| license: mit | |
| tags: [document-layout, yolo26, onnx, doclaynet] | |
| # sc-layout-models | |
| ONNX exports of [Armaggheddon/yolo26-document-layout](https://huggingface.co/Armaggheddon/yolo26-document-layout) | |
| (MIT), used as the default layout detector by `sc_toolkit`'s hybrid PDF parser. | |
| Exported with `tools/export_layout_onnx.py` from the sc_toolkit repo: | |
| imgsz=1280, opset=17, simplified, `names` class metadata embedded. Served | |
| through `onnxruntime` (MIT); `ultralytics` is used at export time only and is | |
| not required to run these files. | |
| | file | variant | size | | |
| |---|---|---| | |
| | `yolo26m_doc_layout.onnx` | m — the sc_toolkit default | ~82 MB | | |
| | `yolo26n_doc_layout.onnx` | n — the cheap option (`HYBRID_LAYOUT_MODEL=superchaintech/sc-layout-models#yolo26n_doc_layout.onnx`) | ~10 MB | | |
| Known deviation from the source weights (measured, sc_toolkit repo, | |
| `docs/layout-model-comparison-2026-08.md`): the ONNX decode agrees with the | |
| ultralytics runtime on ~98% of regions at IoU≥0.5 and can differ on the rest. | |