Ultralytics v8.4.7 is out! (AVIF training + COCO12-Formats) 
Ultralytics v8.4.7 adds AVIF training support 
and introduces a new COCO12-Formats mini-dataset to continuously verify that all supported image formats load end-to-end (especially in CI) 
. Check the full release details in the v8.4.7 GitHub release.
New Features
AVIF training support 
You can now train directly on .avif images—no manual conversion needed 
This expands IMG_FORMATS to include avif via PR #23358 by @glenn-jocher.
More robust decoding for modern formats 
Adds a Pillow-based fallback image reader (_imread_pil) for cases where OpenCV can’t decode AVIF/HEIC reliably across environments 
Included in PR #23358 by @glenn-jocher.
New COCO12-Formats mini-dataset + generator 
A tiny 12-image dataset designed to cover every supported image extension, helping catch format-specific failures early 
Includes YAML + generator script generate_coco12_formats.py, plus docs + dataset index updates 
Shipped in PR #23358 by @glenn-jocher.
Improvements
Training reliability (OBB + Pose26 multi-scale) 
Fixes loss image-size calculation for OBB and Pose26 multi-scale training in PR #23342 by @lmycross.
Cleaner run outputs + better W&B traceability 
Routes outputs under runs/ by default (better for relative projects) and improves W&B run tracking in PR #23026 by @artest08.
More accurate optimizer group logging 
Improves optimizer param-group print formatting in PR #23314 by @Laughing-q.
Dataset download resilience 
Adds retry + validation for NDSON image downloads in PR #23354 by @glenn-jocher.
Multi-scale sampling + docs clarity 
Fixes multi_scale range sampling and clarifies docs in PR #23284 by @artest08.
Bug Fixes
Semantic segmentation masking/indexing 
Fixes semantic segmentation loss indexing issues with overlapping instances in PR #23330 by @lmycross.
Export stability: torch pin for ExecuTorch 
Pins torch < 2.10 to avoid ExecuTorch export incompatibilities in PR #23353 by @onuralpszr.
Docs Updates (big) 
- Fix docs images content disposition in PR #23322 by @glenn-jocher
- Docs images cleanup in PR #23323 by @glenn-jocher
- Ultralytics Platform docs updates in PR #23338 by @glenn-jocher
New Contributors
Welcome @artest08 on their first contribution: PR #23026
Try it out
Update and start training (including AVIF images) in minutes:
pip install -U ultralytics
from ultralytics import YOLO
model = YOLO("yolo26n.pt") # recommended for new projects
model.train(data="coco12-formats.yaml", imgsz=640)
Feedback wanted
If you hit any edge cases with AVIF/HEIC decoding, dataset loading, exports, or CI consistency, please share details (OS, Python, opencv-python version, and a minimal repro). Your reports help keep Ultralytics YOLO solid across environments 
For the complete list of changes, see the v8.4.7 compare view.