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

This paper proposes GTASR, a lightweight one-step Real-World Super-Resolution framework that overcomes the limitations of existing Consistency Models by introducing Trajectory Alignment and Dual-Reference Structural Rectification to eliminate consistency drift and ensure structural coherence.

The Gist

Imagine you have a blurry, low-resolution photo of a beautiful landscape. You want to turn it into a crystal-clear, high-definition masterpiece. This is the job of Super-Resolution (SR).

For a long time, computers were like clumsy artists trying to fix this photo.

• Old methods were like trying to guess the missing puzzle pieces based on simple rules. They were fast, but the results looked smooth and fake, like a plastic toy.
• New "Diffusion" methods are like a master painter who starts with a canvas full of random noise and slowly paints over it, step-by-step, to reveal the image. The results are stunningly realistic, but the process is painfully slow. It might take the computer 50 to 100 "brush strokes" (iterations) to finish one picture.

The Problem: The "One-Step" Dilemma

Researchers wanted to make this process instant—one single brush stroke.

• Some tried to copy the master painter (using a "Teacher" model), but the teacher was so huge and complex that copying them required a supercomputer.
• Others tried a faster technique called Consistency Models. These are like a student who learns to paint the whole picture in one go. It's fast! But the student has two big problems:
  1. The Drifting Compass: As the student learns, they get slightly confused. If they take a wrong turn early on, they keep getting further off-course, ending up with a blurry mess.
  2. The "Geometric Decoupling" (The Ghostly Face): The student gets the colors right, but the structure is wrong. Imagine painting a face where the eyes and mouth are in the right spots, but the nose is floating in the air, or the jawline is wobbly. The pixels match, but the geometry is broken.

The Solution: GTASR (The Smart Architect)

The authors of this paper propose GTASR (Geometric Trajectory Alignment Super-Resolution). Think of GTASR as a Smart Architect who teaches the student artist two new tricks to fix those problems.

Trick 1: The "Full-Path GPS" (Trajectory Alignment)

The Problem: The student's "compass" (the direction they are painting) gets shaky. They might think they are painting a tree, but they are actually drifting toward a bush.
The Fix: Instead of just telling the student, "Paint the final tree," the architect says, "Check your path at every single step."

• The Analogy: Imagine you are hiking to a summit. A normal teacher says, "Just get to the top." If you take a wrong turn halfway up, you might end up at the wrong mountain.
• GTASR's Approach: The architect projects your current position back onto the "correct trail" at every single step. Even if you wander off, the system instantly snaps you back to the right path before you take the next step. This prevents the "Drifting Compass" and ensures the student stays on the right track from start to finish.

Trick 2: The "Double-Check Blueprint" (Dual-Reference Structural Rectification)

The Problem: The student paints the colors perfectly, but the structure is wobbly (the "Ghostly Face"). The computer thinks the pixels are close enough, but the human eye sees the distortion.
The Fix: The architect gives the student two references to check against, not just one.

1. Reference A (The "Real" Path): Compare the student's painting to what a slightly better version of the student would have painted. This ensures the style and flow are consistent.
2. Reference B (The "Ground Truth" Blueprint): Compare the student's painting directly to the original, perfect photo, but specifically looking at the edges and lines (like the outline of a building or the curve of a nose).

• The Analogy: Imagine building a house.
  • Reference A is checking if the walls are straight relative to each other.
  • Reference B is checking if the house matches the architect's original blueprints exactly.
  • GTASR forces the student to satisfy both checks simultaneously. If the walls are straight but the house is in the wrong shape, the architect says, "No, fix the shape!" This stops the "Ghostly Face" and ensures the geometry is rock-solid.

The Result: Fast, Sharp, and Real

By combining these two tricks, GTASR achieves the "Holy Grail" of image processing:

1. Speed: It generates the image in one step (instantly), just like the fastest methods.
2. Quality: It produces images that look as realistic as the slow, multi-step methods.
3. Structure: The details (like fur on an animal or bricks on a wall) are sharp and geometrically correct, not blurry or floating.

In summary: GTASR is like a student artist who, thanks to a smart GPS and a double-check blueprint system, can paint a masterpiece in a single brushstroke, without ever losing their way or messing up the structure. It's fast, efficient, and creates stunningly real images.

Technical Summary

1. Problem Statement

Real-World Image Super-Resolution (Real-ISR) aims to reconstruct high-quality images from low-resolution inputs degraded by complex, unknown factors (noise, blur, compression). While diffusion models have achieved impressive perceptual quality, they suffer from high computational costs due to iterative sampling (often requiring 10–100 steps).

Recent attempts to achieve one-step generation face two main bottlenecks:

1. Distillation-based methods: Leveraging large Text-to-Image (T2I) priors (e.g., Stable Diffusion) achieves one-step inference but requires massive parameter counts (hindering lightweight deployment) and limits the student model's capability to that of the teacher.
2. Consistency Models (CMs): These offer efficient inference without distillation but suffer from two critical limitations:
   • Consistency Drift: The transitive nature of Consistency Training (CT) leads to error accumulation, causing the generated trajectory to drift away from the true image manifold.
   • Geometric Decoupling: While pixel-wise alignment is achieved, the generated images often lack structural coherence (e.g., distorted geometry, incoherent textures) because standard perceptual metrics lack sensitivity to precise local geometric variations.

