README.md

metadata

title: LOCUS-2B Visual Grounding
emoji: 🎯
colorFrom: green
colorTo: purple
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
python_version: '3.12'
short_description: Ground referring expressions with draft-and-verify boxes
startup_duration_timeout: 30m

LOCUS-2B — Locate Once, Check Until Settled

Ground a referring expression in an image and get its bounding box. LOCUS-2B is Qwen3-VL-2B-Instruct with a small box head, trained with the Locus recipe:

1. The image is prefilled once; every expression about it runs as a short branch over the cached prefix.
2. The box head drafts the box in one pass.
3. The draft is rendered as the coordinate tokens the model would write, and the model verifies all of them in parallel by Jacobi iteration until the draft stops changing.

With exact acceptance the returned box is the model's own greedy coordinate-token answer, reached in a few passes instead of one decoding step per token.

Usage

Upload an image and enter one referring expression per line. Pick an output mode:

• Verified, exact — the greedy coordinate-token box (a few verifier passes).
• Verified, typical (α = 0.5) — fewer passes, no longer exactly greedy.
• Head only — the one-pass draft, fastest.

Each row of the results table shows the expression, its box [x0, y0, x1, y1] in pixels, the head's IoU-aware confidence and the number of verifier passes used.

The locus/ directory vendored in this Space is the inference path of the Apache-2.0 LOCUS repository. Example images are from the CC0 pool linoyts/repo-to-space-example-inputs.