JEV_sees

Checked, not fully reviewed. We ran its 24 tests (all pass) and its bus-colour example end to end on CPU, with both vision models running locally and a stand-in answering in place of Jev. No Jev calls, so nothing here says how well Jev reads the facts it is given.

Jev takes text, not images. JEV Sees bridges that on your machine. Florence-2 (base-ft) finds and captions objects, CLIP and plain pixel statistics describe their colour, and a depth camera adds distances. The result is written into a structured state, so one Jev call can ask about every object in the frame. A companion project, JEV Control, acts on the answers with a robot arm.

What left the machine for one question about one photo was a single request of about 18 KB with no image data in it. It held every visible object, not just the bus: its box, centroid, caption, pixel colour statistics and CLIP’s colour probabilities. Faces and shoes in the background went in too, as labelled boxes, with no identities. For the bus the evidence disagreed with itself. The pixel statistics said blue, as the bus is, while CLIP’s top colour was green at 0.52. Which one wins is up to Jev, and the probabilities are there for you to check.

Getting there took two fixes at 8cf89cb, which we reported:

• pip install -e . resolved transformers>=4.49,<5 to 4.57, and Florence-2’s remote code then failed with AttributeError during generation. With transformers==4.49.0 the example ran. Florence-2 loads with trust_remote_code=True from Microsoft’s repository at no pinned revision, which is common but worth knowing.
• Sees() passes its own default API address to the SDK, so TYPESAFE_BASE_URL is ignored. Our first run went to api.typesafe.ai despite it, and was refused for its test key. Sees(base_url=…) works.

Both fixed on 12 October. At 870cde6 the package pins transformers==4.49.0, and the base URL defaults to none, so the SDK falls back to TYPESAFE_BASE_URL. We checked the pin and the code path, and that the SDK sent a request to the address in TYPESAFE_BASE_URL. We did not rerun the vision example.

MIT, 278 stars.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Oct 13, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.