README.md

---
license: mit
tags:
  - onnx
  - image-to-3d
  - 3d-generation
  - rectified-flow
  - diffusion-transformer
  - qtmesheditor
library_name: onnx
pipeline_tag: image-to-3d
---

# TripoSG — ONNX export (staged graphs)

**ONNX re-export of [VAST-AI/TripoSG](https://huggingface.co/VAST-AI/TripoSG)**
(SIGGRAPH 2025 — MIT code + MIT weights): a 1.5B-parameter rectified-flow
diffusion transformer over an SDF VAE, single image → high-fidelity 3D
geometry (≈ commercial Tripo 2.0 quality). All credit for the original
weights goes to VAST-AI-Research.

Exported as **four staged graphs** so the flow loop runs in plain host code —
for **[QtMeshEditor](https://github.com/fernandotonon/QtMeshEditor)**
(`qtmesh generate3d --backend triposg`, GUI Backend dropdown, MCP `backend`
arg), local inference via ONNX Runtime + native marching cubes.

> The files QtMeshEditor downloads at runtime live in the shared
> [`fernandotonon/QtMeshEditor-models`](https://huggingface.co/fernandotonon/QtMeshEditor-models)
> repo under `triposg/`. This repo is the standalone model card + mirror.

## Files

| file | role |
|---|---|
| `triposg_image_encoder.onnx` | DINOv2-224 image encoder (mean/std baked in; CFG unconditional = zeros) |
| `triposg_dit_step.onnx` + `.data` | one DiT flow step (fp32, ~5.8 GB external weights) |
| `triposg_vae_latents.onnx` | VAE latent KV-cache graph — run **once** per generation |
| `triposg_vae_decoder.onnx` | per-point SDF field decoder |

An int8 DiT tier exists in the aggregate repo but is **not recommended**: even
per-channel-quantized, the 1.5B DiT degrades to blobs over the 25-step CFG
flow loop, and dynamic-int8 MatMuls are no faster than fp32 on ARM.

## Inference contract

- **Flow loop (host code):** `σᵢ = 1 − i/N`, timestep `1000·σᵢ`, update
  `x += (σᵢ − σᵢ₊₁)·v` — note the sign is **opposite** of stock diffusers
  `FlowMatchEulerDiscreteScheduler`. CFG as two batch-1 calls, guidance 7.0,
  25 steps default.
- Background removal should composite over **white** (TripoSG's reference
  pipeline), not gray.
- The exported field decoder is **outside-positive** — negate for
  inside-positive marching cubes at iso 0; query bounds ±1.005.
- Query the decoder in chunks of ≤ 8192 points (cross-attention to 2048 kv
  tokens; huge chunks OOM). Open/release the sessions per stage to keep peak
  memory ~1 GB instead of the >4 GB sum.
- Output is already +Y-up.
- **Geometry only** — TripoSG has no colour decoder. QtMeshEditor bakes colour
  by projecting the input photo onto the visible front and filling the rest
  from TripoSR's image-conditioned colour field (see
  [`QtMeshEditor-triposr-onnx`](https://huggingface.co/fernandotonon/QtMeshEditor-triposr-onnx)).

Full measured export contract: `docs/TRIPOSG_EXPORT_NOTES.md` in the
QtMeshEditor repo.

## Reproducing

`scripts/export-triposg-onnx.py` in the QtMeshEditor repo (one-time, offline).

## License

MIT (same as the upstream code and weights). Credit: VAST-AI-Research.