Post
3
Excited to share RAVEN, my first PhD project. Paper, code, and models are all released.
RAVEN is for real-time autoregressive video generation. Instead of simply appending future chunks, we train the model to better remember and use its own generated history, leading to more realistic and natural long-horizon videos.
Technically, RAVEN repacks self-rollouts into interleaved clean historical endpoints and noisy denoising states, aligning training-time attention with inference-time extrapolation.
We also introduce CM-GRPO: by reformulating consistency-model sampling as a conditional Gaussian transition kernel, online RL can directly optimize the sampler transition used at inference.
Project Page: https://yanzuo.lu/raven
Paper: https://arxiv.org/abs/2605.15190
Code: https://github.com/mvp-ai-lab/RAVEN
Model: mvp-lab/RAVEN
RAVEN is for real-time autoregressive video generation. Instead of simply appending future chunks, we train the model to better remember and use its own generated history, leading to more realistic and natural long-horizon videos.
Technically, RAVEN repacks self-rollouts into interleaved clean historical endpoints and noisy denoising states, aligning training-time attention with inference-time extrapolation.
We also introduce CM-GRPO: by reformulating consistency-model sampling as a conditional Gaussian transition kernel, online RL can directly optimize the sampler transition used at inference.
Project Page: https://yanzuo.lu/raven
Paper: https://arxiv.org/abs/2605.15190
Code: https://github.com/mvp-ai-lab/RAVEN
Model: mvp-lab/RAVEN
