Datasets:

dh7
/

imagebench

license: cc-by-4.0
task_categories:
  - text-to-image
language:
  - en
tags:
  - benchmark
  - evaluation
  - text-to-image
  - vlm-eval
  - image-generation
  - diffusion-models
  - leaderboard
pretty_name: ImageBench
size_categories:
  - 1K<n<10K
configs:
  - config_name: verdicts
    data_files: verdicts.csv
  - config_name: scores
    data_files: scores.csv
  - config_name: prompts
    data_files: prompts.csv

ImageBench — 50 Text-to-Image Models Judged by VLMs on 192 Prompts

Reproducibility dataset for imagebench.ai: the 192-prompt V1.2 benchmark, per-(model, prompt) VLM verdicts, and per-model aggregate scores for 50 text-to-image models.

Live leaderboard + every generated image: https://imagebench.ai Methodology: https://imagebench.ai/methodology-v1 Reproducibility repo: https://github.com/dh7/image-bench-ai

What's in this dataset

File	Rows	Description
`prompts.csv`	64 tests × 3 variants	The V1.2 benchmark: prompt variants, category/subcategory/difficulty, VLM judge questions, evaluation criteria, VLM-routing assignment.
`verdicts.csv`	9,600 (50 models × 192 prompts)	Per-image pass/fail verdict, VLM judge reasoning, HPS aesthetic score, VLM model used.
`scores.csv`	50	Per-model aggregate: Overall (0-100), pass rate, aesthetic Estimated Preference Score, per-category pass rates.

Column reference

`verdicts.csv`

model_slug — canonical model identifier (matches URLs on imagebench.ai)
model_name — human-readable name
category / subcategory / difficulty — benchmark taxonomy
prompt_id — matches image filename on imagebench.ai/gallery
prompt — the actual prompt sent to the model
verdict — PASS or FAIL from the VLM judge
vlm — which VLM produced the verdict (Qwen 3.5 122B / Gemini 3.1 Pro / etc.)
judge_response — the VLM's textual reasoning for the verdict
hps_mu — HPSv3 aesthetic score (higher = judged more aesthetically pleasing by human-preference model)
hps_pref — pairwise HPS preference metadata

`scores.csv`

model_slug, model_name
overall — 0-100 blended score (capability + aesthetic preference)
pass_rate — % of prompts passing the VLM verdict
eps — Estimated Preference Score (aesthetic, based on HPSv3)
pass_<category> — pass rate per category (text rendering, spatial reasoning, human realism, truthfulness, studio quality, graphical design)

How this was produced

Prompts: 64 tests × 3 variants, hand-curated across 6 capability categories with binary judge questions per test.
Generation: each model produces 192 images (one per prompt) via its native API — no cherry-picking, no re-runs.
Judging: VLM-routed — Qwen 3.5 122B handles most categories, with per-category routing to specialists (Gemini 3.1 Pro for hands, etc.) based on calibration.
Aesthetic score: HPSv3 (Human Preference Score v3) applied to every generated image.
Overall: 0.5 × normalized pass rate + 0.5 × normalized EPS.

Full methodology + calibration study: https://imagebench.ai/methodology-v1

Known limitations

Text-to-image only — no editing, inpainting, video, long-form coherence.
Prompt bias — the 192-prompt set is weighted toward tests that separate models. "Every model passes" mundane prompts are under-represented, so pass rates are lower than typical real-world use.
VLM judge bias — VLMs have their own systematic errors. We validate on a hands-specific calibration blog: https://imagebench.ai/blog/hands-benchmark-qwen35-122b
Category weighting — Overall gives each category equal weight, but "text rendering" is inherently harder to pass than "professional studio quality". Interpret cross-category comparisons carefully.

Citation

If you use this dataset in research or writing, please cite:

ImageBench: 50 Text-to-Image Models Judged by VLMs on 192 Prompts
https://imagebench.ai

License

CC-BY 4.0 — free to use with attribution.

Feedback

Issues, corrections, or methodology critique: open an issue at https://github.com/dh7/image-bench-ai or leave a comment on this dataset.