Causal Forcing++ distills diffusion models for real-time interactive video
Tsinghua University's Machine Learning Group introduced Causal Consistency Distillation to convert bidirectional diffusion models into frame-level autoregressive models that run in one or two steps. With two steps, it improved VBench Total by 0.1 and VisionReward by 0.335 over the previous best reported results. It halved first-frame latency and cut stage-two training costs to about a quarter. The code supports streaming, controllable rollouts, and action-conditioned world models.
Visit original source