ConCor-1: Vision-Language Grounding as Bidirectional Concept Correspondence

Give an image and its text. ConCor-1 recovers all correspondences between visually referential text spans and instance-level image segments. The text spans are not given as queries — the model decides which parts of the text refer to something visible, groups co-referring mentions ("a skier" … "a person" become one text mask), segments the matching object, and decides how many correspondences exist.

A caption, a referring expression and a category list are all just "the text".

Model: UWGZQ/ConCor-1 · 866.79 M parameters (Qwen3.5-0.8B backbone + 385 bridge tokens + three lightweight heads).
0 1
0 1
0 1
0 1
Examples — validation samples with their own captions, a referring expression, and a category list
Image (optional) Text to ground

How to read the output

  • Masks — one image mask per correspondence, in the same colour as its text mask, so the two sides of a correspondence are easy to link.
  • Grounded text — the predicted text mask of each correspondence. A text mask can be several disjoint spans, which is how repeated and co-referring mentions are grouped. One span can also carry several underlines: with baseball player . baseball glove over an image of two players, the single phrase "baseball player" is claimed by two correspondences, one per person, and each contributes its own coloured underline.
  • Presence — the model's own score for "this is a valid image–text correspondence".
  • Click a mask to emphasise its text mask and table row; click it again to clear. (Gradio exposes only a click event for masks, so this is click-driven rather than hover-driven.)

Thresholds

Defaults are the paper's: presence 0.10, image mask 0.45, text mask 0.45, NMS IoU 0.50. Sliders re-render from cached logits, so they are instant and cost no GPU time. Lower the presence threshold to see more (and noisier) candidates.