Stratified Minute-Scale Rollout Distillation and Self-Resampled Base Stabilization for 60s+ Generation
Extend self-rollout distillation to minute-scale multi-event horizons via stratified late-window supervision and scheduled-sampling base tuning so both base and fast generators remain stable.
Time
Compute level
Resources
What the paper used
Prerequisites
Scope
The Gap
In Section 5.2.1, Section 6.2, and Table 10/11, MovieRoll-Base is trained only up to 15s (361 frames) using WinTF with uniform random history noise perturbation U(0, 0.15), and CuLT-DMD Phase 2 trains only up to 30s (721 frames, 2 events) by accumulating local gradients across all K=9 supervision windows.
“Long-horizon stability also remains a limitation. MovieRoll-Base is insufficiently stable at 60s, so only MovieRoll-Distilled is evaluated at this duration.”
Limitation
MovieRoll-Base collapses or becomes insufficiently stable to evaluate at 60s (1441 frames), and even MovieRoll-Distilled never trains on 60s rollouts or 3+ event transitions, leaving a train-test horizon gap beyond 30s.
Proposed Solution
Stabilize MovieRoll-Base by adding a Stage 3 scheduled-sampling self-rollout phase where the model generates 30s-60s trajectories using AnnealCFG and receives WinTF flow-matching supervision on late-horizon windows conditioned on self-generated history (instead of purely Gaussian-perturbed ground truth); simultaneously, extend CuLT-DMD to 60s (1441 frames, 3-6 events) using stratified window sampling (backpropagating on a random subset of K early, mid, and late supervision windows per rollout while detaching intermediate steps) to keep memory and compute bounded.
Why This Matters
Closing the training-to-inference horizon gap at 60s should make the higher-capacity multi-step base model usable for minute-scale generation and further curb late-horizon drift in the distilled generator.
How to Get Started
How to test it: Test whether self-rollout Stage 3 stabilization enables MovieRoll-Base evaluation at 60s and whether 60s stratified CuLT-DMD improves MovieRoll-Distilled's 1441-frame single- and 6-event HeliosBench TWR and Drifting Naturalness over the 721-frame-trained baseline; stop if 60s base Drifting Naturalness stays at 1 or 60s CuLT-DMD overfits/collapses dynamics.
- Evaluate the released MovieRoll-Base checkpoint at 60s (1441 frames) on 30 HeliosBench prompts to diagnose exact failure modes (e.g., saturation, blur, or repetition).
- Synthesize 10K 6-event 60s prompt sequences using the Streaming Video Prompt Rewriting Skill in Batch/Sequential mode.
- Implement stratified supervision window sampling in the CuLT-DMD training loop so 1441-frame rollouts backpropagate through 9 sampled early/mid/late windows with fixed per-step memory.
- Run a 300-step Stage 3 self-rollout stabilization run on MovieRoll-Base and a 600-step Phase 3 CuLT-DMD run on MovieRoll-Distilled, evaluating at 1441 frames on HeliosBench.
Suggested by AI. Adapt it to your setup.
Has This Been Done?
Works like Self Forcing and End-to-End Training for Autoregressive Video Diffusion via Self-resampling address train-test gaps with self-rollout and resampling, but the extension of CuLT-DMD and scheduled-sampling distillation to 60-second generation using stratified window sampling is open.
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion · 2025 · 759 citations
- End-to-End Training for Autoregressive Video Diffusion via Self-resampling · 2026 · 56 citations
- Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation · 2026 · 7 citations
We searched OpenAlex and arXiv for “minute-scale video diffusion self-rollout distillation exposure bias” and “Stratified Minute-Scale Rollout Distillation and Self-Resampled Base Stabilization for 60s+ Generation” and an AI compared the idea with the top 10 results. A small search can miss relevant work.
Source Paper
MovieRoll: A Systematic Framework for Rolling Out Long-Form VideosYunyang Ge, Weijie Kong, Ryan Xu +9 more · 2026
MovieRoll introduces a base-to-distilled framework for streaming long-form video generation that sustains visual quality, active motion, and prompt continuity up to 60 seconds while bounding historical KV-cache memory. By combining parallel rolling-window fine-tuning, hierarchical layer-wise cache compression, cumulative long-rollout distribution matching distillation, and streaming decoding infrastructure, it outperforms existing streaming baselines on single-event and multi-event HeliosBench evaluations.
Related Ideas
- Closed-Loop Synchronization for Long-Horizon ConsistencyRealWonder: Real-Time Physical Action-Conditioned Video Generation🟠Medium
- Hybrid Sliding-Window and Fixed-State Depth Memory for Deeply Looped TransformersLoopback Transformer: Towards a Unified View of Residual Looped Transformers🔴Hard
- Accelerating 3D-Guided RL Fine-Tuning with Trajectory DistillationGeometry-Guided Reinforcement Learning for Multi-view Consistent 3D Scene Editing🔴Hard
- Infinite-Context Linear State Generalization for Feedforward ReconstructionLoGeR: Long-Context Geometric Reconstruction with Hybrid Memory🟠Medium
- Consistency Distillation for Real-Time InferenceFlashPortrait: 6x Faster Infinite Portrait Animation with Adaptive Latent Prediction🔴Hard