Methods — glossary

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/.

Type tags: warp+inpaint feed-forward autoregressive per-scene-3DGS per-scene-4DGS world-model-memory training-free
MethodWhat it isPaper / arXivBackboneType
Ours — warp + inpaint (Wan2.1-VACE)
Wan2.1-VACE-14BOURS
vace14b
Warp the depth point-cloud to target poses, then VACE inpaints the holes.arXiv:2503.07598VACE, 2025Wan2.1-14B DiT (VACE)warp+inpaint
VACE-14B ft (long-horizon)OURS
vace14b_ft_lh
VACE-14B with a long-horizon LoRA fine-tune for drift/loop trajectories.arXiv:2503.07598ours, FT of VACEWan2.1-VACE-14B + LoRAwarp+inpaint
VACE-14B ft (dynamic)OURS
vace14b_ft_dyn
VACE-14B fine-tuned on dynamic-scene warps for moving content.arXiv:2503.07598ours, FT of VACEWan2.1-VACE-14B + LoRAwarp+inpaint
VACE-14B ft (mixed)OURS
vace14b_ft_mixed
VACE-14B fine-tuned on a mixed static + dynamic + long-horizon corpus.arXiv:2503.07598ours, FT of VACEWan2.1-VACE-14B + LoRAwarp+inpaint
I2V baseOURS
i2v_base
Wan2.1 image-to-video baseline conditioned on the first warped frame.arXiv:2503.20314Wan, 2025Wan2.1-I2V-14Bwarp+inpaint
I2V ftOURS
i2v_ft
Wan2.1-I2V fine-tuned on our warp+inpaint task as an ablation backbone.arXiv:2503.20314ours, FT of Wan-I2VWan2.1-I2V-14B + LoRAwarp+inpaint
Diffusion baselines
GEN3C
gen3c
3D-cache-conditioned video diffusion; renders a point-cloud cache to guide frames.arXiv:2503.03751CVPR 2025 HighlightCosmos / SVD video DiTwarp+inpaint
Voyager
voyager
Long-range world-consistent RGBD video diffusion from one image + camera path.arXiv:2506.04225HunyuanWorld-Voyager, TencentHunyuanVideo DiTwarp+inpaint
FlexWorld
flexworld
V2V diffusion + progressive 3DGS expansion for flexible 360° view synthesis.arXiv:2503.13265FlexWorld, 2025CogVideoX-5B-I2V → 3DGSper-scene-3DGS
ViewCrafter
viewcrafter
Point-cloud-conditioned video diffusion for high-fidelity novel view synthesis.arXiv:2409.02048TPAMI 2025DynamiCrafter video diff.warp+inpaint
FrameCrafter
framecrafter
Our LoRA reproduction / twin of VACE-14B trained on DL3DV.ours, reproWan2.1-VACE-14B + LoRAwarp+inpaint
NVS-Solver
nvs_solver
Zero-shot NVS: modulates SVD sampling with warped input views, no training.arXiv:2405.15364ICLR 2025Stable Video Diffusiontraining-free
ViewExtrapolator
viewextrapolator
Refines artifact-prone rendered views via a redesigned SVD denoising; no fine-tune.arXiv:2411.142082024Stable Video Diffusiontraining-free
TrajectoryCrafter
trajcrafter
Redirects camera trajectory of a monocular video via dual-stream diffusion.arXiv:2503.05638ICCV 2025 OralCogVideoX-I2Vwarp+inpaint
Per-scene reconstruction
Shape-of-Motion
shape_of_motion
Per-scene 4D-GS: fits SE(3) motion bases to one casual monocular video.arXiv:2407.137642024Dynamic 3D Gaussiansper-scene-4DGS
Long-horizon SOTA (world models)
Lyra-2
lyra2
Feed-forward explorable 3D worlds; lifts diffusion video to 3DGS via self-distillation.arXiv:2604.13036Lyra 2.0, NVIDIAWan2.1-14B → 3DGSfeed-forward
Matrix3D
matrix3d
All-in-one photogrammetry: pose, depth and NVS in one multi-modal diffusion model.arXiv:2502.076852025Multi-modal DiTfeed-forward
SpatialMem
spmem
Video world model with geometry-grounded long-term spatial memory to fight forgetting.arXiv:2506.05284NeurIPS 2025Video diffusion + mem.world-model-memory
VMem
vmem
Surfel-indexed view memory retrieves relevant past views for consistent generation.arXiv:2506.18903ICCV 2025 HighlightStable Video Diffusion + mem.world-model-memory
MatrixGame-3
matrixgame3
Real-time streaming interactive world model with long-horizon memory.arXiv:2604.08995Matrix-Game 3.0WanX-derived DiT (DMD)autoregressive