AICA-Bench - Emotion-Guided Generation: leaderboard

AICA-Bench emotion-guided content generation: describe an image so that the text conveys a target emotion, rated 1 to 5 by a scoring model trained on human ratings, rescaled to 20-100. Higher is better.

Metric: Emotion-guided content generation score on a percent scale: the model writes a scene description congruent with a target emotion, scored by the AICA-Bench scoring model (Qwen2.5-VL-7B fine-tuned on human 1-5 ratings) on emotion alignment and descriptiveness, ratings rescaled as s/5 x 100 (20-100), over AICA-Bench (8,086 affective images from nine public emotion datasets with GPT-4o-generated instructions; closed-source models through their APIs at standard settings, open models on A100 GPUs); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 23 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen-VL-Max (AICA-Bench checkpoint unspecified)75.93
2GPT-4o75.73#751
3Gemini 2.5 Pro74.13#276
4GPT-4o Mini74.09#1667
5Ovis2-16B71.56#1594
6Ovis2-8B70.81#2047
7InternVL2.5-8B68.86#2370
8Gemini 2.5 Flash68.19#463
9InternVL3-8B-Instruct67.27#2872
10Qwen-VL-Plus (AICA-Bench checkpoint unspecified)66.86
11InternVL3-14B-Instruct66.5#1801
12Ovis2-2B66.04#3863
13Qwen 2.5 VL 7B Instruct66#1902
14MiniCPM-o-2.664.98#2423
15Qwen 2 VL 7B Instruct64.76#2573

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=aica-bench-emotion-guided-generation · How It Works · Data refreshed daily, snapshot 2026-11-21.