Gemini is the most forgiving of the three on format and the least forgiving on volume. It accepts PNG, JPEG, WEBP, HEIC and HEIF, and one request can carry up to 3,600 image files. That ceiling is generous enough that a real batch job hits it, which means per-file size is the only thing standing between you and a payload you have to chunk by hand. Resolution is where the cost lives. Gemini counts image tokens by tiling: the crop unit is roughly floor(min(width, height) / 1.5), and each dimension is divided by that crop unit and multiplied, so a 960x540 image has a 360 px crop unit and costs 3 by 2, six tiles. Push the pixels up and the tile count climbs with them. The media_resolution parameter then sets the maximum number of tokens allocated per image, which means the pixels you send are a real budget decision rather than a formatting detail. Google's own guidance also asks for correctly rotated, clear, non-blurry input, because upscaling a blurry original adds tiles without adding information. Skwish resizes to an exact dimension, converts between JPG, PNG and WebP, and compresses in the browser with a batch queue and one ZIP.