# HuggingFace ZeroGPU Space — Gradio SDK [cache-bust: 7] # spaces: pre-installed by HF ZeroGPU runtime; DO NOT reinstall from PyPI (breaks GPU proxy) numpy>=2 # Git-pinned installs # NOTE: hmr2, chumpy, skel, nvdiffrast, diso, detectron2 are installed at runtime # in app.py because: # - hmr2/skel declare `chumpy @ git+https://...` as a direct-ref dep; chumpy's # setup.py does `from pip._internal.req import parse_requirements` which fails # in pip>=21 isolated builds. Fix: --no-build-isolation at runtime. # - nvdiffrast/diso/detectron2 require nvcc (NVIDIA CUDA compiler) which is only # available in the ZeroGPU runtime container, not the Docker build stage. # triposg has no setup.py — cloned at runtime in app.py and added to sys.path clip @ git+https://github.com/openai/CLIP.git@d05afc436d78f1c48dc0dbf8e5980a9d471f35f6 # mvadapter and stablenormal use over-pinned deps (opencv-python vs headless, diffusers==0.28.0) # installed at runtime with --no-deps in app.py # Core ML accelerate diffusers>=0.37.0 transformers>=5.0.0 safetensors huggingface_hub peft einops timm xformers # 3D / Mesh trimesh fast-simplification # open3d removed: conflicts with numpy>=2; TexturePipeline is bypassed anyway # pymeshlab: no Python 3.13 wheels — not used in this Space pygltflib # pyrender: no Python 3.13 wheels — not used in this Space moderngl moderngl-window>=3.0.0 # Body model smplx smplpytorch # Pose / Motion ultralytics pyquaternion kornia # Face enhancement insightface onnxruntime-gpu # basicsr/realesrgan/gfpgan/facexlib/face-alignment: installed at runtime in app.py # with --no-build-isolation (their setup.py imports torch at build time, which fails # in pip's isolated build env where torch is not yet installed) # Surface enhancement # stablenormal: installed at runtime with --no-deps (it pins diffusers==0.28.0) controlnet_aux # CV opencv-python-headless scikit-image albumentations # Scientific numpy>=2 scipy scikit-learn pandas # Utils rembg icecream easydict omegaconf yacs gdown pycocotools jaxtyping typeguard sentencepiece spandrel imageio gradio_client # FireRed / GGUF quantization bitsandbytes