Research map / Image generation / Image restoration
DreamUHD: Frequency Enhanced Variational Autoencoder for Ultra-High-Definition Image Restoration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FEVAE-UHD, a frozen frequency-enhanced VAE (FE-VAE) with a wavelet-based adapter that performs UHD image restoration in latent space for dehazing, deblurring, low-light enhancement and demoireing.
Problem
UHD image restoration suffers from inconsistency and detail loss due to downsampling-based resampling, and naive VAEs are large, suffer a domain gap on degraded inputs, and lose high-frequency information in latent encoding.
Approach
Replace learnable resampling with a lightweight Fourier-enhanced VAE pre-trained on clean images, then freeze it and attach a wavelet-based adapter that injects degraded-image features and high-frequency subbands into the encoder and decoder, with a small latent restoration network (Cubic-Mixer) in between.
Key results
- On UHD-Blur deblurring, FEVAE-UHD reaches 29.338 PSNR / 0.8523 SSIM with 1.456M params, about 4 dB above FFTformer.
- On UHD-Haze dehazing, reaches 24.357 PSNR, over 3 dB above MB-TaylorFormer (20.994) with 1.215M params.
- FE-VAE matches the vanilla VAE's stage-1 reconstruction (34.529 vs 35.313 PSNR) with 1.061M params vs 83.614M and 3.437 vs 445.307 GFLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FEVAE-UHD instead of UHDformer | UHD-Haze · UHD-Haze | PSNR: 22.586 -> 24.357 | Helped |
| Use FEVAE-UHD instead of UHDformer | UHD-Haze · UHD-Haze | SSIM: 0.9427 -> 0.9454 | Helped |
| Use FEVAE-UHD instead of UHDformer | UHD-Blur · UHD-Blur | PSNR: 28.821 -> 29.338 | Helped |
| Use FEVAE-UHD instead of UHDDIP | UHD-Blur · UHD-Blur | SSIM: 0.8452 -> 0.8523 | Helped |
| Use FEVAE-UHD instead of WaveMamba | UHD-LL · UHD-LL | PSNR: 27.350 -> 27.729 | Helped |
| Use FEVAE-UHD instead of UHDIP | UHD-LL · UHD-LL | SSIM: 0.9281 -> 0.9287 | Helped |
| Use FEVAE-UHD instead of ESDNet-L | UHDM · UHDM | PSNR: 22.422 -> 23.242 | Helped |
| Use FEVAE-UHD instead of UHDformer | UHDM · UHDM | SSIM: 0.8334 -> 0.8427 | Helped |
| Use FE-VAE resampling rather than the listed alternative resamplers (ablation) | UHD-Blur resampler plus enhancer · UHD-Blur | PSNR: 24.983, 25.672, 27.345, 26.833, 27.819 -> 29.338 | Helped |
| Replace vanilla VAE with FE-VAE (ablation) | FEVAE-UHD · Image reconstruction stage 1 | PSNR: 35.313 -> 34.529 | Hurt |
| Replace the reduced-channel/block VAEs with FE-VAE (ablation) | FEVAE-UHD · Image restoration stage 2, UHD-Blur | PSNR: 25.613 -> 29.338 | Helped |
| Add the FE-block (ablation) | FE-VAE · UHD-Blur | PSNR: 24.548 -> 29.338 | Helped |
Surprising findings
- Standard VAE-based self-reconstruction and lightweight VAE variants lose notable PSNR in the restoration stage (vanilla VAE 30.014 stage-2 PSNR; lightweight VAEs 25.613), showing that a frozen generic VAE alone is insufficient.
Limitations
- Compares only against a limited set of resampling methods and a self-trained lightweight VAE baseline rather than pretrained large VAEs.
- Relies on paired degraded/clean training data per task; separate models per task are evaluated.
Experimental setup
Datasets: UHD-Haze UHD-Blur UHD-LL UHDM
Benchmarks: UHD-Haze UHD-Blur UHD-LL UHDM
Metrics: PSNR SSIM parameter count FLOPs
Baselines: UHD Restormer Uformer DehazeFormer MB-TaylorFormer UHDformer UHDDIP MIMO-Unet++ Stripformer FFTformer FHDe2Net ESDNet
Builds on: Latent diffusion VAE Cubic-Mixer UHDformer Lanczos resampling pixel shuffle
Closest papers by meaning
- UHD-processer: Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts (CVPR 2025 · similarity 0.91)
- Latent Harmony: Synergistic Unified UHD Image Restoration via Latent Space Regularization and Controllable Refinement (NeurIPS 2025 · similarity 0.91)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.89)
- SeNM-VAE: Semi-Supervised Noise Modeling with Hierarchical Variational Autoencoder (CVPR 2024 · similarity 0.87)
- Correlation Matching Transformation Transformers for UHD Image Restoration (AAAI 2024 · similarity 0.86)
- SSC-VAE: Structured Sparse Coding Based Variational Autoencoder for Detail Preserved Image Reconstruction (AAAI 2025 · similarity 0.86)
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.