Recovering an editable design file from a raster image is a common and costly bottleneck in modern design workflows. The challenge lies in recovering multi-modal attributes such as typography, vector geometry, colors, grouping, and layer ordering. Researchers from Hugging Face have introduced ReDesign, an agentic framework that addresses this problem by growing an editable layer hierarchy through the selection and composition of specialized tools across modalities.
To ensure reliability despite imperfect tool outputs, ReDesign employs graceful verification at each expansion step. This mechanism provides local accept, prune, or retry feedback, preventing error accumulation and avoiding large-scale reruns. The framework thus maintains a long decision process while keeping errors in check.
The team also introduced the Figma Edit Replay Benchmark, consisting of 909 raw Figma files and 14,796 controlled edit instructions. This benchmark allows evaluation of editability at scale by replaying edits on reconstructed outputs. Across this benchmark and standard reconstruction metrics, ReDesign achieves strong visual fidelity while delivering the highest editability for layout, color, and text edits.
ReDesign outperforms layered decomposition baselines and serial tool use pipelines, marking a significant step forward in automated design recovery. The data and code are available on Hugging Face Datasets and GitHub.