fix: use original user audio directly in output video for accurate lip-sync (bypass audio VAE decoder); fix _write_audio mono-to-stereo auto-conversion to prevent ValueError on single-channel input

fix: wrap flash_attn_interface import in try/except (OSError/ImportError) - prevents libcudart.so.12 crash at module load time on ZeroGPU; add graceful FA3 fallback to PyTorch SDPA in DEFAULT attention

enhance

guide frame

now with audio support

Update packages/ltx-pipelines/src/ltx_pipelines/distilled.py

LoRA offloading

allow loRA

Upload 100 files

Delete packages

improvements + fixes (#1)

Update packages/ltx-pipelines/src/ltx_pipelines/distilled.py

inital commit