Animation of Historical Images: A Video Diffusion Generation Method Guided by Depth Estimation

Abstract

The use of artificial intelligence and smart algorithms for the restoration, enhancement and creation of videos from historical images continues to grow. To tackle the problems of spatial structure drift, inter-frame flickering, and motion discontinuity in the animation of historical images, we propose a video diffusion generation model based on depth estimation. Based on the video diffusion model, the model includes the historical image preprocessing module, the single-frame depth estimation module, the spatial structure encoding module, and the temporal motion compensation module, using the depth feature as a conditional constraint in the reverse denoising generation of the model. Experimental results show that the proposed method achieves a PSNR of 28.76, an improvement of 1.28 over conventional video diffusion models; SSIM increases to 0.883, LPIPS decreases to 0.108, and FVD decreases to 219.54, with both generation quality and temporal stability outperforming the comparison methods.