daniel/h3-blackwell-runtime
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Return native H3 audio latents

Browse source
This commit is contained in:
Daniel Maddern 2026-08-14 13:57:30 +07:00
parent 59b8302d4e
commit fea69673cd
1 changed files with 6 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
src/h3_blackwell_runtime/sampler.py
View file

@ -52,6 +52,11 @@ def _model_sigma(video_sigma: torch.Tensor) -> torch.Tensor:
return (video_sigma * 1000.0).float() / 1000.0
def _decode_audio_latent(audio_carried: torch.Tensor, *, shift_video: float = 12.0, shift_audio: float = 3.0) -> torch.Tensor:
"""Convert Comfy's AV sampler-carried audio state back to native audio VAE latents."""
return audio_carried * (shift_audio / shift_video)
@torch.inference_mode()
def sample_video_res_multistep(
model,
@ -102,7 +107,7 @@ def sample_video_res_multistep(
f"{time.perf_counter() - step_started:.1f}s, elapsed {elapsed:.1f}s, eta {eta:.1f}s",
flush=True,
)
return (video, audio_carried) if return_audio else video
return (video, _decode_audio_latent(audio_carried)) if return_audio else video
@torch.inference_mode()