Research map / 3D vision / 3D generation

NVComposer: Boosting Generative Novel View Synthesis with Multiple Sparse and Unposed Images

CVPR 2025 · 3D vision › 3D generation · approach: Diffusion-based novel view synthesis · 4 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 NVComposer, a pose-free multi-view generative novel view synthesis model that jointly generates target views and camera poses of unposed condition images via a dual-stream video diffusion model, plus a geometry-aware feature alignment adapter distilled from DUSt3R.

Problem

Generative novel view synthesis from multiple sparse, unposed images currently depends on external multi-view alignment (explicit pose estimation or pre-reconstruction), which fails when views have small overlap or occlusion.

Approach

An image-pose dual-stream diffusion model built on DynamiCrafter that denoises both target-view images and Plücker-ray pose embeddings of condition views (with a separate pose decoding head), combined with a training-only adapter that aligns internal features to DUSt3R pointmaps.

Key results

• On RealEstate10K with 4 input views, NVComposer reaches 22.55 PSNR on easy subset versus 18.48 for ViewCrafter's best (4-view) in the same table.
• On DL3DV, NVComposer PSNR rises monotonically from 15.31 (1 view) to 17.92 (4 views), while ViewCrafter stays around 15.0 to 15.7.
• Implicit pose estimation on hard RealEstate10K pairs gives rotation error 5.86 vs 58.40 for DUSt3R.

Measured effects reported in the paper

Surprising findings

• ViewCrafter's performance drops when going from one to two input views on the hard RealEstate10K subset, attributed to unstable external alignment.
• Implicit pose estimation outperforms DUSt3R on hard sparse-view pairs, though it is comparable on easy pairs.

Limitations

• Relies on DUSt3R features during training, requiring a pretrained dense stereo model
• Evaluation on Objaverse and RealEstate10K subsets is small (e.g., 20 and 100 samples), limiting statistical strength

Experimental setup

Datasets: Objaverse RealEstate10K CO3D DL3DV

Benchmarks: RealEstate10K DL3DV Objaverse

Metrics: PSNR SSIM LPIPS DISTS FID FVD KVD rotation angle error translation error

Baselines: MotionCtrl CameraCtrl DUSt3R ViewCrafter SV3D

Builds on: DynamiCrafter DUSt3R CameraCtrl ViewCrafter

Closest papers by meaning

• NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis (NeurIPS 2026 · similarity 0.90)
• Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.90)
• MVGenMaster: Scaling Multi-View Generation from Any Image via 3D Priors Enhanced Diffusion Model (CVPR 2025 · similarity 0.89)
• UMAMI: Unifying Masked Autoregressive Models and Deterministic Rendering for View Synthesis (NeurIPS 2025 · similarity 0.88)
• MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views (NeurIPS 2024 · similarity 0.88)
• LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias (ICLR 2025 · similarity 0.88)