NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis
NAMVIS generates consistent multi-view images with geometry-conditioned autoregression, beating diffusion baselines and running over 3x faster.
NAMVIS is a diffusion-free model for sparse-view novel view synthesis that generates multi-view images by geometry-conditioned next-scale autoregression. It predicts discrete visual tokens over a few coarse-to-fine scales, sampling tokens in parallel within each scale and across target views. Multi-scale Projective Pose Encoding injects source and target cameras into self- and cross-attention, combined with dense geometry-aware attention. On Objaverse, GSO, and OmniObject3D it improves PSNR, SSIM, and LPIPS over diffusion baselines while running more than three times faster in the reported setting.
- Replaces iterative diffusion with coarse-to-fine next-scale autoregression.
- Tokens within each scale and across views are sampled in parallel.
- Multi-scale Projective Pose Encoding injects camera geometry into attention.
- Beats diffusion baselines on Objaverse, GSO, and OmniObject3D.
- Runs more than three times faster under the same setting.
From the original article
1 min read at huggingface.co
Sparse-view novel view synthesis is a central problem in 3D content creation, but diffusion-based approaches remain limited by iterative denoising, making multi-view generation expensive at inference time. We introduce NAMVIS, a diffusion-free framework that reformulates multi-view image synthesis as geometry-conditioned next-scale autoregression.
By Ramil Khafizov, Ilya Statsenko, Ruslan Rakhimov · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.