MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos
Organizations: Peking University · Tencent Hunyuan · The Hong Kong University of Science and Technology
Abstract
Streaming video generation must sustain quality beyond the training horizon while handling self-generated history and growing cache costs. To address these challenges, we present MovieRoll, a base-to-distilled framework that connects streaming fine-tuning, long-rollout distillation, and long-horizon inference. We first establish a robust multi-step base through Window-wise Teacher Forcing (WinTF), which jointly denoises chunks at progressive noise levels with bidirectional attention within each window and causal access to preceding history. To retain historical context efficiently, hierarchical KV-cache compression (HieraKV) exploits layer-wise attention patterns, retaining full history in shallow layers and sink--recent history in deeper layers. This reduces cache memory by up to 78% while improving generation quality. Building on this base, cumulative long tuning with distribution matching distillation (CuLT-DMD) holds parameters fixed throughout each long self-rollout and averages local gradients before updating. Each update therefore combines supervision across temporal positions while retaining only one local computation graph, enabling efficient long-rollout training of a few-step generator. To extend generation beyond the training horizon, KV recomputation bounds context for both models, while guidance annealing (AnnealCFG) reduces guidance toward the clean endpoint to stabilize MovieRoll-Base. Complementing these modeling advances, FBCA attention and PS-VAE decoding support efficient training and streaming output. On HeliosBench, MovieRoll-Base remains competitive at 30s after training on videos of up to 15s, and MovieRoll-Distilled achieves the highest total scores among the compared methods across single- and multi-event evaluations up to 60s, combining sustained naturalness with active motion.
Figures & tables
| Windows | Implementation | Single Shape | Multiple Shapes | ||||
| Forward | Fwd.+Bwd. | Mem Peak | Preparation | Total Time | Mem Peak | ||
| (ms) | (ms) | (GiB) | (s) | (s) | (GiB) | ||
| 3 | SDPA (cuDNN) | 1678.64 | 6802.64 | 64.19 | 41.76 | 249.83 | 62.14 |
| FlexAttention | 699.81 | 2789.95 | 10.37 | 57.48 | 141.73 | 28.61 | |
| FBCA | 541.44 | 1878.47 | 13.35 | 56.82 | 13.19 | ||
| 5 | SDPA (cuDNN) | OOM | OOM | ||||
| Method | GPUs | Time | Throughput | Speedup | Memory | Reconstruction | Agreement |
| (s) | (FPS) | (GiB) | PSNR (dB) | PSNR (dB) | |||
| VAE (Full Sequence) | 1 | OOM | – | – | – | – | – |
| Temporal Tiling (32 f) | 1 | 219.61 | 6.56 | 37.11 | 38.43 | 42.82 | |
| Spatial Tiling (64 px) | 1 | 273.90 | 5.26 | 20.67 | 38.69 | 43.20 | |
| Temporal + Spatial Tiling (64 f, 256 px) | 1 | 355.11 | 4.06 | 16.15 | 39.91 | 49.15 | |
| PS-VAE | 1 | 176.15 | 8.18 | 28.84 | 40.46 | Ref. |
| Method | Aesthetic | Dynamic | Smoothness | Semantic | Naturalness | Drifting Aesthetic | Drifting Smoothness | Drifting Semantic | Drifting Naturalness | Total Score |
| 241-Frame Evaluation | ||||||||||
| CausVid | 8 | 4 | 9 | 4 | 1 | 7 | 10 | 8 | 2 | 4.578 |
| Self Forcing | 8 | 4 | 10 | 5 | 4 | 7 | 10 | 9 | 1 | 5.628 |
| Causal Forcing | 8 | 10 | 9 | 5 | 4 | 6 | 10 | 8 | 1 | 5.580 |
| Rolling Forcing | 8 | 2 | 10 | 5 | 7 | 8 | 10 | 10 | 2 | 6.630 |
| MemFlow | 8 | 4 | 10 | 5 | 8 | 8 | 10 | 10 | 4 | 7.143 |
| Method | Aesthetic | Dynamic | Smoothness | Semantic | Naturalness | Drifting Aesthetic | Drifting Smoothness | Drifting Semantic | Drifting Naturalness | Total Score |
| 721-Frame (3-Event) Evaluation | ||||||||||
| MemFlow | 8 | 3 | 10 | 3 | 7 | 6 | 10 | 9 | 1 | 5.754 |
| LongLive | 9 | 1 | 10 | 3 | 5 | 6 | 10 | 9 | 1 | 5.214 |
| Echo-Infinity | 8 | 3 | 9 | 3 | 7 | 5 | 10 | 9 | 1 | 5.625 |
| Helios-Base | 8 | 1 | 10 | 2 | 10 | 7 | 10 | 9 | 2 | 6.402 |
| Helios-Distilled | 8 | 3 | 10 | 2 | 7 | 5 | 10 | 9 | 1 | 5.400 |
| Model Variants | Aesthetic | Dynamic | Smoothness | Semantic | Naturalness | Drifting Aesthetic | Drifting Smoothness | Drifting Semantic | Drifting Naturalness | Total Score |
| Autoregressive Formulation (241-Frame Evaluation) | ||||||||||
| Chunk-wise Teacher Forcing | 7 | 4 | 10 | 5 | 9 | 6 | 10 | 9 | 1 | 6.774 |
| Chunk-wise Teacher Forcing with Self-resampling | 7 | 6 | 10 | 5 | 10 | 5 | 10 | 8 | 1 | 6.891 |
