Researchers have fine-tuned an open-source vision-language model to work out how a photo was edited, just by comparing the before and after.
The task is called reverse designing, and it's the subject of a paper posted to arXiv that extends and fine-tunes MiniGPT-4, an existing open-source vision-language model that pairs image understanding with a language model. Given a source image, an edited version, and an optional text description of the edit, the fine-tuned model predicts the edits that were made and their parameters. That's a harder ask than typical vision-language tasks like captioning or visual question answering, since the model has to reason about two images at once and the transformation between them. The authors say their experiments show off-the-shelf VLMs like MiniGPT-4 can be stretched to handle this more complex comparison task, and they've published their code on GitHub.
Most AI image-editing tools run one direction: type an instruction, get an edited photo. Reverse designing runs the tape backward, inferring what happened to an image after the fact, which is useful for auditing suspicious edits, recovering an image's edit history, or building tools that describe the difference between two photos in plain English. It's also a small data point for a bigger trend: general-purpose VLMs built for describing images are increasingly being repurposed for narrower, more technical jobs without training a model from scratch.
This is a paper with a GitHub repo, not a product. Whether reverse designing holds up outside the authors' own test set is the question nobody outside this lab has answered yet.