mlsanityData to 7 Oct 2026

Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion

arXiv 2506.08009 ↗First posted 9 Jun 2025cs.CV cs.AI cs.LGcited by 569

On the map: Diffusion models → Video generation

Abstract (arXiv)

We introduce Self Forcing, a novel training paradigm for autoregressive video diffusion models. It addresses the longstanding issue of exposure bias, where models trained on ground-truth context must generate sequences conditioned on their own imperfect outputs during inference. Unlike prior methods that denoise future frames based on ground-truth context frames, Self Forcing conditions each frame's generation on previously self-generated outputs by performing autoregressive rollout with key-value (KV) caching during training. This strategy enables supervision through a holistic loss at the video level that directly evaluates the quality of the entire generated sequence, rather than relying solely on traditional frame-wise objectives. To ensure training efficiency, we employ a few-step diffusion model along with a stochastic gradient truncation strategy, effectively balancing computational cost and performance. We further introduce a rolling KV cache mechanism that enables efficient autoregressive video extrapolation. Extensive experiments demonstrate that our approach achieves real-time streaming video generation with sub-second latency on a single GPU, while matching or even surpassing the generation quality of significantly slower and non-causal diffusion models. Project website: http://self-forcing.github.io/

Key facts. Cited by 569 papers in the map (data through 7 Oct 2026). Placed in Video generation, one of 14 directions in Diffusion models. Builds on 105 of its 109 references that are in the map. First posted on arXiv 9 Jun 2025.

Builds on

105 of 109 references.

Closest ancestors · 34

Same direction, newest first

  1. 2025Test-Time Training Done Right29 May
  2. 2025Long-Context State-Space Video World Models26 May
  3. 2025Generative Pre-trained Autoregressive Diffusion Transformer12 May
  4. 2025SkyReels-V2: Infinite-length Film Generative Model17 Apr
  5. 2025Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models17 Apr
  6. 2025Long Context Tuning for Video Generation13 Mar
  7. 2025Error Analyses of Auto-Regressive Video Diffusion Models: A Unified Framework12 Mar
  8. 2025AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion10 Mar
  9. 2025Diffusion Adversarial Post-Training for One-Step Video Generation14 Jan
  10. 2024LTX-Video: Realtime Video Latent Diffusion30 Dec
  11. 2024Autoregressive Video Generation without Vector Quantization18 Dec
  12. 2024SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device13 Dec
22 more
  1. 2024From Slow Bidirectional to Fast Autoregressive Video Diffusion Models10 Dec
  2. 2024Ca2-VDM: Efficient Autoregressive Video Diffusion Model with Causal Generation and Cache Sharing25 Nov
  3. 2024ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation27 Oct
  4. 2024MarDini: Masked Autoregressive Diffusion for Video Generation at Scale26 Oct
  5. 2024Progressive Autoregressive Video Diffusion Models10 Oct
  6. 2024Pyramidal Flow Matching for Efficient Video Generative Modeling8 Oct
  7. 2024Redefining Temporal Modeling in Video Diffusion: The Vectorized Timestep Approach4 Oct
  8. 2024OSV: One Step is Enough for High-Quality Image to Video Generation17 Sep
  9. 2024Diffusion Models Are Real-Time Game Engines27 Aug
  10. 2024CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer12 Aug
  11. 2024SF-V: Single Forward Video Generation Model6 Jun
  12. 2024Streaming Video Diffusion: Online Video Editing with Diffusion Models30 May
  13. 2024Looking Backward: Streaming Video-to-Video Translation with Feature Banks24 May
  14. 2024FIFO-Diffusion: Generating Infinite Videos from Text without Training19 May
  15. 2024VidProM: A Million-scale Real Prompt-Gallery Dataset for Text-to-Video Diffusion Models10 Mar
  16. 2023Photorealistic Video Generation with Diffusion Models11 Dec
  17. 2023ARTboldsymbol·V: Auto-Regressive Text-to-Video Generation with Diffusion Models30 Nov
  18. 2023Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets25 Nov
  19. 2023Language Model Beats Diffusion -- Tokenizer is Key to Visual Generation9 Oct
  20. 2023Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models18 Apr
  21. 2022Imagen Video: High Definition Video Generation with Diffusion Models5 Oct
  22. 2022Video Diffusion Models7 Apr