| Window-wise Teacher Forcing | 8 | 5 | 10 | 5 | 10 | 7 | 10 | 9 | 2 | 7.287 |
| Autoregressive Formulation (721-Frame Evaluation) | ||||||||||
| Chunk-wise Teacher Forcing | 7 | 3 | 9 | 4 | 6 | 2 | 10 | 6 | 1 | 5.001 |
| Method | Params | Resolution ( ) | Inference Time (s) | FPS | ||
| E2E | Denoising | E2E | Denoising | |||
| CausVid | 1.3B | 232 | 131 | 6.21 | 11.00 | |
| Self Forcing | 1.3B | 455 | 361 | 3.17 | 3.99 | |
| Causal Forcing | 1.3B | 429 | 336 | 3.36 | 4.29 | |
| Rolling Forcing | 1.3B | 500 | 407 | 2.88 | 3.54 | |
| MemFlow | 1.3B | 338 | 253 | 4.26 | 5.70 | |
| GPUs | FlashAttention-3 | SageAttention | ||||
| Denoising | Total | Throughput | Denoising | Total | Throughput | |
| (s) | (s) | (FPS) | (s) | (s) | (FPS) | |
| 1 | 763 | 948 | 1.52 | 521 | 707 | 2.04 |
| 2 | 423 | 538 | 2.68 | 293 | 407 | 3.54 |
| 4 | 229 | 304 | 4.75 | 158 | 236 | 6.11 |
| 8 | 131 | 185 | 7.77 | 95 | 152 | 9.47 |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Backbone | MMDiT, 8.3B parameters, 54 dual-stream blocks |
| Hidden Size / Heads | 2048 / 16 |
| Patch Size | |
| VAE | causal 3D VAE, spatial and temporal compression, 32 channels |
| Text Encoders | Qwen2.5-VL-7B (up to 1000 tokens), Glyph-ByT5 (up to 256 tokens) |
| Frame Rate | 24 fps |
| Chunk Size | 4 latent frames (16 frames), except 5 latent frames in the first chunk |
| Base Model | Distilled Model | |
| Resolution / Frame Rate | / 24 fps | |
| Chunk Size / Window Size | 4 latent frames / 5 chunks | |
| Sampling Steps | 50 (10 per round) | 5 (1 per round) |
| Timesteps | 50-step schedule, shift 5 | |
| CFG | Per-Frame, Linear | None |
| Denoising Passes per Chunk | 20 | 1 |
| Base Model | Distillation | |||
| Stage 1 | Stage 2 | ODE Init. | CuLT-DMD | |
| Initialization | Bidirectional T2V | Stage 1 (EMA) | Stage 2 (EMA) | ODE Student |
| Training Data | 620K single-event clips | 738K single-event clips + 560K multi-event clips | 1.1K teacher trajectories | 5K single-event prompts (P1) 5K two-event prompts (P2) |
| Frames per Sample | 33–241 | 33–361 | 241 | 241 (P1) / 721 (P2) |
| Duration | 1.4–10s | 1.4–15s | 10s | 10s / 30s |
| Resolution | multi-aspect, pixels | , | , | |
| Phase 1 | Phase 2 | |
| Frames | 241 (61) | 721 (181) |
| Chunks / Supervision Windows | 15 / 3 | 45 / 9 |
| Events per Video | 1 | 2 (switch at frame 129–593) |
| Global Steps / Critic Updates | 600 | 600 |
| Generator Updates | 200 | 200 |
| Generator Schedule | , shift 5, 1 step per round | |
| Metric | Model | Measurement | Dir./Weight |
| Aesthetic | CLIP ViT-L/14 + LAION Head | All-frame aesthetic score divided by and averaged. | / 0.030 |
| Dynamic | Farnebäck Optical Flow | Absolute value after averaging downsampled flow components over pixels and frame pairs. | / 0.030 |
| Motion Smoothness | AMT-S | Interpolation error at omitted frames, normalized as . | / 0.030 |
| Semantic Alignment | ViCLIP (InternVid-10M-FLT) | Cosine similarity between the prompt and eight middle-sampled frames. | / 0.255 |
| Naturalness | Gemini-2.5-Pro | VLM score from 16 uniformly sampled frames, normalized as and averaged over three calls. | / 0.255 |
| Drifting Aesthetic | Same as Aesthetic | Absolute difference between the first and last window scores. | / 0.099 |
| Metric | Rating 10 | Rating 9 | Rating 8 | Rating 7 | Rating 6 | Rating 5 | Rating 4 | Rating 3 | Rating 2 | Rating 1 |
| Aesthetic ( ) | ||||||||||
| Dynamic ( ) | ||||||||||
| Motion Smoothness ( ) | ||||||||||
| Semantic Alignment ( ) | ||||||||||
| Naturalness ( ) | ||||||||||
| Drifting Aesthetic ( ) |
| GPT-5.2 | Gemini-2.5-Pro | |||
| Method | Score | Rank | Score | Rank |
| CausVid | 0.1446 | 11 | 0.0208 | 12 |
| Self Forcing | 0.1596 | 10 | 0.0392 | 11 |
| Causal Forcing | 0.2508 | 9 | 0.1421 | 9 |
| Rolling Forcing | 0.3217 | 8 | 0.3625 | 8 |
| MemFlow | 0.3325 | 7 | 0.4612 | 4 |
| Single-Event | Multi-Event | ||||
| Baseline | 10s | 30s | 60s | 30s | 60s |
| CausVid | 74 | 74 | 78 | – | – |
| Self Forcing | 79 | 81 | 89 | – | – |
| Causal Forcing | 76 | 72 | 69 | – | – |
| Rolling Forcing | 73 | 78 | 73 | – | – |
| MemFlow | 76 | 72 | 82 | 31 | 38 |
Long-Horizon Video Generation
Long-horizon video generation creates coherent, controllable videos over extended durations despite limited context and accumulating errors.
See all 186 papers