Diverse Rare Sample Generation with Pretrained GANs
Authors: Subeen Lee, Jiyeon Han, Soyeon Kim, Jaesik Choi
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of our approach both qualitatively and quantitatively across various datasets and GANs without retraining or fine-tuning the pretrained GANs. Our method can generate diverse versions of rare images utilizing the multi-start method in optimization, without being trapped at the same local optima. Rarity and diversity of generated images and similarity to the initial image can be controlled via a multi-objective optimization framework. We demonstrate the effectiveness of our method with various high-resolution image datasets and GANs, both qualitatively and quantitatively. |
| Researcher Affiliation | Academia | 1Korea Advanced Institute of Science and Technology (KAIST), South Korea 2INEEJI, South Korea EMAIL. All listed email addresses are associated with the '.ac.kr' domain, which indicates an academic institution in South Korea. |
| Pseudocode | No | The paper describes its methodology using equations and diagrams (e.g., Fig. 2), but does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured, code-like steps for its procedures in the main text. |
| Open Source Code | Yes | Code https://github.com/sbrblee/Div Rare Gen |
| Open Datasets | Yes | We validate our proposed method using high-resolution image datasets with a resolution of 1024 1024, including Flickr Faces HQ (FFHQ) (Karras, Laine, and Aila 2019), Animal Faces HQ (AFHQ) (Choi et al. 2020), and Metfaces (Karras et al. 2020a). |
| Dataset Splits | No | The paper primarily discusses the generation of synthetic samples using pre-trained GANs and evaluating them. While it mentions generating '10,000 synthetic samples' or '1,000 initial latent vectors' for evaluation, it does not provide specific training/test/validation splits for any dataset used to train their proposed method or the density estimator. The GANs and feature extractors used are stated to be pre-trained models. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments. It mentions using 'high-resolution image datasets' but no hardware specifications are given. |
| Software Dependencies | No | The paper mentions several software components like 'Style GAN2', 'Style GAN2-ADA', 'VGG16-fc2 architecture', 'Glow architecture', and 'Adam optimizer'. However, it does not provide specific version numbers for any of these software dependencies, which would be necessary for reproducibility. |
| Experiment Setup | Yes | The optimization is performed using the Adam optimizer (Diederik 2014) with a learning rate of 2 10 2, combined with a Step LR scheduler. The best optimization results are recorded when the lowest loss is achieved according to Equation (2). Additional details including computational cost are provided in Appendix B. With our method, we generate ten rare samples for each of 1,000 initial latent vectors from the baselines, using parameters λ1 = 30.0, λ2 = 0.002, σ = 0.1, and k = 1001. With our method, we generate five rare samples for each of 1,000 initial latent vectors from the baseline, using parameters λ1 = 200.0, λ2 = 0.02, σ = 0.01 and k = 100. |