ONNX re-exports of the CC0 SPAN models from Kim2091/PBRify_Remix: predict a tangent-space normal map, roughness map, and height map from a single albedo/diffuse texture. All credit for the weights goes to Kim2091.

Exported for QtMeshEditor (issue #404), powering qtmesh material --generate-pbr, the Material Editor's "Generate PBR maps from diffuse" button, the generate_pbr_maps MCP tool, and the PBR stage of the image-to-3D pipeline.

1×3×H×W float NCHW in [0,1] → 1×3×H×W out. Dynamic H/W (tile large textures with overlapping tiles + feathered seam blending). Opset 18, ~1.6 MB each.

The upstream repo's LICENSE is CC0-1.0 and its README states the models were trained "exclusively on high quality CC0 content from ambientCG". (OpenModelDB's NormalV3 page lists "ambientCG + UltraSharpV2" as the training set — UltraSharpV2 is cc-by-nc-sa-4.0 — which contradicts the author's own statement; we treat the author's explicit repo LICENSE + "exclusively CC0" statement as authoritative. Documented for transparency.)

scripts/export-pbrify-onnx.py in the QtMeshEditor repo (one-time, offline; spandrel load → torch.onnx.export, opset 18, dynamic H/W).