Study maps how much detail vision-language models lose on large images
Researchers measured what happens to small objects when online vision-language services shrink or reject big photos, and tested zooming strategies that recover much of the lost accuracy.
In plain words
Vision-language models answer questions about images, but online services often compress a large picture before the model sees it, so tiny details can vanish. A new study measured this loss across seven models and found that letting the model zoom into regions restores much of the accuracy that shrinking destroys.
What was announced
In a paper on arXiv, the authors report that online vision-language services frequently reduce image size, or refuse very large files, before inference. They give an example where a car spanning 20 pixels in an 8000 by 6000 drone photograph arrives at Claude at roughly 6 pixels, while OpenAI's default setting rejects that photo entirely. To quantify the effect, the team ran two tests: a printed-code eye exam and 601 questions about small objects in real photographs, covering seven models from four vendors. The text a model can read half the time was 6 pixels or fewer on a small image, but a predicted 10 to 13 pixels on a 7680 by 4320 photo at default settings.
Zooming recovered much of the loss. Cutting each photo into as many as nine tiles raised accuracy by 11.1 to 14.5 percentage points for Claude and Gemini, and by preventing rejections it lifted GPT-6.1 Sol from 50.7% to 83.7%. A two-step approach, where the model first names the region it needs, matched or beat tiles on Gemini and came within 2.5 points on Claude while using about 40% of their image tokens. The authors also note a behavioral difference: when detail is lost, Gemini gave a confident wrong answer in 43% to 48% of cases, versus 5% to 9% for Claude. They distilled the results into five rules for when an agent should zoom, and released the eye exam, model answers and code.
What it does not show
- The source is an arXiv abstract, so the full paper's methods and peer-review status are not covered here.
- The reported accuracy gains and error rates come from the authors' own benchmarks and have not been independently confirmed.
- The abstract does not state which image sizes, prompts or pricing conditions apply beyond the tested defaults.
For security and IT teams
- Scope
- Findings cover seven models from four vendors under default service settings, with the eye exam, answers and code released by the authors.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.