Streaming T2V

Key Insight

Most video models generate a whole clip at once, which means you wait for the entire thing before seeing a single frame — fine for a 5-second clip, impossible for an endless or interactive one. Streaming video generation instead emits frames chunk by chunk as it goes, conditioning each new chunk on the ones already produced. This project implements that loop and reuses a KV cache across chunks — storing the attention keys and values from earlier frames so each new chunk does not recompute the whole history — then measures the latency-versus-quality trade-off. The same idea underlies real-time systems like CausVid and Self-Forcing, which distill a slow full-clip diffusion model into a fast autoregressive one that produces frames in time order.