Research map / Image generation / Super-resolution
Eliminating VAE for Fast and High-Resolution Generative Detail Restoration
Yan Wang, Shijie Zhao, Junlin Li, Li 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 |
|---|---|---|---|
| replace the VAE with pixel unshuffle/shuffle (ablation) | GenDR · RealLR250 4K | runtime: 4.92s -> 1.75s | Improved |
| replace the VAE with pixel unshuffle/shuffle (ablation) | GenDR · RealLR250 4K | MUSIQ: 70.96 -> 70.23 | No clear effect |
| remove only the encoder (GenDR-Adc) instead of the whole VAE (ablation) | GenDR · RealLR250 4K | runtime: 4.92s -> 2.69s | Improved |
| add adversarial discriminator in stage II (ablation) | GenDR-Pix · ImageNet-Test | MUSIQ: +0.12 MUSIQ, +0.0533 CLIPIQA | Improved |
| use GenDR-Adc features for adversarial learning instead of GenDR (ablation) | GenDR-Pix · ImageNet-Test | MUSIQ: +2.20 MUSIQ, +0.0307 CLIPIQA | Improved |
| add RandPad augmentation (ablation) | GenDR-Pix · ImageNet-Test | CLIPIQA: +0.0191 CLIPIQA, -0.17 NIQE | Improved |
| add multi-scale frequency (MFS) loss (ablation) | GenDR-Pix · ImageNet-Test | IQA metrics | Improved |
| use PadCFG instead of vanilla CFG (ablation) | GenDR-Pix · RealLR250 4K | MUSIQ: 58.67 -> 59.12 | Improved |
| use PadCFG instead of vanilla CFG (ablation) | GenDR-Pix · RealLR250 4K | CLIPIQA: 0.5005 -> 0.5066 | Improved |
| use PadCFG instead of CFG with ensemble (ablation) | GenDR-Pix · RealLR250 4K | runtime: 3.54s -> 1.75s | Improved |
| disable CFG (ablation) | GenDR-Pix · RealLR250 4K | runtime: 1.75s -> 0.87s | Improved |
| disable CFG (ablation) | GenDR-Pix · RealLR250 4K | MUSIQ: 70.23 -> 68.64 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- GenDR-1 — GenDR: Lighten Generative Detail Restoration (ICLR 2026) · better on 2 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.89)
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.88)
- Noise-Started One-Step Image Super-Resolution via LR-Conditioned SplitMeanFlow and GAN Refinement (NeurIPS 2026 · similarity 0.87)
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.87)
- Epsilon-VAE: Denoising as Visual Decoding (ICML 2025 · similarity 0.87)
- DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation (NeurIPS 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.
