UniEditBench (Image) - Naturalness: leaderboard

UniEditBench (2026) scores image edits on 633 samples across nine editing operations on a 1-5 scale with an evaluator the authors distilled from Qwen3-VL-235B-A22B into Qwen3-VL-8B. This board rates how realistic and artifact-free the result looks, from 1 to 5; higher is better.

Metric: Naturalness score (1-5): the edited image looks realistic and artifact-free; 633 image-editing samples over nine operations, each edit scored on a 1-5 Likert scale by the authors' Qwen3-VL-8B evaluator distilled from Qwen3-VL-235B-A22B (the second entry of each 4B/8B cell); higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1FLUX.1-Kontext-dev4.22
2UniEditBench Hunyuan (checkpoint unspecified)4.15
3Qwen-Image-Edit4.09
4DreamOmni24.02
5Step1X-Edit3.94
6OmniGen23.79#4579
7UniWorld-V23.78
8BAGEL-7B-MoT3.58#4619
9ICEdit3.58
10VAREdit3.51
11LaVida-O2.96
12AnyEdit2.82
13EditAR2.6
14EditMGT1.94
15InstructPix2Pix1.29

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