Funovate-Technology/ai-design-benchmark

From FontVibe — the write-ups are on the FontVibe blog. Mirror of Funovate-Technology/ai-design-benchmark (canonical copy, DOI 10.5281/zenodo.23803802); corrections go to the GitHub repo's issues. Hugging Face's web uploader keeps files flat, so each benchmarks/2026-10-<name>/<file> from GitHub is stored here as <name>__<file> (e.g. llm-designer__designs.jsonl); the analysis scripts are on GitHub.

AI Design Benchmark

Open data from FontVibe's tests of AI models doing design work with text: logos, posters, signs, invitations, thumbnails and packaging. Every brief, every output, every image, every score and the written reason behind each score.

Each benchmark asks one question: a client hands over a brief; can they use what the model delivers, and are they happy with it? The full write-up of each test is on the FontVibe blog; this repository is the raw material behind it.

  3   benchmarks
  8   language models tested as designers   (2026-10-llm-designer)
  5   image models at 8 settings tested      (2026-10-grok-imagine-2, 2026-10-qwen-image-3)
  494 images; every scored one has four scores and a written reason

The benchmarks

Benchmark	What it tests	Article	Data
2026-10-llm-designer	8 language models (GPT-6.1 Terra, GPT-5.6 Terra, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Claude Sonnet 5.5, Gemini 4 Argon, Gemini 3.8 Flash) acting as the designer: reading a brief, proposing directions and writing the image instructions. The image model is fixed. 12 briefs × 3 runs: 288 runs, 269 designs ranked, plus 18 side-note reruns.	GPT-6.1 Terra vs Claude & Gemini: Design Test	benchmarks/2026-10-llm-designer
2026-10-qwen-image-3	5 image models on 10 design jobs with text: Qwen Image 3.0 Pro, GPT Image 2.5 Sunburst, Nano Banana 2.1, Seedream 5.0 Pro, Grok Imagine 2.0	Qwen Image 3.0 vs GPT Image, Nano Banana, Seedream & Grok	benchmarks/2026-10-qwen-image-3
2026-10-grok-imagine-2	4 image models (Grok Imagine 2.0 at Low and Medium, GPT Image 2.5 Sunburst, Nano Banana 2.1, Seedream 5.0 Pro) on 10 design jobs with text	Grok Imagine 2.0 text test	benchmarks/2026-10-grok-imagine-2

How designs are scored

Every design gets four scores from 1 to 10. The overall score is calculated, not judged:

overall = 0.20 × legibility + 0.35 × lettering and finish + 0.25 × information and layout + 0.20 × appeal

The judge plays the client who wrote the brief: is it the thing I asked for; can I use every piece of text as it is; does it have the feeling I wanted; are the details well made; does it impress at first sight. Text the client never gave (founding years, slogans, figures) and placeholders printed as is count against a design. The full standard is in SCORING.md. It covers designs with text; images without text are out of scope.

Reproduce the numbers

python3 scripts/analyze.py benchmarks/2026-10-llm-designer
python3 scripts/verify.py

Standard library only. analyze.py recomputes the tables of the llm-designer benchmark: the per-model table with 95% bootstrap confidence intervals (resampling briefs, fixed seed), cost, time and reliability, and all pairwise comparisons; it writes summary.csv and pairs.csv. The two image-model benchmarks publish every score but have no analysis script. verify.py checks counts, score arithmetic, attempts and that every image URL resolves.

What's in a benchmark folder

2026-10-llm-designer/

File	Contents
briefs.json	The 12 briefs, their layer, language and the text that must appear
models.json	Model IDs, providers, reasoning effort and public prices on the test date
designs.jsonl	One line per run (288): status and failure cause, the three proposed directions, the image instructions, time, tokens, retries and cost of each step, the image, the four scores, the judge's text checks and reasons
attempts.jsonl	Every attempt (300), with the cause of each replaced attempt
opus-budget-2x.jsonl	18 reruns of Claude Opus 5.5 with the output limit doubled (a side note, not in the ranking)
blind-pick.json	One person's anonymised pick per brief (a side note, not in the ranking)
summary.csv, pairs.csv	Per-model table and all pairwise comparisons, written by analyze.py

The image-model folders hold tasks.json (prompts and required text) and images.jsonl (every image generated, which ones were scored, the scores and reasons).

Judges' reasons are in Chinese (reasons_zh, reason_zh), as written at scoring time; the judge's instructions were in Chinese too (SCORING.zh.md). The benchmark folder's README lists how runs were handled, the statistics and the limits.

In the two image-model folders, "scored": false images were extra samples used for consistency checks, and designer-grok2-low images went through FontVibe's designer mode (shown in the article, not scored). Four generations in 2026-10-qwen-image-3 failed (one Nano Banana 2.1, three Seedream 5.0 Pro); their causes were not recorded. Each task and setup has one scored image, so those tables have no confidence intervals.

Images

Images are hotlinked from cos.fontvibe.ai/opensource/ai-design-benchmark/…. Each record also links the design's page on FontVibe (fontvibe_page), where it can be viewed full size and remade. Images are not part of this repository and remain subject to FontVibe's terms.

Attribution

The data is CC BY 4.0; the scripts are MIT. If you publish results, charts or examples from this data, please credit it like this, linking the article of the benchmark you used:

Markdown

Data: [FontVibe AI Design Benchmark](https://fontvibe.ai/blog/gpt-6-1-terra-vs-claude-gemini-design-test) ·
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)

HTML

Data: <a href="https://fontvibe.ai/blog/gpt-6-1-terra-vs-claude-gemini-design-test">FontVibe AI Design Benchmark</a> ·
<a href="https://creativecommons.org/licenses/by/4.0/">CC BY 4.0</a>

Plain text

Data: FontVibe AI Design Benchmark — https://fontvibe.ai/blog/gpt-6-1-terra-vs-claude-gemini-design-test — CC BY 4.0

• Please link the article, not just this repo. The article shows the designs side by side and explains the method; that is what your readers will want.
• rel="nofollow" is fine. We would rather be credited with a nofollow link than not at all.
• A link is not strictly required where it isn't practical, such as in print.

Citing it

See CITATION.cff. GitHub shows a "Cite this repository" button for it. Every release is archived on Zenodo: 10.5281/zenodo.23803802 always points to the latest version.

Who made this

FontVibe is an AI design tool for text effects and lettering, made by Funovate Technology. FontVibe uses some of the models tested here in production, and the llm-designer benchmark ran through FontVibe's own pipeline; both are stated in the articles. FontVibe's production system prompts are not published; researchers who need them can ask at hello@fontvibe.ai. Corrections and reruns are welcome: open an issue or email hello@fontvibe.ai.

Changes

• 1.0.1 (2026-10-31): corrected the causes of the 19 failed llm-designer runs (17 cut off at the limit, 1 invalid image instructions, 1 empty); added attempts.jsonl (every attempt, with provider errors, front-desk reruns, internal retries and cost); confidence intervals now resample briefs, and all 28 pairwise comparisons are published (pairs.csv); cost now also reported as total spend per finished design; price sources and token-field notes in models.json; must_appear in English; the judge's original Chinese instructions (SCORING.zh.md).
• 1.0.0 (2026-10-30): first release.