An image-upscaling method has reported a better balance between reference-based fidelity and perceptual quality on a real-image benchmark. Called UGDiff, it uses uncertainty-guided latent diffusion to adapt how it restores detail. On RealSR, the paper reports the best perception-distortion balance among the compared diffusion-based super-resolution methods when PSNR is weighed against NIQE.
The headline result is not a clean sweep. With gamma = 0.1, the UGDiff setting recorded PSNR 24.1528, SSIM 0.6700, LPIPS 0.3170 and NIQE 5.6148 on RealSR. DiffBIR, using s = 0.3, recorded 24.0969, 0.6707, 0.3279 and 6.2908. That meant higher PSNR and lower LPIPS and NIQE for UGDiff, but slightly lower SSIM.
Different measures told slightly different stories
On synthetic DIV2K-Val, UGDiff at gamma = 0.1 reported PSNR 22.2350, SSIM 0.5471, LPIPS 0.3406 and NIQE 4.7852. DiffBIR at s = 0.15 reported 22.2730, 0.5418, 0.3413 and 5.2346. The paper therefore describes the PSNR values as comparable while reporting higher SSIM and lower LPIPS and NIQE for UGDiff.
The scorecard mixes different kinds of evidence. PSNR and SSIM are used here as reference-based fidelity measures, LPIPS as a reference-based perceptual measure and NIQE as a non-reference perceptual measure.
A sampler that tracks where detail is uncertain
UGDiff's central idea is to estimate uncertainty in the image's latent reconstruction and combine it with the diffusion sampler's posterior variance. The system then adaptively restores high-frequency detail in regions judged uncertain, while aiming to preserve fidelity and reduce its reliance on the high-fidelity image later in sampling.
To train that uncertainty estimate, the paper defines an optimal target from the element-by-element squared difference between the ground-truth latent image and the mean produced by the regression restoration. It trains E-prime with a simple L2 loss, focusing the learning on uncertainty estimation without compromising mean accuracy, according to the method description.
The experimental setup used 4× upsampling, taking 128 × 128 inputs to 512 × 512 outputs. Training for E-prime used LSDIR and the first 10,000 FFHQ face images, with Real-ESRGAN-generated low-quality/high-quality pairs; validation used 100 random DIV2K crops called DIV2K-Ours. The benchmark tests used synthetic DIV2K-Val and real RealSR data.
The implementation builds on DiffBIR and Stable Diffusion v2.1 components and uses BSRNet for restoration. E-prime was trained from scratch with AdamW at a learning rate of 10^-4, batch size 128, crop size 256 × 256, four NVIDIA A100 GPUs and 90,000 iterations. Sampling used spaced DDPM for T = 50 steps.
An ablation of the mean estimator also favored the BSRNet-derived mean. On DIV2K-Ours, it reported PSNR 20.2307 versus 19.5111 for the NLL-trained mean; SSIM 0.4222 versus 0.3844; LPIPS 0.4388 versus 0.4854; and NIQE 4.6974 versus 5.6106.
The balance can be tuned
The method exposes settings that let the authors move along the fidelity-perceptual-quality trade-off. In the DIV2K-Ours guidance ablation, they selected s = 100 because it gave the optimal balance among configurations with gamma = 1.0. With s fixed, the paper describes smaller gamma as a stronger pull toward the mean and as favoring fidelity, with the reverse trade-off for larger gamma.
The visual examples followed that pattern. UGDiff at gamma = 0.5 was reported as visually similar to DiffBIR at s = 0 when PSNR was comparable, while gamma = 0.1 avoided unnecessary high-frequency detail, particularly in flat regions. Supplementary sampling analysis also reported suppressed unnecessary textures in low-uncertainty regions, higher eventual PSNR and an NIQE curve nearly identical to baseline.
A separate ablation compared two uncertainty parameterizations. With s = 100 and gamma = 0.1 on DIV2K-Ours, sigma-u recorded PSNR 20.9637, SSIM 0.4780, LPIPS 0.4095 and NIQE 4.5425. Sigma-u2 recorded 20.7987, 0.4676, 0.4118 and 4.3496. The paper says sigma-u was better on most metrics and selected it, even though sigma-u2 had the lower NIQE in that table.
What the report leaves unresolved
The paper reports point estimates for the comparisons. No confidence intervals or significance tests are reported, so the report does not quantify statistical uncertainty around the differences.
The paper is an arXiv v1 preprint dated 26 Aug 2026. Its experiment description covers DIV2K-Val and RealSR at 4× upsampling, using the listed model components and sampling settings.
The authors interpret the benchmark and ablation results as showing that UGDiff achieves a better perception-distortion balance than the compared diffusion-based methods. The evidence presented consists of image-benchmark metrics, qualitative comparisons and the tested configurations, so the conclusion concerns those tests and settings.
Paper data and sources
Original title: Uncertainty-Guided Latent Diffusion Models for Faithful Super Resolution
Authors: Ren Wang, Yung-Yu Chuang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text