Research map / Image generation / GANs
Efficient Training with Denoised Neural Weights
Yifan Gong, Zheng Zhan, Yanyu Li, Yerlan Idelbayev, Andrey Zharkov, Kfir Aberman, Sergey Tulyakov, Yanzhi Wang, Jian Ren
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 |
|---|---|---|---|
| initialize GAN LoRA weights from diffusion weight generator and fine-tune 20 epochs | GAN image translation · FFHQ concepts avg | FID: 93.86 -> 89.93 (vs E2GAN); 99.20 -> 89.93 (vs Pix2pix) | Improved |
| use generated initialization instead of training from scratch/E2GAN | GAN image translation · FFHQ concepts avg | training time: 659.8 secs -> 43.3 secs (15x vs Pix2pix); 198.5 -> 43.3 secs (4.6x vs E2GAN) | Improved |
| use block size 256 instead of 128 or 512 (ablation) | weight generator · Grey hair / Batik | FID: Grey hair 107.67/130.85 -> 98.35; Batik 128.44/118.31 -> 94.36 | Improved |
| group LoRA down/up and BN of the same layer together rather than per-layer single group (ablation) | weight generator data format · Grey hair / Batik | FID: Grey hair 122.71 -> 98.35; Batik 126.50 -> 94.36 | Improved |
| concatenate grouped weights through channel dimension (Rule 2) vs appending (Rule 1) (ablation) | weight generator data format · Grey hair / Batik | FID: Grey hair 98.35 -> 98.40; Batik 94.36 -> 95.17 | No clear effect |
| train weight generator end-to-end with added conditional GAN loss (ablation) | weight generator · Grey hair, Zombie, Amrita Sher Gil | FID: Grey hair 76.52 -> 77.62; Amrita Sher Gil 95.14 -> 103.87 | Hurt |
| apply LoRA to GAN layers to reduce weights to predict | GAN generator · weight format | number of weights: 7.06M -> 0.22M | Improved |
Closest papers by meaning
- Diffusion-based Neural Network Weights Generation (ICLR 2025 · similarity 0.88)
- Sinusoidal Initialization, Time for a New Start (NeurIPS 2025 · similarity 0.86)
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.86)
- E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 2024 · similarity 0.86)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.85)
- Initializing Models with Larger Ones (ICLR 2024 · similarity 0.85)
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.
