Spaces:
Sleeping
Sleeping
Update clip_slider_pipeline.py
Browse files- clip_slider_pipeline.py +153 -34
clip_slider_pipeline.py
CHANGED
@@ -33,16 +33,12 @@ class CLIPSlider:
|
|
33 |
|
34 |
def find_latent_direction(self,
|
35 |
target_word:str,
|
36 |
-
opposite:str
|
37 |
-
num_iterations: int = None):
|
38 |
|
39 |
# lets identify a latent direction by taking differences between opposites
|
40 |
# target_word = "happy"
|
41 |
# opposite = "sad"
|
42 |
-
|
43 |
-
iterations = num_iterations
|
44 |
-
else:
|
45 |
-
iterations = self.iterations
|
46 |
|
47 |
with torch.no_grad():
|
48 |
positives = []
|
@@ -78,8 +74,6 @@ class CLIPSlider:
|
|
78 |
only_pooler = False,
|
79 |
normalize_scales = False, # whether to normalize the scales when avg_diff_2nd is not None
|
80 |
correlation_weight_factor = 1.0,
|
81 |
-
avg_diff = None,
|
82 |
-
avg_diff_2nd = None,
|
83 |
**pipeline_kwargs
|
84 |
):
|
85 |
# if doing full sequence, [-0.3,0.3] work well, higher if correlation weighted is true
|
@@ -90,14 +84,14 @@ class CLIPSlider:
|
|
90 |
max_length=self.pipe.tokenizer.model_max_length).input_ids.cuda()
|
91 |
prompt_embeds = self.pipe.text_encoder(toks).last_hidden_state
|
92 |
|
93 |
-
if avg_diff_2nd and normalize_scales:
|
94 |
denominator = abs(scale) + abs(scale_2nd)
|
95 |
scale = scale / denominator
|
96 |
scale_2nd = scale_2nd / denominator
|
97 |
if only_pooler:
|
98 |
-
prompt_embeds[:, toks.argmax()] = prompt_embeds[:, toks.argmax()] + avg_diff * scale
|
99 |
-
if avg_diff_2nd:
|
100 |
-
prompt_embeds[:, toks.argmax()] += avg_diff_2nd * scale_2nd
|
101 |
else:
|
102 |
normed_prompt_embeds = prompt_embeds / prompt_embeds.norm(dim=-1, keepdim=True)
|
103 |
sims = normed_prompt_embeds[0] @ normed_prompt_embeds[0].T
|
@@ -109,9 +103,9 @@ class CLIPSlider:
|
|
109 |
|
110 |
# weights = torch.sigmoid((weights-0.5)*7)
|
111 |
prompt_embeds = prompt_embeds + (
|
112 |
-
weights * avg_diff[None, :].repeat(1, self.pipe.tokenizer.model_max_length, 1) * scale)
|
113 |
-
if avg_diff_2nd:
|
114 |
-
prompt_embeds += weights * avg_diff_2nd[None, :].repeat(1, self.pipe.tokenizer.model_max_length, 1) * scale_2nd
|
115 |
|
116 |
|
117 |
torch.manual_seed(seed)
|
@@ -354,22 +348,21 @@ class CLIPSliderXL_inv(CLIPSlider):
|
|
354 |
|
355 |
return images
|
356 |
|
357 |
-
|
358 |
-
class T5SliderFlux(CLIPSlider):
|
359 |
-
|
360 |
def find_latent_direction(self,
|
361 |
target_word:str,
|
362 |
opposite:str,
|
363 |
-
|
364 |
|
365 |
# lets identify a latent direction by taking differences between opposites
|
366 |
# target_word = "happy"
|
367 |
# opposite = "sad"
|
|
|
368 |
if num_iterations is not None:
|
369 |
iterations = num_iterations
|
370 |
else:
|
371 |
iterations = self.iterations
|
372 |
-
|
373 |
with torch.no_grad():
|
374 |
positives = []
|
375 |
negatives = []
|
@@ -378,6 +371,137 @@ class T5SliderFlux(CLIPSlider):
