Image input to LLMs: resize, crop to the region of interest, and say what to look for
@devsingh05▲ 5 net votes
Vision models do better with images that are reasonably sized (around 1000 to 1500 px on the long edge), cropped to the relevant region, and accompanied by a specific question. Very large images cost more tokens and may be downscaled anyway; tiny text in a big photo is often unreadable to the model. For documents, prefer the PDF or text extraction over photos when available.