Research map / Image generation / Super-resolution
Fine-structure Preserved Real-world Image Super-resolution via Transfer VAE Training
Qiaosi Yi, Shuai Li, Rongyuan Wu, Lingchen Sun, Yuhui Wu, Lei Zhang
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 Transfer VAE Training (align latent space with VAE-D8) instead of training VAE-D4 from scratch (ablation) | VAE-D4 Real-ISR · RealSR | PSNR: 25.50 -> 25.83 | Improved |
| add Transfer VAE Training instead of L1 latent alignment (ablation) | VAE-D4 Real-ISR · RealSR | CLIPIQA: 0.6734 -> 0.6924 | Improved |
| add TVT versus training VAE-D4 from scratch (ablation) | VAE-D4 Real-ISR · RealSR | MUSIQ: 67.85 -> 69.26 | Improved |
| use compact VAE-D4 (no middle block, 128/256/256 channels) instead of larger V1 (ablation) | VAE-D4 · RealSR | FLOPs: 3.02 T -> 1.97 T; PSNR 25.84 -> 25.83 | No clear effect |
| use compact VAE-D4 instead of V2 (ablation) | VAE-D4 · RealSR | SSIM: 0.7603 -> 0.7619 | No clear effect |
| replace VAE-D8 with VAE-D4 (full UNet) (ablation) | one-step Real-ISR · RealSR | SSIM: 0.7419 -> 0.7632 | Improved |
| replace VAE-D8 with VAE-D4 (full UNet) (ablation) | one-step Real-ISR · RealSR | Q-Align: 3.592 -> 3.797 | Improved |
| add skip connections to VAE-D8 (ablation) | VAE-D8 Real-ISR · RealSR | Q-Align: 3.592 -> 3.482 | Hurt |
| use 8-channel latent VAE-D8 instead of 4-channel (ablation) | VAE-D8 Real-ISR · RealSR | MUSIQ: 68.86 -> 67.79 | Hurt |
| use compute-efficient UNet instead of 25% channel-pruned UNet (ablation) | VAE-D4 Real-ISR · RealSR | SSIM: 0.7446 -> 0.7619 | Improved |
| use compute-efficient UNet instead of full UNet with VAE-D4 (ablation) | VAE-D4 Real-ISR · RealSR | FLOPs: 2.59 T -> 1.97 T | Improved |
| replace VAE-D8 with adapted VAE-D4 for reconstruction | VAE · Urban100 (VAE reconstruction) | PSNR: 25.29 -> 32.35 | Improved |
Closest papers by meaning
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.88)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.88)
- Edge-SD-SR: Low Latency and Parameter Efficient On-device Super-Resolution with Stable Diffusion via Bidirectional Conditioning (CVPR 2025 · similarity 0.87)
- One-Step Effective Diffusion Network for Real-World Image Super-Resolution (NeurIPS 2024 · similarity 0.87)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.87)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · 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.
