Research map / Image generation / Autoregressive and latent generation
RestoreVAR: Visual Autoregressive Generation for All-in-One Image Restoration
Sudarshan Rajagopalan, Kartik Narayan, Vishal M. Patel
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 latent refiner transformer (LRT) to convert discrete to continuous latents (ablation) | RestoreVAR · RESIDE test | PSNR: 21.71 -> 24.67 | Improved |
| use LRT instead of HART diffusion refiner (ablation) | RestoreVAR · RESIDE test | PSNR: 23.48 -> 24.67 | Improved |
| feed final transformer block output z to LRT via cross-attention (ablation) | LRT · RESIDE test | PSNR: 21.23 -> 24.67 | Improved |
| add LRT refiner (ablation) | RestoreVAR · perceptual quality (MUSIQ/CLIPIQA) | MUSIQ / CLIPIQA: 66.56/0.472 -> 66.45/0.481 | No clear effect |
| condition on continuous latent of degraded image instead of discrete multi-scale latents (ablation) | RestoreVAR · validation accuracy | validation accuracy | Improved |
| add PatchGAN discriminator to VAE decoder fine-tuning (ablation) | VAE decoder · reconstruction quality | sharpness (qualitative) | Improved |
| replace absolute positional embeddings with 2D RoPE (ablation) | RestoreVAR transformer · 512x512 fine-tuning | validation accuracy | Improved |
| fine-tune VAE decoder on continuous latents only | VAR VQVAE decoder · reconstruction (1000 samples) | PSNR (dB): 22.59 -> 28.14 | Improved |
| use VAR-based RestoreVAR instead of LDM-based AiOR methods | generative AiOR · RESIDE | PSNR: 24.48 -> 24.67 | Improved |
| use RestoreVAR instead of LDM-based methods for inference speed | AiOR · inference time per image | time (s): 8.477 -> 0.201 | Improved |
| compare RestoreVAR to LDM-based methods on unseen-degradation perceptual metrics | RestoreVAR · unseen/mixed degradations average | MUSIQ: 56.767 (AutoDir) vs 56.217 | No clear effect |
| compare RestoreVAR to non-generative methods on pixel fidelity | RestoreVAR · RESIDE | PSNR: 32.02 (PromptIR) vs 24.67 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 4 on most shared results.
- AutoDIR — AutoDIR: Automatic All-in-One Image Restoration with Latent Diffusion (ECCV 2024) · better on 6 of 7 shared results
- PixWizard — PixWizard: Versatile Image-to-Image Visual Assistant with Open-Language Instructions (ICLR 2025) · better on 1 of 2 shared results
- PromptIR — PromptIR: Prompting for All-in-One Image Restoration (NeurIPS 2023) · better on 0 of 2 shared results
- Diff-Plugin — Diff-Plugin: Revitalizing Details for Diffusion-based Low-level Tasks (CVPR 2024) · better on 1 of 1 shared result
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
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.91)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.90)
- Visual Autoregressive Modeling for Image Super-Resolution (ICML 2025 · similarity 0.90)
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.90)
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.89)
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.89)
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.