|
|
378 |
subject = random.choice(SUBJECTS)
|
379 |
pos_prompt = f"a {medium} of a {target_word} {subject}"
|
380 |
neg_prompt = f"a {medium} of a {opposite} {subject}"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
381 |
|
382 |
pos_toks = self.pipe.tokenizer_2(pos_prompt,
|
383 |
return_tensors="pt",
|
@@ -412,9 +536,7 @@ class T5SliderFlux(CLIPSlider):
|
|
412 |
seed = 15,
|
413 |
only_pooler = False,
|
414 |
normalize_scales = False,
|
415 |
-
correlation_weight_factor = 0
|
416 |
-
avg_diff = None,
|
417 |
-
avg_diff_2nd = None,
|
418 |
**pipeline_kwargs
|
419 |
):
|
420 |
# if doing full sequence, [-0.3,0.3] work well, higher if correlation weighted is true
|
@@ -453,34 +575,31 @@ class T5SliderFlux(CLIPSlider):
|
|
453 |
prompt_embeds = self.pipe.text_encoder_2(toks.to(self.device), output_hidden_states=False)[0]
|
454 |
dtype = self.pipe.text_encoder_2.dtype
|
455 |
prompt_embeds = prompt_embeds.to(dtype=dtype, device=self.device)
|
456 |
-
|
457 |
-
if avg_diff_2nd is not None and normalize_scales:
|
458 |
denominator = abs(scale) + abs(scale_2nd)
|
459 |
scale = scale / denominator
|
460 |
scale_2nd = scale_2nd / denominator
|
461 |
if only_pooler:
|
462 |
-
prompt_embeds[:, toks.argmax()] = prompt_embeds[:, toks.argmax()] + avg_diff * scale
|
463 |
-
if avg_diff_2nd
|
464 |
-
prompt_embeds[:, toks.argmax()] += avg_diff_2nd * scale_2nd
|
465 |
else:
|
466 |
normed_prompt_embeds = prompt_embeds / prompt_embeds.norm(dim=-1, keepdim=True)
|
467 |
sims = normed_prompt_embeds[0] @ normed_prompt_embeds[0].T
|
468 |
|
469 |
weights = sims[toks.argmax(), :][None, :, None].repeat(1, 1, prompt_embeds.shape[2])
|
470 |
-
print("weights", weights.shape)
|
471 |
|
472 |
standard_weights = torch.ones_like(weights)
|
473 |
|
474 |
weights = standard_weights + (weights - standard_weights) * correlation_weight_factor
|
475 |
prompt_embeds = prompt_embeds + (
|
476 |
-
weights * avg_diff * scale)
|
477 |
-
|
478 |
-
if avg_diff_2nd is not None:
|
479 |
prompt_embeds += (
|
480 |
-
weights * avg_diff_2nd * scale_2nd)
|
481 |
|
482 |
torch.manual_seed(seed)
|
483 |
images = self.pipe(prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds,
|
484 |
-
**pipeline_kwargs).images
|
485 |
|
486 |
return images
|
|
|
33 |
|
34 |
def find_latent_direction(self,
|
35 |
target_word:str,
|
36 |
+
opposite:str):
|
|
|
37 |
|
38 |
# lets identify a latent direction by taking differences between opposites
|
39 |
# target_word = "happy"
|
40 |
# opposite = "sad"
|
41 |
+
|
|
|
|
|
|
|
42 |
|
43 |
with torch.no_grad():
|
44 |
positives = []
|
|
|
74 |
only_pooler = False,
|
75 |
normalize_scales = False, # whether to normalize the scales when avg_diff_2nd is not None
|
76 |
correlation_weight_factor = 1.0,
|
|
|
|
|
77 |
**pipeline_kwargs
|
78 |
):
|
79 |
# if doing full sequence, [-0.3,0.3] work well, higher if correlation weighted is true
|
