Image prep for LLMs

What it does

Open an image, choose the model it is for, and prepare it. The tool:

• resizes it to the size the model sees, so the provider does not resize it again and no bytes are sent for pixels it would discard. You can choose a long edge of your own instead.
• writes it as WebP, JPEG or PNG, at the quality you choose;
• shows the size, the tokens, the file size and the metadata before and after;
• lets you download it or copy it as a data: URI.

Everything runs in your browser. The image is never uploaded. A shared link holds the settings only, never the image.

What it removes

The image is redrawn and written again, and only the pixels survive. EXIF, which in a phone photo often includes the GPS position where it was taken, is removed, along with XMP and colour profiles. The table shows what the original carried and what the prepared image carries: nothing.

The orientation EXIF records is applied first, so the prepared image is upright even without it.

How it sizes

For Anthropic's models, "what the model sees" is the size Anthropic's own resize rule gives: the largest aspect-preserving size within the model's tier. That is 1568 px and 1568 tokens on the standard tier, and 2576 px and 4784 tokens on the high-resolution tier. The Image token calculator explains the rule. Tokens depend on pixels, not on file size, so a smaller file does not cost fewer tokens. It is faster to send, and it is easier to keep under the 10 MB, as base64, that a request's image may be. The tool warns when a prepared image is still over.

For requests of more than 20 images, Anthropic advises keeping each side at 2000 px or less. The "For requests of more than 20 images" example prepares images to that size. Through Salyro, though, one request carries at most 20 images, so a larger set is split across requests of 20 or fewer.

What it does not know

• OpenAI's, Gemini's and xAI's image rules. Their pages could not be read when this tool was built, so only models with a documented rule are offered.
• How your model reads a heavily compressed image. Lower quality can blur text, so check a prepared screenshot before you rely on it.
• Animation: a GIF is read as its first frame, which is also what Claude uses.
• A GIF's metadata. Metadata is looked for in JPEG, PNG and WebP files. In a GIF it is not looked for, and the tool says so. The prepared image carries none either way, because it is written from pixels only.

Sources

The sizes follow Anthropic's vision page and its page on how Claude resizes images, on platform.claude.com, as recorded in Reference Data.