Joint Geometric and Trajectory Consistency Learning for One-Step Real-World Super-Resolution
| Authors | Chengyan Deng et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2602.24240 |
| Download | |
| Categories | cs.CV |
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.
Engineering Breakdown
Plain English
This paper tackles the problem of real-world image super-resolution using one-step diffusion models, which is challenging because standard diffusion requires many iterations and consistency models suffer from accuracy issues. The authors identify two specific problems: consistency drift that accumulates during training, and 'Geometric Decoupling' where pixel-level alignment is achieved but structural coherence fails. They propose a method called GT (Geometric and Trajectory consistency) that jointly constrains both the geometric structure and the generative trajectory to produce high-quality super-resolved images efficiently. The approach aims to match the perceptual quality of multi-step diffusion models while maintaining the computational efficiency of one-step generation.
Core Technical Contribution
The core novelty is a dual-constraint framework that simultaneously enforces geometric consistency (structural preservation) and trajectory consistency (smooth generative paths) during training of one-step diffusion models. The authors introduce the concept of 'Geometric Decoupling' as a failure mode specific to consistency models—where the model achieves pixel-wise alignment with targets but loses high-level structural information. Rather than scaling up model parameters or relying on distillation from massive T2I priors, GT uses lightweight consistency models with explicit geometric regularization to prevent this decoupling. The approach directly addresses the transitive error accumulation problem in consistency model training by constraining the generation trajectory itself, not just the endpoint.
How It Works
The system takes a low-resolution input image and processes it through a lightweight consistency model that generates a high-resolution output in a single forward pass. The training process enforces two complementary losses: (1) a trajectory consistency loss that ensures the generative path follows a smooth, valid path through the diffusion process, and (2) a geometric consistency loss that preserves structural features like edges, shapes, and spatial relationships from both the input image and high-quality reference outputs. The geometric component likely uses perceptual losses or structural similarity metrics to compare intermediate representations, while the trajectory component constrains how the model moves through latent space during generation. These constraints work together to prevent both drift accumulation (trajectory loss) and structural collapse (geometric loss), allowing the model to maintain coherence while operating with limited parameters and single-step inference.
Production Impact
For production image processing pipelines, this approach offers a dramatic reduction in latency—replacing multi-step diffusion (10-50 iterations) with a single forward pass while maintaining competitive perceptual quality. Teams building real-time super-resolution systems (mobile apps, video processing, surveillance) would benefit from the lightweight parameter count, which enables deployment on edge devices and reduces memory overhead compared to distillation-based competitors. The explicit geometric constraints mean the model naturally preserves important structural details like text, faces, and fine textures, reducing hallucination artifacts that plague standard consistency models. Integration would require replacing the diffusion sampling loop with a single inference call, simplifying deployment complexity; however, the method's effectiveness depends on access to clean reference images during training, which may require careful dataset curation for domain-specific applications.
Limitations and When Not to Use This
The paper's approach assumes high-quality paired training data (low-resolution input with ground-truth high-resolution reference) is available, which limits applicability to domains where such pairs are expensive or infeasible to obtain. The geometric consistency loss relies on hand-crafted structural metrics or perceptual losses, which may not capture all forms of semantic coherence—some applications might need task-specific geometric constraints that aren't addressed here. The method doesn't address whether trajectory constraints generalize across different image domains or degradation types; if trained on natural images, it may fail on medical, synthetic, or highly stylized content. Additionally, there's no discussion of how the approach scales to 4x or 8x magnification factors, only implicit focus on 2-4x typical super-resolution, and the paper likely doesn't explore robustness to out-of-distribution noise or compression artifacts.
Research Context
This work builds directly on consistency models (Song et al., 2023) which enabled one-step generation but suffered from training instability and quality degradation compared to diffusion. It competes with recent distillation-based super-resolution approaches that leverage large pretrained text-to-image models (like Stable Diffusion), but offers a lighter-weight alternative with explicitly better structural preservation. The research contributes to the broader trend of trading multi-step inference for single-step inference through better training objectives rather than model scaling, relevant to the wider efficiency-focused diffusion research. The geometric decoupling observation is a novel failure mode characterization that opens opportunities for structure-aware consistency training in other generative tasks beyond super-resolution.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::