Research map / Image generation / GANs
Turning Waste into Wealth: Leveraging Low-Quality Samples for Enhancing Continuous Conditional Generative Adversarial Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dual-NDA, a negative data augmentation scheme for CcGANs that adds label-inconsistent real images and NIQE-filtered low-quality generator samples as negatives in a modified vicinal discriminator loss, improving SFID, NIQE and label consistency on UTKFace and Steering Angle.
Problem
CcGANs produce low-quality and label-inconsistent fake images when training data is sparse or imbalanced, and vanilla NDA (designed for unconditional/class-conditional GANs) does not capture the negative samples CcGANs actually produce.
Approach
Dual-NDA builds two negative sample types tailored to CcGANs: Type I are real images with labels far from the conditioning label, and Type II are visually unrealistic generator samples selected by NIQE filtering; both enter a new soft vicinal discriminator loss with weights lambda_1 and lambda_2.
Key results
- On Steering Angle 64x64, Dual-NDA reaches SFID 1.114 vs 1.334 for CcGAN w/o NDA and NIQE 1.738 vs 1.784.
- On UTKFace 128x128, Dual-NDA achieves SFID 0.361 and NIQE 1.081, the best among compared methods, including class-conditional GANs and diffusion models.
- Vanilla NDA degrades CcGAN performance in most settings, e.g., NIQE rises from 1.784 to 1.994 on Steering Angle 64x64.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Dual-NDA Type I and Type II negatives | CcGAN (SVDL+ILI) · UTKFace 64×64, UTKFace 128×128, Steering Angle 64×64, Steering Angle 128×128 | SFID: 0.413 (0.155) -> 0.396 (0.153); 0.367 (0.123) -> 0.361 (0.127); 1.334 (0.531) -> 1.114 (0.503); 1.689 (0.443) -> 1.390 ( | Helped |
| Add Dual-NDA Type I and Type II negatives | CcGAN (SVDL+ILI) · UTKFace 64×64, UTKFace 128×128, Steering Angle 64×64, Steering Angle 128×128 | NIQE: 1.733 (0.189) -> 1.678 (0.183); 1.113 (0.033) -> 1.081 (0.042); 1.784 (0.065) -> 1.738 (0.055); 2.411 (0.100) -> 2.135 ( | Helped |
| Add conventional NDA negatives | CcGAN · UTKFace and Steering Angle | SFID | Hurt |
| Vary q1 across the tested values (ablation) | Dual-NDA · Steering Angle 64×64 | SFID: 1.125 (0.510); 1.114 (0.503); 1.161 (0.540); 1.187 (0.558) | No clear effect |
| Vary q2 across the tested values (ablation) | Dual-NDA · Steering Angle 64×64 | SFID: 1.107 (0.506); 1.193 (0.592); 1.150 (0.494); 1.174 (0.563); 1.114 (0.503) | No clear effect |
| Use ADCGAN or ReACGAN instead of CcGAN to generate Type II negatives (ablation) | Dual-NDA · Steering Angle 64×64 | NIQE: 1.738 (0.055) -> 1.832 (0.055); 1.738 (0.055) -> 1.800 (0.078) | Hurt |
| Add Type I negatives only (ablation) | CcGAN (SVDL+ILI) · Steering Angle 64×64 | Label Score | Helped |
| Add Type I negatives only (ablation) | CcGAN (SVDL+ILI) · Steering Angle 64×64 | NIQE | Hurt |
| Add Type II negatives only (ablation) | CcGAN (SVDL+ILI) · Steering Angle 64×64 | NIQE | Helped |
| Add Type II negatives only (ablation) | CcGAN (SVDL+ILI) · Steering Angle 64×64 | Label Score | Helped |
| Add Type I and Type II negatives together (ablation) | CcGAN (SVDL+ILI) · Steering Angle 64×64 | SFID: 1.334 (0.531) -> 1.114 (0.503) | Helped |
| Vary the combined weight λ1+λ2 within 0.2 to 0.7 (ablation) | Dual-NDA · Steering Angle 64×64 | SFID | No clear effect |
Surprising findings
- Vanilla NDA consistently fails to help CcGANs and generally worsens visual quality.
- Type II negatives (visually unrealistic images) improve label consistency as well, since poor-quality images are often label-inconsistent.
Limitations
- Dual-NDA does not uniformly win: on some metrics (e.g., Label Score on UTKFace 64x64 and Diversity) baselines or ablations are better.
- Reliance on NIQE filtering and several hyperparameters (lambda_1, lambda_2, q_1, q_2) tuned per dataset.
- Type II samples are generated from a pre-trained CcGAN, so quality depends on that generator; ADCGAN/ReACGAN generators hurt NIQE.
Experimental setup
Datasets: UTKFace Steering Angle
Benchmarks: UTKFace Steering Angle
Metrics: SFID NIQE Diversity Label Score
Baselines: ReACGAN ADCGAN ADM-G CFG CcGAN (SVDL+ILI) w/o NDA CcGAN w/ NDA
Builds on: CcGAN Negative Data Augmentation ReACGAN ADCGAN SAGAN
Closest papers by meaning
- CcGAN: Continuous Conditional Generative Adversarial Networks for Image Generation (ICLR 2021 · similarity 0.92)
- Robust Conditional Generative Adversarial Networks (ICLR 2019 · similarity 0.89)
- A Unified View of cGANs with and without Classifiers (NeurIPS 2021 · similarity 0.88)
- Rebooting ACGAN: Auxiliary Classifier GANs with Stable Training (NeurIPS 2021 · similarity 0.88)
- Diversity-Sensitive Conditional Generative Adversarial Networks (ICLR 2019 · similarity 0.88)
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · similarity 0.88)
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.