The content introduces SARGD as a training-free method for artifact-free super-resolution. It addresses over-smoothing issues and improves image fidelity through Reality-Guided Refinement (RGR) and Self-Adaptive Guidance (SAG). Extensive experiments demonstrate superior results compared to existing methods, reducing sampling steps by 2×. The study includes detailed methodology, experimental setups, comparisons with state-of-the-art methods, ablation studies on denoising strategies, realistic latent update approaches, impact of artifact detection, and inference steps analysis.
Sang ngôn ngữ khác
từ nội dung nguồn
arxiv.org
Thông tin chi tiết chính được chắt lọc từ
by Qingping Zhe... lúc arxiv.org 03-26-2024
https://arxiv.org/pdf/2403.16643.pdfYêu cầu sâu hơn