Edit-Compass (Chinese): leaderboard

Edit-Compass (2026) tests instruction-based image editing on 2,388 instances with Chinese instructions in six categories (general editing, dynamic manipulation, world knowledge, visual reasoning, multi-image and complex tasks), judged by Gemini 3.1 Pro with rubric prompts on a 1-5 scale. The overall combines instruction awareness, visual consistency and visual quality per category; higher is better.

Metric: Overall score (1-5) of image editing under Chinese instructions, Gemini-3.1-Pro judge with rubric prompts rates instruction awareness, visual consistency and visual quality on a 1-5 scale; the per-category overall is a weighted geometric mean, set to the lowest dimension when that is at most 1; 2,388 instances in six categories (models without multi-image input print no multi-image scores); higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 29 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Nano Banana Pro3.95
2Nano Banana 23.71
3Seedream 4.53.23
4Wan2.7-Image3.21
5Qwen-Image-Edit-25112.73
6LongCat-Image-Edit2.68
7EMU3.52.63
8JoyAI-Image-Edit2.63
9FLUX.2-dev2.6
10UniWorld-V22.53
11Step1X-Edit-v1p22.53
12Skywork UniPic 3.02.47
13Bagel2.1
14Bagel-Think2.1
15ChronoEdit1.9

Interactive version: theaggregate.ai/benchmark?slug=edit-compass-chinese · How It Works · Data refreshed daily, snapshot 2026-11-21.