NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis

TL;DR

NAMVIS introduces a diffusion-free method for synthesizing multiple novel views from sparse input images by using a coarse-to-fine autoregressive process over discrete visual tokens. It leverages multi-scale projective pose encoding to integrate camera geometry into attention mechanisms, enabling parallel token generation across views and scales. The approach achieves higher image quality and faster inference than diffusion-based methods.

Why It Matters

The method significantly reduces inference time for multi-view synthesis, making 3D content creation more efficient and scalable. By eliminating iterative denoising, it enables real-time applications in virtual environments, robotics, and 3D modeling without sacrificing visual fidelity.

Key Findings

• NAMVIS replaces iterative diffusion with a parallelizable, next-scale autoregressive token prediction process.
• Multi-scale projective pose encoding effectively integrates camera pose information into both self- and cross-attention layers at all resolutions.
• The model achieves superior PSNR, SSIM, and LPIPS scores while running over 3 times faster than diffusion baselines on standard benchmarks.