Major Projects
Concrete research with real results. Each project links to our published work and open-source implementations.
3D Consistency
Depth map integration and point cloud conditioning for spatially coherent world generation. See our work on depth-aware diffusion decoders.
Long-term Spatiotemporal
Solving temporal flickering through frame-pair training and optical flow conditioning. Achieving smooth, consistent video generation.
Embedded Devices
60+ FPS on RTX 4090, 30 FPS on mobile. Our models run anywhere from Steam Deck to iPhone, no cloud required.
Model Distillation
ODE regression enables 3.5x speedup by pruning 50% of layers with minimal quality loss. Production-ready optimization.
Novel Architectures
Diffusion transformers for decoding, hierarchical VAE-GANs, and self-forcing for KV cache optimization.
Human Evaluation
OWL Eval platform reveals what metrics miss. Open-source tools for understanding human perception of AI videos.