Sys.Intel // Generators/Weekly snapshot

The best AI image generators, ranked.

78 text-to-image models scored from 2.4M blind, side-by-side human votes on Arena. Refreshed every week — so you always know which generators an AI detector has to keep up with.

78
Models ranked
2.4M
Human votes
Sep 7, 2026
Data as of
8
Categories
  1. #1OpenAIPreliminary
    gpt-image-2.5-sunburst
    Proprietary · 1,198 votes
    1440
    ±21
  2. #2OpenAIPreliminary
    gpt-image-2.5-flare
    Proprietary · 1,149 votes
    1428
    ±21
  3. #3OpenAI
    gpt-image-2 (medium)
    Proprietary · 31,995 votes
    1400
    ±6
Text-to-Image Arena · Cartoon, Anime & Fantasy

Cartoon, Anime & Fantasy

Stylised illustration, anime and fantasy art.

2,437,796 votes · 78 models · as of Sep 7, 2026

Cartoon, Anime & Fantasy rankings

RankModelScore
1
OpenAI·Proprietary
1440±21
2
OpenAI·Proprietary
1428±21
3
OpenAI·Proprietary
1400±6
4
Microsoft AI·Proprietary
1348±10
5
SpaceXAI·Proprietary
1322±19
6
Reve·Proprietary
1310±11
7
Meta·Proprietary
1286±8
8
Alibaba·Proprietary
1276±10
9
Reve·Proprietary
1275±9
10
Bytedance·Proprietary
1271±7
11
Microsoft AI·Proprietary
1270±6
12
1263±7
13
1258±9
14
OpenAI·Proprietary
1245±5
15
1242±5
16
1238±7
17
Ideogram·Ideogram open model
1195±6
18
Luma AI·Proprietary
1194±9
19
Alibaba·Proprietary
1191±9
20
Luma AI·Proprietary
1190±7
21
1183±11
22
Microsoft AI·Proprietary
1180±7
23
Nvidia·OpenMDW 1.1
1179±9
24
Recraft·Proprietary
1177±19
25
SpaceXAI·Proprietary
1176±4

Showing 25 of 78 models · data as of Sep 7, 2026

Source: Arena
Statistics

How settled is the top 20?

Arena scores come with a 95% confidence interval. Where intervals overlap, the order between two models is not statistically settled — that is exactly what the “rank spread” column in the table expresses.

Arena score with 95% confidence interval

Top 20 models · hover a rank for details

115012001250130013501400145015001234567891011121314151617181920RANK
#1gpt-image-2.5-sunburstOpenAI1440 ±2114191460 · 1,198 votes

Average win rate against every opponent

Ties excluded · as published by Arena · a model can out-rank another yet win less often, because it has faced stronger opponents

  1. #1
    gpt-image-2.5-sunburst
    86.2%
  2. #2
    gpt-image-2.5-flare
    76.3%
  3. #3
    gpt-image-2 (medium)
    75.0%
  4. #4
    mai-image-2.6
    60.6%
  5. #9
    reve-2.0
    60.5%
  6. #6
    reve-2.1
    59.1%
  7. #5
    grok-imagine-image-2.0 (low)
    59.1%
  8. #7
    muse-image
    53.2%
  9. #16
    gemini-3-pro-image-preview (nano-banana-pro)
    47.7%
  10. #13
    gemini-3.1-flash-lite-image (nano-banana-2-lite)
    47.5%
  11. #12
    gemini-3.1-flash-image (nano-banana-2) [web-search]
    45.7%
  12. #8
    qwen-image-3.0-pro
    45.4%
  13. #11
    mai-image-2.5
    45.2%
  14. #10
    seedream-5.0-pro
    44.7%
  15. #14
    gpt-image-1.5-high-fidelity
    42.1%
  16. #15
    gemini-3-pro-image-2k (nano-banana-pro)
    41.9%
  17. #19
    qwen-image-2.0-pro-2026-06-22
    34.4%
  18. #17
    ideogram-4.0-quality
    34.1%
  19. #18
    uni-1.1-max
    31.9%
  20. #20
    uni-1.1
    31.1%
Why we track this

Every model on this board is a model our detector has to catch.

Sightova builds AI image detection. The generators people prefer are the generators that end up in fraudulent claims, fake profiles and manipulated news — so we watch the same leaderboard the labs do, and we make it public. Right now that means keeping pace with gpt-image-2.5-sunburst from OpenAI.

Photorealism is the threat

Switch to the Photorealistic or Portraits category: those rankings show which models produce the images most likely to pass as real photographs.

Metadata-independent

Our detection reads pixels, not watermarks or EXIF, so a new generator climbing this board is a training signal, not a blind spot.

Human-judged, not benchmarked

Arena ranks by anonymous side-by-side votes from real people. It measures what people actually prefer — and therefore what gets used.

Kept weekly

Snapshots are archived every week. Movement badges show who climbed, fell or entered since the last release.

About the data

How the scores are made

On Arena, a visitor types a prompt, two anonymous models each generate an image, and the visitor votes for the better one. Millions of those pairwise votes are fitted with a Bradley–Terry style model to produce each model's Arena score.

The ± figure is a 95% confidence interval; the rank spread is the range of ranks a model could hold once that uncertainty is taken into account. Preliminary marks models added in the last release, which still have comparatively few votes.

Sightova republishes the numbers as published — no re-weighting, no editorial adjustments. Scores are rounded to whole numbers for display.

Attribution

Rankings, scores, confidence intervals and vote counts are data from Text-to-Image Arena · Cartoon, Anime & Fantasy by Arena Intelligence, data as of Sep 7, 2026. © Arena Intelligence. Sightova is not affiliated with, sponsored by or endorsed by Arena; we mirror the board weekly with attribution. Model and lab names are trademarks of their respective owners.

View the live board on arena.ai
FAQ
As of Sep 7, 2026, gpt-image-2.5-sunburst from OpenAI holds the top Arena score (1440 ±21). Check the rank spread before calling it a clear winner: when confidence intervals overlap, the top few models are statistically tied.
The highest-ranked model with a non-proprietary license is ideogram-4.0-quality (Ideogram open model) at #17. Use the “Open weights” filter above to see them all — note that several licenses are non-commercial.
Every week. Arena publishes a new release with a vote cutoff date; we capture it, archive it, and show movement badges relative to the previous release. The exact cutoff is printed under the table and in the attribution box.
Because the generators people prefer are the generators whose output ends up in fraud, misinformation and fake identities. Tracking this board is part of how we prioritise what our detector is trained and tested against — and publishing it keeps us honest about what is out there.
Our detector is trained on output from the major generator families and is evaluated continuously against new releases. See the testing methodology for how we measure accuracy, or upload an image to try it.