mlsanityData to 7 Oct 2026

Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression

arXiv 2512.05081 ↗First posted 4 Dec 2025cs.CVcited by 72

On the map: Diffusion models → Video generation

also in 1 more
Abstract (arXiv)

Recent advances in autoregressive video diffusion have enabled real-time frame streaming, yet existing solutions still suffer from temporal repetition, drift, and motion deceleration. We find that naively applying StreamingLLM-style attention sinks to video diffusion leads to fidelity degradation and motion stagnation. To overcome this, we introduce Deep Forcing, which consists of two training-free mechanisms that address this without any fine-tuning. Specifically, 1) Deep Sink dedicates half of the sliding window to persistent sink tokens and re-aligns their temporal RoPE phase to the current timeline, stabilizing global context during long rollouts. 2) Participative Compression performs importance-aware KV cache pruning that preserves only tokens actively participating in recent attention while safely discarding redundant and degraded history, minimizing error accumulation under out-of-distribution length generation. Together, these components enable over 12x extrapolation (e.g. 5s-trained to 60s+ generation) with better imaging quality than LongLive, better aesthetic quality than RollingForcing, almost maintaining overall consistency, and substantial gains in dynamic degree, all while maintaining real-time generation. Our results demonstrate that training-free KV-cache management can match or exceed training-based approaches for autoregressively streaming long-video generation.

Key facts. Cited by 72 papers in the map (data through 7 Oct 2026). Placed in Video generation, one of 14 directions in Diffusion models. Builds on 31 of its 34 references that are in the map. First posted on arXiv 4 Dec 2025.

Builds on

31 of 34 references.

Closest ancestors · 11

Same direction, newest first

  1. 2025Self-Forcing++: Towards Minute-Scale High-Quality Video Generation2 Oct
  2. 2025Rolling Forcing: Autoregressive Long Video Diffusion in Real Time29 Sep
  3. 2025SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer29 Sep
  4. 2025Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion9 Jun
  5. 2025Sparse VideoGen2: Accelerate Video Generation with Sparse Attention via Semantic-Aware Permutation24 May
  6. 2025Error Analyses of Auto-Regressive Video Diffusion Models: A Unified Framework12 Mar
  7. 2024LTX-Video: Realtime Video Latent Diffusion30 Dec
  8. 2024From Slow Bidirectional to Fast Autoregressive Video Diffusion Models10 Dec
  9. 2024Pyramidal Flow Matching for Efficient Video Generative Modeling8 Oct
  10. 2024CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer12 Aug
  11. 2023Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets25 Nov

Other references · 20

  1. 2025Qwen3 Technical Reportcited by 8,195
  2. 2021RoFormer: Enhanced Transformer with Rotary Position Embeddingcited by 4,581
  3. 2025Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilitiescited by 3,655
  4. 2019The Curious Case of Neural Text Degenerationcited by 2,881
  5. 2025Wan: Open and Advanced Large-Scale Video Generative Modelscited by 2,536
  6. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,115
  7. 2023VBench: Comprehensive Benchmark Suite for Video Generative Modelscited by 1,755
  8. 2024SnapKV: LLM Knows What You are Looking for Before Generationcited by 848
  9. 2023H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Modelscited by 805
  10. 2024Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusioncited by 634
  11. 2024Movie Gen: A Cast of Media Foundation Modelscited by 614
  12. 2023InternVid: A Large-scale Video-Text Dataset for Multimodal Understanding and Generationcited by 488
8 more
  1. 2025MAGI-1: Autoregressive Video Generation at Scalecited by 262
  2. 2025LongLive: Real-time Interactive Long Video Generationcited by 237
  3. 2025History-Guided Video Diffusioncited by 165
  4. 2025VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Modelscited by 112
  5. 2025MotionStream: Real-Time Video Generation with Interactive Motion Controlscited by 71
  6. 2024D2O: Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Modelscited by 30
  7. 2025Dialogue Without Limits: Constant-Sized KV Caches for Extended Responses in LLMscited by 20
  8. 2025LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Modelscited by 9

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 1; 2026: 71.

72 papers.

