Skip to content
Models.

LMArena Image Edit what it measures, how it is scored and who leads it

Arenas · LMArena · Arena rating

Leaderboard

Scores from LMArena Image Edit, published Sep 21, 2026

RankModelArena rating
1GPT Image 2.5 SunburstOpenAI · as “gpt-image-2.5-sunburst”1,526
2GPT Image 2.5 FlareOpenAI · as “gpt-image-2.5-flare”1,482
3GPT Image 2OpenAI · as “gpt-image-2 (medium)”1,461
4Grok Imagine Image 2.0SpaceXAI · as “grok-imagine-image-2.0 (low)”1,430
5MAI Image 2.6Microsoft · as “mai-image-2.6”1,429
6Muse ImageMeta · as “muse-image”1,402
7MAI Image 2.5Microsoft · as “mai-image-2.5”1,400
8Seedream 5.0 ProByteDance · as “seedream-5.0-pro”1,394
9Gemini 3 Pro ImageGoogle · as “gemini-3-pro-image-2k (nano-banana-pro)”1,390
10Grok Imagine Image QualitySpaceXAI · as “grok-imagine-image-quality (20260519)”1,390
11GPT Image FidelityOpenAI · as “chatgpt-image-latest-high-fidelity (20251216)”1,388
12Gemini 3.1 Flash ImageGoogle · as “gemini-3.1-flash-image (nano-banana-2) [web-search]”1,388
13Reve 2.1Reve · as “reve-2.1”1,375
14GPT Image 1.5 FidelityOpenAI · as “gpt-image-1.5-high-fidelity”1,370
15Qwen Image 2.1Alibaba · as “qwen-image-2.1”1,367
16Reve 2.0Reve · as “reve-2.0”1,358
17Uni 1.1Luma AI · as “uni-1.1-max”1,334
18Grok Imagine ImageSpaceXAI · as “grok-imagine-image”1,329
19Gemini 3.1 Flash Lite ImageGoogle · as “gemini-3.1-flash-lite-image (nano-banana-2-lite)”1,314
20Qwen Image 2.0 ProAlibaba · as “qwen-image-2.0-pro-2026-06-22”1,304
21Hunyuan Image 3.0Tencent · as “hunyuan-image-3.0-instruct”1,302
22Wan2.7 Image ProAlibaba · as “wan2.7-image-pro”1,302
23Seedream 4.5ByteDance · as “seedream-4.5”1,302
24Wan2.7 ImageAlibaba · as “wan2.7-image”1,301
25Seedream 5.0 LiteByteDance · as “seedream-5.0-lite”1,294

About this benchmark

Task
Edit an image from an instruction; people vote for the better edit.
Dataset
Crowd-sourced images, instructions and votes.
Method
Pairwise votes fitted with a Bradley-Terry model.
Metric
Arena rating
Organization
LMArena
Official leaderboard
arena.ai/leaderboard/image-edit

Sources: each benchmark's paper (arXiv) and its official site or leaderboard. Scores are the boards captured here every day, exactly as published.

New #1s by email

Saturdays, only in weeks when a leaderboard has a new #1.

Double opt-in. Unsubscribe any time.