Selection, Aggregation, and Enhancement: Trajectory Consistent Diffusion Model for Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Diffusion models show potential for image super-resolution (ISR).
• Existing methods often underutilize pretrained backbones and lack sampling trajectory constraints, leading to reduced structural consistency and detail.
• This limits the fidelity and detail of reconstructed images.

Purpose of the Study:

• To introduce a novel Trajectory Consistent Diffusion Model (TCDM) for image super-resolution.
• To enhance structural consistency and fine details in super-resolved images.
• To optimize the sampling process using lightweight components and inference-time strategies while keeping the diffusion backbone frozen.

Main Methods:

• Dynamic Semantic Selection (DSS): Records early intermediates, matches them to upsampled low-resolution features, and reconditions sampling to minimize conditioning-noise scale mismatch.
• Cross-Step Aggregation Guidance (CAG): Aggregates current state features with selected intermediates to ensure trajectory-level consistency in noise prediction.
• Frequency Enhancement Adapter (FE-Adapter): A plug-and-play module that injects frequency-domain cues during training to improve high-frequency perception and preserve global structures.

Main Results:

• TCDM achieves high-fidelity, detail-rich image reconstructions.
• Demonstrates strong structural fidelity and competitive no-reference perceptual quality across multiple ISR benchmarks.
• Offers a favorable trade-off between image fidelity and perceptual quality.

Conclusions:

• TCDM effectively addresses limitations of current diffusion-based ISR methods by optimizing the sampling trajectory.
• The proposed DSS, CAG, and FE-Adapter mechanisms contribute to improved structural consistency and detail preservation.
• TCDM presents a promising approach for high-quality image super-resolution.