Other references · 71

  1. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
  2. 2022Scalable Diffusion Models with Transformerscited by 5,931
  3. 2022Flow Matching for Generative Modelingcited by 5,426
  4. 2024Qwen2.5 Technical Reportcited by 4,713
  5. 2017Neural Discrete Representation Learningcited by 4,650
  6. 2023Mamba: Linear-Time Sequence Modeling with Selective State Spacescited by 4,616
  7. 2022Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flowcited by 3,550
  8. 2025Wan: Open and Advanced Large-Scale Video Generative Modelscited by 2,536
  9. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,115
  10. 2023Consistency Modelscited by 1,778
  11. 2023VBench: Comprehensive Benchmark Suite for Video Generative Modelscited by 1,755
  12. 2024HunyuanVideo: A Systematic Framework For Large Video Generative Modelscited by 1,597
38 more
  1. 2023ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillationcited by 1,207
  2. 2021Variational Diffusion Modelscited by 1,145
  3. 2023One-step Diffusion with Distribution Matching Distillationcited by 978
  4. 2022CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformerscited by 973
  5. 2024Genie: Generative Interactive Environmentscited by 706
  6. 2018Which Training Methods for GANs do actually Converge?cited by 672
  7. 2021Maximum Likelihood Training of Score-Based Diffusion Modelscited by 670
  8. 2021VideoGPT: Video Generation using VQ-VAE and Transformerscited by 651
  9. 2024Improved Distribution Matching Distillation for Fast Image Synthesiscited by 648
  10. 2024Autoregressive Image Generation without Vector Quantizationcited by 644
  11. 2024Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusioncited by 634
  12. 2024Movie Gen: A Cast of Media Foundation Modelscited by 614
  13. 2015Sequence Level Training with Recurrent Neural Networkscited by 568
  14. 2017MoCoGAN: Decomposing Motion and Content for Video Generationcited by 523
  15. 2024FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precisioncited by 507
  16. 2023VideoPoet: A Large Language Model for Zero-Shot Video Generationcited by 479
  17. 2022Phenaki: Variable Length Video Generation From Open Domain Textual Descriptioncited by 475
  18. 2024Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Modelcited by 468
  19. 2023Learning Interactive Real-World Simulatorscited by 440
  20. 2025Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Modelscited by 385
  21. 2016Generating Videos with Scene Dynamicscited by 376
  22. 2016SeqGAN: Sequence Generative Adversarial Nets with Policy Gradientcited by 375
  23. 2024Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillationcited by 276
  24. 2022Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformercited by 244
  25. 2020Infinite Nature: Perpetual View Generation of Natural Scenes from a Single Imagecited by 239
  26. 2018The relativistic discriminator: a key element missing from standard GANcited by 234
  27. 2023Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Modelscited by 232
  28. 2016Professor Forcing: A New Algorithm for Training Recurrent Networkscited by 214
  29. 2016Temporal Generative Adversarial Nets with Singular Value Clippingcited by 198
  30. 2025Unified Video Action Modelcited by 195
  31. 2022The Parallelism Tradeoff: Limitations of Log-Precision Transformerscited by 191
  32. 2019Scaling Autoregressive Video Modelscited by 189
  33. 2024Flex Attention: A Programming Model for Generating Optimized Attention Kernelscited by 180
  34. 2017Decomposing Motion and Content for Natural Video Sequence Predictioncited by 168
  35. 2025History-Guided Video Diffusioncited by 165
  36. 2024Real-Time Video Generation with Pyramid Attention Broadcastcited by 140
  37. 2017Unsupervised Learning of Disentangled Representations from Videocited by 138
  38. 2025GameFactory: Creating New Games with Generative Interactive Videoscited by 124

50 most cited of 71 All 71 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 106; 2026: 463.

569 papers.

