Ancient murals, vital cultural assets, are frequently marred by weathering and human impact. While deep‑learning restoration shows promise, judging its quality is difficult when no intact reference image exists. We present a dual evaluation framework that unites (1) traditional reference‑based metrics and (2) a CLIP‑driven no‑reference metric. A paired dataset was built by applying 400 synthetic damage masks to high‑resolution murals, enabling PSNR and SSIM assessment. For real‑world cases lacking references, CLIP similarity between restorations and prompts such as “undamaged mural” supplies an objective score. LoRA‑fine‑tuned Stable Diffusion excels on both tracks, yielding higher PSNR/SSIM and closer CLIP alignment with “undamaged” descriptions. This combined scheme offers a rigorous, widely applicable benchmark for mural restoration research.