Image Generation Benchmarks¶
Text-to-image artificial intelligence (AI) models are judged differently from language models, because there is no answer key for a picture. The field's standard is the arena: people see two images generated from the same prompt, without knowing which model made which, and vote for the better one. Depending on the site, the voters are members of the public or a recruited panel. Votes become Elo-style ratings, a close relative of the system used to rank chess players.
That measures human preference at scale, which is useful for general illustration. The caveat is that preference rewards what looks good, not what is accurate, so look past the overall rank for:
- Prompt adherence: whether the image contains what was asked for, in the arrangement asked for.
- Text rendering: legible, correctly spelled labels, which vary widely between models. Some leaderboards now rank text rendering as its own category.
- Anatomy: counts and spatial relations (fingers, joints, teeth, ribs), worth checking in every image.
- Style range: whether a model can produce a clean schematic as well as a photograph.
A model can win on beauty and lose on every one of those.
The short narrated film How to Read a Leaderboard, on the Language Model Benchmarks page, covers image and video arenas as well.
Image leaderboards also rank image editing (changing a supplied picture by instruction) separately from generating one from text, so check which task a ranking covers.
How an arena ranks models
One prompt
"a heart, labeled"
Image A
model name hidden
Image B
model name hidden
A person votes
better one wins
Many votes
become Elo-style ratings, like those that rank chess players
Blind at the vote, honest in aggregate: it measures preference at scale.
Preference rewards what looks good, which is not always what is accurate.
The Leaderboards Worth Knowing¶
-
Artificial Analysis: Text-to-Image
Rankings from blind head-to-head votes, with proprietary and open-weights models ranked together and an uncertainty range beside each score.
-
Artificial Analysis: Image Models
The companion view, adding generation speed and price per 1,000 images to the quality ratings, with charts of quality against each.
-
Arena: Text-to-Image
The arena that popularized blind-vote ranking, applied to image generation, run by the Arena team (formerly LMArena).
Medical Images Are a Different Question¶
No public leaderboard ranks general-purpose image generators on the medical accuracy of what they draw, and the peer-reviewed evidence says the gap between pretty and correct is wide:
- A 2025 study in the Journal of Pediatric Ophthalmology and Strabismus had general text-to-image models depict common pediatric eye pathologies and scored them against human medical illustration: the generated images rated poorly overall and worse on pathological accuracy specifically (Jong et al.).
- A clinical perspective in Cureus draws the practical conclusion: generated imagery may eventually enrich patient communication and teaching, but inaccuracy and bias call for caution, and it should support clinical judgment, never replace it (Javan et al., 2024).
- The technical literature on medical image synthesis and translation is advancing quickly, with its own evaluation metrics distinct from preference arenas; a 2025 review in Medical Image Analysis maps that landscape (Chen et al., 2025).
- For teaching figures specifically, the AI-Generated Images in Teaching playbook covers the anatomy evaluations, what to reach for instead, and when a flawed image is defensible in a session.
The practical guidance for the American University of Antigua College of Medicine (AUACOM) follows directly: treat general-purpose image models as illustration tools, not anatomy references. Before a generated image goes in front of students, it is worth checking its anatomy and pathology, yourself or with a colleague who knows the area, since pathological accuracy is where the studies above found generated images weakest. Labeling it as AI-generated tells students what they are looking at.
To run open image models yourself, see Running Models Locally. Image generation tools are listed under Image Generation in the tools directory. For language model rankings, see Language Model Benchmarks; for video generation, the Video Generation Benchmarks.
Content last reviewed October 12, 2026. Review dates are maintained as described on the About page.