Research map / Image generation / GANs
Content-Aware GAN Compression
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use content-aware pruning and distillation instead of standard ℓ1-out pruning and non-content-aware distillation | StyleGAN2 · FFHQ, 256px | FID: 8.9 -> 7.9 | Helped |
| Use content-aware pruning and distillation instead of standard ℓ1-out pruning and non-content-aware distillation | StyleGAN2 · FFHQ, 1024px | FID: 8.1 -> 7.6 | Helped |
| Use content-aware pruning and distillation instead of standard ℓ1-out pruning and non-content-aware distillation | SN-GAN · CIFAR-10, 32px, 0.42B FLOPs | IS: 8.21±0.11 -> 8.31±0.08 | Helped |
| Replace low-activation channel pruning with ℓ1-out pruning (ablation) | StyleGAN2 · FFHQ, 256px, 22.3B FLOPs | FID: 7.9 -> 5.4 | Helped |
| Replace training the pruned network from scratch with ℓ1-out channel pruning (ablation) | StyleGAN2 · FFHQ, 256px, 22.3B FLOPs | FID: 8.1 -> 5.4 | Helped |
| Remove knowledge distillation (ablation) | StyleGAN2 compressed with ℓ1-out pruning · FFHQ, 256px, 1.9B FLOPs | FID: 12.5 -> 15.1 | Hurt |
| Add output-only norm-based distillation to LPIPS perceptual distillation (ablation) | StyleGAN2 compressed with ℓ1-out pruning · FFHQ, 256px, 1.9B FLOPs | FID: 13.6 -> 12.5 | Helped |
| Replace output-only norm-based distillation with intermediate norm-based distillation while retaining LPIPS (ablation) | StyleGAN2 compressed with ℓ1-out pruning · FFHQ, 256px, 1.9B FLOPs | FID: 12.5 -> 13.6 | Hurt |
| Replace VGG style-transfer loss with LPIPS perceptual distillation (ablation) | StyleGAN2 compressed with ℓ1-out pruning · FFHQ, 256px, 1.9B FLOPs | FID: 14.6 -> 13.6 | Helped |
| Use content-aware distillation rather than distilling over all spatial locations (ablation) | StyleGAN2 · FFHQ | FID: null | Helped |
| Set the pruning content of interest to the image background rather than the foreground face (ablation) | StyleGAN2 · FFHQ, 256px | image quality: null | Hurt |
| Increase the content-aware distillation weights beyond λ=γ=3 (ablation) | StyleGAN2 compressed with content-aware distillation · FFHQ, 256px, 11× acceleration | FID: null | Hurt |
Closest papers by meaning
- GAN Compression: Efficient Architectures for Interactive Conditional GANs (CVPR 2020 · similarity 0.93)
- Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold (AAAI 2024 · similarity 0.91)
- Singular Value Scaling: Efficient Generative Model Compression via Pruned Weights Refinement (AAAI 2025 · similarity 0.90)
- E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 2024 · similarity 0.89)
- The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS 2024 · similarity 0.89)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 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.