Sort:

By area:

  1. 2026Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generationcited by 159
  2. 2026Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusioncited by 28
  3. 2026LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generationcited by 25
  4. 2026Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Modelscited by 23
  5. 2026ShotStream: Streaming Multi-Shot Video Generation for Interactive Storytellingcited by 22
  6. 2026LiveWorld: Simulating Out-of-Sight Dynamics in Generative Video World Modelscited by 20
  7. 2026MemRoPE: Training-Free Infinite Video Generation via Evolving Memory Tokenscited by 20
  8. 2026Relax Forcing: Relaxed KV-Memory for Consistent Long Video Generationcited by 20
  9. 2026Anchor Forcing: Anchor Memory and Tri-Region RoPE for Interactive Streaming Video Diffusioncited by 14
  10. 2026PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inferencecited by 14
  11. 2026WorldKV: Efficient World Memory with World Retrieval and Compressioncited by 14
  12. 2026Past- and Future-Informed KV Cache Policy with Salience Estimation in Autoregressive Video Diffusioncited by 13
38 more
  1. 2026HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoisingcited by 13
  2. 2026ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPUcited by 13
  3. 2026Pathwise Test-Time Correction for Autoregressive Long Video Generationcited by 12
  4. 2026Mode Seeking meets Mean Seeking for Fast Long Video Generationcited by 12
  5. 2026LongLive-RAG: A General Retrieval-Augmented Framework for Long Video Generationcited by 12
  6. 2026OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generatorscited by 12
  7. 2026Addressable Memory for Video World Modelscited by 11
  8. 2026Grounding World Simulation Models in a Real-World Metropoliscited by 10
  9. 2026Forcing-KV: Hybrid KV Cache Compression for Efficient Autoregressive Video Diffusion Modelscited by 10
  10. 2025LongVie 2: Multimodal Controllable Ultra-Long Video World Modelcited by 9
  11. 2026TetherCache: Stabilizing Autoregressive Long-Form Video Generation with Gated Recall and Trusted Alignmentcited by 7
  12. 2026Efficient Video Diffusion Models: Advancements and Challengescited by 6
  13. 2026A2RD: Agentic Autoregressive Diffusion for Long Video Consistencycited by 6
  14. 2026RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPOcited by 6
  15. 2026Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memorycited by 6
  16. 2026Echo-Infinity: Learning Evolving Memory for Real-Time Infinite Video Generationcited by 6
  17. 2026Video Generation Models as World Models: Efficient Paradigms, Architectures and Algorithmscited by 5
  18. 2026Head Forcing: Long Autoregressive Video Generation via Head Heterogeneitycited by 5
  19. 2026DySink: Dynamic Frame Sinks for Autoregressive Long Video Generationcited by 5
  20. 2026FadeMem: Distance-Aware Memory Consolidation for Autoregressive Video Diffusioncited by 5
  21. 2026Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillationcited by 5
  22. 2026Ring Forcing: Towards Precise Long-Term Memory for Autoregressive Video Diffusioncited by 5
  23. 2026Pyramid Forcing: Head-Aware Pyramid KV Cache Policy for High-Quality Long Video Generationcited by 4
  24. 2026UniTemp: Unlocking Video Generation in Any Temporal Order via Bidirectional Distillationcited by 4
  25. 2026TempAct: Advancing Temporal Plausibility in Autoregressive Video Generation via Planner-Executor RLcited by 4
  26. 2026Vidu S1: A Real-Time Interactive Video Generation Modelcited by 4
  27. 2026EMFormer: Efficient Multi-Scale Transformer for Accumulative Context Weather Forecastingcited by 3
  28. 2026VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusioncited by 3
  29. 2026HeadCast: Casting Attention Heads for Efficient Autoregressive Video Generationcited by 3
  30. 2026FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoringcited by 3
  31. 2026DCARL: A Divide-and-Conquer Framework for Autoregressive Long-Trajectory Video Generationcited by 2
  32. 2026Registers Matter for Pixel-Space Diffusion Transformerscited by 2
  33. 2026AdaState: Self-Evolving Anchors for Streaming Video Generationcited by 2
  34. 2026DecMem: Towards Minute-Long Consistent World Generation with Decoupled Memorycited by 2
  35. 2026Lip Forcing: Few-Step Autoregressive Diffusion for Real-time Lip Synchronizationcited by 2
  36. 2026Surprise Forcing: What to Remember, When to Skip in Long Video Generationcited by 2
  37. 2026StreamHOI: Interaction-aware Temporal Memory Adaptation for Streaming HOI Video Generationcited by 2
  38. 2026Ms. Forcing: Efficient Streaming Video Generation with Multi-Scale Patchification and Attentioncited by 2

50 most cited of 72 All 72 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2026Memory Forcing: Attendable Mid-Horizon History for Streaming Video Generation
  2. 2026In-Distribution Forcing for Long Video Generation at Test Time
  3. 2026Sparse Forcing: Native Trainable Sparse Attention for Real-time Autoregressive Diffusion Video Generation
  4. 2026Flex-Forcing: Towards a Unified Autoregressive and Bidirectional Video Diffusion Model
  5. 2026Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion
  6. 2026MV-Forcing: Long Multi-View Video Generation via 4D-Grounded Spatio-Temporal Self-Forcing
  7. 2026Echo-Forcing: A Scene Memory Framework for Interactive Long Video Generation
  8. 2026Self Gradient Forcing: Native Long Video Extrapolation