F3D NVS Benchmark Evaluation results for 9 Novel View Synthesis (NVS) methods across 3 benchmarks , stored as NPZ files containing predicted and ground truth RGB images with camera parameters. Models Model Type Reference DA3RAE Feature space diffusion (DA3 backbone) Ours DA3RAE DINO Feature space diffusion (DINO backbone) Ours (ablation) DA3RAE VAE Feature space diffusion (VAE backbone) Ours (ablation) VGGT Feature space diffusion (VGGT backbone) Ours (ablation) MVGenMaster Multi view generation GitHub NVComposer Novel view composition GitHub Matrix3D 3D aware generation GitHub CAMEO Attention distillation NVS — CAT3D Naive attention NVS — Benchmarks Benchmark Setting Samples Description mipnerf360 easy 2 view, easy 200 MipNeRF 360 scenes, easy camera intervals re10k medium 2 view, medium 200 RealEstate10K, medium difficulty dl3dv medium 2 view, medium 200 DL3DV dataset, medium difficulty All settings use 2 conditioning views ( cond num=2 ) and 8 total views (2 ref + 6 target). Directory Structure Example: NPZ Format Each .npz file contains one evaluation sample. Load with: Common Fields Field Shape Range Description gt rgb (V, 3, H, W) [0, 1] float32 Ground truth RGB (all views) p…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy