Research map / 3D vision / 3D generation

NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis

NeurIPS 2026 · 3D vision › 3D generation · approach: Diffusion-based novel view synthesis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NAMVIS, a diffusion-free next-scale autoregressive transformer (initialized from Infinity 2B pruned to 1B) that synthesizes sparse-view target images with multi-scale projective pose encoding and dual global/dense source conditioning.

Problem

Sparse-view zero-shot novel view synthesis with diffusion models is slow because of iterative denoising, especially when generating multiple target views.

Approach

Reformulates multi-view synthesis as geometry-conditioned next-scale autoregression over discrete visual tokens, injecting multi-scale ProPE camera encodings into target self-attention and source-to-target cross-attention, plus a global [SOS] token and dense cross-attention to frozen VQ-VAE source features.

Key results

• On Objaverse, NAMVIS 1B reaches 22.485 PSNR, 0.861 SSIM, 0.091 LPIPS, beating EscherNet (19.078 PSNR, 0.147 LPIPS) and other diffusion baselines.
• On zero-shot GSO NAMVIS reaches 21.715 PSNR vs 18.878 for EscherNet; on OO3D 21.098 PSNR vs 17.767.
• NAMVIS takes 0.6 s per target view in the 1-to-1 setting, about 3.3x faster than Wonder3D (2.0 s) at 256x256.

Surprising findings

• LoRA tuning with rank 256 performs substantially worse than full fine-tuning (18.912 vs 21.360 PSNR in ablation).
• Removing ProPE caching raises 3-to-3 runtime from 0.96 s to 1.28 s, more than removing KV caching.

Limitations

• Trained and evaluated only at 256x256 resolution; 512x512 requires further training.
• Trained on object-centric Objaverse-XL, so may not transfer to complex real-world scenes.
• Full block-causal attention over previous target scales is memory-heavy at high resolution.

Experimental setup

Datasets: Objaverse-XL Objaverse GSO OmniObject3D

Benchmarks: Objaverse GSO OmniObject3D

Metrics: PSNR SSIM LPIPS COLMAP reconstructable points COLMAP registration coverage inference time

Baselines: Zero-1-to-3 Zero-1-to-3 XL Wonder3D SyncDreamer EscherNet

Builds on: Infinity VAR ProPE Zero-1-to-3 LoRA

Closest papers by meaning

• UMAMI: Unifying Masked Autoregressive Models and Deterministic Rendering for View Synthesis (NeurIPS 2025 · similarity 0.92)
• MVGenMaster: Scaling Multi-View Generation from Any Image via 3D Priors Enhanced Diffusion Model (CVPR 2025 · similarity 0.91)
• MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024 · similarity 0.90)
• NVComposer: Boosting Generative Novel View Synthesis with Multiple Sparse and Unposed Images (CVPR 2025 · similarity 0.90)
• Novel View Synthesis from A Few Glimpses via Test-Time Natural Video Completion (NeurIPS 2025 · similarity 0.89)
• Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.89)