[CVPR'2025] EVOS: Efficient Implicit Neural Training via EVOlutionary Selector

Project Page | Paper | Code | Supplementary | Checkpoints

Weixiang Zhang, Shuzhao Xie, Chengwei Ren, Siyi Xie, Chen Tang, Shijia Ge, Mingzi Wang, Zhi Wang*
Tsinghua University
*: Corresponding Author

This is the official PyTorch implementation of "EVOS: Efficient Implicit Neural Training via EVOlutionary Selector" (accepted by CVPR 2025).

Overview

Abstract. We propose EVOlutionary Selector (EVOS), an efficient training paradigm for accelerating Implicit Neural Representation (INR). Unlike conventional INR training that feeds all samples through the neural network in each iteration, our approach restricts training to strategically selected points, reducing computational overhead by eliminating redundant forward passes. Specifically, we treat each sample as an individual in an evolutionary process, where only those fittest ones survive and merit inclusion in training, adaptively evolving with the neural network dynamics. While this is conceptually similar to Evolutionary Algorithms, their distinct objectives (selection for acceleration vs. iterative solution optimization) require a fundamental redefinition of evolutionary mechanisms for our context. In response, we design sparse fitness evaluation, frequency-guided crossover, and augmented unbiased mutation to comprise EVOS. These components respectively guide sample selection with reduced computational cost, enhance performance through frequency-domain balance, and mitigate selection bias from cached evaluation. Extensive experiments demonstrate that our method achieves approximately 48%-66% reduction in training time while ensuring superior convergence without additional cost, establishing state-of-the-art acceleration among recent sampling-based strategies.

Quick Start

Clone Repository

git clone https://github.com/zwx-open/EVOS-INR
cd EVOS-INR

Enviroment Setup

Tested Enviroments:
- Ubuntu 20.04 with PyTorch 1.12.1 & CUDA 11.3.1 on RTX 3090

Run Demo

Demo: Fit DIV2K/test/00.png with SIREN + EVOS(step-wise scheduler) (5k epochs; ~5minutes)

python demo.py 

Reproducing Results from the Paper

Table 1. Comparison of sampling strategies under fixed iterations.

To reproduce the results presented in Table 1, uncomment the corresponding code section in run.py (approximately line 41). All baseline have been pre-configured for replication purposes.

if __name__ == "__main__":
   
    '''Reproduce Table_1 (Constant Scheduler)'''
    # param_idxs = [
    #     "full",
    #     "random",
    #     "egra",
    #     "expansive",
    #     "soft",
    #     "nmt_incre", 
    #     # "nmt_dense",
    #     "evos",
    #     "evos_wo_cfs"
    # ]
    # gpu_list = [0,1,2,3,4,5,6,7]
    # run_tasks("001", param_idxs, gpu_list)

    '''Reproduce Table_1 (Step-wise Scheduler)'''
    param_idxs = [
        "random",
        "egra",
        "expansive",
        "nmt_incre", 
        "evos",
        "evos_wo_cfs",
    ]
    gpu_list = [0,1,2,3,4,5,6,7]
    run_tasks("002", param_idxs, gpu_list)

python run.py

If only a single GPU is available, the experiments can be conducted sequentially by:

    param_idxs = [
        "random",
        "egra",
        "expansive",
        "nmt_incre", 
        "evos",
        "evos_wo_cfs",
    ]
    gpu_list = [0]
    run_tasks("002", param_idxs, gpu_list)

The baseline configurations are implemented in _by_sampler() within manager.py, which can be customized to suit specific experimental needs.

More Flexible Way

If you prefer a more flexible way to run this code, please refer to demo.py:

def demo(use_cuda=0):
    args = [
        "--model_type",
        "siren",
        "--input_path",
        "./data/div2k/test_data/00.png",
        "--eval_lpips",
        "--log_epoch",
        "500",
        "--num_epochs",
        "5000",
        "--use_ratio",
        "0.2",
        "--strategy",
        "evos",
        "--sample_num_schedular",
        "step",
        "--tag",
        "evos",
        "--lr",
        "0.0001",
        "--up_folder_name",
        "000_demo_evos_stepwith",
    ]
    os.environ["CUDA_VISIBLE_DEVICES"] = str(use_cuda) 
    script = "python main.py " + " ".join(args)
    print(f"Running: {script}")
    os.system(script)


if __name__ == "__main__":
    demo(0)

Checkpoints

Chekckpoints can be found in here. We thank @WillYao-THU for providing reproduced results and checkpoints in his environment.

Acknowledgments

We thank the authors of the following works for releasing their codebases:

• INT
• Soft Mining

Additional Related Research

Welcome to explore our related research. The source code for all works has been available.

• (AAAI'2025) Enhancing Implicit Neural Representations via Symmetric Power Transformation | [paper] | [project] | [code] |
• (ICME'2025) Expansive Supervision for Neural Radiance Fields| [paper] | [code] |
• Recent Progress of Implicit Neural Representations | [code]

Citation

Please consider leaving a ⭐ and citing our paper if you find this project helpful:

@article{evos-inr,
  title={EVOS: Efficient Implicit Neural Training via EVOlutionary Selector},
  author={Zhang, Weixiang and Xie, Shuzhao and Ren, Chengwei and Xie, Siyi and Tang, Chen and Ge, Shijia and Wang, Mingzi and Wang, Zhi},
  journal={arXiv preprint arXiv:2412.10153},
  year={2024}
}