EmoStyle: Emotion-Driven Image Stylization (CVPR 2026)

Jingyuan Yang, Zihuan Bai, Hui Huang^*
Shenzhen University
Affective Image Stylization (AIS) aims to stylize user-input images to match the desired emotion. This task is inherently challenging due to the complex mapping between emotions and visual styles to evoke the desired emotional response.

Fig 1. Affective Image Stylization with EmoStyle, aiming to transform user-provided images through artistic stylization to evoke specific emotional responses. It requires only emotion words as prompts, eliminating the need for reference images or detailed text descriptions.

Preliminary

You can download the pretrained EmoStyle models from here.

Fig 2. Construction process of EmoStyleSet. Given artworks from ArtEmis, after generation and filtering, each triplet contains a content image, a target emotion, and a stylized image.

Other pretrained models like SigLIP, USO can be found in the Hugging Face Model Hub.

Quick Start

Requirements

git clone https://github.com/JingyuanYY/EmoStyle.git
cd EmoStyle

Install the requirements

conda create -n emostyle python=3.9 -y
conda activate emostyle

## install torch
## recommended version:
pip install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu124

pip install -r requirements.txt

EmoStyle

Fig 3. Overview of EmoStyle. We introduce an Emotion–Content Reasoner to integrate emotion and content features, and a Style Quantizer to map continuous queries to discrete style prototypes, generating stylized images with faithful emotion and preserved content.

Inference

Please make sure the pretrained models are downloaded and put them in the weights folder. And you can specify the checkpoint path, emotion, and save path by modifying arguments in inference_emostyle.py or add arguments in the command line.

python inference_emostyle.py --eval_json_path "path-to-your-inference-json"

Training

Your can choose to train EmotionContentReasoner or StyleQuantizer by specifying the --optimize_target argument with 'transformer' or 'codebook'. And before training, please make sure the EmoStyleSet and specify the --emotion argument with the emotion you want to train.

accelerate launch train_emostyle.py --json_file "path-to-training-json"

Results

Qualitative Results

Fig 4. Comparison with the state-of-the-art methods, where EmoStyle surpasses others on emotion fidelity and aesthetic appeal.

Quantitative Results

Table 1. Comparisons with the state-of-the-art methods on style transfer, image editing and AIM methods.

Method	CLIP ↑	DINO ↑	SG ↓	Emo-A ↑	SD ↓
LSAST	0.551	0.747	2.231	12.50	11.28
CLIPStyler	0.709	0.769	3.001	12.60	19.89
InST	0.569	0.679	2.016	21.22	11.48
OmniStyle	0.710	0.813	2.615	12.80	11.90
IP2P	0.708	0.729	3.459	24.34	12.76
LEDITS++	0.687	0.807	2.637	15.97	13.11
EmoEditor	0.686	0.761	2.744	14.88	13.65
EmoEdit	0.597	0.545	2.245	12.60	28.83
CLVA	0.727	0.789	2.030	14.99	9.49
AIF	0.712	0.780	2.625	12.99	8.48
EmoEdit	0.718	0.842	1.976	33.36	7.59

Table 2. User preference study. The numbers indicate the percentage of participants who vote for the result.

Method	Aesthetic Perception ↑	Emotion fidelity ↑	Balance ↑
CLVA	8.50±12.13%	0.81±2.16%	1.19±7.07%
InST	2.50±4.86%	29.63±2.56%	1.34±5.53%
AIF	9.08±9.42%	5.09±2.22%	7.76±9.91%
EmoEdit	79.92±21.24%	64.47±4.55%	89.70±14.48%

As shown in Fig.5, you can adjust the guidance scale to achieve different results.

Fig 5. Ablation study on image guidance scale. EmoStyle can progressively edit an image towards different emotional polarities.

You can also use textual descriptionsto achieve emotion-aware text-to-image generation.

Fig 6. EmoStyle can be extended to emotion-aware text-to-image generation, producing semantically faithful, emotionally expressive and aesthetically appealing stylized results.

Citation

If you find this work useful, please kindly cite our paper:

@article{yang2025emostyle,
  title={EmoStyle: Emotion-Driven Image Stylization},
  author={Yang, Jingyuan and Bai, Zihuan and Huang, Hui},
  journal={arXiv preprint arXiv:2512.05478},
  year={2025}
}