Repository navigation
LibreYOLO v1.8.0, Axelera Metis export, optical flow, document OCR, eleven new generators and tracking on exported models.
Highlights
- Axelera Metis export and runtime:
export(format="axelera")for 23 families, each size with a measured accuracy label against FP32 (#1045). - Optical flow: a new
flowtask with SEA-RAFT, RAFT, NeuFlow v2, WAFT, FlowSeek, DPFlow, UniMatch and MEMFOF (#1001). - Document OCR and layout: DeepSeek-OCR, PaddleOCR-VL, GLM-OCR, Falcon OCR and PP-DocLayoutV3 with reading order (#995, #996, #999, #1005, #1007).
- LibreGen: image to mesh (TripoSR, CRM, TRELLIS.2, Direct3D-S2), image to Gaussians (TripoSplat, TRELLIS, LGM), text to image (PRX, PRXPixel), and Speridlabs Iris-Z and Open Vision Banana for depth and perception (#993-#1049).
- More families: InstantHMR body meshes, RiO-DETRv2 oriented boxes, SAM 3.1, MoGe-3, Video Depth Anything, Falcon Perception, Qwen3-VL-Embedding.
- Tracking on every exported backend, plus a
libreyolo trackcommand (#1031). - Faster RF-DETR training, about 20% per epoch in our tests, with the optional
libreyolo[rfdetr-train]GPU matcher (#1044, #1052). - Grayscale training with
channels: 1in the dataset YAML (#1015).
Changed (read before upgrading)
results[0]returns the whole single-image result;metrics.box.mapand friends exist (#1039).SAMPLE_IMAGEis the Guggenheim Bilbao photo;parkour.jpgis no longer shipped (#1034).- Relative
path:in a dataset YAML resolves against the YAML's folder (#1040). - Fresh RF-DETR training uses upstream's loss weights;
early_stopping=Truenow means patience 10 (#1044). val()numbers change for several exported families after fixes (#1045).
Fixed
- RF-DETR training falls back to SciPy when torch-hungarian cannot build (#1052).
- LibreGen downloads a snapshot once under concurrent loads on Windows (#1050).
- Grayscale and RGB runs on different threads no longer mix their dataset channel count (#1055).
- Many more; see the full changelog.
Full changelog: https://github.com/LibreYOLO/libreyolo/blob/release/CHANGELOG.md
Docs: https://www.libreyolo.com/docs/start/upgrade
pip install --upgrade libreyolo