| Date | Update |
|---|---|
| 2026-08-15 | Added official DeepSeek Harness support for coding agents |
| 2026-08-14 | Initial public release |
Learn a reusable DesignHarness around fixed models, then turn one paper into an editable poster, slides, webpage, and narrated + captioned video.
Paper · arXiv:2608.13560 ↗ · Dataset · PosterBench ↗ · Dataset · PosterBench-mini ↗
✦ Explore the AutoDesign story ↗ · Open the Demo Page ↗
Demos · Quickstart · Paper Suite · Methodology · PosterBench · Human Evaluation · Outputs
These are real outputs, not mockups. AutoDesign turned its own paper into the paper's Figure 2 poster, a 24-slide formal academic talk, a complete editorial research webpage, and a six-minute 1080p conference video.
 Poster · AutoDesign Figure 2 of the paper: an information-dense, editable academic poster made by AutoDesign for itself. Open full poster PDF ↗ |
 Slides · AutoDesign A complete 24-slide formal academic conference talk made by AutoDesign for itself. Open full slide deck PDF ↗ |
 Webpage · AutoDesign An editorial research experience that turns the paper's method, evidence, results, and limitations into an interactive story. Download full landing page ↗ |
 Video · AutoDesign A six-minute 1080p conference video introducing Meta-Harness Optimization, DesignHarness, and PosterBench. Watch MP4 ↗ |
Prerequisites: Node.js 22+ and ffmpeg/ffprobe.
curl -fsSL https://designanything.ai/install.sh | bash
autodesign startThe launcher installs under ~/.local/share/autodesign, keeps state under
~/.autodesign, serves the bundled web/dist, and opens a browser. Run
autodesign doctor to check the installed runtime. Existing
~/.designanything state is migrated with a compatibility symlink. If the
hosted endpoint is unavailable, use the source setup below.
Requirements: Python 3.10+, uv, Node.js 22+,
npm, and ffmpeg/ffprobe for Video.
uv sync
uv run python scripts/install_playwright_browsers.py
cd runtime/video && npm ci --omit=dev
cd ../../web && npm installConfigure provider keys in .env or enter them in the Web UI Settings drawer.
Do not replace an existing .env during an update.
Start the backend:
uv run uvicorn scripts.web_server:app --reload --port 8000In another terminal, start the frontend:
cd web
npm run devOpen localhost:5173. Backend health is available at
/api/health.
Upload one PDF and choose Paper All-in-One to launch the poster, slides, webpage, and narrated-video tracks together.
uv --cache-dir .uv-cache run python -m autodesign run \
"Create a dense academic conference poster from the attached paper." \
--from-file /absolute/path/to/paper.pdf \
--template cvpr-landscapeInspect final/poster.html, final/preview.png, the final manifest, and
run_events.jsonl. A file can exist after a fallback, so terminal status and
validation feedback remain part of the result.
Use a visual reference
uv --cache-dir .uv-cache run python -m autodesign run \
"Create a paper poster using the reference's visual system." \
--from-file /absolute/path/to/paper.pdf \
--reference-poster /absolute/path/to/reference.pngReference posters transfer visual systems only. Their text, claims, logos, QR codes, figures, tables, and links never become paper evidence.
Finish the paper once. Paper All-in-One packages the same source into everything that usually comes next: a promotional webpage, conference slides, an academic poster, and a narrated video with timed subtitles. No need to rebuild the paper's story for each format.
Generate the complete paper suite ↗
Follow a guided local walkthrough: configure the Workbench, launch Paper All-in-One, inspect the run, and enter each editable canvas. You can also try the online demo in your browser; for the complete, most reliable experience, we recommend installing AutoDesign locally.
Guided local walkthrough · Paper All-in-One → editable canvases
AutoDesign-readme-demo-tutorial.mp4
Open the academic poster wall
Claude 4.8 authoring route
Codex GPT-5.5 xhigh authoring route
- The whole paper journey, in one workflow. Build the promotional webpage, talk deck, conference poster, and narrated + captioned video from the same source instead of restarting four times.
