ReDesign uses agentic tool decomposition to recover editable designs from images
kaist-ai · hf · 2026-07-29
ReDesign reconstructs editable design hierarchies from images
KAIST AI and collaborators present ReDesign, an agentic framework for recovering editable design files from raster images by composing specialized tools across modalities.
- The system incrementally grows an editable layer hierarchy instead of doing a one-shot conversion.
- It uses graceful verification at each step to locally accept, prune, or retry tool outputs, reducing error accumulation in long decomposition chains.
- To measure editability at scale, the authors introduce Figma Edit Replay Benchmark with 909 raw Figma files and 14,796 controlled edit instructions.
- On reconstruction metrics and the new benchmark, ReDesign reports strong visual fidelity and the highest editability for layout, color, and text edits versus layered decomposition baselines and serial tool pipelines.
More from coding & agent
- GitHub repo curates 100+ libraries for LLM apps, RAG, agents, and testing — tom_doerr · 2026-07-29
- Claude and Codex open-source programs reached an estimated $77,685 in API value — rudrank · 2026-07-29
- Why production AI agents drift after passing every eval — Diligent_Response_30 · 2026-07-29
- MCP stateless means self-contained requests, not memory loss — prdeepakbabu · 2026-07-29
- An AI-agent integration idea for Foilwick emerged during a demo recording — tristanbob · 2026-07-29
- AgentOS turns agents into a production API and MCP server with memory and tracing — pritisinghhhh · 2026-07-29