AIM-Bench (Affective Image Manipulation) - Negative Emotion Accuracy: leaderboard
Metric: Emotion accuracy (%) on the items whose target is a negative emotion (disgust, sadness, fear, anger): Gemini-2.5-Pro classifies each edited image into the 8 Mikels emotions, 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.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Step1X-Edit (AIM-Bench checkpoint unspecified) | 70.1 |
| 2 | UniWorld-V2 | 58.1 |
| 3 | Seedream 4.0 | 56.74 |
| 4 | BAGEL-7B-MoT | 56.61 |
| 5 | Flux-kontext-max | 55 |
| 6 | Qwen-Image-Edit-Plus (AIM-Bench checkpoint unspecified) | 53.75 |
| 7 | Qwen-Image-Edit-2509 | 50.76 |
| 8 | SeedEdit 3.0 | 50.53 |
| 9 | FLUX.1 Kontext [dev] | 49.13 |
| 10 | Flux-kontext-pro | 47.71 |
Interactive version: theaggregate.ai/benchmark?slug=aim-bench-affective-image-manipulation-negative-emotion-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-07.