| --- |
| license: mit |
| tags: |
| - onnx |
| - image-to-3d |
| - 3d-reconstruction |
| - triplane |
| - qtmesheditor |
| library_name: onnx |
| pipeline_tag: image-to-3d |
| --- |
| |
| # TripoSR — ONNX export |
|
|
| **ONNX re-export of [stabilityai/TripoSR](https://huggingface.co/stabilityai/TripoSR)** |
| (Tripo AI + Stability AI — MIT code + MIT weights): single-image → 3D |
| reconstruction via a triplane radiance field. All credit for the original |
| weights goes to Tripo AI and Stability AI. |
|
|
| Exported for **[QtMeshEditor](https://github.com/fernandotonon/QtMeshEditor)** |
| (epic #764), powering `qtmesh generate3d`, the Inspector's **AI: Image → 3D** |
| section, and the `generate_mesh_from_image` MCP tool — local inference via |
| ONNX Runtime, with surface extraction done by a native marching-cubes |
| implementation (upstream's `torchmcubes` is torch/GPU-only). |
|
|
| > The files QtMeshEditor downloads at runtime live in the shared |
| > [`fernandotonon/QtMeshEditor-models`](https://huggingface.co/fernandotonon/QtMeshEditor-models) |
| > repo under `triposr/`. This repo is the standalone model card + mirror. |
|
|
| ## Files |
|
|
| | file | role | |
| |---|---| |
| | `triposr_encoder.onnx` | fp32 encoder (best quality, ~1.7 GB): `image [1,3,512,512]` → `scene_codes [1,3,40,64,64]` (triplane) | |
| | `triposr_encoder_int8.onnx` | int8 encoder tier (~430 MB, slight quality loss) | |
| | `triposr_decoder.onnx` | per-point field decoder: `scene_codes` + `points [1,P,3]` → `density [1,P,1]`, `color [1,P,3]` | |
|
|
| An fp16 tier is deliberately absent: TripoSR's attention has a hardcoded |
| Cast-to-float32 that the ONNX fp16 converters cannot rewrite. |
|
|
| ## Inference contract |
|
|
| - Input image: 512×512 RGB in `[0,1]`, subject isolated (upstream composites |
| over gray-128 and re-pads the foreground to a 0.85 ratio; QtMeshEditor uses |
| U²-Net for background removal first). |
| - Query the decoder in **chunks** — generating the full `res³` grid up front |
| OOMs at high resolutions. |
| - Surface = marching cubes on `density − threshold` at iso 0, threshold 25.0, |
| within radius 0.87. The field is inside-positive; emit flipped winding |
| (`v0,v2,v1`) to keep faces outward. |
| - Output is not +Y-up: bake −90°X then +90°Y to stand the model upright |
| facing forward. |
| - The decoder's colour output can also be used as an image-conditioned colour |
| field for texture baking (per-texel queries after a UV unwrap). |
|
|
| ## Reproducing |
|
|
| `scripts/export-triposr-onnx.py` in the QtMeshEditor repo (one-time, offline; |
| `transformers==4.35.0`, `torchmcubes` stubbed, frozen ViT positional |
| encoding; emits the int8 variant unless `--no-quant`). |
|
|
| ## License |
|
|
| MIT (same as the upstream code and weights). Credit: Tripo AI + Stability AI. |
|
|
