Understanding Generative AI Capabilities in Everyday Image Editing Tasks
Mohammad Reza Taesiri, Brandon Collins, Logan Bolton, Viet Dac Lai, Franck Dernoncourt, Trung Bui, Anh Totti Nguyen
2025-05-23
Summary
This paper talks about how well generative AI models, like GPT-4o, can handle common image editing tasks, and what kinds of edits they are good or bad at compared to humans.
What's the problem?
Even though AI is getting better at editing pictures, it still has trouble with simple, very precise changes and doesn't always do what people expect for basic tasks, while it does better with more creative or open-ended requests.
What's the solution?
The researchers studied over 83,000 image editing requests and compared how AI and humans performed, also looking at how both people and AI models judged the results, to figure out where AI editors succeed and where they fall short.
Why it matters?
This matters because it helps us understand the limits of current AI for everyday image editing, showing where improvements are needed and making sure people know when to trust AI or stick with human editors.
Abstract
Analysis of 83k image editing requests reveals that AI editors, including GPT-4o, struggle with low-creativity tasks and precise editing, while performing better on open-ended tasks, and human and VLM judges differ in their preferences for AI versus human edits.