ReDesign turns flat screenshots back into editable design layers

ReDesign is an open-source tool that reverses flat images back into editable project structures. When you only have a flat PNG screenshot or the original source file is lost, it reconstructs real editable typography, vector shapes, RGBA layers, and stacking order into a structured JSON hierarchy.

Once exported, you can tweak individual text layers, recolor vector shapes, reposition graphic components, or reorder elements just as you would inside Figma or Photoshop.

How it deconstructs designs

ReDesign treats any design as a hierarchical layer tree, expanding it top-down from the root image until every leaf becomes an atomic, fully editable element.

A Vision-Language Model controller inspects each region and orchestrates five distinct tool actions. For text areas, it combines PaddleOCR, font recognition, Hi-SAM, and LaMa inpainting to extract editable typography while restoring the background behind it. For multi-layer visuals, it triggers Qwen-Image-Layered to produce depth-aware RGBA layers.

It also uses connected-component analysis to separate disjoint items, uses GroundingDINO alongside SAM 2 to detect and segment foreground subjects, and runs VTracer to convert flat shape leaves into clean vector paths. After every split, an automated verifier checks the output, pruning broken branches or retrying with another tool whenever a split fails.

Hardware specs and setup

The project code is live on GitHub, accompanied by the Figma-909 benchmark dataset released under a CC BY 4.0 license.

The main hardware bottleneck is the layered diffusion backbone. Qwen-Image-Layered requires roughly 55 GB of VRAM in bf16, split between roughly 39 GB for the transformer and 16 GB for the text encoder. You can run it on a single 80 GB GPU, shard it across two 40 GB cards, or distribute it across four 24 GB consumer GPUs. Auxiliary vision tools like SAM 2 take another 10 to 16 GB of memory, while PaddleOCR runs purely on the CPU.

The agentic controller itself connects through any OpenAI-compatible API endpoint. The default setup targets gemini-3-flash-preview via an environment file, but you can swap it for GPT-4o or your own local vLLM instance to keep runs private.

Why it matters

If you deal with lost vector assets, archived marketing banners, or client mockups trapped in flattened exports, ReDesign provides a genuinely practical rescue path. Rather than hallucinating a fuzzy raster reinterpretation with generative diffusion, it recovers real text properties, vector contours, and clean separation masks that design workflows actually demand.

We think the smart move here is avoiding the temptation to build an unwieldy all-in-one neural net. By letting a lightweight planner orchestrate best-of-class modular tools like SAM 2, PaddleOCR, and Qwen-Image-Layered, ReDesign beats single-model baselines across the board on the Figma and Crello benchmarks. The only real catch is the hefty GPU memory footprint required to host the multi-model pipeline locally.