Spaces:
Running
on
T4
Running
on
T4
DeepBeepMeep
commited on
Commit
·
03085c8
1
Parent(s):
e420cd0
optimization for i2v with CausVid
Browse files- hyvideo/modules/models.py +1 -2
- wan/image2video.py +23 -24
- wan/modules/model.py +3 -4
hyvideo/modules/models.py
CHANGED
@@ -492,8 +492,7 @@ class MMSingleStreamBlock(nn.Module):
|
|
492 |
return img, txt
|
493 |
|
494 |
class HYVideoDiffusionTransformer(ModelMixin, ConfigMixin):
|
495 |
-
|
496 |
-
def preprocess_loras(model_filename, sd):
|
497 |
if not "i2v" in model_filename:
|
498 |
return sd
|
499 |
new_sd = {}
|
|
|
492 |
return img, txt
|
493 |
|
494 |
class HYVideoDiffusionTransformer(ModelMixin, ConfigMixin):
|
495 |
+
def preprocess_loras(self, model_filename, sd):
|
|
|
496 |
if not "i2v" in model_filename:
|
497 |
return sd
|
498 |
new_sd = {}
|
wan/image2video.py
CHANGED
@@ -330,8 +330,11 @@ class WanI2V:
|
|
330 |
'current_step' :i,
|
331 |
})
|
332 |
|
333 |
-
|
334 |
-
|
|
|
|
|
|
|
335 |
if audio_proj == None:
|
336 |
noise_pred_cond, noise_pred_uncond = self.model(
|
337 |
[latent_model_input, latent_model_input],
|
@@ -347,13 +350,7 @@ class WanI2V:
|
|
347 |
if self._interrupt:
|
348 |
return None
|
349 |
else:
|
350 |
-
noise_pred_cond = self.model(
|
351 |
-
[latent_model_input],
|
352 |
-
context=[context],
|
353 |
-
audio_scale = None if audio_scale == None else [audio_scale],
|
354 |
-
x_id=0,
|
355 |
-
**kwargs,
|
356 |
-
)[0]
|
357 |
if self._interrupt:
|
358 |
return None
|
359 |
|
@@ -377,22 +374,24 @@ class WanI2V:
|
|
377 |
return None
|
378 |
del latent_model_input
|
379 |
|
380 |
-
|
381 |
-
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
|
388 |
-
|
389 |
-
|
|
|
|
|
|
|
|
|
|
|
390 |
else:
|
391 |
-
noise_pred_uncond
|
392 |
-
|
393 |
-
noise_pred = noise_pred_uncond + guide_scale * (noise_pred_cond - noise_pred_uncond)
|
394 |
-
else:
|
395 |
-
noise_pred = noise_pred_uncond + guide_scale * (noise_pred_noaudio - noise_pred_uncond) + audio_cfg_scale * (noise_pred_cond - noise_pred_noaudio)
|
396 |
noise_pred_uncond, noise_pred_noaudio = None, None
|
397 |
temp_x0 = sample_scheduler.step(
|
398 |
noise_pred.unsqueeze(0),
|
|
|
330 |
'current_step' :i,
|
331 |
})
|
332 |
|
333 |
+
if guide_scale == 1:
|
334 |
+
noise_pred = self.model( [latent_model_input], context=[context], audio_scale = None if audio_scale == None else [audio_scale], x_id=0, **kwargs, )[0]
|
335 |
+
if self._interrupt:
|
336 |
+
return None
|
337 |
+
elif joint_pass:
|
338 |
if audio_proj == None:
|
339 |
noise_pred_cond, noise_pred_uncond = self.model(
|
340 |
[latent_model_input, latent_model_input],
|
|
|
350 |
if self._interrupt:
|
351 |
return None
|
352 |
else:
|
353 |
+
noise_pred_cond = self.model( [latent_model_input], context=[context], audio_scale = None if audio_scale == None else [audio_scale], x_id=0, **kwargs, )[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
354 |
if self._interrupt:
|
355 |
return None
|
356 |
|
|
|
374 |
return None
|
375 |
del latent_model_input
|
376 |
|
377 |
+
if guide_scale > 1:
|
378 |
+
# CFG Zero *. Thanks to https://github.com/WeichenFan/CFG-Zero-star/
|
379 |
+
if cfg_star_switch:
|
380 |
+
positive_flat = noise_pred_cond.view(batch_size, -1)
|
381 |
+
negative_flat = noise_pred_uncond.view(batch_size, -1)
|
382 |
+
|
383 |
+
alpha = optimized_scale(positive_flat,negative_flat)
|
384 |
+
alpha = alpha.view(batch_size, 1, 1, 1)
|
385 |
+
|
386 |
+
if (i <= cfg_zero_step):
|
387 |
+
noise_pred = noise_pred_cond*0. # it would be faster not to compute noise_pred...
|
388 |
+
else:
|
389 |
+
noise_pred_uncond *= alpha
|
390 |
+
if audio_scale == None:
|
391 |
+
noise_pred = noise_pred_uncond + guide_scale * (noise_pred_cond - noise_pred_uncond)
|
392 |
else:
|
393 |
+
noise_pred = noise_pred_uncond + guide_scale * (noise_pred_noaudio - noise_pred_uncond) + audio_cfg_scale * (noise_pred_cond - noise_pred_noaudio)
|
394 |
+
|
|
|
|
|
|
|
395 |
noise_pred_uncond, noise_pred_noaudio = None, None
|
396 |
temp_x0 = sample_scheduler.step(
|
397 |
noise_pred.unsqueeze(0),
|
wan/modules/model.py
CHANGED
@@ -589,8 +589,7 @@ class MLPProj(torch.nn.Module):
|
|
589 |
|
590 |
|
591 |
class WanModel(ModelMixin, ConfigMixin):
|
592 |
-
|
593 |
-
def preprocess_loras(model_filename, sd):
|
594 |
|
595 |
first = next(iter(sd), None)
|
596 |
if first == None:
|
@@ -634,8 +633,8 @@ class WanModel(ModelMixin, ConfigMixin):
|
|
634 |
print(f"Lora alpha'{alpha_key}' is missing")
|
635 |
new_sd.update(new_alphas)
|
636 |
sd = new_sd
|
637 |
-
|
638 |
-
if
|
639 |
new_sd = {}
|
640 |
# convert loras for i2v to t2v
|
641 |
for k,v in sd.items():
|
|
|
589 |
|
590 |
|
591 |
class WanModel(ModelMixin, ConfigMixin):
|
592 |
+
def preprocess_loras(self, model_filename, sd):
|
|
|
593 |
|
594 |
first = next(iter(sd), None)
|
595 |
if first == None:
|
|
|
633 |
print(f"Lora alpha'{alpha_key}' is missing")
|
634 |
new_sd.update(new_alphas)
|
635 |
sd = new_sd
|
636 |
+
from wgp import test_class_i2v
|
637 |
+
if not test_class_i2v(model_filename):
|
638 |
new_sd = {}
|
639 |
# convert loras for i2v to t2v
|
640 |
for k,v in sd.items():
|