WeDetect-tiny — browser ONNX (int8)
Browser-ready int8 ONNX export of WeDetect-tiny, an open-vocabulary object detector
(text-prompted boxes). Two towers run client-side via onnxruntime-web (WebGPU, WASM fallback):
| file | size | role |
|---|---|---|
wedetect_tiny_language_int8.onnx |
~279 MB | XLM-R text encoder → prompt features (once per prompt set) |
wedetect_tiny_vision_int8.onnx |
~38 MB | image encoder + detection head (per image) |
Pipeline: prompts → XLM-R tokenizer (Xenova/xlm-roberta-base) → language.onnx → text features;
image → letterbox 640 → vision.onnx → decoded xyxy + sigmoid scores → score filter + class-agnostic
NMS + rescale.
License — GPL-3.0
WeDetect is GPL-3.0. These ONNX weights are a derived export of the upstream GPL model and are therefore distributed under GPL-3.0-only. Using them keeps you within GPL terms.
Upstream / attribution: WeDetect by WeChatCV — https://github.com/WeChatCV/WeDetect
(GPL-3.0). This repository redistributes an ONNX/int8 export of that model with no change to its
license. Corresponding source (model definition + training/export pipeline) lives in the upstream
project; the browser export/decoder is in the momoto-plugins repo (also GPL-3.0):
https://github.com/sweetlhare/momoto-plugins (src/copyleft/wedetect.js).
How it's used
Loaded at runtime by the momoto WeDetect plugin (momoto-plugins/src/copyleft/wedetect.js),
which points meta.assets at the two ONNX URLs here. The plugin is opt-in and never bundled into a
closed build — GPL bytes are fetched from this public GPL host at the user's choice, keeping the
closed core clear of copyleft. Self-hosters can override the asset URLs.
Resolve URLs (once uploaded):
https://huggingface.co/sweetlhare/wedetect-tiny-onnx/resolve/main/wedetect_tiny_vision_int8.onnxhttps://huggingface.co/sweetlhare/wedetect-tiny-onnx/resolve/main/wedetect_tiny_language_int8.onnx