ADORE: Adaptive Diffusion Optimized Restoration for AI-Generated Facial Imagery

Junxue Li, Hong Chen, Guanglei Qi

Abstract: We introduce ADORE (Adaptive Diffusion Optimized Restoration), a pioneering solution that addresses facial distortion issues in diffusion-based, language-guided image generation. ADORE enhances facial quality based on image characteristics and style, improving the visual fidelity of AI-generated images. It also mitigates boundary distortions during the face-background fusion process, offering a novel approach to address instability issues by using generative models for image restoration. Rigorous experiments validate ADORE’s proficiency in achieving high-quality, style-consistent facial restorations. ADORE supports text-driven, fine-tuned facial refinement, leveraging the model’s open-domain synthesis capability. As the first method tailored to enhance facial generation quality in text-to-image models, with its versatility and innovative solutions, ADORE successfully addresses a pressing issue and paves new avenues in image generation.