Dynamic Resolution Diffusion Models for Efficient Image Restoration
A new research paper proposes dynamic resolution diffusion models (DMs) to accelerate image restoration tasks. Existing DMs operating in high-dimensional pixel space suffer from high computational overhead, while latent-space methods require repeated encoder-decoder inference, often resulting in slower runtime. The proposed approach projects data into lower-dimensional subspaces using dynamic resolution, fine-tuning pre-trained DMs for dynamic resolution priors. It adapts DPS and DAPS, two pixel-space methods for general image restoration, into this framework to improve computational efficiency. The paper is available on arXiv under identifier 2605.14267.
Key facts
- The paper proposes dynamic resolution diffusion models for image restoration.
- Existing pixel-space DMs have high computational overhead.
- Latent-space methods require repeated encoder-decoder inference.
- The proposed method projects data into lower-dimensional subspaces.
- It fine-tunes pre-trained DMs for dynamic resolution priors.
- It adapts DPS and DAPS methods into the framework.
- The paper is on arXiv with ID 2605.14267.
- The goal is to accelerate inference in image restoration.
Entities
Institutions
- arXiv