Imagera AI - AI content creation platform for generating images, cloning voices, and enhancing videos. Privacy Policy | Terms

Sign in
Illustration: two phones held up side by side to compareIllustration

Imagera Arena · Methodology

How the arena ranks models.

Same prompts, real runs, blind pairs, every number shown with its margin.

Vote nowFrom a prompt to a rank.

By the numbers

What the boards stand on.

Image models
Models
22
Prompts
30
Real outputs
660
Judge verdicts
12,000
People votes
0
Video models
Models
10
Prompts
6
Real outputs
60
Judge verdicts
540
People votes
0
Image editing
Models
19
Prompts
8
Real outputs
152
Judge verdicts
2,736
People votes
0

Step by step

From a prompt to a rank.

  1. Same prompts, real runs

    Every model renders the same 44 prompts through Imagera itself, at the settings its card opens with. Clips are asked for at 5 seconds. Nothing is retouched or picked: a run that fails counts in the model's failed-runs figure.

  2. Who competes

    Image: every general text-to-image model on Imagera released in the last 15 months. Image editing: every image editing model released in the same window, each given the same input photo and instruction. Video: the 10 text-to-video models people run most on Imagera. Style presets and vector-only variants are left out. New models join at the next benchmark run.

  3. Blind pairs

    You see two outputs of one prompt with no names, re-framed to one size. A vote counts once per pair per visitor, only on a pair we served you, after the pair has been on screen for a moment, and up to 100 votes a day.

  4. The rating

    All votes are fitted at once with a Bradley-Terry model and shown on the Elo scale, with each board's average at 1000. The bar's whiskers are a 95% bootstrap interval. A model is ranked after 30 votes in a board; a people's board ranks once 40 different visitors have voted on it.

  5. The AI judge

    Alongside people, an AI vision model judges every pair blind, twice, with the sides swapped, against the criteria below. Its verdicts are a separate board and are never mixed with people’s votes. The judge is a Gemini model, and Gemini models also compete on the video board: compare its board with people’s votes.

    • Prompt adherence: every requested element is present and correct; when text is requested, it is spelled exactly as written in the prompt.
    • Image quality: fine detail, correct anatomy and hands, no artifacts, warping or smearing.
    • Composition and lighting: framing, balance and light that suit the brief.

  6. Speed and price

    Median time is measured from submit to result on our own runs. The table’s price is the model’s card price at its default settings today; each output in the gallery shows what that run was charged (our runs used the faster lane, which can cost more).

Reading a score

How to read a rating.

  • Relative, not absolute

    A rating says how a model does against the other models on the same board. Each board's average is 1000; 100 points apart means the higher model wins about 64% of blind pairs.

  • Overlap means not settled

    When two models’ 95% intervals overlap, the board does not say one beats the other, and the page says “within the margin” instead of “leads”.

  • Never across boards

    An image rating and a video rating come from different prompts and different opponents: compare models within one board, never between boards.

  • Categories are their own boards

    A model that wins product shots can lose on posters: each category is fitted from its own prompts only.

What is judged

The criteria, board by board.

Image models

  1. Prompt adherence: every requested element is present and correct; when text is requested, it is spelled exactly as written in the prompt.
  2. Image quality: fine detail, correct anatomy and hands, no artifacts, warping or smearing.
  3. Composition and lighting: framing, balance and light that suit the brief.

Video models

  1. Prompt adherence: every requested element is present and correct; when text is requested, it is spelled exactly as written in the prompt.
  2. Image quality: fine detail, correct anatomy and hands, no artifacts, warping or smearing.
  3. Composition and lighting: framing, balance and light that suit the brief.
  4. Motion quality and temporal consistency: natural, physically plausible motion; people, objects and style stay the same from frame to frame, with no flicker or morphing.

Image editing

  1. Instruction adherence: every requested change is made; when text is requested, it is spelled exactly as written in the instruction.
  2. Preservation: what the instruction did not ask to change stays as in the INPUT photo — the same face and identity, the same product, the same composition.
  3. Image quality: fine detail, correct anatomy and hands, no artifacts, warping or smearing.
  4. Realism of the edit: new parts match the light, perspective and edges of the rest of the picture.

Rules

The rules in numbers.

Votes a model needs in a board before it is ranked30
Different visitors a people's board needs before it ranks anyone40
Bootstrap resamples behind each 95% interval200
Every board's average rating1,000
Seconds every video model is asked for5
Votes a visitor can cast a day100
Minutes between refits of every board10
Time a pair on the image board is on screen before a vote counts2.5 s
Time a pair on the video board is on screen before a vote counts6.0 s
Time a pair on the image editing board is on screen before a vote counts3.0 s

Limits

What the ratings do not cover.

  • Settings beyond each model’s defaults: a model may do better with tuning, a longer prompt or a seed search.
  • Features the prompts do not ask for: reference images, audio, very long clips, batch runs.
  • Cost at every setting: the card price is the default setting’s price today; a run at other settings costs what its card says then.
  • Taste outside the prompt set: the boards measure these prompts, chosen to cover common commercial jobs, not every use.

Disclosure

Who runs the arena.

Imagera runs this arena and sells runs of every model on it, including models that carry Imagera names. That is why every output is our own unedited run at the model’s default settings, every model gets the same prompts, names are hidden in every vote, the judge’s family is named, and every output, vote count and interval is public on the boards.

Questions

Methodology FAQ.

Why blind pairs instead of scores?

Choosing the better of two pictures is easier and more consistent than scoring one on a scale, and hiding the names keeps brand reputation out of the vote.

Why a Bradley-Terry model and not running Elo?

Running Elo depends on the order votes arrive in; fitting all votes at once does not, so the same votes always give the same board. The result is shown on the familiar Elo scale.

What does the 95% interval mean?

It is the range the rating falls in across 200 resamples of the prompts. When two models' intervals overlap, the board does not say one beats the other.

Why is there an AI judge?

People take time to cover every pair; the AI judge compares every pair blind, twice with the sides swapped, so a board has a ranking from day one. Its verdicts are a separate board and never mixed with people’s votes.

Can a model provider game the board?

The outputs are our own runs at the model’s default settings, every run is shown (failed runs included), names are hidden in the vote, and each visitor’s votes are capped.

The boards

See the rankings.

Each board has its own page: the full table, the price chart, every category and every output.

Illustration: a creator pinning favourite prints to a wallIllustration

Imagera Arena

Pick a winner. Then make your own.

Vote nowOpen the Sandbox