vLLM Semantic Router The Decision Box Gallery

Find one figure in a crowded masterpiece. Box questions on Decision 3 (d3 and d3‑flash), live on the free Decision API.

How the search works

  1. I. Tiles. A whole painting read at once is read small, and one figure in a crowd is a few pixels. So the page cuts the work into tiles, each read at full resolution, and asks one yes / no question per tile: does it show the figure? One more asks it of the whole work at once, for a figure larger than a tile. If no tile and not the whole work shows it, the search ends here: not found.
  2. II. Zoom. The tile with the most likely “yes” is enlarged by a quarter on every side, so a figure on its border is not lost.
  3. III. Box. One Box question on that tile. The Decision API answers it with 65 yes / no sub-questions on an outlined grid that zooms into the figure, and the box is mapped back onto the whole work. A box below 10% confidence means not found, and the search ends.
  4. IV. Edge checks. Four yes / no questions in one request: is the figure cut off by the top, bottom, left or right edge of the box? Every edge that cuts it off moves out and the four are asked again, for up to six rounds; a round with no cut edge ends the search.

The confidence is the model’s own estimate that its box overlaps the true one by at least half. Every step is a live request from your browser to decision.vllm-sr.ai; nothing is precomputed. d3 (27B) and d3-flash (9B) are the two Decision 3 models that answer Box questions; Compare runs both on the same tiles and draws both boxes. If the models are busy, the page quietly waits its turn.

Your own picture

Choose a picture, say what to find, then press Find it. Large pictures go through all four steps; small ones start at the box.

Run details

No search yet.

The works on view