ECCV 2026

Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution

StrSR is a one-step adversarial distillation framework that brings diffusion transformer priors to real-world image super-resolution while suppressing DiT-specific periodic artifacts.

Jingkai Wang1,* Yixin Tang1,* Jue Gong1 Jiatong Li1 Shu Li2 Libo Liu2 Jianliang Lan2 Yutong Liu1,† Yulun Zhang1,†

1Shanghai Jiao Tong University   2Shenzhen Transsion Holdings Co., Ltd.
*Equal contribution   †Corresponding authors

Paper arXiv Code
Low-quality input image
LQ input
TSD-SR super-resolution result
TSD-SR
FluxSR super-resolution result
FluxSR
StrSR super-resolution result
StrSR (ours)

Abstract

Diffusion transformer (DiT) architectures show great potential for real-world image super-resolution (Real-ISR). However, their computationally expensive iterative sampling necessitates one-step distillation. Existing one-step distillation methods struggle with Real-ISR on DiT: they suffer from fundamental trajectory mismatch and generate severe grid-like periodic artifacts.

To tackle these challenges, we propose StrSR, a novel one-step adversarial distillation framework featuring spectral and trajectory regularization. An asymmetric discriminative distillation architecture bridges the trajectory gap, while frequency distribution matching suppresses DiT-specific periodic artifacts caused by high-frequency spectral leakage. Extensive experiments demonstrate state-of-the-art performance in Real-ISR, across both quantitative metrics and visual perception.

Method

StrSR combines dual-encoder conditioning, one-step DiT generation, asymmetric adversarial distillation, and frequency-domain regularization in a unified Real-ISR framework.

Training pipeline of StrSR
Overall training pipeline of StrSR.

Trajectory Regularization

An asymmetric discriminator based on CLIP-ConvNeXt helps align the one-step restoration trajectory with high-quality image distributions without heavy progressive distillation.

Spectral Regularization

Frequency distribution matching constrains both amplitude and phase statistics, reducing the grid-like periodic artifacts that appear in one-step DiT restoration.

Dual-Encoder Conditioning

A VLM encoder captures semantic guidance while a VAE encoder preserves low-level spatial structure, giving the DiT generator both global priors and local restoration cues.

Results

The paper evaluates StrSR on synthetic and real-world Real-ISR benchmarks, including DIV2K-Val, RealSR, and RealLQ250. Larger versions of the figures are shown below.

Quantitative Comparisons
Quantitative results on DIV2K-Val
Table 1: Quantitative results on the DIV2K-Val synthetic benchmark.
Quantitative results on RealSR
Table 2: Quantitative results on the RealSR benchmark.
Quantitative results on RealLQ250
Table 3: Quantitative results on the RealLQ250 benchmark.
Visual Comparisons
Visual comparison on DIV2K-Val
Figure 7: Visual comparison for x4 image SR on DIV2K-Val.
Visual comparison on RealSR
Figure 8: Visual comparison on RealSR.
Visual comparison on RealLQ250
Figure 9: Visual comparison on RealLQ250.

BibTeX

@inproceedings{wang2026strsr,
  title={Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution},
  author={Wang, Jingkai and Tang, Yixin and Gong, Jue and Li, Jiatong and Li, Shu and Liu, Libo and Lan, Jianliang and Liu, Yutong and Zhang, Yulun},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}