Research map / Image generation / GANs
Nickel and Diming Your GAN: A Dual-Method Approach to Enhancing GAN Efficiency via Knowledge Distillation
Sangyeop Yeo, Yoojin Jang, Jaejun Yoo
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 DINO plus CLIP kernels (DiME) over DINO alone with global features (ablation) | DiME · FFHQ 256 StyleGAN2, 98.92% compression | FID: 19.60 -> 18.32 | Improved |
| use CLIP alone as embedding kernel instead of DINO (ablation) | DiME · FFHQ 256 StyleGAN2, 98.92% compression | FID: 20.75 -> 152.15 | Hurt |
| add precomputed global teacher features to DINO kernel (ablation) | DiME · FFHQ 256 StyleGAN2, 98.92% compression | FID: 20.75 -> 19.60 | Improved |
| add NICKEL (teacher generator to student discriminator distillation) to DiME (ablation) | DiME · FFHQ 256 StyleGAN2, 98.92% compression | FID: 18.32 -> 15.93 | Improved |
| add NICKEL to DiME at 90.73% compression | DiME · FFHQ 256 StyleGAN2, 90.73% compression | FID: 8.39 -> 7.43 | Improved |
| replace ITGC with DiME distillation | StyleGAN2 compression · FFHQ 256 StyleGAN2, 98.92% compression | FID: 22.23 -> 18.32 | Improved |
| use NICKEL & DiME at extreme 99.69% compression instead of ITGC | StyleGAN2 compression · FFHQ 256 StyleGAN2, 99.69% compression | FID: 164.92 -> 29.38 | Improved |
| apply NICKEL & DiME instead of CAGC on BigGAN CIFAR-10 | BigGAN compression · CIFAR-10 BigGAN, 90.45% compression | FID: 26.32 -> 26.66 | No clear effect |
| apply NICKEL & DiME at 74.96% compression on AFHQ instead of the full model | StyleGAN2 compression · AFHQ 512 | FID: 3.01 -> 3.17 | Mixed |
Closest papers by meaning
- Content-Aware GAN Compression (CVPR 2021 · similarity 0.92)
- GAN Compression: Efficient Architectures for Interactive Conditional GANs (CVPR 2020 · similarity 0.91)
- Online Multi-Granularity Distillation for GAN Compression (ICCV 2021 · similarity 0.91)
- AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks (ICML 2020 · similarity 0.90)
- Discriminator-Cooperated Feature Map Distillation for GAN Compression (CVPR 2023 · similarity 0.90)
- Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold (AAAI 2024 · 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.
