KAIST's ReDesign rebuilds editable design files from raster images

Recovering an editable Figma file from a flat PNG has been one of design's ugliest manual chores, and a training-free agentic system out of KAIST AI and collaborators at Helmholtz Munich and Korea University claims a workable answer. The paper, posted on Hugging Face's paper stream, calls the system ReDesign, and the core move is to stop treating raster-to-editable as an end-to-end learning problem and treat it as a tree-growth problem instead. A vision-language model agent walks the image from root canvas down to leaves, picking specialised tools at each expansion: OCR, layered decomposition, connected component labeling, detect-and-segment, and vectorisation.

The interesting twist is what the authors call graceful verification. Instead of validating only the final JSON, a verifier evaluates every parent-to-children split and can accept it, prune bad children, or retry with different tool settings. That local repair is what keeps errors from cascading down the tree and forcing an expensive full rerun. It also happens to be embarrassingly parallel, since sibling nodes do not depend on each other, and the paper reports up to a 7.1 times speedup over a comparable serial ReAct-style tool agent.

To measure editability the team built the Figma Edit Replay Benchmark: 909 raw Figma files, averaging 2.62 times more elements per sample than the older Crello dataset, paired with 14,796 controlled edit instructions and roughly 15 edits per design. The protocol replays the same instruction on the ground-truth file and the reconstructed one, then compares the rendered results, including OCR text recall for text edits. Across layout, color, and text, ReDesign posts the best editability numbers against Qwen-Image-Layered, LayerD, VTracer, and a serial tool-using agent, and the authors also compare against Nano Banana 2, which they say struggles with spatial edits and can shift aspect ratios even on local instructions.

The honest caveats are the ones the paper itself flags. It is training-free but leans on a commercial VLM (Gemini-3-flash by default, with GPT-5 mini as a robustness check), so cost per reconstruction is a function of API pricing that is not disclosed. It is one benchmark, built by the same team, and best editability still means an approximate reconstruction, not a lossless recovery of the source. The authors also raise the copyright angle openly, noting that a system like this lowers the barrier to modifying branded or copyrighted graphics and should be paired with usage policies and provenance mechanisms.

The forward-looking bit is the structural argument. If the pattern holds, agentic pipelines that impose a tree shape on their output and verify locally may generalise well beyond design reconstruction to any long-horizon task where a serial agent's early mistakes silently poison everything downstream.