AI Image Arena / About the data

Last published Nov 17, 2026

About the AI Image Arena data

The AI Image Arena pages mirror LMArena's public image leaderboards - 14 boards and 103 models - so you can compare image generators and editors in one place, alongside guides for fashion and e-commerce imagery.

Where the numbers come from

  • LMArena (arena.ai) collects blind pairwise votes: two anonymous models answer the same prompt and the voter picks the better image. Names are shown only after the vote.
  • LMArena publishes the resulting leaderboards as an open dataset on Hugging Face (lmarena-ai/leaderboard-dataset) under CC BY 4.0. We republish every board unchanged, with attribution.
  • Two boards - Image Edit › Single Image › Product, Branding & Commercial Design and › Photorealistic - are shown on arena.ai but not split out in the dataset; we transcribe them from the arena.ai pages on the same publish date.

How to read a board

  • Score: LMArena's Bradley-Terry rating (Elo-like scale). Higher means people preferred the model's images more often.
  • ± CI: the 95% confidence interval. Narrow intervals come from many votes.
  • Rank spread: every rank a model could plausibly hold given the intervals - computed by us from LMArena's published bounds. Models with overlapping intervals are statistically tied.
  • Votes: how many head-to-head votes the model has on that board.

Updates

  • We re-sync from LMArena's dataset on every site deploy, and the board date always shows LMArena's publish date.
  • Flashlook does not run the votes, choose the prompts or change any score. For questions about methodology, see LMArena.

Source and license

Rankings © LMArena, published as lmarena-ai/leaderboard-dataset under CC BY 4.0. Flashlook is not affiliated with LMArena.