NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis

How it works

NAMVIS tokenizes the source views and generates target-view tokens coarse-to-fine, with camera geometry injected into attention at every scale.

NAMVIS architecture overview — NAMVIS predicts target views autoregressively, refining view-conditioned tokens from coarse to fine.

Multi-scale ProPE

Projective pose encoding injects the source and target camera transformations into target-view self-attention and source-to-target cross-attention at every generation scale.

Dual-path conditioning

A pooled source representation provides global conditioning, while dense, geometry-aware cross-attention preserves appearance details from the input views.

Comparison with prior methods

All methods are evaluated on the same source observations and target camera for direct visual comparison.

Input viewZero-1-to-3Zero-1-to-3 XLSyncDreamerEscherNetNAMVISGT

Zero-1-to-3-XL BUILD_A_ZOO generated view

Zero-1-to-3-XL backpack_022 generated view

Zero-1-to-3 3D_Dollhouse_Swing generated view

Zero-1-to-3-XL 3D_Dollhouse_Swing generated view

SyncDreamer 3D_Dollhouse_Swing generated view

EscherNet 3D_Dollhouse_Swing generated view

NAMVIS 3D_Dollhouse_Swing generated view

Quantitative comparison

Better perceptual quality and faster inference

Method	PSNR ↑	LPIPS ↓	Time ↓
Zero-1-to-3	16.935	0.209	2.6s
Zero-1-to-3 XL	17.199	0.194	2.6s
SyncDreamer	17.698	0.203	16.6s
EscherNet	18.574	0.157	2.1s
NAMVIS 1B	21.766	0.102	0.6s

PSNR and LPIPS are averaged across Objaverse, GSO, and OmniObject3D (OO3D). Time is seconds per target view in the 1-to-1 setting.

Full quantitative results

Method	Objaverse			GSO			OO3D			Time
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	Time
Zero-1-to-3	19.097	0.816	0.152	15.614	0.749	0.257	16.094	0.768	0.219	2.6s
Zero-1-to-3 XL	19.341	0.827	0.136	15.936	0.747	0.236	16.320	0.767	0.209	2.6s
SyncDreamer	19.368	0.826	0.156	16.961	0.781	0.246	16.766	0.793	0.207	16.6s
EscherNet	19.078	0.816	0.147	18.878	0.799	0.156	17.767	0.780	0.167	2.1s
NAMVIS 1B	22.485	0.861	0.091	21.715	0.843	0.111	21.098	0.845	0.104	0.6s

Results are averaged over nine source-target configurations: 1-to-1, 1-to-2, 1-to-3, 2-to-1, 2-to-2, 2-to-3, 3-to-1, 3-to-2, and 3-to-3 at 256 × 256 resolution.

Generated views as turntable videos

Objaverse sample 01

Objaverse sample 02

Backpack 003

Backpack 016

Chair 002

JBL speaker

Boat shoe

Patch cord

Toasty orca

Toy animal