Joint Geometric and Trajectory Consistency Learning for One-Step Real-World Super-Resolution

Chengyan Deng; Zhangquan Chen; Li Yu; Kai Zhang; Xue Zhou; Wang Zhang

cs.CVFeb 27, 2026

Joint Geometric and Trajectory Consistency Learning for One-Step Real-World Super-Resolution

Authors: Chengyan Deng, Zhangquan Chen, Li Yu, Kai Zhang, Xue Zhou, Wang Zhang

Organizations: University of Electronic Science and Technology of China · Tsinghua University · Nanjing University

Abstract

Diffusion-based Real-World Image Super-Resolution (Real-ISR) achieves impressive perceptual quality but suffers from high computational costs due to iterative sampling. While recent distillation approaches leveraging large-scale Text-to-Image (T2I) priors have enabled one-step generation, they are typically hindered by prohibitive parameter counts and the inherent capability bounds imposed by teacher models. As a lightweight alternative, Consistency Models offer efficient inference but struggle with two critical limitations: the accumulation of consistency drift inherent to transitive training, and a phenomenon we term "Geometric Decoupling" - where the generative trajectory achieves pixel-wise alignment yet fails to preserve structural coherence. To address these challenges, we propose GTASR (Geometric Trajectory Alignment Super-Resolution), a simple yet effective consistency training paradigm for Real-ISR. Specifically, we introduce a Trajectory Alignment (TA) strategy to rectify the tangent vector field via full-path projection, and a Dual-Reference Structural Rectification (DRSR) mechanism to enforce strict structural constraints. Extensive experiments verify that GTASR delivers superior performance over representative baselines while maintaining minimal latency. The code and model will be released at https://github.com/Blazedengcy/GTASR.

Figures & tables

Figure 1: The pipeline of the proposed GTASR. We employ a two-stage training scheme. In Stage I, we train the online model $f_{\theta}$ using the standard $\mathcal{L}_{\mathrm{CT}}$ , augmented by our proposed $\mathcal{L}_{\mathrm{TA}}$ , where the reference parameters $\theta^{-}$ are updated at every step using the stop-gradient online parameters. In Stage II, we address geometric decoupling by introducing a target model $f_{\theta^{\prime}}$ initialized with the pre-trained weights from Stage I, with its parameters updated via periodic synchronization to ensure temporal alignment. We feed the intermediate states $\hat{x}_{t^{\prime}}$ and $x_{t^{\prime}}$ into $f_{\theta^{\prime}}$ to obtain the trajectory endpoints ( $x^{\mathrm{fake}}_{0}$ and $x^{\mathrm{real}}_{0}$ ), which are directly utilized for $\mathcal{L}_{\mathrm{DTM}}$ while their structure maps extracted via the Sobel operator are utilized for the proposed DRSR objectives ( $\mathcal{L}_{\mathrm{Stab}}$ and $\mathcal{L}_{\mathrm{Rect}}$ ). Finally, the gradients derived from $\mathcal{L}_{\mathrm{CT}}$ , $\mathcal{L}_{\mathrm{DTM}}$ , $\mathcal{L}_{\mathrm{Stab}}$ , and $\mathcal{L}_{\mathrm{Rect}}$ are backpropagated to $f_{\theta}$ to jointly enhance perceptual realism and structural integrity.

Figure 2: Visual analysis of consistency across timesteps. (Left) Baseline CT suffers from severe trajectory drift and detail loss as $t$ increases, particularly in the eye region. (Right) After introducing TA strategy effectively preserves sharp features even at large time steps.

Figure 3: Visual comparison of different strategies. (a) The baseline $\mathcal{L}_{\mathrm{CT}}$ fails to model fine details effectively. (b) Integrating $\mathcal{L}_{\mathrm{TP}}$ with $\mathcal{L}_{\mathrm{CT}}$ guides the reconstruction direction; however, the metallic grille remains overly blurred and lacks definition. (c) Ours restores intricate mesh textures with superior clarity.

Figure 4: Visualization of Geometric Decoupling. Evaluated on 3,000 ImageNet-Test images, the Baseline (blue) achieves positional convergence (low y-axis error) but fails to maintain structural stability (high x-axis variance). In contrast, GTASR (red) minimizes errors on both metrics, effectively resolving the decoupling.

Datasets	Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	NIQE $\downarrow$	LIQE $\uparrow$	TOPIQ $\uparrow$
ImageNet-Test	BSRGAN	24.42	0.6585	0.2585	0.5812	54.70	0.3866	6.08	3.90	0.6078
	Real-ESRGAN	24.04	0.6650	0.2540	0.5234	52.54	0.3671	6.07	3.83	0.5545
	StableSR-200	20.73	0.4650	0.3927	0.6263	56.85	0.4272	8.32	3.49	0.5818
	ResShift-15	24.95	0.6741	0.2377	0.5823	53.11	0.4168	6.90	3.45	0.5877
	ResShift-4	25.02	0.6833	0.2074	0.5993	52.10	0.3888	7.33	3.40	0.5679
	SinSR-1	24.70	0.6635	0.2187	0.6079	53.53	0.4152	6.29	3.58	0.5950

Table 1: Quantitative comparison on ImageNet-Test and RealSR. The best and second-best results are highlighted in bold and underline .

