How the benchmark works

What is tested

Every model we track answers the same set of prompts, grouped by the job they stand for. The categories are: Typography & Posters, Portraits & Faces, Product & Packshot, Food & Drink, Landscape & Nature, Architecture & Interiors, Illustration & Anime, Logo & Graphic Design, Fashion & Editorial, 3D & Renders, Animals & Wildlife, Sci-Fi & Fantasy, Infographics & UI, Cars & Vehicles, Macro & Close-ups, Counting & Composition, Hands & Gestures, Glass, Water & Reflections, Signs & World Scripts.

Each prompt has its own page with every model's output side by side (all prompts), and the leaderboard ranks the models on their average score.

How each model is run

• Every model runs every prompt once, through the Wiro API, with the model's default settings. The first successful output is the one published: no second attempt for a better result, no cherry-picking.
• The aspect ratio comes from the prompt's category, so every model answers the same frame.
• A run that returns nothing, or an empty file, is tried again. A model that still returns nothing is marked as failed on that prompt and left out of its average.

How outputs are scored

Each output is rated from 0 to 5, in half steps, on 7 criteria. The scale is fixed rather than a curve against the other models: 5 is flawless on that criterion, 4 is good with one small imperfection, and anything with a clearly visible flaw is 3 or below.

• Prompt adherence25%
• Style fidelity13%
• Quality & detail15%
• Realism & anatomy15%
• Composition & framing12%
• Look & lighting10%
• Text rendering10%

A criterion that does not apply - text rendering when the prompt asks for no words in the image - is skipped and its weight is shared by the rest. A result's score is the weighted average of its criteria; a model's score on the leaderboard is the average over every prompt it was scored on.

Scores are given by an AI judge - a vision model working from the written rubric - not by the companies behind the models. The judge looks at the full image and never at the model's name. Every image is published next to its score, so anyone can check a score against the picture.

Limits

• One run per prompt: these models are random, and a second run can come out better or worse.
• An AI judge can be wrong about a single output. A model's average over many prompts is far more reliable than any one score.
• Providers sometimes update a model behind the same name. Every result carries the date it was run.
• Speed and price are not part of the score.

How it stays current

A notable new model is added within about 48 hours of reaching Wiro and is run on every prompt. New prompts are added only when every existing prompt has a result from every model.

Download the data

• Leaderboard (CSV) - one row per model: score, prompts scored, the average on each criterion and its best category.
• Every result (CSV) - one row per prompt and model: status, score, each criterion, the judge's note and the dates.

The scores and rankings are licensed under CC BY 4.0: use, share and adapt them, with credit to Image Models. Images and videos remain subject to the terms of the model that produced them.

How to cite

Wiro AI (2026). Image Models benchmark: AI image model leaderboard, updated 30 October 2026 [Data set]. https://imagemodels.ai/leaderboard/. Licensed under CC BY 4.0.

@misc{imagemodels2026,
  author       = {{Wiro AI}},
  title        = {Image Models benchmark: AI image model leaderboard},
  year         = {2026},
  howpublished = {\url{https://imagemodels.ai/leaderboard/}},
  note         = {Data set, updated 2026-10-30. CC BY 4.0}
}

Spotted a wrong score or a broken result? Tell us and we will look at it.