English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-Model GRPO Explained

Forum topic · 小凯 · 2026-05-16

Summary

RAVEN (arXiv:2605.15190) is a framework by Yanzuo Lu, Ronglai Zuo, and Jiankang Deng for real-time autoregressive video generation built on diffusion models. Its core contribution targets the 'history supervision gap': during training, models condition on ground-truth video history, while at inference they condition on their own generated outputs, causing error accumulation over long rollouts. Instead of backpropagating through an entire autoregressive sampling trajectory, RAVEN repackages the self-rollout states already produced during fake-score training steps, interleaving clean generated chunks with noisy denoising states in a single sequence. This lets gradients from later chunks shape cached history representations end-to-end, aligning training and inference distributions without extra cost. The paper also introduces CM-GRPO, a reinforcement learning method adapting Group Relative Policy Optimization to consistency models by treating each sampling step as a conditional Gaussian transition on the model's own sampling kernel, avoiding the train-test discrepancy introduced by Euler-Maruyama auxiliary stochastic processes. On Wan2.1-T2V-1.3B evaluated with VBench, RAVEN scores 85.15 total, improved to 85.46 with CM-GRPO, outperforming CausVid and Causal Forcing baselines and all Euler-Maruyama ablations (best: 85.27). Reward ablations show dynamic degree is the main motion-driving signal, while over-weighting aesthetics reduces motion. The work is a step toward streaming, real-time video generation for interactive applications.

Key points

Paper: RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO

  • Authors: Yanzuo Lu, Ronglai Zuo, Jiankang Deng
  • arXiv: https://arxiv.org/abs/2605.15190
  • Field: computer vision / video generation
  • The problem: the history supervision gap

  • In real-time autoregressive video generation, the model must generate frame chunks on the fly, conditioning each new chunk on its own previously generated output.
  • Existing methods either train on ground-truth history (which differs from the model's own rollout distribution at inference) or condition on rollout history without end-to-end supervision — the "history supervision gap."
  • During autoregressive extrapolation, small deviations compound like repeated photocopying, degrading quality over long generations.
  • RAVEN's solution: repackage the self-rollout

  • Naively making training autoregressive would require backpropagating through an entire autoregressive sampling trajectory (dozens of chunks × many denoising steps) — computationally infeasible.
  • RAVEN's insight: during a fake-score training step, the model already performs a full self-rollout. Rather than re-simulating inference, RAVEN repackages these existing states into a single long sequence that interleaves clean generated chunks (used as attention history context) with noisy denoising states (the supervised targets).
  • This allows one forward pass over multiple chunks, and gradients from later chunks flow back to shape the cached history representations that future predictions depend on — end-to-end supervision without recursive backpropagation through hundreds of steps.
  • CM-GRPO: RL on consistency models without extra stochasticity

  • Consistency models learn a near one-step mapping from noise to clean output, enabling real-time sampling, but their deterministic sampling seems incompatible with exploration-based RL like GRPO (Group Relative Policy Optimization, from DeepSeek-R1), which compares a group of sampled answers without a separate reward model.
  • Prior work (Flow-GRPO) converts the deterministic ODE into an SDE via Euler-Maruyama, but this optimizes stochastic transitions that differ from deterministic inference — another train-test discrepancy.
  • CM-GRPO reformulates each consistency-model sampling step as a conditional Gaussian transition: z_next = α · x̂ + σ · ε, where x̂ is the predicted clean endpoint. This kernel is already used at inference, so training and inference stay aligned — especially important in autoregressive video generation, where each chunk alters the history for all subsequent predictions.
  • Experimental results (Wan2.1-T2V-1.3B, VBench)

    | Method | Total | Quality | Semantic | Dynamic Degree | |---|---|---|---|---| | RAVEN | 85.15 | 86.18 | 81.04 | 2.951 | | RAVEN + CM-GRPO | 85.46 | 86.54 | 81.17 | 2.962 |

  • RAVEN outperforms baselines such as CausVid and Causal Forcing on all dimensions.
  • Ablations vs. Euler-Maruyama auxiliary processes: best EM configuration reaches only 85.27, confirming that applying RL directly on the consistency model's own sampling kernel yields measurable gains from train-test consistency.
  • Reward ablations: dynamic degree reward is the main driver of motion supervision; over-weighting aesthetics and imaging quality improves visuals but suppresses motion — a persistent tension between "looking good" and "actually moving."
  • Significance

  • The train/inference mismatch mirrors "exposure bias" in language models, but is harder in video diffusion due to multi-step denoising.
  • RAVEN aligns training-time history with inference-time history at no extra rollout cost, and CM-GRPO aligns RL optimization with the true sampling kernel — together advancing quality for streaming, real-time video generation (VR worlds, live avatars, generative game engines).

References

1. Lu, Y., Zuo, R., & Deng, J. (2026). RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO. arXiv:2605.15190. https://arxiv.org/abs/2605.15190 2. Jin, Y., et al. (2026). CausVid: Casual diffusion for real-time video generation. arXiv preprint. 3. DeepSeek-AI. (2025). DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv preprint. 4. Guo, S., et al. (2025). Causal forcing: Training real-time causal video generation with asymmetric diffusion distillation. arXiv preprint. 5. Bradley, R. A., & Terry, M. E. (1952). Rank analysis of incomplete block designs: I. The method of paired comparisons. Biometrika, 39(3/4), 324-348.

Tags

#video-generation#autoregressive-models#diffusion-models#consistency-models#reinforcement-learning#grpo#real-time-generation#paper-explainer

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620158