Update pipeline.py
Browse files- pipeline.py +6 -7
pipeline.py
CHANGED
@@ -518,9 +518,7 @@ class AnimateDiffPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdap
|
|
518 |
)
|
519 |
|
520 |
# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
|
521 |
-
def prepare_latents(
|
522 |
-
self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None, smooth_weight=0.5
|
523 |
-
):
|
524 |
shape = (
|
525 |
batch_size,
|
526 |
num_channels_latents,
|
@@ -543,9 +541,7 @@ class AnimateDiffPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdap
|
|
543 |
latents = latents * self.scheduler.init_noise_sigma
|
544 |
return latents
|
545 |
|
546 |
-
def prepare_latents_consistent(
|
547 |
-
self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
|
548 |
-
):
|
549 |
shape = (
|
550 |
batch_size,
|
551 |
num_channels_latents,
|
@@ -563,7 +559,6 @@ class AnimateDiffPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdap
|
|
563 |
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
564 |
|
565 |
# blend each frame with the surrounding N frames making sure to wrap around at the end
|
566 |
-
smooth_steps = 3
|
567 |
for i in range(num_frames):
|
568 |
blended_latent = torch.zeros_like(latents[:, :, i])
|
569 |
for s in range(-smooth_steps, smooth_steps + 1):
|
@@ -817,6 +812,8 @@ class AnimateDiffPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdap
|
|
817 |
init_image_strength: Optional[float] = 1.0,
|
818 |
init_noise_correlation: Optional[float] = 0.0,
|
819 |
latent_mode: Optional[str] = "normal",
|
|
|
|
|
820 |
):
|
821 |
r"""
|
822 |
The call function to the pipeline for generation.
|
@@ -997,6 +994,8 @@ class AnimateDiffPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdap
|
|
997 |
device,
|
998 |
generator,
|
999 |
latents,
|
|
|
|
|
1000 |
)
|
1001 |
|
1002 |
|
|
|
518 |
)
|
519 |
|
520 |
# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
|
521 |
+
def prepare_latents(self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None):
|
|
|
|
|
522 |
shape = (
|
523 |
batch_size,
|
524 |
num_channels_latents,
|
|
|
541 |
latents = latents * self.scheduler.init_noise_sigma
|
542 |
return latents
|
543 |
|
544 |
+
def prepare_latents_consistent(self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None,smooth_weight=0.5,smooth_steps=3):
|
|
|
|
|
545 |
shape = (
|
546 |
batch_size,
|
547 |
num_channels_latents,
|
|
|
559 |
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
560 |
|
561 |
# blend each frame with the surrounding N frames making sure to wrap around at the end
|
|
|
562 |
for i in range(num_frames):
|
563 |
blended_latent = torch.zeros_like(latents[:, :, i])
|
564 |
for s in range(-smooth_steps, smooth_steps + 1):
|
|
|
812 |
init_image_strength: Optional[float] = 1.0,
|
813 |
init_noise_correlation: Optional[float] = 0.0,
|
814 |
latent_mode: Optional[str] = "normal",
|
815 |
+
smooth_weight: Optional[float] = 0.5,
|
816 |
+
smooth_steps: Optional[int] = 3,
|
817 |
):
|
818 |
r"""
|
819 |
The call function to the pipeline for generation.
|
|
|
994 |
device,
|
995 |
generator,
|
996 |
latents,
|
997 |
+
smooth_weight,
|
998 |
+
smooth_steps,
|
999 |
)
|
1000 |
|
1001 |
|