I made an open-source Jev for object detection ๐
retained text
TypeSafe recently added a Box question type to Jev: you describe something in an image in plain words and get back one bounding box with a confidence. Itโs closed and API-only, so I built an open version: **LOCUS** (Locate Once, Check Until Settled). The weights are open (2B and 0.8B) and run on as little as X GB of VRAM. **The problem:** VLMs normally write a box out as text, one token at a time. Thatโs precise but slow, about 18 steps per box. A small regression head can predict the whole box in one pass, which is fast, but the edges come out loose. **The idea:** use both. The head makes a quick draft, then the same model checks every draft token in parallel and fixes the ones it disagrees with, repeating until nothing changes. The final box is exactly what the model would have decoded on its own, in about 5.6 steps instead of 18. **Results** (RefCOCO / RefCOCO+ / RefCOCOg): * Same weights, Acc@0.9 goes from 54.9 โ 59.1 with verification * \+4.52 Acc@0.9 over Jag (strongest open regression-head model) at the same size and training budget * 1.5x faster than vLLM decoding the exact same tokens **Use it on your own data:** the released models are trained on RefCOCO, but you can fine-tune them on your own dataset (images, a description and a box for each). The repo covers full fine-tuning and LoRA, so a smaller GPU works too. **Limitations:** a pure regression head is still faster, and since itโs trained only on RefCOCO it struggles with very small objects. Fine-tuning on your own domain is the fix there. * Project page: [https://muhammad-talha-ad.github.io/LOCUS/](https://muhammad-talha-ad.github.io/LOCUS/) * Code: [https://github.com/muhammad-talha-ad/LOCUS](https://github.com/muhammad-talha-ad/LOCUS) * Models: [https://huggingface.co/MTalha2001/LOCUS-2B](https://huggingface.co/MTalha2001/LOCUS-2B) Not affiliated with TypeSafe. Since Jevโs Box is API-only, I couldnโt evaluate it under the same conditions, so all comparisons are against open models. Happy to answer questions.
document fetched 2026-10-28 00:23 ยท canonical https://reddit.com/r/computervision/comments/1xglj1i/i_made_an_opensource_jev_for_object_detection