# `diffusers` from **main**, pinned to a commit: the MiniMax-H3 integration and its LoRA mixin are both merged, and
# this Space's blocks subclass the merged block classes. `c401daa14` is the commit the inpainting blocks were
# verified against — bitwise against stock `ref2va` on an all-ones mask, and on real weights end to end.
--extra-index-url https://download.pytorch.org/whl/cu130
diffusers @ git+https://github.com/huggingface/diffusers.git@c401daa1428b386d487b8c69bab147a094ea3c87
torch==2.11.0
torchvision==0.26.0
# A source soundtrack that is not already at the audio VAE's 32 kHz is resampled with `torchaudio`. Inpainting hits
# this on nearly every request, since camera and phone audio is 44.1/48 kHz.
torchaudio==2.11.0
# 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 few-step LoRA is a PEFT adapter. Without this, `load_lora_weights` raises ModuleNotFoundError at startup, the
# Space logs one line about it and then quietly runs the base model at whatever step count was asked for — which is
# how "8 steps · turbo" ran for a while as 8 steps of no turbo at all.
peft
# diffusers pins <2.
huggingface-hub==1.24.0
# SAM 3's mask post-processing — NMS, hole filling, sprinkle removal — is skipped entirely without `kernels`, and the
# masks are visibly rougher for it. 0.14.1 coexists with transformers 5.8.0 (verified on the box, where the two
# together produced the clean tracks); this Space selects its attention backend by name rather than through a Hub
# kernel, so the incompatibility that kept earlier siblings unpinned does not apply here.
kernels==0.14.1
# PyAV decodes the source clip, the mask clip and the reference media.
av
# Optional pose conversion for a motion reference. mediapipe needs `protobuf<5`, which is why it is worth pinning
# nothing else against protobuf here; the Space reports at startup whether it resolved to a usable combination.
# DWPose was the alternative and is not viable: 43 s per frame on CPU onnxruntime.
mediapipe
# Dilating a segmentation by tens of pixels, per frame. `cv2.dilate` with an ellipse kernel is one call where PIL
# would need a loop of MaxFilters.
opencv-python-headless
pillow
numpy
requests
safetensors>=0.8.0