DrawAI: Agentic Benchmark and Workflow for Making Raster Images Editable
This paper introduces DrawAI, a comprehensive framework comprising the DrawAI-Bench benchmark and the DrawAI-Flow agentic workflow, designed to reconstruct raster images into structured, editable graphics by effectively balancing visual fidelity with semantic editability.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are looking at a beautiful, high-definition painting of a futuristic city. It looks amazing, right? But now, imagine trying to change just one thing: maybe you want to swap the red spaceship for a blue one, or move a skyscraper to the other side of the street. If that picture is just a standard digital photo (a "raster image"), you can't just grab the spaceship and move it. To the computer, the spaceship isn't a separate object; it's just a specific pattern of colored dots mixed in with the sky and the buildings. To fix it, you'd have to paint over the whole thing or start the picture from scratch. This is the frustrating reality of most images created by modern AI: they are beautiful but "flat," making them impossible to edit without ruining the whole picture.
This paper lives in the world of Artificial Intelligence and Computer Vision, specifically focusing on how we can make AI-generated images not just pretty, but useful. The key idea is turning a flat picture into a "layered" file, like a digital scrapbook where every sticker, word, and shape is a separate piece you can pick up and move. The researchers are trying to solve a tricky puzzle: how do you take a flat photo and magically rebuild it as a set of editable parts without losing the original look? They call this "image-to-editable reconstruction." It matters because as AI gets better at making visuals for school projects, science papers, and presentations, we need those visuals to be flexible enough for humans to tweak and improve, not just stare at.
The Paper's Big Idea: Turning Flat Photos into Editable Lego Sets
The researchers behind this paper, from universities in Beijing and Tsinghua, noticed a major gap. AI can now draw amazing pictures, but those pictures are stuck in "flat mode." If you want to fix a typo or move a chart, you're out of luck unless you redraw the whole thing. So, they created a new system called DrawAI to fix this. Think of DrawAI as a super-smart digital architect that looks at a flat photo and says, "I know this looks like a solid wall, but I'm going to rebuild it out of Lego bricks so you can take it apart later."
To prove their system works, they didn't just guess; they built a giant testing ground called DrawAI-Bench. Imagine a massive obstacle course with four different types of challenges: scientific diagrams, presentation slides, posters, and flowcharts. They filled this course with 40 images (10 real and 10 AI-generated images per domain) and created a strict scoring system with 39 different rules to grade the results. The rules check two main things:
- Fidelity: Does the rebuilt version look exactly like the original? (Did the colors match? Is the text readable?)
- Editability: Can you actually edit it? (Is the text a real text box you can click? Is the arrow a real line you can move?)
The paper found that these two goals often fight each other. If you just copy the whole image as a single block, it looks perfect (high Fidelity) but you can't edit anything (zero Editability). If you break it into too many tiny pieces, you can edit everything, but the picture might look messy or lose its original style. The challenge is finding the sweet spot.
How They Did It: The Two-Step Robot Team
Instead of asking one AI to do the whole job at once (which often leads to mistakes), the team built a two-step workflow called DrawAI-Flow. They treated the problem like a construction project with two specialized workers:
- The Parser Agent (The Detective): First, this agent looks at the flat photo and uses special tools (like a magnifying glass for shapes and a scanner for text) to figure out what's in the picture. It doesn't just guess; it creates a detailed "blueprint" or plan. It decides, "Okay, this part is a text box, this is a circle, and this is a photo that needs to be cropped."
- The Reconstruction Agent (The Builder): This agent takes the blueprint and starts building. But here's the cool part: instead of just spitting out a final image, it writes code (like a recipe) to draw the picture. It draws a little bit, checks if it looks right, and if it's wrong, it fixes the code and tries again. It keeps looping—draw, check, fix—until the picture is perfect and fully editable.
What They Found: It's Not Just About the AI Model
The team tested 13 different AI models (from big names like OpenAI, Anthropic, Google, and others) using 5 different "harnesses" (which are like different toolkits or operating systems that help the AI run). Here is what they discovered:
- No Single Winner: There wasn't one AI model that was the best at everything. Some models were great at making the picture look real (Fidelity) but terrible at making it editable. Others were great at editing but made the picture look a bit off. For example, one model called GPT-5.6 Sol got a 94.0 score for looking like the original but only an 85.1 score for being editable. Another model, Claude Opus 4.8, was the opposite, getting a 91.6 for editability but a lower score for look-alike.
- The Toolkit Matters More Than You Think: The "harness" (the tool used to run the AI) made a huge difference. Using the right toolkit could boost a model's score by over 13 points. It's like having a master carpenter with a dull saw versus a sharp one; the carpenter (the AI model) is the same, but the result changes completely based on the tools.
- The Workflow is the Secret Sauce: When they used their special two-step workflow (DrawAI-Flow) instead of just letting the AI try to do it in one go, the results got much better. Specifically, the Editability score jumped by 17.6 points on average. The workflow helped turn recognized content into actual, separate objects you can click and move.
- Text is the Hardest Part: Even the best systems struggled with text. While they could make images and shapes easy to edit, turning text into a real, editable text box was still a weak spot for many models.
- Cost vs. Quality: They also looked at how much this cost. They found that spending more money didn't always mean a better result. Some expensive setups were actually worse than cheaper ones because they used the wrong tools or made the AI do unnecessary work.
The Bottom Line
The paper concludes that making an image editable isn't just about having a smarter AI model. It's about having a complete system that includes a good model, the right tools to run it, and a smart step-by-step plan. You can't just ask an AI to "make this editable" and expect magic. You need a detective to plan it and a builder to code it, checking their work along the way.
While the system isn't perfect yet (especially with complex tables and dense posters), it shows a clear path forward. By breaking the problem down and using code to rebuild the images, we can finally turn those flat, unchangeable AI pictures into flexible, working drafts that humans can actually use and improve. The future of AI art isn't just about making pretty pictures; it's about making pictures you can play with.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.