb, pooled_base
return weighted_emb, None
def encode_token_weights_g(model, token_weight_pairs):
return model.clip_g.encode_token_weights(token_weight_pairs)
def encode_token_weights_l(model, token_weight_pairs):
l_out, _ = model.clip_l.encode_token_weights(token_weight_pairs)
return l_out, None
def encode_token_weights(model, token_weight_pairs, encode_func):
if model.layer_idx is not None:
if hasattr(model.cond_stage_model, 'set_clip_options'):
model.cond_stage_model.set_clip_options({"layer": model.layer_idx})
else:
print(f"[ComfyUI_ADV_CLIP_emb] ComfyUI is outdated.")
model.cond_stage_model.clip_layer(model.layer_idx)
model_management.load_model_gpu(model.patcher)
return encode_func(model.cond_stage_model, token_weight_pairs)
def prepareXL(embs_l, embs_g, pooled, clip_balance):
l_w = 1 - max(0, clip_balance - .5) * 2
g_w = 1 - max(0, .5 - clip_balance) * 2
if embs_l is not None:
return torch.cat([embs_l * l_w, embs_g * g_w], dim=-1), pooled
else:
return embs_g, pooled
def advanced_encode(clip, text, token_normalization, weight_interpretation, w_max=1.0, clip_balance=.5, apply_to_pooled=True):
tokenized = clip.tokenize(text, return_word_ids=True)
if isinstance(clip.cond_stage_model, (SDXLClipModel, SDXLRefinerClipModel, SDXLClipG)):
embs_l = None
embs_g = None
pooled = None
if 'l' in tokenized and isinstance(clip.cond_stage_model, SDXLClipModel):
embs_l, _ = advanced_encode_from_tokens(tokenized['l'],
token_normalization,
weight_interpretation,
lambda x: encode_token_weights(clip, x, encode_token_weights_l),
w_max=w_max,
return_pooled=False)
if 'g' in tokenized:
embs_g, pooled = advanced_encode_from_tokens(tokenized['g'],
token_normalization,
weight_interpretation,
lambda x: encode_token_weights(clip, x, encode_token_weights_g),
w_max=w_max,
return_pooled=True,
apply_to_pooled=apply_to_pooled)
return prepareXL(embs_l, embs_g, pooled, clip_balance)
else:
return advanced_encode_from_tokens(tokenized['l'],
token_normalization,
weight_interpretation,
lambda x: (clip.encode_from_tokens({'l': x}), None),
w_max=w_max)
def advanced_encode_XL(clip, text1, text2, token_normalization, weight_interpretation, w_max=1.0, clip_balance=.5, apply_to_pooled=True):
tokenized1 = clip.tokenize(text1, return_word_ids=True)
tokenized2 = clip.tokenize(text2, return_word_ids=True)
embs_l, _ = advanced_encode_from_tokens(tokenized1['l'],
token_normalization,
weight_interpretation,
lambda x: encode_token_weights(clip, x, encode_token_weights_l),
w_max=w_max,
return_pooled=False)
embs_g, pooled = advanced_encode_from_tokens(tokenized2['g'],
token_normalization,
weight_interpretation,
lambda x: encode_token_weights(clip, x, encode_token_weights_g),
w_max=w_max,
return_pooled=True,
apply_to_pooled=apply_to_pooled)
gcd_num = gcd(embs_l.shape[1], embs_g.shape[1])
repeat_l = int((embs_g.shape[1] / gcd_num) * embs_l.shape[1])
repeat_g = int((embs_l.shape[1] / gcd_num) * embs_g.shape[1])
return prepareXL(embs_l.expand((-1,repeat_l,-1)), embs_g.expand((-1,repeat_g,-1)), pooled, clip_balance)