OpenJev-Vision
Research toolkit that encodes an image once and answers several typed questions from a shared distribution, plus OpenJev-0.6B text models.
Overview
OpenJev-Vision is an open research toolkit for visual probabilistic decisions: encode an image once and answer several typed questions from a shared probability distribution. The release includes 8,192 original synthetic images, a 2,960-image Oxford-IIIT Pet subset and 1,680 CLEVR-4 images; synthetic scenes use a small CNN plus a supplied prior, and public images use frozen DINOv2 features with trained readouts. Questions use declared event semantics and a limited English parser, so it is not a general visual assistant.
Three-seed results include calibration and negative findings: the direct joint predictor degrades on new dependency structures, and independent attribute heads beat joint and binding heads on unseen CLEVR-4 combinations. The repository also keeps the earlier text-decision work, OpenJev-0.6B, which returns probabilities for choice, noul and ordinal score without decoding. It is independent of TypeSafe.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is a given event true of the scene? | true · false |
| Choice | Which attribute value applies to an object?(inferred) | — |
How it works
- Download trained readouts.
- Encode the image once with the CNN or DINOv2.
- Answer several declared questions from the shared distribution.
- Evaluate calibration across seeds.
Strengths
- Reproducible datasets and readouts.
- Reports negative results honestly.
- Shared encoding for multiple questions.
Caveats
- Fixed event vocabularies; not a general visual model.
- Limited English query parser.
Getting started
git clone https://github.com/IamBusy/OpenJev-Vision.git && cd OpenJev-Vision && uv sync --frozen --extra vision --extra dev && uv run --no-sync openjev-vision downloadAt a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch, CLI
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Encodes an image once and answers several typed questions from the shared distribution. Ships synthetic scenes, trained readouts, a dataset and reproducible evaluations.
MrJev/awesome-jev
Frequently asked questions
Is it a general image assistant?
No, it uses fixed event vocabularies and a limited parser.
What datasets are included?
Synthetic scenes, an Oxford-IIIT Pet subset and CLEVR-4 images.
What about text decisions?
The repo keeps OpenJev-0.6B for text choice, noul and score.
Related projects
Research code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPythonPaper introducing PixelJev, a native-image decision interface on small open multimodal models with few-shot adaptation.
PaperresearchFree hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPython- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython Joke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.
ProjectresearchPython