ScaleResfusion: Residual Rectified Flow for Image Restoration
ScaleResfusion is an advanced diffusion framework designed for restoring real-world images, utilizing pre-trained text-to-image rectified-flow models. Its main breakthrough, Residual Rectified Flow, incorporates a residual term R into the Standard Rectified Flow. Rather than initiating from pure Gaussian noise, it employs a residual transport path that originates from noisy, low-quality images and includes a precise acceleration point. This innovation tackles two significant issues: methods that start with Gaussian noise tend to be slow and less accurate regarding the degraded input, while residual-based approaches require training from the ground up and fail to utilize pre-trained generative priors. By harnessing modern pre-trained generative priors, ScaleResfusion enhances both speed and fidelity.
Key facts
- ScaleResfusion is a scalable diffusion framework for real-world image restoration.
- It is built on pre-trained text-to-image rectified-flow models.
- The core method is Residual Rectified Flow, which introduces a residual term R into Standard Rectified Flow.
- It starts from noisy low-quality images instead of pure Gaussian noise.
- It uses a residual transport path with an exact acceleration point.
- It addresses slowness and lack of faithfulness in Gaussian noise methods.
- It avoids training from scratch, enabling use of pre-trained generative priors.
- The paper is on arXiv with ID 2607.25275.
Entities
Institutions
- arXiv