ViewFusion addresses the challenge of maintaining multi-view consistency in novel-view synthesis using diffusion models. By integrating an auto-regressive mechanism, it leverages previously generated views to guide the generation process. The approach extends single-view conditioned models to work in multi-view settings without additional fine-tuning. Experimental results demonstrate the effectiveness of ViewFusion in generating consistent and detailed novel views. The method offers several advantages, including improved image quality by incorporating more information from all available views, flexibility in setting adaptive weights for conditional images based on their relative view distance, and training-free integration into pre-trained diffusion models.
Vers une autre langue
à partir du contenu source
arxiv.org
Questions plus approfondies