Research map / Image generation / Super-resolution

Auto-Encoded Supervision for Perceptual Image Super-Resolution

CVPR 2025 · Image generation › Super-resolution · approach: Perceptual and real-world SR · 6 citations (OpenAlex)

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 AESOP, a reconstruction loss that computes L_p distance in the output space of an auto-encoder pretrained with pixel loss, so GAN-based perceptual SR is supervised only on the fidelity-bias component with weight 1 instead of a small pixel-loss coefficient.

Problem

Pixel-level L_p loss in GAN-based perceptual super-resolution penalizes perceptual variance (texture) and causes blurring, forcing prior work to use small loss coefficients or low-pass filters that weaken fidelity guidance.

Approach

Replace the pixel-space L_p loss with L_p computed after a pretrained auto-encoder (decoder output space) that acts as a differentiable fidelity-bias estimator; the AE's LR bottleneck removes randomness-driven perceptual variance, so only the regressable fidelity-bias error is penalized and a full weight of 1 can be used.

Key results

• On the RRDB backbone, AESOP improves LPIPS and DISTS and PSNR over LDL across the benchmarks (e.g., DIV2K-val LPIPS 0.0977 to 0.0936 and PSNR 29.079 vs 28.819 in ablation-table settings).
• Ablation on DIV2K-val: full AESOP reaches DISTS 0.0518 and PSNR 29.079 versus LDL baseline DISTS 0.0526 and PSNR 28.819.
• On real-world sets, AESOP improves NIQE on RealSRv3 from 4.6790 to 4.2337 and MANIQA from 0.3662 to 0.4136 versus Real-ESRGAN with pixel loss.

Measured effects reported in the paper

Surprising findings

• Using a full loss weight of 1 for the AE-space loss, rather than the usual small coefficient of about 0.01, improves both fidelity and perceptual metrics.
• Low-pass filtering the pixel loss hurts performance relative to the AE-based loss, indicating that regressable high-frequency fidelity components are lost by LPF.

Limitations

• Relies on the quality of the pretrained auto-encoder; an SRResNet-based AE gives a slight performance drop.
• Restricted to GAN-based SR; diffusion-based SR is out of scope.
• Does not improve the perceptual losses themselves, which are kept unchanged.

Experimental setup

Datasets: DIV2K Flickr2K DF2K

Benchmarks: Set14 Manga109 General100 Urban100 DIV2K-val BSD100 LSDIR RealSRv3 DRealSR

Metrics: PSNR SSIM LPIPS DISTS NIQE MANIQA AE-PSNR LR-PSNR

Baselines: ESRGAN SPSR LDL CALGAN SwinIR+GAN Real-ESRGAN Bicubic

Builds on: LDL ESRGAN SPSR SwinIR RRDBNet

Closest papers by meaning

• Training Generative Image Super-Resolution Models by Wavelet-Domain Losses Enables Better Control of Artifacts (CVPR 2024 · similarity 0.88)
• Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · similarity 0.88)
• Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · similarity 0.88)
• When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution (NeurIPS 2026 · similarity 0.87)
• Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment (CVPR 2025 · similarity 0.87)
• SRSR: Enhancing Semantic Accuracy in Real-World Image Super-Resolution with Spatially Re-Focused Text-Conditioning (NeurIPS 2025 · similarity 0.87)