How luca-saggese/hunyuan3d-2mini-turbo works
Project type and purpose
• It can be packaged in Docker for GPU-accelerated inference and loads pretrained models from Hugging Face.
File structure
• Top-level files:
• gradio_app.py / gradio_app_diffloading.py – entry points for the UI.
• text2image.py – auxiliary script for 2D image conditioning.
• Dockerfile / .dockerignore – containerization.
• requirements.txt – Python dependencies.
• Core library (hy3dgen):
• rembg.py – background removal component.
• texgen/ – all texture generation pipelines, including:
· custom_rasterizer (pure-Python wrappers + C++/CUDA kernels).
· differentiable_renderer (mesh processing + Python/C++ bridge).
· hunyuanpaint (UNet-based painting).
· pipelines.py – orchestration of texgen stages.
· utils/ – mesh/image utilities (alignment, simplification, UV warping, etc.).
• External model wheels, assets (env maps, example prompts, HTML templates).
README for additional insights
• Describes two key acceleration techniques: Lightning Vectset Decoder and Progressive Flow Distillation.
• Shows usage via Docker + Gradio and code snippets for enabling FlashVDM on Hunyuan3D-2 models.
• Lists supported models (Fast vs. Turbo, various size families).
Essential details to capture
• The two acceleration techniques inside the VDM flow.
• The interplay between Python orchestration and C++/CUDA kernels.
• Model loading from Hugging Face and caching.
• Docker/GPU boundary.