Text-to-Image Models Fail to Count Objects Past Ten
Original title: Same Layout, Different Count: What Number Words Do Past Ten in Text-to-Image Models
Five open models follow number words only up to ten objects; SD3.5-M and FLUX.1-dev ignore counts beyond that.
Researchers fixed object type and starting noise while varying the number of objects in text prompts for five open text-to-image models. Models accurately follow the number for small crowds but reach a limit around ten. SD3.5-M and FLUX.1-dev stop following the number soon after ten, while FLUX.2 and Qwen-Image follow it loosely up to twenty-five. Redrawing only finer noise components moves the count two-thirds as far as changing the number word. People detect the encoded count in 59% of images when shown as a noised grid versus 4% when only told.
- Models follow counts accurately up to ten objects
- SD3.5-M and FLUX.1-dev ignore counts beyond ten
- FLUX.2 and Qwen-Image follow loosely to twenty-five
- Finer noise redrawing shifts count without changing layout
- Count detectable in 59% of images when encoded as grid
From the original article
1 min read at arxiv.orgAsk SD3. 5-Medium for twelve of an object and then, from the same starting noise, for twenty-five. The picture keeps its layout, but the count changes about as much as it would under a new random seed, and in no consistent direction.
By Liang Yue · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.