Some consistency-related fixes for HunyuanVideo #10835
Conversation
a-r-r-o-w
changed the title
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| latents = self.prepare_latents( | ||
| batch_size * num_videos_per_prompt, | ||
| num_channels_latents, | ||
| height, | ||
| width, | ||
| num_latent_frames, | ||
| num_frames, |
There was a problem hiding this comment.
is this change just for make code is less confusing? (we were passing latents_frame as num_frames before)
There was a problem hiding this comment.
So, in prepare_latents, we calculate the latent height and latent width from the actual height/width BUT expect the num_frames to mean number of latent frames.
With this change, we make the interface uniform. prepare_latents now takes pixel-space width/height/num_frames and performs the latent related calculations in itself. We need to make these interfaces uniform a bit otherwise it leads to inconvenient behavior, so I will try to keep this behaviour consistent across all new implementations and gradually introduce changes to other pipelines
| latents = self.prepare_latents( | ||
| batch_size * num_videos_per_prompt, | ||
| num_channels_latents, | ||
| height, | ||
| width, | ||
| num_latent_frames, | ||
| num_frames, |
No description provided.