Live browser demo: one image encoding, several probability queries
The OpenJev-Vision live demo is now available:
https://huggingface.co/spaces/IamBusy/OpenJev-Vision-Demo
The demo runs the published synthetic-joint seed-17 CNN inside your browser. Reveal more visual evidence, watch the estimates change, then ask another event question without rerunning the image encoder. No account or model-service key is needed.
The ONNX export includes the published temperature calibration. Its predictions were checked against the PyTorch model on both bundled views; the maximum absolute posterior difference was below 0.000001. The examples come from one fixed development episode.
The full release also includes separate DINOv2-based Pets and CLEVR-4 experiments, downloadable weights, data and three-seed evaluations. The failures are included: the direct joint predictor degrades on unseen dependencies, and independent attributes beat the tested joint and binding heads on unseen color/shape pairs.
Scope: fixed vocabularies and explicit event semantics. This is an experimental toolkit, not a general-purpose VLM. It is an independent project inspired by Jev and is not affiliated with TypeSafe.
I’d especially welcome reproducible failure cases and suggestions for evaluating generalization to new event structures.