Figure 5: Visual comparisons of different methods on two synthetic examples of the ImageNet-Test dataset.

Methods	RealLQ250						RealSet65
	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	NIQE $\downarrow$	LIQE $\uparrow$	TOPIQ $\uparrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	NIQE $\downarrow$	LIQE $\uparrow$	TOPIQ $\uparrow$
BSRGAN	0.5690	63.51	0.3514	4.54	3.31	0.5386	0.6162	65.58	0.3888	4.58	3.72	0.5809
Real-ESRGAN	0.5435	62.52	0.3565	4.13	3.34	0.5079	0.5990	63.22	0.3893	4.41	3.60	0.5365
StableSR-200	0.4001	48.83	0.2694	5.84	2.46	0.4405	0.4492	48.75	0.3098	5.74	2.91	0.4644
ResShift-15	0.6006	58.92	0.3898	6.29	3.11	0.4939	0.6496	61.16	0.4079	6.16	3.34	0.5397
ResShift-4	0.5835	56.20	0.3480	7.34	2.80	0.4420	0.6274	59.39	0.3621	6.74	3.16	0.4951

Table 2: Quantitative results of models on two real-world datasets. The best and second best results are highlighted in bold and underline .

Figure 6: Visual comparisons of different methods on two examples of real-world datasets. Please zoom in for more details.

Methods	Runtime (s) $\downarrow$	LPIPS $\downarrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$	TOPIQ $\uparrow$
StableSR-200	11.21	0.3927	0.4272	0.6263	0.5818
ResShift-15	0.93	0.2377	0.4168	0.5823	0.5877
UPSR-5	0.35	0.2464	0.4695	0.6149	0.6491
SinSR-1	0.12	0.2187	0.4152	0.6079	0.5950
CTMSR-1	0.08	0.1969	0.4857	0.6912	0.6793
GTASR-1 (Ours)	0.08	0.1916	0.5826	0.7475	0.7423

Table 3: Computational efficiency and performance comparison. We report Runtime for

128\times 128

input on a single RTX 4090 GPU and present perceptual metrics evaluated on ImageNet-Test.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	LIQE $\uparrow$	TOPIQ $\uparrow$
w/o $\mathcal{L}_{\mathrm{TA}}$	24.65	0.2073	0.6060	55.38	0.3830	3.86	0.5939
w/ $\mathcal{L}_{\mathrm{TA}}$	25.05	0.1856	0.6291	58.40	0.4457	4.05	0.6433

Table 4: Ablation study of the Stage-I training objectives on ImageNet-Test.

$\mathcal{L}_{\mathrm{DTM}}$	$\mathcal{L}_{\mathrm{Stab}}$	$\mathcal{L}_{\mathrm{Rect}}$	PSNR $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	LIQE $\uparrow$	TOPIQ $\uparrow$
✓			24.76	0.2019	0.6965	58.91	0.5070	3.77	0.6863
✓	✓		24.00	0.1929	0.7351	63.43	0.5676	4.42	0.7349
✓		✓	24.62	0.2000	0.7040	60.41	0.5403	3.89	0.6953
✓	✓	✓	24.01	0.1916	0.7475	65.09	0.5826	4.47	0.7423

Table 5: Ablation study of the Stage-II training objectives on ImageNet-Test.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	Runtime (s) $\downarrow$	Params (M) $\downarrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$	LIQE $\uparrow$	TOPIQ $\uparrow$
StableSR-200	11.21	1928	0.3927	56.85	0.4272	0.6263	3.49	0.5818
ResShift-15	0.93	174	0.2377	53.11	0.4168	0.5823	3.45	0.5877
ResShift-4	0.37	174	0.2074	52.10	0.3888	0.5993	3.40	0.5679
UPSR-5	0.35	122	0.2464	59.20	0.4695	0.6149	4.00	0.6491
SinSR-1	0.12	174	0.2187	53.53	0.4152	0.6079	3.58	0.5950
CTMSR-1	0.08	172	0.1969	60.17	0.4857	0.6912	4.08	0.6793

Appendix

Table 6: Computational efficiency and performance comparison. We report Runtime and Parameters (Params) measured with an input size of

128\times 128

on a single RTX 4090 GPU. We also present perceptual metrics evaluated on ImageNet-Test . Bold and underlined indicate the best and second-best performance, respectively.

Methods	Runtime (s) $\downarrow$	Params (M) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	NIQE $\downarrow$	LIQE $\uparrow$	TOPIQ $\uparrow$
AddSR-4	0.70	2280	24.23	0.7032	0.3141	0.5127	62.64	0.4116	6.19	3.38	0.5535
OSEDiff-1	0.43	1775	25.15	0.7341	0.2921	0.6682	69.08	0.4717	5.64	4.06	0.6253
GTASR-1 (Ours)	0.08	172	25.92	0.7471	0.3041	0.6962	67.31	0.4930	5.33	3.70	0.6780

Appendix

Table 7: Quantitative comparison on RealSR with center-cropped

128\times 128

LR inputs. The best and second-best results are highlighted in bold and underline (”-N” behind the method name represents the number of inference steps).

