Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Time-Aware One Step Diffusion Network for Real-World Image Super-Resolution

Created by
  • Haebom

Author

Tainyi Zhang, Zheng-Peng Duan, Peng-Tao Jiang, Bo Li, Ming-Ming Cheng, Chun-Le Guo, Chongyi Li

Outline

This paper proposes a time-aware single-step diffusion network (TADSR) to overcome the limitation of existing single-step real-image super-resolution (Real-ISR) methods, which fail to properly utilize the generative prior knowledge of a stable diffusion model (SD). Existing methods utilize SDs at fixed time steps, failing to fully utilize the different generative prior knowledge of SDs across different noise-injected time steps. TADSR introduces a time-aware VAE encoder to project images into different latent features at different time steps. Through dynamic changes in time steps and latent features, the learning model better aligns with the input pattern distribution of the pre-trained SD. Furthermore, the time-aware VSD loss function bridges the gap between the learning model and the SD time steps, providing consistent generative prior knowledge guidance. Consequently, TADSR achieves state-of-the-art performance and controllable super-resolution in a single step. It also offers the advantage of controlling the trade-off between fidelity and realism by varying the time step conditions.

Takeaways, Limitations

•
Takeaways:
◦
Achieving state-of-the-art real-world image super-resolution performance in a single step.
◦
Provides tunability between fidelity and realism through time step control.
◦
Solving the fixed time step problem of Limitations of existing methods.
◦
Generation of effective pre-trained SDs using a time-aware VAE encoder and a time-aware VSD loss function: leveraging prior knowledge.
•
Limitations:
◦
Lack of analysis on the computational cost and memory efficiency of the proposed method.
◦
Lack of generalization performance evaluation on diverse types of image datasets.
◦
Further research is needed on optimal strategies for time step control.
◦
Further explanation and justification of the design of the temporally aware VAE encoder and temporally aware VSD loss function are needed.
👍
Made with Slashpage