|
|
84 |
max_length=self.pipe.tokenizer.model_max_length).input_ids.cuda()
|
85 |
prompt_embeds = self.pipe.text_encoder(toks).last_hidden_state
|
86 |
|
87 |
+
if self.avg_diff_2nd and normalize_scales:
|
88 |
denominator = abs(scale) + abs(scale_2nd)
|
89 |
scale = scale / denominator
|
90 |
scale_2nd = scale_2nd / denominator
|
91 |
if only_pooler:
|
92 |
+
prompt_embeds[:, toks.argmax()] = prompt_embeds[:, toks.argmax()] + self.avg_diff * scale
|
93 |
+
if self.avg_diff_2nd:
|
94 |
+
prompt_embeds[:, toks.argmax()] += self.avg_diff_2nd * scale_2nd
|
95 |
else:
|
96 |
normed_prompt_embeds = prompt_embeds / prompt_embeds.norm(dim=-1, keepdim=True)
|
97 |
sims = normed_prompt_embeds[0] @ normed_prompt_embeds[0].T
|
|
|
103 |
|
104 |
# weights = torch.sigmoid((weights-0.5)*7)
|
105 |
prompt_embeds = prompt_embeds + (
|
106 |
+
weights * self.avg_diff[None, :].repeat(1, self.pipe.tokenizer.model_max_length, 1) * scale)
|
107 |
+
if self.avg_diff_2nd:
|
108 |
+
prompt_embeds += weights * self.avg_diff_2nd[None, :].repeat(1, self.pipe.tokenizer.model_max_length, 1) * scale_2nd
|
109 |
|
110 |
|
111 |
torch.manual_seed(seed)
|
|
|
348 |
|
349 |
return images
|
350 |
|
351 |
+
class CLIPSliderFlux(CLIPSlider):
|
|
|
|
|
352 |
def find_latent_direction(self,
|
353 |
target_word:str,
|
354 |
opposite:str,
|
355 |
+
num_iterations: int = None):
|
356 |
|
357 |
# lets identify a latent direction by taking differences between opposites
|
358 |
# target_word = "happy"
|
359 |
# opposite = "sad"
|
360 |
+
|
361 |
if num_iterations is not None:
|
362 |
iterations = num_iterations
|
363 |
else:
|
364 |
iterations = self.iterations
|
365 |
+
|
366 |
with torch.no_grad():
|
367 |
positives = []
|
368 |
negatives = []
|
|
|
371 |
subject = random.choice(SUBJECTS)
|
372 |
pos_prompt = f"a {medium} of a {target_word} {subject}"
|
373 |
neg_prompt = f"a {medium} of a {opposite} {subject}"
|
374 |
+
pos_toks = self.pipe.tokenizer(pos_prompt,
|
375 |
+
padding="max_length",
|
376 |
+
max_length=self.pipe.tokenizer_max_length,
|
377 |
+
truncation=True,
|
378 |
+
return_overflowing_tokens=False,
|
379 |
+
return_length=False,
|
380 |
+
return_tensors="pt",).input_ids.cuda()
|
381 |
+
neg_toks = self.pipe.tokenizer(neg_prompt,
|
382 |
+
padding="max_length",
|
383 |
+
max_length=self.pipe.tokenizer_max_length,
|
384 |
+
truncation=True,
|
385 |
+
return_overflowing_tokens=False,
|
386 |
+
return_length=False,
|
387 |
+
return_tensors="pt",).input_ids.cuda()
|
388 |
+
pos = self.pipe.text_encoder(pos_toks).pooler_output
|
389 |
+
neg = self.pipe.text_encoder(neg_toks).pooler_output
|
390 |
+
positives.append(pos)
|
391 |
+
negatives.append(neg)
|
392 |
+
|
393 |
+
positives = torch.cat(positives, dim=0)
|
394 |
+
negatives = torch.cat(negatives, dim=0)
|
395 |
+
|
396 |
+
diffs = positives - negatives
|
397 |
+
|
398 |
+
