AIM-Bench (Affective Image Manipulation): leaderboard

AIM-Bench (2026) tests instruction-based image editing models on affective image manipulation: editing 800 EmoSet photos so that they evoke a target emotion with target valence, arousal and dominance. The overall score in percent averages normalized instruction-following, image-quality and emotion metrics, higher is better.

Metric: Overall score (%): the mean of the eight Table 1 metrics (CLIP-T, VIEScore SC and PQ, aesthetic score, GPT-4o ground-truth consistency, emotion accuracy and F1, VAD distance) after normalizing each to 0-1 with fixed empirical bounds the paper does not print, over the 800 AIM-Bench affective image editing items (EmoSet source images, a target Mikels emotion with target valence-arousal-dominance coordinates and an editing instruction; 8 emotion categories, 5 editing types), each model run once with its official default settings; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Seedream 4.048.94
2Qwen-Image-Edit-250947.09
3UniWorld-V241.57
4Step1X-Edit (AIM-Bench checkpoint unspecified)41.17
5FLUX-Kontext-Max41.14
6Qwen-Image-Edit-Plus (AIM-Bench checkpoint unspecified)40.25
7SeedEdit 3.038.79
8FLUX-Kontext-Pro38.25
9FLUX.1-Kontext-dev33.12
10OmniGen232.53#4625
11BAGEL-7B-MoT28.59#4668
12DreamOmni224.47
13InstructPix2Pix12.11

Interactive version: theaggregate.ai/benchmark?slug=aim-bench-affective-image-manipulation · How It Works · Data refreshed daily, snapshot 2026-11-22.