# The exact set inference/sample.py was validated with: on 4 jobs its rendered frames match our reference
# worker's. The mp4 is not bit-repeatable even so: PyAV's default libx264 encode and the audio VAE's decoder
# vary run to run, so compare decoded frames before encoding, not files.
# diffusers is pinned to a commit rather than a release: the `minimax_h3` modular pipeline is newer
# than any tag. That commit has the plain `ref2va` workflow only; the keyframe pin the adapters use
# is added by sample.py itself, so no fork or patched install is needed.
# Also needs an `ffmpeg` binary on PATH (we used 4.4.2) to read the soundtrack.
torch==2.9.1
git+https://github.com/huggingface/diffusers@d6726f3
transformers==5.14.1
huggingface_hub==1.27.0
safetensors==0.8.0
peft==0.18.0
accelerate==1.12.0
tokenizers==0.22.1
av==16.1.0
numpy==2.3.5
Pillow==12.2.0