What it is
A structure-conditioned video renderer that turns many kinds of scene proxies into realistic video in 4 steps using a frozen video backbone plus two LoRAs. Instead of training a renderer per engine or condition type, it feeds arbitrary structural videos through the backbone's native reference path with frame-aligned positions and plain-text descriptions, then post-trains a 4-step acceleration LoRA with structure-guided distribution matching.
Key results
- 4-step rendering is 12.3x faster than 50-step sampling
- SOTA on both in-distribution and 16-type generalization benchmarks
- Generalizes to 11 unseen structure types, including skeletons and point clouds
A single reference-pathway control interface can generalize across heterogeneous and even unseen structural video representations.