avg_diff = diffs.mean(0, keepdim=True)
|
399 |
+
return avg_diff
|
400 |
+
|
401 |
+
def generate(self,
|
402 |
+
prompt = "a photo of a house",
|
403 |
+
scale = 2,
|
404 |
+
scale_2nd = 2,
|
405 |
+
seed = 15,
|
406 |
+
normalize_scales = False,
|
407 |
+
avg_diff = None,
|
408 |
+
avg_diff_2nd = None,
|
409 |
+
**pipeline_kwargs
|
410 |
+
):
|
411 |
+
# if doing full sequence, [-0.3,0.3] work well, higher if correlation weighted is true
|
412 |
+
# if pooler token only [-4,4] work well
|
413 |
+
|
414 |
+
with torch.no_grad():
|
415 |
+
text_inputs = self.pipe.tokenizer(
|
416 |
+
prompt,
|
417 |
+
padding="max_length",
|
418 |
+
max_length=77,
|
419 |
+
truncation=True,
|
420 |
+
return_overflowing_tokens=False,
|
421 |
+
return_length=False,
|
422 |
+
return_tensors="pt",
|
423 |
+
)
|
424 |
+
|
425 |
+
text_input_ids = text_inputs.input_ids
|
426 |
+
prompt_embeds = self.pipe.text_encoder(text_input_ids.to(self.device), output_hidden_states=False)
|
427 |
+
|
428 |
+
# Use pooled output of CLIPTextModel
|
429 |
+
prompt_embeds = prompt_embeds.pooler_output
|
430 |
+
pooled_prompt_embeds = prompt_embeds.to(dtype=self.pipe.text_encoder.dtype, device=self.device)
|
431 |
+
|
432 |
+
# Use pooled output of CLIPTextModel
|
433 |
+
|
434 |
+
text_inputs = self.pipe.tokenizer_2(
|
435 |
+
prompt,
|
436 |
+
padding="max_length",
|
437 |
+
max_length=512,
|
438 |
+
truncation=True,
|
439 |
+
return_length=False,
|
440 |
+
return_overflowing_tokens=False,
|
441 |
+
return_tensors="pt",
|
442 |
+
)
|
443 |
+
toks = text_inputs.input_ids
|
444 |
+
prompt_embeds = self.pipe.text_encoder_2(toks.to(self.device), output_hidden_states=False)[0]
|
445 |
+
dtype = self.pipe.text_encoder_2.dtype
|
446 |
+
prompt_embeds = prompt_embeds.to(dtype=dtype, device=self.device)
|
447 |
+
if avg_diff_2nd is not None and normalize_scales:
|
448 |
+
denominator = abs(scale) + abs(scale_2nd)
|
449 |
+
scale = scale / denominator
|
450 |
+
scale_2nd = scale_2nd / denominator
|
451 |
+
|
452 |
+
pooled_prompt_embeds = pooled_prompt_embeds + avg_diff * scale
|
453 |
+
if avg_diff_2nd is not None:
|
454 |
+
pooled_prompt_embeds += avg_diff_2nd * scale_2nd
|
455 |
+
|
456 |
+
torch.manual_seed(seed)
|
457 |
+
images = self.pipe(prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds,
|
458 |
+
**pipeline_kwargs).images
|
459 |
+
|
460 |
+
return images
|
461 |
+
|
462 |
+
def spectrum(self,
|
463 |
+
prompt="a photo of a house",
|
464 |
+
low_scale=-2,
|
465 |
+
low_scale_2nd=-2,
|
466 |
+
high_scale=2,
|
467 |
+
high_scale_2nd=2,
|
468 |
+
steps=5,
|
469 |
+
seed=15,
|
470 |
+
normalize_scales=False,
|
471 |
+
**pipeline_kwargs
|
472 |
+
):
|
473 |
+
|
474 |
+
images = []
|
475 |
+
for i in range(steps):
|
476 |
+
scale = low_scale + (high_scale - low_scale) * i / (steps - 1)
|
477 |
+
scale_2nd = low_scale_2nd + (high_scale_2nd - low_scale_2nd) * i / (steps - 1)
