ai_detect

LOTA: Bit-Planes Guided AI-Generated Image Detection,   ICCV25_Published,   arXiv

ICCV 2025 | 东南大学等提出LOTA：火眼金睛新范式，从“噪声”中秒速揪出AI生成图，准确率达98.9%

ICCV2025 | LOTA：从底层噪声中识别伪造真相

LOTA: Bit-Planes Guided AI-Generated Image Detection

(PS: 以上报道为自媒体编辑撰写，我们虽未向任何平台投稿，但欢迎翻译/宣传/讨论该工作，注明出处即可，谢谢)

Contributors: Hongsong Wang ([email protected]), Renxi Cheng ([email protected])

Southeast University, Nanjing, China

📰 News

• 🚨 The whole ICCV 2025 papers with Codes are summarized on ICCV2025_ABSTRACT/
• 🚨 Paper Portal for Top Conferences in the Field of Artificial intelligence: CV_Paper_Portal

Hightlights

<p align="center"> <img src="images/intro_00.png" width="55%"> <br> <b>Figure 1: Comparison of least bit-planes between real images and AI-generated images.</b> </p> <p align="center"> <img src="images/method_00.png" width="70%"> <br> <b>Figure 2: Overview of our method.</b> </p>

Novel solution for AI-generated image detection: We innovatively address AI-generated image detection based on bit-planes, and propose an efficient approach for noisy representation extraction.

Efficient pipeline design: We propose a simple yet effec tive pipeline with three modules: noise generation, patch selection and classification. We design a heuristic strategy called maximum gradient patch selection and introduce two effective classifiers: noise-based classifier and noise guided classifier. Our approach operates at millisecond level, nearly a hundred times faster than current methods.

Exceedingly superior performance: Extensive exper iments demonstrate the effectiveness of LOTA, which achieves 98.9% ACC onGenImage, showing great cross-generator generalization capability and outperforming existing mainstream methods by more than 11.9%.

Dataset

We use GenImage for training and evaluation, which can be downloaded online. GenImage is composed of 8 subsets (BigGAN, Midjourney, Wukong, Stable_Diffusion_v1.4, Stable_Diffusion_v1.5, ADM, GLIDE, VQDM), each of which contains fake images and real images from ImageNet. Additionally, each subset are invided into training dataset and validating dataset, and we train LOTA on the training dataset of one subset (e.g., Stable_Diffusion_v1.5) and evaluate on the validating dataset of all subsets.

Training

You can train the LOTA model on Stable_Diffusion_v1.5 of GenImage by running the following command:

python train.py --choice=[0, 0, 0, 0, 1, 0, 0, 0]
                --image_root='Path/to/GenImage'
                --save_path='Path/to/saved_weights'
                --bit_mode='scaling'
                --patch_size=32
                --patch_mode='random'

Evaluation

You can evaluate LOTA on all subsets of GenImage by running the following command:

python test.py  --choice=[1, 1, 1, 1, 1, 1, 1, 1]
                --load='Path/to/saved_weights'
                --image_root='Path/to/GenImage'
                --bit_mode='scaling'
                --patch_size=32
                --patch_mode='max'

Additionally, we provide the pretrained weights trained on Stable_Diffusion_v1.4 (code: imjw) and Stable_Diffusion_v1.5 (code: a942) for evaluation. You can download the weights and easily evaluate LOTA on GenImage.

Results

We provide the evaluation results in Results, which contains path_to_testing_images, true_label, predict_prob_real, and predict_prob_fake of all testing images in GenImage.

Acknowledgments

This repository borrows partially from CNNDetection, PatchCraft and SSP. Thanks for their work sincerely.

@InProceedings{Wang_2025_ICCV,
    author    = {Wang, Hongsong and Cheng, Renxi and Zhang, Yang and Han, Chaolei and Gui, Jie},
    title     = {LOTA: Bit-Planes Guided AI-Generated Image Detection},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {17246-17255}
}