Since we first introduced fine-tuning on GPT‑4o, hundreds of thousands of developers have customized our models using text-only datasets to improve performance on specific tasks. However, for many cases, fine-tuning models on text alone doesn’t provide the performance boost expected.

Vision fine-tuning follows a similar process to fine-tuning with text—developers can prepare their image datasets to follow the proper format⁠(opens in a new window) and then upload that dataset to our platform. They can improve the performance of GPT‑4o for vision tasks with as few as 100 images, and drive even higher performance with larger volumes of text and image data.

We collaborated with a small group of trusted partners to better understand the real-world applications of vision fine-tuning. We’re excited to showcase a few examples of what they built.

Grab⁠(opens in a new window), a leading food delivery and rideshare company, turns street-level imagery collected from their drivers into mapping data used to power GrabMaps⁠(opens in a new window), a mapping service enabling all of their Southeast Asia operations. Using vision fine-tuning with only 100 examples, Grab taught GPT‑4o to correctly localize traffic signs and count lane dividers to refine their mapping data. As a result, Grab was able to improve lane count accuracy by 20% and speed limit sign localization by 13% over a base GPT‑4o model, enabling them to better automate their mapping operations from a previously manual process.