GPT-4 Vision: Best Image Size and Format for Accuracy

A surprising number of teams send 8 MB phone screenshots to GPT-4 Vision and wonder why responses are slow. The truth is that every major vision model resizes your image internally before reasoning over it — sending the original costs upload time, bandwidth, and sometimes accuracy.

The internal resize, summarized

• GPT-4 Vision: tiles images into 512×512 patches; high-detail mode caps at ~2048×2048
• Claude 3.5 Sonnet: long side capped around 1568 pixels
• Gemini 1.5/2.0: each image becomes 258 tokens regardless of size, so resolution beyond a threshold buys nothing

The practical rule: aim for ~1024 pixels on the long side and under 1 MB. Above that you're paying for pixels the model will throw away.

Format choice, by content type

Text-heavy screenshots → PNG

If the image is a screenshot of code, a contract, an error message, or an article, use PNG. Lossless compression preserves the crisp anti-aliased edges that make characters readable. Better still, run it through Screenshot to Text first and send the extracted text — it uses fewer tokens and is more accurate for dense type.

Photos and natural scenes → JPEG

For real-world photos, JPEG at quality 80 is virtually indistinguishable from the original to a vision model and a fraction of the size.

Motion or UI animation → GIF

Vision models accept GIFs but reject video files. To show a model an animation, a UI flow, or a moment from a video, convert with Video to GIF for AI Vision. Aim for under 1 MB total — the tool's target-size mode will iteratively shrink dimensions, frame rate, and palette to hit your budget.

The accuracy paradox

Counter-intuitively, downscaling can improve answers. A massive image gets aggressively resized inside the model, but JPEG artefacts from the resize can corrupt fine details. Pre-resizing to the model's native target with a clean Lanczos filter avoids that.

When to skip vision entirely

For dense text content, OCR plus a text prompt is almost always:

• Cheaper (no per-image token cost)
• Faster (no image upload)
• More accurate on long passages
• Compatible with every model, including text-only ones

Use Screenshot to Text for one-off screenshots, or Batch Document Extractor for whole folders.