New AI Turns Chart Screenshots Back Into Editable Data

Charts are everywhere — in reports, slide decks, news articles and academic papers — but they usually arrive frozen as pictures. You can look at them, but you can't update them, reuse the numbers, or drop them into your own work without painstakingly retyping every value and re-picking every color by hand. Two researchers, Yuki Ueno and Aditeya Pandey, set out to automate that chore with a system called ChartRevive.

The tool is built on what researchers call "multimodal large language models" — AI that can look at an image and read text at the same time, like a very fast intern who understands both pictures and language. Before building anything, the team ran a bake-off: five of these vision-capable AI models, tested on five common chart types (think bar charts, line charts, pie charts). The results were split. Labels and text came out reliably. Numbers and exact positions — how tall a bar really is, where a line crosses another — remained hard. GPT-5.4 won overall and became the engine behind the tool.

ChartRevive is not fully automatic, and that's deliberate. It's what the authors call a "mixed-initiative" system: the AI makes a first pass, then hands control to you. You see the original image and the rebuilt chart stacked on top of each other, so mismatches jump out. Fix a wrong number, and the chart redraws instantly. The idea is that the AI handles the boring 90%, and you spend your attention only where it stumbled.

The honest limitation: this is early-stage research, presented as a workshop poster, not a product you can download today. And because the AI struggles most with the numbers themselves, the very thing you'd want it to get right is the thing most likely to be wrong. A misread figure in a business report or a news graphic isn't a small thing — it's a false claim with a picture attached. So the promise is real and the time savings are real, but the checking step isn't optional.