A unified post-training and real-time inference framework for accelerated video generation, including sparse attention, distillation recipes, and deployable demos.

Native multimodal video generation for complex motion, prompt following, and controllable generation across text, image, video, and audio conditions.

A line of work on sparse, tiled, and quantization-aware attention mechanisms for faster and cheaper video diffusion transformers, with wide-scale adoption across industry, publicly and privately.

A real-time video generation interface built on FastVideo for steering and revising generated videos through natural-language vibe directing.

A sparse distillation recipe for FastVideo that releases fast video generation models, training recipes, and datasets for reproducible acceleration work.

Led the project to pretrain a 1.1B parameter Llama model on 3T tokens with a compact, widely reused training codebase.

Co-led the initial development and release of LMMs-Eval, a one-for-all evaluation package for large multimodal models with broad task coverage and reproducible evaluation workflows.

Training recipes for million-token context extension and long-context transfer from language models to vision-language models.