README.md

---
license: other
license_name: per-model
license_link: https://github.com/fernandotonon/QtMeshEditor/blob/master/THIRD_PARTY_AI_MODELS.md
tags:
- onnx
- gguf
- pbr
- texture
- normal-map
- 3d
- rigging
- image-to-3d
- qtmesheditor
- qtmesh
- qtmesh-cloud
library_name: onnx
---

# QtMeshEditor — AI models

The models used by [QtMeshEditor](https://github.com/fernandotonon/QtMeshEditor)'s
AI-assisted authoring features. **This repo is what the app downloads from at
runtime** (each model on first use, then it runs locally/offline).

**Licenses are per model** — see the table and each dedicated repo. The
dedicated repos carry the full model cards (I/O contracts, provenance,
reproduction scripts) for anyone who wants the converted weights standalone.

| folder / files | feature | dedicated repo (full card) | license |
|---|---|---|---|
| `1x-PBRify_*.onnx` | PBR maps from albedo | [QtMeshEditor-pbrify-onnx](https://huggingface.co/fernandotonon/QtMeshEditor-pbrify-onnx) | CC0-1.0 |
| `RealESRGAN_x{2,4}plus.onnx` | texture upscaling | [QtMeshEditor-realesrgan-onnx](https://huggingface.co/fernandotonon/QtMeshEditor-realesrgan-onnx) | BSD-3-Clause |
| `unirig/` | auto-rig skeleton prediction | [QtMeshEditor-unirig-onnx](https://huggingface.co/fernandotonon/QtMeshEditor-unirig-onnx) | MIT |
| `skintokens/` | ML skin-weight prediction | [QtMeshEditor-skintokens-onnx](https://huggingface.co/fernandotonon/QtMeshEditor-skintokens-onnx) | MIT |
| `triposr/` | image → 3D (triplane) | [QtMeshEditor-triposr-onnx](https://huggingface.co/fernandotonon/QtMeshEditor-triposr-onnx) | MIT |
| `triposg/` | image → 3D (rectified-flow DiT) | [QtMeshEditor-triposg-onnx](https://huggingface.co/fernandotonon/QtMeshEditor-triposg-onnx) | MIT |
| `inbetween/rmib.onnx` | animation in-betweening (ours) | [QtMeshEditor-rmib-inbetween](https://huggingface.co/fernandotonon/QtMeshEditor-rmib-inbetween) | CC-BY-4.0 |
| `motion/` | text-to-motion + clip library (ours) | [QtMeshEditor-t2m](https://huggingface.co/fernandotonon/QtMeshEditor-t2m) | CC0-1.0 |
| `segment/meshseg.onnx` | mesh part segmentation (ours) | [QtMeshEditor-mesh-segmentation](https://huggingface.co/fernandotonon/QtMeshEditor-mesh-segmentation) | CC-BY-4.0 |
| `rembg/u2net.onnx` | background removal | [QtMeshEditor-u2net-onnx](https://huggingface.co/fernandotonon/QtMeshEditor-u2net-onnx) | Apache-2.0 |
| `caption/SmolVLM-500M-*.gguf` | image captioning | [QtMeshEditor-smolvlm-gguf](https://huggingface.co/fernandotonon/QtMeshEditor-smolvlm-gguf) | Apache-2.0 |

## PBR map synthesis

`1x-PBRify_NormalV3.onnx`, `1x-PBRify_RoughnessV2.onnx`, `1x-PBRify_Height.onnx`
generate tangent-space normal / roughness / height maps from a single albedo
texture. ONNX re-exports of the CC0 SPAN models from
**[Kim2091/PBRify_Remix](https://github.com/Kim2091/PBRify_Remix)** — all
credit to Kim2091. I/O: `1×3×H×W` float `[0,1]` → `1×3×H×W`, dynamic H/W.

## Texture upscaling

`RealESRGAN_x2plus.onnx`, `RealESRGAN_x4plus.onnx` — 2×/4× super-resolution.
ONNX re-exports of **Real-ESRGAN**
([xinntao](https://github.com/xinntao/Real-ESRGAN), BSD-3-Clause). Credit: xinntao.

## Auto-rig skeleton prediction (UniRig)

`unirig/{encoder,decoder,embed}.onnx` — autoregressive skeleton prediction for
unrigged meshes. ONNX re-export of the skeleton stage of
**[VAST-AI/UniRig](https://huggingface.co/VAST-AI/UniRig)** (SIGGRAPH 2025,
MIT code + weights). Credit: VAST-AI-Research.

## ML skin weights (SkinTokens / TokenRig)

`skintokens/` — five ONNX graphs + manifest; QtMeshEditor's **default
skinner**. ONNX re-export of **VAST-AI SkinTokens/TokenRig** (MIT code +
weights, Qwen3-0.6B backbone). `decoder.onnx.data` holds the LM weights as
external data (ORT can't parse the >1.6 GB single-file proto). Credit:
VAST-AI-Research.

## Image → 3D

- `triposr/` — **TripoSR** (Tripo AI + Stability AI, MIT): triplane encoder
  (fp32 + int8 tiers) + per-point density/colour decoder.
- `triposg/` — **TripoSG** (VAST-AI, SIGGRAPH 2025, MIT): DINOv2 image
  encoder, rectified-flow DiT step graph (fp32 external weights; the int8
  tier here is deprecated — it degrades to blobs over the CFG flow loop),
  VAE latent + field-decoder graphs. Geometry-only; colour comes from
  TripoSR's colour field.
- `rembg/u2net.onnx` — **U²-Net** saliency for background removal
  (Apache-2.0, the rembg model).

## Animation in-betweening (RMIB) — trained by us

`inbetween/rmib.onnx` — fills the gap between two keyframes. **Trained from
scratch** on the permissive
[CMU MoCap database](http://mocap.cs.cmu.edu); beats slerp by >2× on held-out
CMU motion. License: CC-BY-4.0.

## Text-to-motion — trained by us

`motion/t2m.onnx` + `motion/t2m-vocab.json` — **v8.0**, a flow-matching DiT
(21.7M params) mapping a text keyword to a 22-joint world-frame clip, over a
30-action vocab. Trained on the curated template clips in
`motion/motion-library-v2.json` (which is also the fallback for prompts outside
the vocab). v8.0 fixed the backwards-facing problem inherited from the CMU
corpus and a walk defect where the bend sat in the ankle rather than the knee;
`motion/t2m-v61.onnx` is kept for rollback. Full details and metrics:
[QtMeshEditor-t2m](https://huggingface.co/fernandotonon/QtMeshEditor-t2m).
License: CC0-1.0.

## Mesh part segmentation — trained by us

`segment/meshseg.onnx` — per-point head/torso/arm/leg labels
(PointNet++-style). Trained on synthetic bodies we own + CC0 rigged
characters (Quaternius). 94.7% per-vertex accuracy on rig-truth eval.
License: CC-BY-4.0.

## Image captioning

`caption/SmolVLM-500M-Instruct-Q8_0.gguf` + `mmproj` — quantized
**SmolVLM-500M-Instruct** (HuggingFaceTB, Apache-2.0) for llama.cpp-based
captioning. Credit: Hugging Face TB.

---

These models power the AI-assisted authoring features in
**[QtMeshEditor](https://github.com/fernandotonon/QtMeshEditor)** and its
companion **QtMesh Cloud** ([qtmesh.dev](https://qtmesh.dev)). Provenance and
licensing decisions are documented in the QtMeshEditor repo's
`THIRD_PARTY_AI_MODELS.md`.