All 20 methods in this benchmark, grouped by family. Ours rows (green accent) are warp+inpaint with Wan2.1-VACE-14B and its fine-tuned / I2V variants. arXiv ids are clickable; “—” = no verified paper (e.g. internal repro). Run ids in monospace match runs/<id>_out/.
| Method | What it is | Paper / arXiv | Backbone | Type |
|---|---|---|---|---|
| Ours — warp + inpaint (Wan2.1-VACE) | ||||
Wan2.1-VACE-14BOURSvace14b | Warp the depth point-cloud to target poses, then VACE inpaints the holes. | arXiv:2503.07598VACE, 2025 | Wan2.1-14B DiT (VACE) | warp+inpaint |
VACE-14B ft (long-horizon)OURSvace14b_ft_lh | VACE-14B with a long-horizon LoRA fine-tune for drift/loop trajectories. | arXiv:2503.07598ours, FT of VACE | Wan2.1-VACE-14B + LoRA | warp+inpaint |
VACE-14B ft (dynamic)OURSvace14b_ft_dyn | VACE-14B fine-tuned on dynamic-scene warps for moving content. | arXiv:2503.07598ours, FT of VACE | Wan2.1-VACE-14B + LoRA | warp+inpaint |
VACE-14B ft (mixed)OURSvace14b_ft_mixed | VACE-14B fine-tuned on a mixed static + dynamic + long-horizon corpus. | arXiv:2503.07598ours, FT of VACE | Wan2.1-VACE-14B + LoRA | warp+inpaint |
I2V baseOURSi2v_base | Wan2.1 image-to-video baseline conditioned on the first warped frame. | arXiv:2503.20314Wan, 2025 | Wan2.1-I2V-14B | warp+inpaint |
I2V ftOURSi2v_ft | Wan2.1-I2V fine-tuned on our warp+inpaint task as an ablation backbone. | arXiv:2503.20314ours, FT of Wan-I2V | Wan2.1-I2V-14B + LoRA | warp+inpaint |
| Diffusion baselines | ||||
GEN3Cgen3c | 3D-cache-conditioned video diffusion; renders a point-cloud cache to guide frames. | arXiv:2503.03751CVPR 2025 Highlight | Cosmos / SVD video DiT | warp+inpaint |
Voyagervoyager | Long-range world-consistent RGBD video diffusion from one image + camera path. | arXiv:2506.04225HunyuanWorld-Voyager, Tencent | HunyuanVideo DiT | warp+inpaint |
FlexWorldflexworld | V2V diffusion + progressive 3DGS expansion for flexible 360° view synthesis. | arXiv:2503.13265FlexWorld, 2025 | CogVideoX-5B-I2V → 3DGS | per-scene-3DGS |
ViewCrafterviewcrafter | Point-cloud-conditioned video diffusion for high-fidelity novel view synthesis. | arXiv:2409.02048TPAMI 2025 | DynamiCrafter video diff. | warp+inpaint |
FrameCrafterframecrafter | Our LoRA reproduction / twin of VACE-14B trained on DL3DV. | —ours, repro | Wan2.1-VACE-14B + LoRA | warp+inpaint |
NVS-Solvernvs_solver | Zero-shot NVS: modulates SVD sampling with warped input views, no training. | arXiv:2405.15364ICLR 2025 | Stable Video Diffusion | training-free |
ViewExtrapolatorviewextrapolator | Refines artifact-prone rendered views via a redesigned SVD denoising; no fine-tune. | arXiv:2411.142082024 | Stable Video Diffusion | training-free |
TrajectoryCraftertrajcrafter | Redirects camera trajectory of a monocular video via dual-stream diffusion. | arXiv:2503.05638ICCV 2025 Oral | CogVideoX-I2V | warp+inpaint |
| Per-scene reconstruction | ||||
Shape-of-Motionshape_of_motion | Per-scene 4D-GS: fits SE(3) motion bases to one casual monocular video. | arXiv:2407.137642024 | Dynamic 3D Gaussians | per-scene-4DGS |
| Long-horizon SOTA (world models) | ||||
Lyra-2lyra2 | Feed-forward explorable 3D worlds; lifts diffusion video to 3DGS via self-distillation. | arXiv:2604.13036Lyra 2.0, NVIDIA | Wan2.1-14B → 3DGS | feed-forward |
Matrix3Dmatrix3d | All-in-one photogrammetry: pose, depth and NVS in one multi-modal diffusion model. | arXiv:2502.076852025 | Multi-modal DiT | feed-forward |
SpatialMemspmem | Video world model with geometry-grounded long-term spatial memory to fight forgetting. | arXiv:2506.05284NeurIPS 2025 | Video diffusion + mem. | world-model-memory |
VMemvmem | Surfel-indexed view memory retrieves relevant past views for consistent generation. | arXiv:2506.18903ICCV 2025 Highlight | Stable Video Diffusion + mem. | world-model-memory |
MatrixGame-3matrixgame3 | Real-time streaming interactive world model with long-horizon memory. | arXiv:2604.08995Matrix-Game 3.0 | WanX-derived DiT (DMD) | autoregressive |