Research map / Image generation / GANs
Towards Discovery and Attribution of Open-World GAN Generated Images
Sharath Girish, Saksham Suri, Sai Saketh Rambhatla, Abhinav Shrivastava
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add OOD detection, merge, refine and iterations to the single-iteration pipeline (k=209) (ablation) | GAN discovery pipeline · GAN discovery (20 sources, k=209) | Avg. Purity / NMI: 0.795 -> 0.861 purity, 0.609 -> 0.724 NMI | Improved |
| add the merge step after clustering (ablation) | GAN discovery pipeline · GAN discovery, iteration 1 stages | number of clusters, Avg. Purity: 512 -> 391 clusters, 0.793 -> 0.792 purity, 0.682 -> 0.689 NMI | Improved |
| add the SVM refine step, evaluated on discovered samples (ablation) | GAN discovery pipeline · GAN discovery, iteration 1 stages | Avg. Purity / NMI: 0.792 -> 0.849 purity, 0.689 -> 0.838 NMI | Improved |
| add the refine step, evaluated on all samples (ablation) | GAN discovery pipeline · GAN discovery, iteration 1 stages | Avg. Purity: 0.792 -> 0.772 purity, 0.689 -> 0.720 NMI | Hurt |
| run more iterations (1 to 4) (ablation) | GAN discovery pipeline · GAN discovery by iteration | Avg. Purity / sources discovered: 0.772 -> 0.861 purity, 16/20 -> 20/20 sources | Improved |
| add one GAN per dataset to the labeled set (ablation) | GAN discovery pipeline · GAN discovery (labeled set variants) | Avg. Purity: 0.497 -> 0.897 | Improved |
| add a second GAN per dataset to the labeled set (ablation) | GAN discovery pipeline · GAN discovery (labeled set variants) | Avg. Purity: 0.897 -> 0.954 | Improved |
| use WTA hash distance instead of cosine distance for OOD (ablation) | OOD detection step · GAN discovery, 4 iterations | net OOD accuracy: 89.33 -> 90.65 at iteration 1, 88.27 -> 90.33 at iteration 4 | Improved |
| add JPEG and blur augmentations to the method (ablation) | GAN discovery pipeline · GAN discovery (20 sources) | Avg. Purity / NMI: 0.8216 -> 0.7696 purity, 0.6552 -> 0.6249 NMI | Hurt |
| use cluster majority voting for real/fake detection instead of per-sample prediction | real/fake classifier · Real/fake detection, original setup | accuracy: 92.56% -> 98.62% | Improved |
| use cluster majority voting for real/fake detection in new real plus new GANs setup | real/fake classifier · Real/fake detection, new real + new GANs | accuracy: 89.09% -> 83.33% | Hurt |
Closest papers by meaning
- Attributing Fake Images to GANs: Learning and Analyzing GAN Fingerprints (ICCV 2019 · similarity 0.93)
- RepMix: Representation Mixing for Robust Attribution of Synthesized Images (ECCV 2022 · similarity 0.89)
- T-GD: Transferable GAN-generated Images Detection Framework (ICML 2020 · similarity 0.89)
- Are GANs Created Equal? A Large-Scale Study (NeurIPS 2018 · similarity 0.89)
- Small-GAN: Speeding up GAN Training using Core-Sets (ICML 2020 · similarity 0.88)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