- Editable by default. HTML, native text, tables, and named assets remain available for revision instead of being flattened into one image.
- Source-grounded. Claims, figures, and tables retain provenance beside the run; a reference can transfer style, never evidence.
- Optimizes the system, not model weights. Complete trajectories expose recurring failures, while meta-harness optimization improves one reusable DesignHarness component at a time.
- Inspectable and local-first. Events, manifests, candidates, validation feedback, and final files stay available on your machine.
A design harness is the system around a fixed LLM or MLLM that turns a multimodal source into a human-facing artifact through an execution trajectory. A meta-harness improves that surrounding system. AutoDesign therefore learns from complete rollouts while keeping the underlying model weights fixed. Before autonomous optimization, an evaluator coding agent uses human-annotated reference artifacts across seven quality dimensions to implement a fixed optimization-time evaluator. It combines rule-based checks with VLM judgments and remains distinct from the frozen PosterBench protocol used for final system comparison.
Autonomous outer-loop iterations evolve the harness through rollout, evaluation, one-component update proposals, and acceptance. After autonomous optimization reaches a plateau, optional Human-in-the-loop guidance can redirect the search and further improve production poster quality.
| Loop | What it improves | Evidence and update |
|---|---|---|
| Inner loop · artifact generation | One editable artifact under a fixed design harness | A Designer revises the artifact; a Critic returns feedback; their interactions form an execution trajectory |
| Outer loop · harness optimization | The reusable design harness across tasks | The MetaHarnessOptimizer analyzes trajectories, evaluator scores, the persistent optimization record, and optional human guidance |
Every outer-loop iteration follows four stages: rollout → evaluation → update proposal → acceptance. The optimizer acts as a planner and code editor, updates exactly one harness component, and retains the candidate only when training performance improves without reducing performance on an independent development set. Development trajectories are hidden from the update proposer.
Human-in-the-loop guidance is optional. A user can give the planner observations or high-level directions to redirect a stalled search; explicit human input can also correct a systematic evaluator bias. Without guidance, the outer loop runs autonomously.
| Component | Elements optimized by the meta-harness |
|---|---|
| Context and Memory | Multimodal source management, task prompts, skills, reusable assets, and persistent revision state |
| Tools and Specifications | Tools and editable-artifact specifications for layout, typography, and provenance |
| Execution Runtime | The workspace and runtime for authoring, rendering, validating, and exporting |
| Orchestration | Task routing, attempt budgets, loop control, candidate selection, fallback, and finalization |
| Evaluation and Feedback | Rule-based validation, model-based critique, and localized revision feedback |
Meta-harness optimization yields DesignHarness, the reusable artifact-producing system. Its four stages are source ingestion, iterative artifact generation and revision, validation with dual critics, and finalization. Paper metadata, claims, figures, tables, and source locations become provenance-aware context; a coding-agent Designer edits native HTML; a rule-based validator and VLM critic return localized feedback; and the best valid candidate is made self-contained for delivery.
The current implementation permits up to 12 refinement attempts. Blocking checks cover unsafe or missing assets, broken provenance, severe overflow or overlap, and required typography or layout constraints. If no candidate passes within the budget, the retained attempt history supports a constrained fallback before the same finalization stage.
The latest paper traces one poster run through five selected attempts. The critic identifies a clipped analysis lane at A1; A3 restores the fit, A5 refits the header, A6 rescales evidence, and A9 preserves the repaired composition and is accepted. The trajectory shows that diagnostics drive localized edits while valid layout and source-derived content survive across revisions.
PosterBench evaluates a 100-paper large set and a fixed 10-paper small set across AI/ML, biomedicine and health, climate and earth environment, economics and policy, and physics and astronomy. Every output is rendered to a common poster format before scoring.
The metadata-only manifests are released on Hugging Face as
YaxinLuo/PosterBench
and
YaxinLuo/PosterBench-mini.
They can be downloaded or loaded directly with datasets without
redistributing the underlying paper PDFs.
