xinyun/ComfyUI
1
0
Fork 0
mirror of https://git.datalinker.icu/comfyanonymous/ComfyUI synced 2026-11-12 02:52:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

images to video latents

Browse Source 
# Check if the number of frames needed matches the current frames
This commit is contained in:
u 2024-08-30 16:24:04 +08:00
parent 1dc10dcb16
commit f942ecae69
1 changed files with 19 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
comfy_extras/nodes_cogvidex_model.py
View File

@ -87,6 +87,7 @@ def prepare_extra_step_kwargs(
return extra_step_kwargs return extra_step_kwargs
def prepare_latents( def prepare_latents(
cogvideo_pipeline: CogVideoXPipeline, cogvideo_pipeline: CogVideoXPipeline,
batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
@ -211,6 +212,7 @@ def cogvideox_sampler(
cogvideo_pipeline._num_timesteps = len(timesteps) cogvideo_pipeline._num_timesteps = len(timesteps)
# 5. Prepare latents. # 5. Prepare latents.
current_frames = latents.shape[1] if latents is not None else (num_frames - 1) // cogvideo_pipeline.vae_scale_factor_temporal + 1
latent_channels = cogvideo_pipeline.transformer.config.in_channels latent_channels = cogvideo_pipeline.transformer.config.in_channels
latents = prepare_latents( latents = prepare_latents(
cogvideo_pipeline, cogvideo_pipeline,
@ -229,10 +231,24 @@ def cogvideox_sampler(
timesteps, num_inference_steps = get_timesteps(cogvideo_pipeline, num_inference_steps, denoise_strength, device) timesteps, num_inference_steps = get_timesteps(cogvideo_pipeline, num_inference_steps, denoise_strength, device)
latent_timestep = timesteps[:1] latent_timestep = timesteps[:1]
noise_latents = randn_tensor(latents.shape, generator=generator, device=device, dtype=cogvideo_pipeline.vae.dtype) # images to video latents
# Check if the number of frames needed matches the current frames
frames_shape = (
batch_size,
(num_frames - 1) // cogvideo_pipeline.vae_scale_factor_temporal + 1,
latent_channels,
height // cogvideo_pipeline.vae_scale_factor_spatial,
width // cogvideo_pipeline.vae_scale_factor_spatial,
)
noise_latents = randn_tensor(frames_shape, generator=generator, device=device, dtype=cogvideo_pipeline.vae.dtype)
frames_needed = frames_shape[1]
if frames_needed > current_frames:
repeat_factor = frames_needed - current_frames
additional_frame = torch.randn((latents.size(0), repeat_factor, latents.size(2), latents.size(3), latents.size(4)), dtype=latents.dtype, device=latents.device)
latents = torch.cat((latents, additional_frame), dim=1)
elif frames_needed < current_frames:
latents = latents[:, :frames_needed, :, :, :]
latents = cogvideo_pipeline.scheduler.add_noise(latents, noise_latents, latent_timestep) latents = cogvideo_pipeline.scheduler.add_noise(latents, noise_latents, latent_timestep)
latents = latents.to(cogvideo_pipeline.transformer.dtype) latents = latents.to(cogvideo_pipeline.transformer.dtype)
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline