QtMeshEditor — AI models
The models used by 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 | CC0-1.0 |
RealESRGAN_x{2,4}plus.onnx | texture upscaling | QtMeshEditor-realesrgan-onnx | BSD-3-Clause |
unirig/ | auto-rig skeleton prediction | QtMeshEditor-unirig-onnx | MIT |
skintokens/ | ML skin-weight prediction | QtMeshEditor-skintokens-onnx | MIT |
triposr/ | image → 3D (triplane) | QtMeshEditor-triposr-onnx | MIT |
triposg/ | image → 3D (rectified-flow DiT) | QtMeshEditor-triposg-onnx | MIT |
inbetween/rmib.onnx | animation in-betweening (ours) | QtMeshEditor-rmib-inbetween | CC-BY-4.0 |
motion/ | text-to-motion + clip library (ours) | QtMeshEditor-t2m | CC0-1.0 |
segment/meshseg.onnx | mesh part segmentation (ours) | QtMeshEditor-mesh-segmentation | CC-BY-4.0 |
rembg/u2net.onnx | background removal | QtMeshEditor-u2net-onnx | Apache-2.0 |
caption/SmolVLM-500M-*.gguf | image captioning | 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 — 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, 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 (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; 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.
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 and its
companion QtMesh Cloud (qtmesh.dev). Provenance and
licensing decisions are documented in the QtMeshEditor repo's
THIRD_PARTY_AI_MODELS.md.