Research map / Video / Video generation
RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RAVEN, a training-time-test distillation framework that repacks self rollouts into interleaved clean history and noisy denoising states, and CM-GRPO, an RL method applying GRPO directly to the consistency sampler's Gaussian transition kernel, improving causal video diffusion on VBench.
Problem
Causal autoregressive video diffusion distillation suffers a mismatch between training-time history distributions and inference-time history, and history representations receive no end-to-end supervision; RL for few-step consistency generators also relies on auxiliary Euler-Maruyama stochastic proce
Approach
Repack each self rollout into an interleaved sequence of clean chunk endpoints and noisy denoising states so later chunk losses supervise history representations, and define GRPO directly on the consistency sampling step as a conditional Gaussian transition whose mean is the predicted clean endpoint.
Key results
- RAVEN reaches 85.15 VBench total score vs 84.96 for Causal Forcing, with dynamic degree 2.951 vs 2.669.
- RAVEN + CM-GRPO achieves the best total score of 85.46 and dynamic degree 2.962 among compared methods.
- In the policy-interface ablation, CM-GRPO (total 85.46) outperforms all Euler-Maruyama variants (best 85.27) on top of RAVEN.
Surprising findings
- Teacher Forcing gives the highest dynamic degree (3.000) but the lowest total score (82.64) in the training-time test ablation.
- Replacing real prefixes in Diffusion Forcing with self-rollout endpoints raises dynamic degree but lowers total score (83.30), indicating unsupervised history alignment redistributes error.
Limitations
- Reference-policy KL regularization is derived but not applied, since the bidirectional teacher cannot be sampled through the consistency interface.
- Evaluation is limited to 5-second short videos with a 1.3B base model and VBench-based metrics.
- Dynamic degree in VBench is sensitive to camera jitter, motivating an extra VLM-based evaluation.
Experimental setup
Datasets: VidProM OpenVidHD-0.4M
Benchmarks: VBench
Metrics: VBench total score quality score semantic score dynamic degree
Baselines: CausVid LongLive Rolling Forcing Self Forcing Reward Forcing Causal Forcing Teacher Forcing Diffusion Forcing
Builds on: Self Forcing CausVid Flow-GRPO GRPO Causal Forcing
Closest papers by meaning
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.91)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.90)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.88)
- Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation (NeurIPS 2025 · similarity 0.87)
- Generative Pre-trained Autoregressive Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.