|
478 |
+
image = self.generate(prompt, scale, scale_2nd, seed, normalize_scales, **pipeline_kwargs)
|
479 |
+
images.append(image[0].resize((512,512)))
|
480 |
+
|
481 |
+
canvas = Image.new('RGB', (640 * steps, 640))
|
482 |
+
for i, im in enumerate(images):
|
483 |
+
canvas.paste(im, (640 * i, 0))
|
484 |
+
|
485 |
+
return canvas
|
486 |
+
class T5SliderFlux(CLIPSlider):
|
487 |
+
|
488 |
+
def find_latent_direction(self,
|
489 |
+
target_word:str,
|
490 |
+
opposite:str):
|
491 |
+
|
492 |
+
# lets identify a latent direction by taking differences between opposites
|
493 |
+
# target_word = "happy"
|
494 |
+
# opposite = "sad"
|
495 |
+
|
496 |
+
|
497 |
+
with torch.no_grad():
|
498 |
+
positives = []
|
499 |
+
negatives = []
|
500 |
+
for i in tqdm(range(self.iterations)):
|
501 |
+
medium = random.choice(MEDIUMS)
|
502 |
+
subject = random.choice(SUBJECTS)
|
503 |
+
pos_prompt = f"a {medium} of a {target_word} {subject}"
|
504 |
+
neg_prompt = f"a {medium} of a {opposite} {subject}"
|
505 |
|
506 |
pos_toks = self.pipe.tokenizer_2(pos_prompt,
|
507 |
return_tensors="pt",
|
|
|
536 |
seed = 15,
|
537 |
only_pooler = False,
|
538 |
normalize_scales = False,
|
539 |
+
correlation_weight_factor = 1.0,
|
|
|
|
|
540 |
**pipeline_kwargs
|
541 |
):
|
542 |
# if doing full sequence, [-0.3,0.3] work well, higher if correlation weighted is true
|
|
|
575 |
prompt_embeds = self.pipe.text_encoder_2(toks.to(self.device), output_hidden_states=False)[0]
|
576 |
dtype = self.pipe.text_encoder_2.dtype
|
577 |
prompt_embeds = prompt_embeds.to(dtype=dtype, device=self.device)
|
578 |
+
if self.avg_diff_2nd and normalize_scales:
|
|
|
579 |
denominator = abs(scale) + abs(scale_2nd)
|
580 |
scale = scale / denominator
|
581 |
scale_2nd = scale_2nd / denominator
|
582 |
if only_pooler:
|
583 |
+
prompt_embeds[:, toks.argmax()] = prompt_embeds[:, toks.argmax()] + self.avg_diff * scale
|
584 |
+
if self.avg_diff_2nd:
|
585 |
+
prompt_embeds[:, toks.argmax()] += self.avg_diff_2nd * scale_2nd
|
586 |
else:
|
587 |
normed_prompt_embeds = prompt_embeds / prompt_embeds.norm(dim=-1, keepdim=True)
|
588 |
sims = normed_prompt_embeds[0] @ normed_prompt_embeds[0].T
|
589 |
|
590 |
weights = sims[toks.argmax(), :][None, :, None].repeat(1, 1, prompt_embeds.shape[2])
|
|
|
591 |
|
592 |
standard_weights = torch.ones_like(weights)
|
593 |
|
594 |
weights = standard_weights + (weights - standard_weights) * correlation_weight_factor
|
595 |
prompt_embeds = prompt_embeds + (
|
596 |
+
weights * self.avg_diff * scale)
|
597 |
+
if self.avg_diff_2nd:
|
|
|
598 |
prompt_embeds += (
|
599 |
+
weights * self.avg_diff_2nd * scale_2nd)
|
600 |
|
601 |
torch.manual_seed(seed)
|
602 |
images = self.pipe(prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds,
|
603 |
+
**pipeline_kwargs).images
|
604 |
|
605 |
return images
|