Scaling Image and Video Generation via Test-Time Evolutionary Search

📖 Introduction

We propose Evolutionary Search (EvoSearch), a novel and generalist test-time scaling framework applicable to both image and video generation tasks. EvoSearch significantly enhances sample quality through strategic computation allocation during inference, enabling Stable Diffusion 2.1 to exceed GPT4o, and Wan 1.3B to outperform Wan 14B model and Hunyuan 13B model with 10✖️ fewer parameters.

🧠 Method

EvoSearch introduces a novel perspective that reinterprets the denoising trajectory as an evolutionary path, where both the initial noise $x_T$ and the intermediate state $x_t$ can be evolved towards higher-quality generation, actively expanding the exploration space beyond the constraints of the pre-trained model's distribution. Different from classic evolutionary algorithms that optimize a population set in a fixed space, EvoSearch considers dynamically moving forward the evolutionary population along the denoising trajectory starting from $x_T$ (i.e., Gaussian noises).

🛠️ Instructions

Set up the environment

git clone https://github.com/tinnerhrhe/EvoSearch-codes.git
cd EvoSearch-codes
conda create -n evosearch python=3.9.0 -y
conda activate evosearch
pip install -r requirements.txt

You can try EvoSearch on Stable Diffusion 2.1 via the following commands.

cd SD2.1
python evosearch.py --guidance_reward ImageReward --evolution_schedule 0 20 30 40 --population_size_schedule 42 26 26 26 30

Guidance reward function is pre-defined for search, while evolution schedule and population size schedule can be freely chosen, considering the available test-time computation budget.

You can try EvoSearch on a more advanced T2I model Flux.1-dev via the following commands.

cd FLUX
python evosearch.py --guidance_reward ImageReward --evolution_schedule 5 20 30 40 --population_size_schedule 14 7 7 7 10

For video generation, please first download the guidance reward model and T2V model following the instructions. Then you can play EvoSearch via the following commands.

cd Wan
python evosearch.py --evolution_schedule 5 20 30 45 --population_size_schedule 10 5 5 5 5

📮 TODO

More codes will come soon! Stay tuned!

• Release the codes of EvoSearch on Hunyuan for video generation.
• Release the codes of EvoSearch on FLUX.1-dev for image generation.
• Release the codes of EvoSearch on Wan for video generation.
• Release codes of evosearch on Stable Diffusion 2.1 for a quick example.

👍 Acknowledgement

We thank the following open-source works:

Stable Diffusion

FLUX

Wan-Video

HunyuanVideo

FK-Diffusion-Steering

Diffusion Alignment as Sampling

📜 Citation

If you find our work helpful, please kindly cite our paper:

@misc{he2025scaling,
    title={Scaling Image and Video Generation via Test-Time Evolutionary Search},
    author={Haoran He and Jiajun Liang and Xintao Wang and Pengfei Wan and Di Zhang and Kun Gai and Ling Pan},
    year={2025},
    eprint={2505.17618},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

🌟 Star History