OpenJev Vision v0.1 — experimental research weights
Small trained readouts for OpenJev's visual research track. The dataset contains original synthetic scenes and attributed public-image subsets.
Contents
Seven fixed-seed-17 checkpoints: three synthetic-scene CNNs (joint, independent, and learned visual evidence with exact finite fusion), one Pets breed head, and three CLEVR-4 color/shape heads (joint, independent and low-rank binding). Every checkpoint was selected by development NLL and calibrated on a disjoint calibration split. All three training seeds are reported in the code repository.
Synthetic checkpoints include their complete small CNN. Public-image checkpoints contain only trained heads and train-fitted feature normalization. They require the separately downloaded, frozen facebook/dinov2-small backbone at revision ed25f3a31f01632728cabb09d1542f84ab7b0056, and its exact image processor. They are not Transformers AutoModel checkpoints or general-purpose VLMs.
Load
Install OpenJev with its vision extra, then:
from openjev.vision.hub import download_models
paths = download_models("artifacts/openjev-vision-v0.1")
The returned directories can be used with SceneModel or PublicVisionModel. See the repository's visual research guide for complete image/question examples. Model file SHA256 hashes are recorded in FILE_MANIFEST.json.
Measured behavior and limitations
The learned joint synthetic model improves compound probabilities over its own factorized marginals in distribution, but its posterior prediction degrades substantially on unseen dependency topologies. The learned-evidence/exact-fusion control is much stronger in the known simulator; it also has privileged observation-category training supervision.
On the 740-image balanced Pets subset, the seed-17 frozen-feature breed head scores 93.24% accuracy; this is not the official full-dataset benchmark. On 80 held-out color/shape-composition images in CLEVR-4, seed-17 joint, independent and binding heads score 0%, 63.75% and 56.25%, respectively. The low-rank binding head does not outperform the independent baseline. These negative results are part of the release.
Query composition is exact probability arithmetic over a fixed ontology. Free-form language understanding, general scene grounding, and arbitrary real-world uncertainty calibration have not been established. Synthetic inputs are 64x192 three-slot images with a supplied 64-world prior. Pets/CLEVR-4 heads operate on their separate declared taxonomies. DINOv2 pretraining overlap with the public datasets cannot be excluded.
Training and licenses
Synthetic CNNs use original Apache-2.0 rendered data and exact observation-model targets. Public heads use resized Oxford-IIIT Pet (CC BY-SA 4.0) and CLEVR-4 (CC BY 4.0) subsets, with attribution and license notices in the dataset card. Original source data retain their respective licenses. We distribute the original synthetic CNN weights under Apache-2.0, the Pets heads under CC BY-SA 4.0, and the CLEVR-4 heads under CC BY 4.0. See LICENSES.md. No source images or third-party backbone weights are included. DINOv2's own license applies to its separate backbone. Do not use this model bundle to relabel the source data.