Methods	RealLQ250						RealSet65
	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	NIQE $\downarrow$	LIQE $\uparrow$	TOPIQ $\uparrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	NIQE $\downarrow$	LIQE $\uparrow$	TOPIQ $\uparrow$
AddSR-4	0.5732	64.40	0.3681	4.44	3.26	0.5486	0.5917	63.77	0.3903	5.22	3.49	0.5586
OSEDiff-1	0.6722	69.556	0.4250	3.98	3.90	0.6075	0.7041	68.65	0.4699	4.79	4.08	0.6137
GTASR-1 (Ours)	0.7355	70.90	0.4870	4.32	3.79	0.7047	0.7491	69.49	0.4989	4.45	3.97	0.6937

Appendix

Table 8: Quantitative results of models on two real-world datasets. The best and second best results are highlighted in bold and underline .

$\mathcal{L}_{\mathrm{CT}}$	$\mathcal{L}_{\mathrm{TA}}$	$\mathcal{L}_{\mathrm{DTM}}$	$\mathcal{L}_{\mathrm{Stab}}$	$\mathcal{L}_{\mathrm{Rect}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	NIQE $\downarrow$	LIQE $\uparrow$	TOPIQ $\uparrow$
✓		✓	✓	✓	24.24	0.6540	0.2040	0.7036	61.94	0.4770	5.87	4.20	0.7036
✓	✓	✓	✓	✓	24.01	0.6520	0.1916	0.7475	65.09	0.5826	5.83	4.47	0.7423

Appendix

Table 9: Impact of the Trajectory Alignment (TA) strategy on restoration quality on ImageNet-Test. The best and second best results are highlighted in bold and underline .

Figure 7: Visual comparison of different methods on a synthetic dataset. Please zoom in for more details.

Figure 8: Visual comparison of different methods on a synthetic dataset. Please zoom in for more details.

Figure 9: Visual comparison of different methods on a synthetic dataset. Please zoom in for more details.

Figure 10: Visual comparison of different methods on a synthetic dataset. Please zoom in for more details.

Figure 11: Visual comparison of different methods on real-world datasets. Please zoom in for more details.

Figure 12: Visual comparison of different methods on real-world datasets. Please zoom in for more details.

Figure 13: Visual comparison of different methods on real-world datasets. Please zoom in for more details.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 248 papers

teasr: training-efficient any-step diffusion transformer for real-world image super-resolution
Jun 15, 2026Xiang Gao, Chenxin Zhu, Yushun Fang +2Diffusion Transformer Diffusion Model Distillation
Shanghai Jiao Tong University, China
DriftSR: One-Step Real-World Image Super-Resolution via Distribution Drifting
Oct 3, 2026Wei Zhu, Kai Zhang, Yu Zheng +4Image Super-Resolution Diffusion Models for Image Restoration
Nanjing University of Science and Technology · Huawei · Nanjing University +1
Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution
Mar 6, 2026Jingkai Wang, Yixin Tang, Jue Gong +6Diffusion Transformer Diffusion Model Distillation
Shanghai Jiao Tong University, China · Shenzhen Transsion Holdings Co., Ltd., China

Jun 15, 2026J/K move · Enter open · S save

teasr: training-efficient any-step diffusion transformer for real-world image super-resolution

Xiang Gao, Chenxin Zhu, Yushun Fang, Qiang Hu+1

Shanghai Jiao Tong University, China

Diffusion Transformer Diffusion Model Distillation Diffusion Sampling Self-Distillation Diffusion Model Inference Acceleration Image Super-Resolution

Open paper AskPDF arXiv

Diffusion models excel in Real-World Image Super-Resolution (Real-ISR) due to their powerful generative priors but suffer from slow iterative sampling. Although existing one-step distillation methods accelerate inference, they typically require auxiliary teacher models that inflate training memory and restrict scalability to large-scale architectures. Furthermore, these fixed-step models lack the flexibility to trade off speed for quality. In this paper, we propose TEASR, a training-efficient any-step diffusion framework for Real-ISR that enables both one-step and multi-step restoration within a unified model. Our key idea is to perform self-adversarial distillation within a single diffusion model, eliminating the need for auxiliary teachers or discriminators. Specifically, we propose a timestep-aware rectification strategy that stabilizes one-step generation across noise levels. These two designs further enables the distillation of 20B-parameter diffusion models on a single GPU, significantly improving training efficiency. Moreover, we introduce a dual-branch diffusion transformer with decoupled timestep condition to separate the current noise state and the denoising target to enhance sampling quality. Extensive experiments demonstrate that TEASR supports seamless any-step sampling and consistently outperforms state-of-the-art methods across multiple datasets.

Abstract

Image Super-Resolution

Explore similar work

teasr: training-efficient any-step diffusion transformer for real-world image super-resolution

DriftSR: One-Step Real-World Image Super-Resolution via Distribution Drifting

Spectral and Trajectory Regularization for Diffusion Transformer Super-Resolution