Foot Anime Yolo

--- license: agpl-3.0 library_name: ultralytics pipeline_tag: object-detection base_model: Ultralytics/YOLO11 datasets: - deepghs/danbooru2024 - deepghs/danbooru2024-sfw metrics: - mAP tags: - yolo - yolo11 - ultralytics - object-detection - anime - foot-detection - adetailer - comfyui model-index: - name: foot_anime_yolo11m_v3 results: - task: type: object-detection dataset: name: Anime foot held-out (generated, 100 images / 185 feet) type: custom metrics: - type: mAP50 value: 0.81 name: mAP@50 - type: mAP50-95 value: 0.59 name: mAP@50-95 --- # Anime foot detector (YOLO11m) A small YOLO11m detector that finds feet in anime and illustration images. It has a single class, `foot`. The main use is image generation cleanup: pair it with an ADetailer or Impact Pack workflow in ComfyUI so a detail or inpaint pass can fix feet, which diffusion models often render badly. It was built for the Anima text-to-image model, but it works on anime-style imagery in general, so it should transfer to other anime or illustration generators without retraining. Three versions are provided. They share the same architecture and differ only in training. `v3` is the current production model. ## Which one to use - `foot_anime_yolo11m_v3.pt` is the production model and the one to use. On a held-out test set it has the best box accuracy across every image type and the broadest coverage, including harder cases like stockings and unusual poses. - `foot_anime_yolo11m_v2.pt` is the previous production model, very close behind, and a little better at finding feet in plain, clearly visible shots. A good alternative. - `foot_anime_yolo11m_v1.pt` is the first and weakest version, kept for reference. Start with v3. If it misses a foot in a simple image, try v2. ## Models | file | size | input | role | |---|---|---|---| | `foot_anime_yolo11m_v1.pt` | 40 MB | 640 | seed (reference) | | `foot_anime_yolo11m_v2.pt` | 40 MB | 640 | previous production | | `foot_anime_yolo11m_v3.pt` | 40 MB | 640 | production (recommended) | All are YOLO11m (about 20M parameters), fine-tuned from the COCO-pretrained `yolo11m.pt`. ## Files and formats Each version ships in two formats: - `.pt` is the standard Ultralytics PyTorch checkpoint, and what ComfyUI and `YOLO(...)` load directly. It is a pickle, so the Hub lists its imports; those are the normal torch and Ultralytics i...