The seven dimensions are Faithfulness, Coverage, Density, Visual Evidence, Layout, Readability, and Aesthetics, weighted 10/10/15/10/20/25/10. Programmatic evidence and source-conditioned VLM judgments are aggregated first; then the strictest active ceiling for severe layout damage, insufficient presentation viability, confirmed visible failure, or protected render integrity is applied to each poster.
AutoDesign achieves the two highest PosterBench Scores. With Claude Code and Claude 4.8 fixed, it scores 78.32, exceeding Claude Design by 7.45 points and OpenDesign by 8.87 points.
| Rank | Score | System | Design harness | Coding agent | Model |
|---|---|---|---|---|---|
| 1 | 78.32 | AutoDesign | DesignHarness | Claude Code | Claude 4.8 |
| 2 | 77.97 | AutoDesign | DesignHarness | Codex | GPT-5.5 |
| 3 | 73.37 | Codex | — | Codex | GPT-5.5 |
| 4 | 70.87 | Claude Design | Claude Design | Claude Code | Claude 4.8 |
| 5 | 70.01 | Claude Code | — | Claude Code | Claude 4.8 |
| 6 | 69.45 | OpenDesign | OpenDesign | Claude Code | Claude 4.8 |
| 7 | 62.17 | OpenDesign | OpenDesign | Codex | GPT-5.5 |
| 8 | 61.14 | Doubao | — | Claude Code | Seed 2.1 |
| 9 | 56.71 | PosterGen | — | — | Claude 4.8 |
| 10 | 52.22 | GLM | — | Claude Code | GLM 5.2 |
| 11 | 51.46 | Kimi | — | Claude Code | Kimi K2.7 |
| 12 | 49.09 | Any2Poster | — | — | Claude 4.8 |
| 13 | 46.01 | DeepSeek | — | Claude Code | DeepSeek V4 Pro |
| 14 | 44.61 | Paper2Poster | — | — | Claude 4.8 |
Open the Small-Scale Benchmark Main Track · fixed 10-paper subset
| Rank | Score | System | Design harness | Coding agent | Model |
|---|---|---|---|---|---|
| 1 | 81.46 | AutoDesign | DesignHarness | Codex | GPT-5.5 |
| 2 | 75.87 | Codex | — | Codex | GPT-5.5 |
| 3 | 74.56 | AutoDesign | DesignHarness | Claude Code | Claude 4.8 |
| 4 | 70.36 | OpenDesign | OpenDesign | Claude Code | Claude 4.8 |
| 5 | 69.55 | Claude Code | — | Claude Code | Claude 4.8 |
| 6 | 66.83 | Claude Design | Claude Design | Claude Code | Claude 4.8 |
| 7 | 60.58 | OpenDesign | OpenDesign | Codex | GPT-5.5 |
| 8 | 57.20 | Kimi | — | Claude Code | Kimi K2.7 |
| 9 | 54.01 | Doubao | — | Claude Code | Seed 2.1 |
| 10 | 51.82 | PosterGen | — | — | Claude 4.8 |
| 11 | 50.32 | GLM | — | Claude Code | GLM 5.2 |
| 12 | 46.88 | Any2Poster | — | — | Claude 4.8 |
| 13 | 42.06 | Paper2Poster | — | — | Claude 4.8 |
| 14 | 34.73 | DeepSeek | — | Claude Code | DeepSeek V4 Pro |
Each controlled track varies one factor while holding the others fixed.
| Rank | Design Harness Track Fixed: Claude Code + Claude 4.8 |
Score | Coding Harness Track Fixed: AutoDesign + GLM 5.2 |
Score | Model Track Fixed: AutoDesign + Claude Code |
Score |
|---|---|---|---|---|---|---|
| 1 | AutoDesign | 74.56 | Kimi Code | 82.31 | Claude 4.8 | 74.56 |
| 2 | OpenDesign | 70.36 | ZCode | 69.53 | Seed 2.1 Pro | 71.83 |
| 3 | Claude Design | 66.83 | OpenCode | 67.87 | Kimi K2.7 | 70.12 |
| 4 | — | — | Claude Code | 64.33 | GLM 5.2 | 64.33 |
| 5 | — | — | — | — | LongCat 2.0 | 55.13 |
| 6 | — | — | — | — | DeepSeek V4 Pro | 54.29 |
Across seven matched model–coding-agent configurations, attaching DesignHarness improves every PosterBench Score by +5.01 to +19.56 points. Native Codex–GPT-5.5 rises from 75.87 to 81.46 (+5.59); Claude Code–Kimi K2.7 rises from 57.20 to 70.12 (+12.92); and the largest gain is +19.56 for Claude Code–DeepSeek V4 Pro.
