Image generation
Open diffusion, flow-matching, and transformer/DiT models—focused on correctness, throughput, memory efficiency, compilation, and accelerator utilization.
Open-source generative media research
Compute for the next generation of visual AI.
OpenRoot Compute explores, benchmarks, and optimizes open-source image and video generation models across modern AI accelerators.
Explore our focus →Video generation
Open video models with an emphasis on long-sequence attention, memory pressure, distributed execution, and efficient inference and training.
Hardware-aware systems
TPU, Intel Gaudi, and GPU research spanning compilers, kernels, precision, graph execution, sharding, communication, and multi-accelerator scaling.
Research direction
Make leading open generative-media models run efficiently across more compute platforms.
Our direction is reproducible benchmarks, practical optimization notes, and upstreamable open-source improvements wherever possible.