Requirements¶
Requirements range from contract-only operation to full generative runs.
Software¶
| Layer | Requirement |
|---|---|
| Core runtime | Python 3.11 through 3.13 with the exact resolution recorded in uv.lock |
| Documentation | mkdocs from the docs extra |
| USD and material verification | usd-core and MaterialX from the validation extra |
| Mesh conditioning | trimesh plus networkx, scipy and fast-simplification from the mesh extra |
| Segmentation priors | numpy and opencv-python-headless from the vision extra; torch plus checkpoints for the separately gated SAM option |
| Isaac load gate | a local Isaac Sim installation, pointed at by AFB_ISAAC_SIM_ROOT |
| Backend installs | git, pip and for Linux-oriented backends a WSL or Linux host |
uv sync --frozen --all-extras installs the locked validation, mesh, vision, documentation and development toolchain. The dry-run pipeline, the agent loop in dry mode, the library and all contract checks run on CPU only. afb capabilities reports the options this machine can serve and plans installs for the gaps.
GPU expectations per reconstruction backend¶
| Backend | Model reference | Minimum VRAM |
|---|---|---|
| trellisv2 | microsoft/TRELLIS.2-4B | 24 GB |
| hunyuan3d | tencent/Hunyuan3D-2 | 16 GB |
| hunyuan3d-mv | tencent/Hunyuan3D-2mv | 16 GB |
| dust3r and video-multiview | naver/DUSt3R ViT-L 512 dpt | 12 GB |
| triposg | VAST-AI/TripoSG | 8 GB |
| partcrafter | wgsxm/PartCrafter | 8 GB |
Values mirror configs/reconstruction-backends.json, which is the machine-readable source of truth.
Model dependency table¶
| Role | Lane | Default model | Configured by |
|---|---|---|---|
| Planner and stage reasoning | nvidia_nim | nvidia/llama-3.1-nemotron-nano-8b-v1 | AFB_LLM_MODEL, NVIDIA_API_KEY |
| VLM sign-off reviewer | nvidia_nim | nvidia/llama-3.1-nemotron-nano-vl-8b-v1 | AFB_VISION_MODEL |
| Vision reasoning | openai | gpt-5.5 | AFB_VISION_MODEL, OPENAI_API_KEY |
| Texture image generation | openai, local_flux or hf_flux_schnell | FLUX.1 family on the local and hosted lanes | AFB_TEXTURE_MODEL, AFB_IMAGE_GENERATION_MODEL |
| Local OpenAI-compatible lane | local, openai_compatible | llama3.2:1b | AFB_LOCAL_MODEL, AFB_LOCAL_BASE_URL |
Provider lanes, key envs and defaults live in configs/provider-policy.json; the reviewer defaults live in configs/vlm-review-policy.json. Model weights for reconstruction backends carry their own licences, catalogued in THIRD_PARTY_NOTICES.md at the repository root.
Disk and network¶
- Backend checkouts and weights land under
.cache/afb/backends/(override withAFB_BACKEND_INSTALL_ROOT); budget 20 to 60 GB per generative backend with weights. - Library downloads land under
library/downloads/, https only, capped at 512 MB per file. - Project workspaces grow with render and texture evidence; keep
projects/on fast local storage.