Research map / Image generation / GANs
RG-GAN: Dynamic Regenerative Pruning for Data-Efficient 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 RG-GAN, which combines layer-wise dynamic pruning with recurrent and augmented weight regeneration during GAN training to improve data-efficient image generation and reduce parameters/FLOPs.
Problem
GAN training with limited data is unstable and prone to losing important weights when sparsity is introduced during training; existing GAN lottery-ticket methods need costly train-prune-retrain cycles.
Approach
Dynamic layer-wise pruning with learned thresholds, paired with recurrent regeneration (pruned weights can be reactivated when their gradients show importance) and augmented regeneration (restoring pruned weights to a dense network when FID worsens), applied jointly to generator and discriminator.
Key results
- On CIFAR-10 with 10% data, RG-ProGAN and RG-StyleGAN2 improve FID by 5.78% and 2.08% over their base models.
- On FFHQ 1k images at 1024x1024, RG-StyleGAN2 reduces FID from 40.24 to 32.48.
- On full CIFAR-10, Sparse_RG-SNGAN reaches FID 15.64 versus 17.74 for SNGAN, with fewer FLOPs than GAN tickets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dynamic pruning and regeneration | StyleGAN2 · CIFAR-10 (32×32), 10% data | FID: 36.45 -> 31.69 | Helped |
| add dynamic pruning and regeneration | ProGAN · Tiny-ImageNet (64×64), 10% data | FID: 83.27 -> 76.99 | Helped |
| add dynamic pruning and regeneration | StyleGAN2 · FFHQ (1024×1024), 1k images | FID: 40.24 -> 32.48 | Helped |
| add dynamic pruning and regeneration | StyleGAN2 · Obama, G-Cat, Panda, Animal Face Cat, Animal Face Dog (256×256) | FID: 86.6 -> 71.0; 51.3 -> 36.3; 34.2 -> 19.9; 81.9 -> 71.2; 157.6 -> 130.3 | Helped |
| add Differentiable Augmentation to RG-StyleGAN2 | RG-StyleGAN2 · Obama, G-Cat, Panda, Animal Face Cat, Animal Face Dog (256×256) | FID: 71.0 -> 44.8; 36.3 -> 27.3; 19.9 -> 12.2; 71.2 -> 42.1; 130.3 -> 56.7 | Helped |
| add recurrent regeneration to dynamic pruning (ablation) | SNGAN · CIFAR-10 | FID | Helped |
| add augmented regeneration in both generator and discriminator (ablation) | SNGAN with dynamic pruning and recurrent regeneration · CIFAR-10 | FID | Helped |
| add dynamic pruning (ablation) | SNGAN · CIFAR-10 | FID | Helped |
| add augmented regeneration (ablation) | SNGAN with dynamic pruning and recurrent regeneration · CIFAR-10 | FID | Helped |
| add recurrent regeneration (ablation) | SNGAN with dynamic pruning · CIFAR-10 | FID | Helped |
Limitations
- Theoretical framework for regeneration in GAN is absent
- Evaluated only on image generation with unconditional GANs
- Augmented regeneration relies on heuristic FID-based trigger (n=5 weighted average)
Experimental setup
Datasets: CIFAR-10 Tiny-ImageNet FFHQ Obama Grumpy-Cat Panda AnimalFace Cat AnimalFace Dog
Benchmarks: CIFAR-10 Tiny-ImageNet FFHQ few-shot 256x256 datasets (Obama, Grumpy-Cat, Panda, Animal Face Cat/Dog)
Metrics: FID number of real images (#RI) FLOPs
Baselines: ProGAN StyleGAN2 FastGAN SNGAN GAN tickets (IMP) Re-GAN
Builds on: Re-GAN (dynamic pruning during GAN training) GAN tickets / lottery ticket hypothesis Differentiable Augmentation FastGAN StyleGAN2
Closest papers by meaning
- Re-GAN: Data-Efficient GANs Training via Architectural Reconfiguration (CVPR 2023 · similarity 0.92)
- MSG-GAN: Multi-Scale Gradients for Generative Adversarial Networks (CVPR 2020 · similarity 0.91)
- Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold (AAAI 2024 · similarity 0.90)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.90)
- MS$^3$D: A RG Flow-Based Regularization for GAN Training with Limited Data (ICML 2024 · similarity 0.89)
- Singular Value Scaling: Efficient Generative Model Compression via Pruned Weights Refinement (AAAI 2025 · similarity 0.89)
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.