requirements.txt

# `diffusers` itself is not installed: the MiniMax-H3 integration branch's `src/diffusers` tree is vendored into this
# Space as a top-level `diffusers/` package and imported from the working directory (`sync.sh` re-syncs it). What
# follows is everything that tree imports. Nothing is quantized in this deployment, so there is no `torchao`.
--extra-index-url https://download.pytorch.org/whl/cu130
torch==2.11.0
torchvision==0.26.0
# A reference soundtrack that is not already at the audio VAE's 32 kHz is resampled with `torchaudio`, which is the
# only thing in the `ref2va` path that needs it — and it is easy to miss, because a 32 kHz reference skips the
# resample entirely. Both halves need it: the conditioner's `setup` step prepares the same waveforms this one does.
torchaudio==2.11.0
# Pinned to the version the MiniMax-H3 parity work was verified against: the Qwen3-VL processor decides the vision
# patch count, so a different minor changes the conditioning.
transformers==5.8.0
accelerate==1.14.0
# The vendored tree needs `get_cached_repo_tree`, which landed in 1.23; diffusers pins <2.
huggingface-hub==1.24.0
gradio==6.20.0
spaces==0.51.1
# No `kernels` pin on purpose: the Hub attention backends want `kernels>=0.12.3`, and that version breaks
# transformers 5.8.0 at import. cuDNN attention (`_native_cudnn`) is faster than the SDPA default and needs nothing.
# PyAV: muxing the generated soundtrack onto the frames (`encode_video`).
av
pillow
numpy
requests
safetensors>=0.8.0
filelock
regex
importlib_metadata