Research code that decomposes a flat raster design image into an editable JSON hierarchy of text, shapes, and image layers.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
ReDesign is the official implementation of an ECCV 2026 paper, "Recovering Editable Design Structures from Images via Agentic Decomposition". It takes a single flat raster image and reconstructs an editable design: text with typography attributes, vector shapes with fill and stroke, images, groups and z-order, exported as a JSON hierarchy.
It treats a design as a tree of layers, starting with the whole image as the root. A vision-language model controller examines each region and picks one tool-backed action: extract text (PaddleOCR, font recognition, Hi-SAM, LaMa inpainting), fork into layers (Qwen-Image-Layered), split by connected-component analysis, detect and segment (GroundingDINO, SAM 2, inpainting), or vectorize (VTracer). The tree is expanded breadth-first, and a modular verifier accepts, prunes, or retries each split until every leaf is an atomic element.
It is a Python research repository, not a hosted app. Setup uses a conda environment and a post-install script, and checkpoints are downloaded by script, with one (Hi-SAM's text-segmentation head) requiring a manual download. The controller needs an OpenAI-compatible chat-completions endpoint (OpenAI API, a proxy, or a self-hosted vLLM server) and an API key. Qwen-Image-Layered needs about 55 GB of GPU memory. The repo also provides the Figma-909 benchmark (909 Figma frames, CC BY 4.0) and evaluation code. The GitHub license is reported as NOASSERTION, so the exact license terms should be checked in the repository.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Turns a flat image into editable layers, like a screenshot becoming a Figma or Photoshop project. It orchestrates PaddleOCR, Qwen Image Layered, Grounding DINO and SAM 2. Code is under 4 GB but needs an OpenAI API key by default.
Closest neighbours by meaning, not just by category.
Research · Stability AI
Stability AI research method that fine-tunes Qwen-Image-Layered for image layer decomposition using VLM-scored RL.
Model · Ideogram
Ideogram's model for generating editable SVG vector graphics from text prompts or by converting existing images.
Model · OpenAI
OpenAI's image generation and editing model in ChatGPT and the API, with Sunburst and Flare variants.
Model · OpenAI
OpenAI image generation and editing model with GIF output, SVG export and image markup, in Flare and Sunburst variants.
Model · Bria
Bria's open-weights image generation and natural-language editing model with multi-reference support, non-commercial license.
Model · Reve
Closed, paid image generation and editing model from Reve with 4K output, text rendering and reference-based generation.