2. Methodology: GTASR

The authors propose GTASR (Geometric Trajectory Alignment Super-Resolution), a lightweight, one-step framework that addresses the above limitations through a two-stage training paradigm.

Core Components

A. Trajectory Alignment (TA)

• Goal: Mitigate consistency drift inherent in standard CT.
• Mechanism: Instead of relying solely on adjacent step consistency (which accumulates errors), GTASR introduces a Full-Path Projection strategy.
• Process:
  1. The model predicts a clean image $\hat{x}_0$ from a noisy state $x_t$.
  2. This prediction is re-projected back onto the noisy manifold at various time steps $t$ using the forward diffusion operator $Q(\hat{x}_0, y_0, t)$.
  3. A loss ($L_{TA}$) is computed between these re-projected states and the ground-truth noisy states.
• Effect: This forces the tangent vector field of the Probability Flow ODE (PF-ODE) to align with the ground truth across the entire trajectory, correcting the evolution direction and preventing detail loss at high noise levels.

B. Dual-Reference Structural Rectification (DRSR)

• Goal: Resolve "Geometric Decoupling" where pixel alignment fails to preserve structural integrity.
• Mechanism: The authors derive a theoretical upper bound for structural error, identifying two necessary conditions: minimizing the Consistency Gap (local trajectory stability) and minimizing Target Bias (global structural accuracy).
• Process:
  1. Stability Loss ($L_{Stab}$): Uses a frozen reference model to measure the structural discrepancy (via Sobel operator) between the "fake" trajectory (generated from prediction) and the "real" trajectory. This ensures local geometric consistency.
  2. Rectification Loss ($L_{Rect}$): Directly aligns the structural derivatives (Sobel maps) of the generated image with the Ground Truth ($x_0$).
• Effect: These dual constraints enforce strict structural fidelity, ensuring that high-frequency details (like edges and textures) are recovered without geometric distortion.

Training Strategy

• Stage I: Trains the online model from scratch using Consistency Training ($L_{CT}$) augmented by Trajectory Alignment ($L_{TA}$) to establish a stable, drift-free trajectory.
• Stage II: Fine-tunes the model using Distribution Trajectory Matching ($L_{DTM}$) combined with the Dual-Reference Structural Rectification losses ($L_{Stab}$ and $L_{Rect}$) to refine perceptual quality and structural details.

3. Key Contributions

1. Novel Paradigm: Introduction of GTASR, a one-step Real-ISR method that avoids heavy T2I distillation while outperforming existing consistency-based approaches.
2. Trajectory Alignment (TA): A strategy that rectifies the tangent vector field via full-path projection, effectively solving the error accumulation problem in Consistency Training.
3. Dual-Reference Structural Rectification (DRSR): A mechanism that bridges the gap between pixel-wise alignment and geometric coherence by enforcing structural constraints via Sobel-based stability and rectification losses.
4. Efficiency: Achieves state-of-the-art perceptual quality with minimal latency (single-step inference) and a compact model size (~172M parameters), making it suitable for real-time deployment.

4. Experimental Results

The method was evaluated on synthetic (ImageNet-Test) and real-world datasets (RealSR, RealLQ250, RealSet65).

• Perceptual Quality: GTASR significantly outperforms baselines (including CTMSR, SinSR, ResShift, and StableSR) on no-reference metrics.
  • On ImageNet-Test, GTASR achieves a MANIQA score of 0.5826 (vs. 0.4857 for CTMSR) and TOPIQ of 0.7423.
  • On RealLQ250, it achieves CLIPIQA 0.7355 and TOPIQ 0.7047, surpassing the next best methods by a large margin.
• Structural Integrity: Visual comparisons show GTASR recovers intricate details (e.g., fur, architectural grids, eye details) with superior clarity and geometric correctness compared to the blurring or artifacts seen in other methods.
• Efficiency:
  • Inference Time: 0.08 seconds per image (RTX 4090), which is ~140x faster than StableSR-200 (11.21s) and faster than other one-step methods like SinSR (0.12s).
  • Parameters: 172M parameters, significantly smaller than T2I-distilled models (often >1B parameters).
• Ablation Studies: Confirm that removing $L_{TA}$ causes severe performance drops due to drift, and removing $L_{Stab}$/$L_{Rect}$ leads to structural inconsistencies.

5. Significance

GTASR represents a significant advancement in the field of Real-World Super-Resolution by breaking the traditional trade-off between inference speed, model size, and perceptual quality.

• Lightweight Deployment: By eliminating the need for massive T2I teacher models, GTASR makes high-quality, one-step super-resolution feasible for edge devices and real-time applications.
• Theoretical Insight: The paper provides a rigorous theoretical analysis of "Geometric Decoupling" and "Consistency Drift," offering a new perspective on how to stabilize consistency training through trajectory projection and structural constraints.
• Practical Impact: The ability to generate realistic, structurally coherent images in a single step with minimal computational overhead opens new possibilities for real-time video enhancement, medical imaging, and mobile photography.

The code and models are released at: https://github.com/Blazedengcy/GTASR.