Does GPT-4 Vision Preview Rescale Images Proportionally to 512×512?

Below is a concise explanation based on the documentation:

How Low-Res Mode Works

When you use the low mode, the API receives a version of your image that is 512 × 512 in size and is represented using a limited token budget. This means that regardless of your original image’s dimensions or aspect ratio, the backend will convert it into a 512 × 512 image for processing.

What It Means for Your Image Proportions

The documentation doesn’t detail whether the resizing is done by proportionally scaling the image to fit within a 512 × 512 square (using padding or letterboxing when necessary) or by performing a direct transformation (which could distort non-square images). In practice:

• If you rely on precise proportions: You may want to preprocess your images so they’re already sized (or fit within a square canvas with appropriate padding) to ensure critical visual details remain as intended.
• If you’re flexible with proportions: You can simply submit the image and let the backend resize it according to its own internal process.

Summary

Since the internal resizing behavior (preserving aspect ratio vs. stretching/cropping) isn’t explicitly described, pre-resizing your images (if proportions matter to your use case) is a safer approach. This way, you can guarantee that the image’s visual layout aligns with your expectations when processed by the API.

If you need further clarification on preserving details or maintaining the aspect ratio, it may help to experiment with your specific images to see how the backend resizing impacts the final output.