We're open-sourcing the first publicly available blood detection model: dataset, weights, and CLI [P] [R]

BloodshotNet marks a significant milestone in content moderation as the first publicly available open-source model for detecting blood in images and videos. Developed by a Reddit user, it addresses the critical need for front-line filtering in Trust & Safety, protecting users and human reviewers from graphic content. The release includes a comprehensive dataset of over 23,000 annotated images sourced from forensic scenes, UFC footage, horror/gore movies, and surgical content, with a hard-negative slice to minimize false positives. Model weights are available in YOLO26 small and nano variants under the AGPL-3.0 license, alongside a CLI that enables analysis of single images, folders, or videos with just two lines of setup via uv. The small model delivers approximately 0.8 precision and 0.6 recall, running at over 40 frames per second even on CPU, making it practical for real-time video moderation.

The development process revealed key insights: while recall appears modest, the model performs effectively in video contexts, reliably catching blood in high-contrast action and gore scenes. For borderline cases, a sliding window over 5-10 second clips provides robust scene-level detection without requiring per-frame perfection. Interestingly, open-vocabulary text-prompt models like YOLO-E struggled significantly, likely due to filtered training data and blood's irregular patterns that resist text-based description. YOLO26 with ProgLoss and STAL training techniques proved superior, particularly for detecting small objects like tiny droplets. While transformer architectures could theoretically handle fluid dynamics better, annotated video datasets for blood detection are virtually nonexistent, making YOLO26 the pragmatic choice. Future plans include expanding the dataset with more cinematic content, training a medium variant, and developing OpenVINO INT8 exports for faster edge inference.