| --- |
| license: agpl-3.0 |
| tags: |
| - object-detection |
| - yolo |
| - yolo-world |
| - open-vocabulary |
| - onnx |
| pipeline_tag: object-detection |
| --- |
| |
| # YOLOv8-World v2 — ONNX with dynamic text prompts |
|
|
| ONNX exports of Ultralytics YOLOv8-World v2 (`yolov8s-worldv2.pt`, `yolov8l-worldv2.pt`) |
| with a **dynamic `txt_feats` input** so class prompts can be re-encoded at runtime |
| (zero-shot open-vocabulary detection), rather than being baked into the graph. |
| |
| ## Files |
| |
| | File | Size | Params | |
| |---|---|---| |
| | `yolov8s-worldv2.onnx` | 48.8 MB | 12.7 M | |
| | `yolov8l-worldv2.onnx` | 178.8 MB | 46.8 M | |
| |
| ## Input schema |
| |
| Two named inputs (all axes dynamic): |
| |
| - `images` — `float32[batch, 3, height, width]`, normalized to `[0, 1]`, RGB, imgsz=640 recommended. |
| - `txt_feats` — `float32[batch, num_classes, 512]`, CLIP ViT-B/32 text embeddings, |
| L2-normalized (though the graph re-normalizes internally). |
| |
| ## Output |
| |
| - `output0` — `float32[batch, 4 + num_classes, num_anchors]` |
| (YOLO head: 4 box regressors + per-class logits; anchors depend on input HxW). |
| |
| Post-processing: apply sigmoid to the class channels, then standard NMS. |
| |
| ## Text embedding |
| |
| Use CLIP ViT-B/32 (`openai/clip-vit-base-patch32`) to encode each class prompt |
| into a 512-dim vector, stack into `(1, N, 512)`. |
| |
| ## Export |
| |
| Exported from Ultralytics 8.4.37 with a custom `torch.onnx.export` path that |
| patches `WorldDetect.forward` to use `torch.split` sized from `text.shape[1]` |
| (so N stays dynamic). Opset 18. Simplified with onnxslim. |
| |
| ## License |
| |
| AGPL-3.0 (inherits from Ultralytics YOLOv8-World weights). |
| |
