AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-27 17:22:53 +03:00
Code Activity

Fix bug when calculating audio RoPE coords

Browse Source
This commit is contained in:
Daniel Gu
2026-01-07 12:14:25 +01:00
parent 249ae1f853
commit 040c1188d9
3 changed files with 3 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/diffusers/models/transformers/transformer_ltx2.py
View File

@@ -761,11 +761,9 @@ class LTX2AudioVideoRotaryPosEmbed(nn.Module):
"""
# 1. Generate coordinates in the frame (time) dimension.
audio_duration_s = num_frames / fps
latent_frames = int(audio_duration_s * self.audio_latents_per_second)
# Always compute rope in fp32
grid_f = torch.arange(
start=shift, end=latent_frames + shift, step=self.patch_size_t, dtype=torch.float32, device=device
start=shift, end=num_frames + shift, step=self.patch_size_t, dtype=torch.float32, device=device
)
# 2. Calculate start timstamps in seconds with respect to the original spectrogram grid

2
src/diffusers/pipelines/ltx2/pipeline_ltx2.py
View File

@@ -689,7 +689,7 @@ class LTX2Pipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMix
latents_per_second = (
float(sampling_rate) / float(hop_length) / float(self.audio_vae_temporal_compression_ratio)
)
latent_length = int(duration_s * latents_per_second)
latent_length = round(duration_s * latents_per_second)
if latents is not None:
return latents.to(device=device, dtype=dtype), latent_length

2
src/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py
View File

@@ -749,7 +749,7 @@ class LTX2ImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoL
latents_per_second = (
float(sampling_rate) / float(hop_length) / float(self.audio_vae_temporal_compression_ratio)
)
latent_length = int(duration_s * latents_per_second)
latent_length = round(duration_s * latents_per_second)
if latents is not None:
return latents.to(device=device, dtype=dtype), latent_length