Dynamic Resolution

Key Insight

A plain VLM squashes every image down to one fixed square (e.g. 336×336), which throws away the fine print in a document or a dense chart; AnyRes fixes this by tiling the picture at its native aspect ratio into several sub-images, encoding each tile separately, and handing all of their image tokens to the LLM. Because more tiles means more tokens means more detail preserved, this is exactly the change that lifts OCR (Optical Character Recognition)-heavy benchmarks — tasks where the answer hides in small text the squashed-down image literally cannot resolve. The trade-off to verify is cost: each extra tile adds image tokens the LLM must process, so AnyRes buys accuracy on detail-dense images at the price of a longer, slower sequence.