AIM-Bench (Affective Image Manipulation) - VAD Distance: leaderboard
AIM-Bench affective image editing scored by the Euclidean distance between the valence, arousal and dominance that GPT-4o reads from the edited image and the target values. Lower is better.
Metric: Mean Euclidean distance between the valence-arousal-dominance coordinates GPT-4o predicts for the edited image and the target coordinates (0 or more; the coordinate scale is not stated), 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; lower is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 13 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seedream 4.0 | 2.54 | |
| 2 | Qwen-Image-Edit-2509 | 2.57 | |
| 3 | UniWorld-V2 | 2.62 | |
| 4 | Step1X-Edit (AIM-Bench checkpoint unspecified) | 2.63 | |
| 5 | SeedEdit 3.0 | 2.65 | |
| 6 | FLUX-Kontext-Max | 2.67 | |
| 7 | Qwen-Image-Edit-Plus (AIM-Bench checkpoint unspecified) | 2.72 | |
| 8 | FLUX-Kontext-Pro | 2.73 | |
| 9 | OmniGen2 | 2.75 | #4579 |
| 10 | FLUX.1-Kontext-dev | 2.77 | |
| 11 | BAGEL-7B-MoT | 2.83 | #4619 |
| 12 | DreamOmni2 | 2.88 | |
| 13 | InstructPix2Pix | 3.39 |
Interactive version: theaggregate.ai/benchmark?slug=aim-bench-affective-image-manipulation-vad-distance · How It Works · Data refreshed daily, snapshot 2026-11-21.