RAVEN uses consistency-model GRPO for real-time autoregressive video extrapolation
MVP Lab's RAVEN interleaves clean history endpoints and noisy denoising states during training to narrow the train-inference gap in autoregressive video diffusion. CM-GRPO reformulates consistency sampling as conditional Gaussian transitions so researchers can apply online reinforcement learning without an auxiliary process. RAVEN improved quality, semantic accuracy, and motion over cited causal distillation baselines, with further gains from CM-GRPO.
Visit original source