MonarchRT vlatest
Paid
Video Generation
Efficient Attention
MonarchRT is a research method from Carnegie Mellon University that enables true real-time video generation by parameterizing attention maps in video diffusion transformers as sparse Monarch matrices. It achieves up to 95% effective attention sparsity with no quality loss, unlocking 16 FPS real-time video generation on a single RTX 5090.
Monarch matrix attention factorization — sparsely parameterizes 3D video attention maps using structured Monarch matrices that respect spatiotemporal block alignment, achieving 95% effective sparsity with no quality loss
1.4–11.8× speedup over FlashAttention (-2, -3, -4) kernels via custom Triton kernel implementation — first method to achieve true real-time 16 FPS video generation with Self-Forcing on a single RTX 5090
Tiled Monarch parameterization solves monotonic compute-accuracy trade-off and 1-iteration + finetuning approach reduces iterative refinement overhead for real-time few-step diffusion models
PricingFree (Open Source, Apache-2.0)
Versionlatest