PRODUCTION SCENARIO
An inspection app sends high-resolution photos to a deployed multimodal model with the default image settings and asks it to read tiny component labels. The responses often miss the small text.
Which request change is most likely to help?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Set detail to high on the image_url part
With detail set to high the model activates high-res mode: it views the low-resolution image first, then generates detailed 512x512 segments with double the token budget. auto is already the default and is what the app is getting now, so it changes nothing. Downscaling to 512x512 destroys the small text before the model sees it. Tiling by hand reproduces what high-res mode already does internally, at the cost of losing whole-image context.