On the fixed 10-paper subset, the observed Pareto frontier runs from LongCat 2.0 (55.13 at $0.27/poster) through Doubao Seed 2.1 Pro (71.83 at $2.75) and Claude 4.8 (74.56 at $7.63) to GPT-5.5 (81.46 at $10.02). Doubao reaches 88% of the GPT-5.5 score at 27% of its normalized designer-only API cost.
For the executable protocol, data preparation, score ownership, record-level ceilings, and reproduction commands, see the PosterBench evaluation guide.
The fully system-blind study collected 936 responses from 11 volunteer reviewers: 933 ranking judgments and three skips. AutoDesign has the highest Bradley–Terry estimate at 64.0%, with a 55.2–77.8% 95% interval. Its tie-adjusted empirical preference is 61.3% against Claude Code, 63.1% against OpenDesign, and 67.6% against Claude Design.
PosterBench is positively, though imperfectly, associated with human preference (r = 0.34, 95% interval 0.22–0.44). Agreement with the PosterBench-preferred poster rises from 51.9% for 0–3-point gaps to 74.4% when the score gap is at least 20 points.
The current DesignHarness already produces pilot paper-to-slide, paper-to-webpage, and paper-to-conference-video artifacts, but PosterBench formally validates academic posters only. Slides, webpages, and videos still need medium-specific source–output data, evaluators, rendering and validation gates, and communication objectives before their research claims match the poster pipeline.
Longer term, AutoDesign aims toward multimodal-in, multimodal-out agentic design: integrating papers, visual evidence, code, data, and human guidance to create medium-specific outputs. Open research problems include better component selection, evaluator evolution anchored by frozen tasks and human audits, and combining harness optimization with model post-training.
We welcome researchers, designers, and engineers to contribute new design harnesses, refinement workflows, evaluators, and artifact capabilities.
Contribute on GitHub ↗ · Explore the project ↗
The Web UI provides Paper All-in-One generation, model and provider settings, progress streaming, cancel and retry, server-backed history, and direct editing for supported HTML-first artifacts.
Start the interactive CLI with:
uv --cache-dir .uv-cache run python -m autodesign| Use case | Primary output |
|---|---|
| Academic paper poster | final/poster.html, final/preview.png, optional PDF |
| Slide deck | final/deck.html, final/deck.pdf, slide previews |
| Landing or project page | final/index.html, final/preview.png |
| Video | Editable HyperFrames project, narrated MP4 with AAC audio, transcript, and timed SRT/VTT subtitles |
| Creative poster | HTML/PNG, with legacy PSD/SVG paths where supported |
| Research reproduction handoff | OpenResearch project, session, and report links |
Single-run output lives under out/runs/<run_id>/; EvaData batch output lives
under out/eva_poster_batches/<batch_id>/. Both locations are ignored by Git.
The canonical Python module and installed launcher are autodesign. The
design_anything module, design-anything console command, designanything
launcher, and DESIGN_ANYTHING_* environment variables are deprecated
compatibility aliases. New configuration and automation should use
AUTODESIGN_*.
AutoDesign is made possible by the open-source community. We are especially grateful to:
- HyperFrames for the HTML-first video runtime, composition linting, and MP4 rendering.
- KaTeX for offline mathematical typesetting in portable HTML artifacts.
- html-ppt-skill for the MIT-licensed deck-authoring reference assets adapted in this repository.
MIT. Bundled third-party assets retain their own licenses; see Third-Party Notices for details.