Personalized Restoration via Dual-Pivot Tuning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Generative diffusion models act as priors for natural image manifold adherence in image restoration.
• Personalized priors are crucial for accurately reconstructing individual-specific facial features in facial image restoration.

Purpose of the Study:

• To propose Dual-Pivot Tuning, a novel two-stage approach for personalizing blind facial image restoration systems.
• To maintain general prior integrity while achieving high fidelity to individual identity and degraded image attributes.

Main Methods:

• A two-stage tuning process: first, tuning the diffusion model around a fixed textual pivot.
• Second, tuning a guiding network generically, using the personalized diffusion model as a fixed pivot.
• Extensive experiments on images of widely recognized individuals.

Main Results:

• Dual-Pivot Tuning effectively personalizes blind restoration systems.
• The approach yields results with natural appearance and high fidelity to identity and image attributes.
• Personalized priors achieve superior identity fidelity and outperform state-of-the-art generic priors in overall image quality.

Conclusions:

• Dual-Pivot Tuning offers a simple yet effective method for personalized facial image restoration.
• This technique balances personalization with the preservation of general image prior integrity.
• The proposed method advances the state-of-the-art in personalized image restoration.