MAPGR: Multi-Agent Prompt-Guided Residual Diffusion for ancient mural restoration

Digital inpainting provides a non-destructive approach to ancient mural conservation, but reconstructed content must remain grounded in verifiable historical evidence under the minimum intervention principle. Existing diffusion-based restoration models often rely on unconstrained generative priors, which can introduce implausible content in severely degraded regions. We propose Multi-Agent Prompt Guided Restoration (MAPGR), an evidence-driven framework that reformulates mural restoration as a multi-stage reasoning process. MAPGR constrains generation with diagnostic evidence and incorporates a dual-constraint diffusion architecture with a Semantic Modulation Block (SMB) and Mask-Guided Self-Attention (MG-SA) for stylistic and spatial control. Experiments on the DUNHUANG and DhMurals datasets demonstrate that MAPGR achieves strong restoration performance while improving interpretability and historical plausibility, offering an evidence-driven computational framework for ancient mural conservation.