The creative model reference

AI image editing model comparison

Explore reference results for instruction-driven image edits and transformations.

September 2026 benchmark snapshot · Curated by Oakgen. Data source and limits.

image / Image Editing

The scores, in perspective.

7 models · 95% intervals

Swipe to compare all models →

Oakgen Arena · v1.0

11,045

GPT Image 2

21,037

Seedream 5.0 Pro

31,035

Nano Banana Pro

41,032

Nano Banana 2

51,000

Grok Imagine 2.0Anchor

6992

Qwen Image 3.0

7955

Flux.2 Pro

↥ 95% confidence interval · Anchor interval not reported.7 ranked models

Relative scores on a zero-based display scale. Compare within this board only. Values are rounded for display.

September 2026 benchmark snapshot · Curated by Oakgen. Data source and limits.

Published reference / v1.0

Image Editing leaderboard

Image Editing image rankings · Benchmark index 1.0
Rank	Model	Overall score	95% confidence interval	Vote appearances	Distinct judges
01	GPT Image 2OpenAI Criterion scores Creativity / variation1,081.995% CI 1,037.4 to 1,128.3Reused fit from master-image Precise editing1,089.195% CI 1,016 to 1,159.5 Prompt adherence1,073.295% CI 993.4 to 1,155.2 Prompt adherence1,041.595% CI 1,010.7 to 1,074.8Reused fit from master-image Reference adherence1,035.395% CI 987 to 1,083.2Reused fit from master-image Style adaptation970.795% CI 897.2 to 1,040.8 Subjective aesthetics1,030.195% CI 987.2 to 1,071.2Reused fit from master-image	1,045.2LEAD	1,015.9 to 1,078	2,574	Not reported
02	Seedream 5.0 ProByteDance Criterion scores Creativity / variation1,102.695% CI 1,050.6 to 1,156.5Reused fit from master-image Precise editing1,094.595% CI 1,012.9 to 1,170.9 Prompt adherence1,030.795% CI 953.9 to 1,117.1 Prompt adherence1,022.295% CI 982.2 to 1,060.8Reused fit from master-image Reference adherence1,036.795% CI 988.2 to 1,080Reused fit from master-image Style adaptation966.995% CI 857.7 to 1,065.7 Subjective aesthetics1,042.895% CI 990.5 to 1,095.7Reused fit from master-image	1,036.8#2	1,000.5 to 1,073.4	2,595	Not reported
03	Nano Banana ProGoogle Criterion scores Creativity / variation1,032.295% CI 978.2 to 1,084.8Reused fit from master-image Precise editing1,096.595% CI 1,030.5 to 1,168.7 Prompt adherence1,036.795% CI 978.8 to 1,096.1 Prompt adherence1,018.395% CI 978.4 to 1,058.1Reused fit from master-image Reference adherence1,002.295% CI 948.1 to 1,052.2Reused fit from master-image Style adaptation1,007.495% CI 921.3 to 1,096.8 Subjective aesthetics981.195% CI 930.5 to 1,027.6Reused fit from master-image	1,035.3#3	1,004.8 to 1,067.7	2,605	Not reported
04	Nano Banana 2Google Criterion scores Creativity / variation972.695% CI 913.5 to 1,028.8Reused fit from master-image Precise editing1,05695% CI 977.1 to 1,140.6 Prompt adherence1,03495% CI 964.6 to 1,106.9 Prompt adherence977.395% CI 937.4 to 1,017.6Reused fit from master-image Reference adherence990.895% CI 945.5 to 1,039.8Reused fit from master-image Style adaptation1,064.795% CI 993.1 to 1,137.2 Subjective aesthetics1,000.495% CI 954 to 1,052.4Reused fit from master-image	1,031.7	1,001.4 to 1,062.9	2,607	Not reported
05	Grok Imagine 2.0xAI Criterion scores Creativity / variation1,00095% CI Anchor · interval not reportedReused fit from master-image Precise editing1,00095% CI Anchor · interval not reported Prompt adherence1,00095% CI Anchor · interval not reported Prompt adherence1,00095% CI Anchor · interval not reportedReused fit from master-image Reference adherence1,00095% CI Anchor · interval not reportedReused fit from master-image Style adaptation1,00095% CI Anchor · interval not reported Subjective aesthetics1,00095% CI Anchor · interval not reportedReused fit from master-image	1,000Anchor	Anchor · interval not reported	2,593	Not reported
06	Qwen Image 3.0Alibaba Criterion scores Creativity / variation997.195% CI 949.8 to 1,047.1Reused fit from master-image Precise editing1,02795% CI 958.8 to 1,091.1 Prompt adherence1,000.195% CI 929.2 to 1,068.9 Prompt adherence915.195% CI 880.5 to 949.6Reused fit from master-image Reference adherence973.295% CI 921.3 to 1,028.5Reused fit from master-image Style adaptation980.595% CI 903.5 to 1,053.9 Subjective aesthetics979.495% CI 928.4 to 1,035.8Reused fit from master-image	991.6	959.5 to 1,021.4	2,592	Not reported
07	Flux.2 ProBlack Forest Labs Criterion scores Creativity / variation909.695% CI 857.1 to 961.2Reused fit from master-image Precise editing947.895% CI 877.9 to 1,017.9 Prompt adherence98895% CI 896 to 1,085.1 Prompt adherence940.995% CI 907.8 to 976.6Reused fit from master-image Reference adherence881.195% CI 833.9 to 927.5Reused fit from master-image Style adaptation966.395% CI 883.2 to 1,038.5 Subjective aesthetics977.795% CI 921.3 to 1,028.8Reused fit from master-image	955.3	919.2 to 987	2,604	Not reported

Ranks are the source’s published positions. Vote appearances are not unique people and must not be added across boards. Scores are rounded for display.

September 2026 benchmark snapshot · Curated by Oakgen. Data source and limits.

From comparison to creation

Choose for your brief.

Compare before and after images side by side. Look for unintended changes outside the edited area, especially faces, hands and product geometry. Test both a small correction and a larger transformation before choosing a workflow.

Compare your own image prompt

A source listing does not confirm current availability in Oakgen.

How this board is composed

Criteria and weights

Style adaptation23.3%
Precise editing23.3%
Prompt adherence23.3%
Prompt adherence7.5%
Subjective aesthetics7.5%
Creativity / variation7.5%
Reference adherence7.5%

Read the methodology