Advanced NS-Net: Improving Generalizable AI Generated Image Detection via Learned Semantic Null-Space Projections

Authors: Sachi Deshmukh, Shiwani Mishra, Neel Rambhia.

This repository provides an improved and extensible implementation of NS-Net (Yan et al. 2025). AI-generated images produced by GANs and diffusion models have become nearly indistinguishable from real photographs. Traditional supervised detectors fail to generalize to unseen generation models. NS-Net (Null Space Network) showed that:

Semantic information in CLIP embeddings suppresses artifact cues — and these semantic components must be removed.

Advanced NS-NET improves the original model and provides a powerful approach for detecting AI-generated images by removing semantic information from CLIP embeddings using NULL-space projection. Due to the unavailability of the codebase for NS-Net, we replicated it and compared it with our implemented Advanced NS-Net. Note: Due to the unavailability of the code base for NS-NET, we have replicated it for our analysis and comparison.

Advanced NS-Net Architecture

For details regarding the architecture, viewers can refer to our paper or presentation.

Contents of the Repositiory

• Implementation of the original NS-NET model (src->Original_NSNET.ipynb)
• Implementation of our Advanced NS-NET model (src->Advanced_NSNET.ipynb)
• Project Report cum Paper on Advanced NS-NET
• Presentation slides on Advanced NS-NET
• Architecture of Advanced NS-NET

The codes can be found in the src folder. Note, incase Github is not able to render/ display the .ipynb files correctly, you may download and view the files or visit here to open them on Google Colab.

Status and Future Prospects

• Due to limited compute resources, we trained our model on only 4000 images (2000 Real and 2000 Fake) of DALLE Recognition Dataset from Kaggle, with 2 epochs. Testing was done on 200 images.
• We have met and presented our work to Prof. Amit Sethi and discussed about the possiblities of extending our work to benchmarked datasets (provided in the original paper).
• Notable datasets include AIGIBench, UniversalFakeDetect, GenImage
• If we get significantly good results, we will aim to submit our work to a conference.