Design Anything, Design at Scale

Xuchao Ren1,*, Siyuan Feng1,*, Jing Huang1,2,*, Yanhua Yao1,3, Jing Wang1, Lingyu Alina Wang1, Hangjie Yuan4, Sai Ma1,†, Qiong Wu1,†, Tao Feng1,†

1Academy of Arts & Design, Tsinghua University2Beijing Institute of Fashion Technology3Senseable City Lab, Massachusetts Institute of Technology4DAMO Academy, Alibaba Group

* Equal contribution† Corresponding authors

Paper — not yet availableGitHubDataset — not yet releasedExplore all results in Gemus

Design Anything keeps a brief’s design intent in a graph outside the model, so an agent stays on brief across 19 design domains and projects of up to 24 artifacts.

Figure 1. Design Anything outperforms the best same-model baseline at every scale.

Abstract

As multimodal resources are transformed into media outputs, agentic AI is expected to stay aligned with design intent over long horizons. Designers follow briefs specifying goals, constraints, and deliverables, with later work building on earlier decisions. Recent work reuses design experience across tasks, yet alignment among artifacts within one project remains challenging. Agents therefore need to preserve and apply this intent across domains and as projects grow. We introduce Design Anything, a graph-based harness that preserves design intent outside the model and binds deliverables to the workflow steps that produce them through an intent–execution graph. A task contract retains constraints and decisions, selected designs become explicit inputs to later generations, and a deterministic controller uses bindings and execution records to check output types and counts, identify outstanding deliverables, and guide subsequent turns. We also introduce DesignAnything-Bench, with 228 tasks across 19 design domains, including 38 project families at scales of 6, 14, and 24 artifacts. With the LLM held fixed, Design Anything satisfies 85.0% and 92.5% of the brief's factual constraints with DeepSeek V4.1 Flash High and GPT-5.6 Sol High, respectively, exceeding the strongest same-model baselines by 15.4% and 14.4%. These gains reach 25.1% and 23.0% on 24-artifact tasks. Under each model, Design Anything leads or ties in constraint satisfaction in 18 of 19 domains. Professional designers also prefer its complete deliverable sets over those of same-model baselines in a blind evaluation.

Qualitative examples

Pick a project to see what Design Anything delivered and the workflow graph its agent built.

Product design · 24 artifacts

Design Anything (DeepSeek)
  • 18 pages
Open in Gemus
1 / 14

One brief, every system

Every system received the same animation brief; switch between them to compare what each delivered, slot by slot.

Open in Gemus
  1. Spec.
  2. Chars.
  3. Shots
  4. Color
  5. Keyframe S01
  6. Keyframe S03
  7. Keyframe S04
  8. Keyframe S06
  9. Storyboard: all
  10. Storyboard: Act 1
  11. Storyboard: Act 2
  12. Storyboard: Act 3
  13. Video
  14. Deck
  • Content error (selected examples, not an exhaustive audit)
  • Not credited (a file may still exist)
  • Video, shown by one frame
  • Keyframe S06: Pip still holds the circle case and Rowan the star case in S06, contrary to returning them.
Figure 6. Same-task comparison across five configurations.

As projects grow

The same architecture project at three sizes: Design Anything next to another system on the same model, artifact by artifact.

Model
Compared with
Project size

Constraint satisfaction: Benchmark average, 6 artifacts (38 tasks)

Design AnythingGPT-5.6 Sol High
97.4%
6 of 6 credited
  • Daytime commons
  • Evening community
  • Site plan
  • Floor plans
  • Site + building section
  • Four elevations
CodexGPT-5.6 Sol High
85.1%
6 of 6 credited
  • Daytime commons
  • Evening community
  • Site plan
  • Floor plans
  • Site + building section
  • Four elevations

Aerial imagery: USDA NAIP via USGS The National Map (image m_3708963_se_16_060_20190912, acquired 12 September 2019), public domain — https://imagery.nationalmap.gov/arcgis/rest/services/USGSNAIPImagery/ImageServer ; study boundary and annotations by the Design Anything benchmark authors. No government endorsement is implied.

Only the Design Anything (GPT) column is drawn in the paper’s figure. The other columns are each system’s scored run of the same brief, showing the file the evaluator matched to each requested artifact.

Figure 8. Design Anything (GPT-5.6 Sol High) on RIVER COMMONS with 6, 14, and 24 requested artifacts.

Method

Design Anything keeps the brief’s design intent in a graph outside the model and binds each deliverable to the workflow step that produces it.

Figure 3. Overview of Design Anything.

Results

With the same LLM, Design Anything leads the other systems on constraint satisfaction overall and leads or ties in nearly every domain.

Figure 9. The two Design Anything configurations have the highest estimated designer preference.
Show the full results table
Table 1. Results on DesignAnything-Bench (228 tasks).
SystemDelivery completionConstraint satisfactionReference fidelityIdentity consistencySet consistencyTechnical qualityAestheticsText renderingDomain criteriaDesigner preference
(a) DeepSeek V4.1 Flash High
Design Anything (DeepSeek)96.685.075.684.587.988.487.193.092.060.7
OpenDesign95.269.655.075.288.385.182.380.685.521.8
JarvisHub87.668.049.668.565.480.980.284.784.249.1
(b) GPT-5.6 Sol High
Design Anything (GPT)97.392.579.390.287.788.788.593.892.574.0
Codex98.178.171.375.389.888.987.994.390.844.4

Citation

@misc{ren2026designanything,
  title  = {Design Anything, Design at Scale},
  author = {Ren, Xuchao and Feng, Siyuan and Huang, Jing and Yao, Yanhua and Wang, Jing and Wang, Lingyu Alina and Yuan, Hangjie and Ma, Sai and Wu, Qiong and Feng, Tao},
  year   = {2026}
}