A novel method has been introduced to improve the perceptual quality of burst super-resolution (SR) images by integrating burst SR with diffusion models. Unlike conventional approaches that commence diffusion processes from random noise, this method initiates from an intermediate step, focusing on reconstructing detailed textures and boundaries effectively.

This technique leverages hierarchical burst SR features to condition the reverse diffusion process, ensuring optimization for the burst SR task. The integration of feature extraction and alignment modules from the deterministic burst SR method, Burstormer, further enhances the efficacy of the proposed method. These modules facilitate the extraction of features from burst frames and their alignment using mechanisms such as Burst Feature Attention (BFA) and Reference-Based Feature Enrichment (RBFE).

Moreover, the fusion module incorporates Style Feature Transform (SFT) for conditioning with low-resolution (LR) features, thereby improving the reconstruction process. Training involves the utilization of linear or sigmoid noise schedulers for the diffusion model, focusing on reconstructing detailed appearances near the initial steps for enhanced SR performance.