Design Anything, Design at Scale

Design Anything, Design at Scale. By Xuchao Ren and ten others, from Tsinghua University, the Beijing Institute of Fashion Technology, and two other institutions. Published on the seventeenth of October. Its code repository is worked out: about fifty-five source files, with a readme and a licence.

Designers work from briefs. A brief sets out goals, constraints and deliverables, and later work in a project builds on decisions taken earlier. As agentic AI turns multimodal resources into media outputs, it is expected to stay aligned with that design intent over long horizons. Recent work reuses design experience from one task to another, but keeping the artifacts within a single project aligned with each other remains hard. An agent has to preserve the intent and keep applying it, across domains and as a project grows from a handful of pieces to many.

Design Anything is a graph-based harness that keeps design intent outside the model. It does not rely on the model remembering what was decided. Instead it binds each deliverable to the workflow steps that produce it, through what the authors call an intent-execution graph. The intent lives in a structure the harness maintains, and the model works against that structure.

Three components carry the intent forward. A task contract retains the constraints and the decisions already made. Designs that have been selected become explicit inputs to later generations, so a choice made early is passed on rather than reconstructed. And a deterministic controller uses the bindings and the execution records to check the types and counts of outputs, to identify which deliverables are still outstanding, and to guide the turns that follow.

To measure all this, the authors also introduce DesignAnything-Bench. It holds two hundred and twenty-eight tasks across nineteen design domains. Among them are thirty-eight project families, built at three scales: six artifacts, fourteen artifacts, and twenty-four artifacts per project.

The comparisons keep the language model fixed, so that only the harness differs. With DeepSeek V4.1 Flash High, Design Anything satisfies eighty-five percent of the brief's factual constraints. With GPT-5.6 Sol High, it satisfies ninety-two point five percent. Those figures exceed the strongest baselines using the same model by fifteen point four percent and fourteen point four percent respectively.

The margins widen as projects get bigger. On the twenty-four-artifact tasks, the gains reach twenty-five point one percent with the first model, and twenty-three percent with the second. Under each model, Design Anything leads or ties in constraint satisfaction in eighteen of the nineteen domains. In a blind evaluation, professional designers also preferred its complete sets of deliverables over those of the same-model baselines.

The argument is about where memory should sit. Asking a model to hold a long project's intent in mind becomes harder as the project grows, and the gains here are largest on the longest projects. Moving the intent into a contract, a graph and a controller that checks the outputs leaves the model free to design, while the harness keeps track of what the brief still requires.