Sort:

By area:

  1. 2026World Action Models are Zero-shot Policiescited by 411
  2. 2025LongLive: Real-time Interactive Long Video Generationcited by 237
  3. 2025Rolling Forcing: Autoregressive Long Video Diffusion in Real Timecited by 209
  4. 2025Self-Forcing++: Towards Minute-Scale High-Quality Video Generationcited by 176
  5. 2025Matrix-game 2.0: An open-source real-time and streaming interactive world modelcited by 174
  6. 2026Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generationcited by 159
  7. 2025WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modelingcited by 152
  8. 2026Advancing Open-source World Modelscited by 144
  9. 2026DreamDojo: A Generalist Robot World Model from Large-Scale Human Videoscited by 117
  10. 2025Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcingcited by 111
  11. 2025SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformercited by 91
  12. 2025RELIC: Interactive Video World Model with Long-Horizon Memorycited by 89
38 more
  1. 2025Geometry Forcing: Marrying Video Diffusion and 3D Representation for Consistent World Modelingcited by 86
  2. 2026Matrix-Game 3.0: Real-Time and Streaming Interactive World Model with Long-Horizon Memorycited by 86
  3. 2025Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistencycited by 81
  4. 2025Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rolloutcited by 78
  5. 2025Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillationcited by 75
  6. 2025Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compressioncited by 72
  7. 2025MotionStream: Real-Time Video Generation with Interactive Motion Controlscited by 71
  8. 2025Stable Video Infinity: Infinite-Length Video Generation with Error Recyclingcited by 68
  9. 2025Large Video Planner Enables Generalizable Robot Controlcited by 67
  10. 2025Hunyuan-GameCraft-2: Instruction-following Interactive Game World Modelcited by 58
  11. 2026Helios: Real Real-Time Long Video Generation Modelcited by 56
  12. 2026SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformercited by 50
  13. 2026Context Forcing: Consistent Autoregressive Video Generation with Long Contextcited by 47
  14. 2025Memory Forcing: Spatio-Temporal Memory for Consistent Scene Generation on Minecraftcited by 46
  15. 2026WoVR: World Models as Reliable Simulators for Post-Training VLA Policies with RLcited by 46
  16. 2025End-to-End Training for Autoregressive Video Diffusion via Self-Resamplingcited by 45
  17. 2026Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generationcited by 44
  18. 2025StreamDiT: Real-Time Streaming Text-to-Video Generationcited by 43
  19. 2025FlashVSR: Towards Real-Time Diffusion-Based Streaming Video Super-Resolutioncited by 40
  20. 2025MemFlow: Flowing Adaptive Memory for Consistent and Efficient Long Video Narrativescited by 40
  21. 2025VideoSSM: Autoregressive Long Video Generation with Hybrid State-Space Memorycited by 39
  22. 2026RISE: Self-Improving Robot Policy with Compositional World Modelcited by 39
  23. 2026Solaris: Building a Multiplayer Video World Model in Minecraftcited by 38
  24. 2026Infinite Worlds with Versatile Interactionscited by 37
  25. 2026Lyra 2.0: Explorable Generative 3D Worldscited by 36
  26. 2025YoNoSplat: You Only Need One Model for Feedforward 3D Gaussian Splattingcited by 34
  27. 2026HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worldscited by 34
  28. 2025MultiShotMaster: A Controllable Multi-Shot Video Generation Frameworkcited by 33
  29. 2026DreamX-World 1.0: A General-Purpose Interactive World Modelcited by 33
  30. 2025OmniNWM: Omniscient Driving Navigation World Modelscited by 32
  31. 2026MIND: Benchmarking Memory Consistency and Action Control in World Modelscited by 31
  32. 2025FlashWorld: High-quality 3D Scene Generation within Secondscited by 30
  33. 2026Native Video-Action Pretraining for Generalizable Robot Controlcited by 30
  34. 2026LoL: Longer than Longer, Scaling Video Generation to Hourcited by 29
  35. 2026Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusioncited by 28
  36. 2025StreamDiffusionV2: A Streaming System for Dynamic and Interactive Video Generationcited by 27
  37. 2026Transition Matching Distillation for Fast Video Generationcited by 27
  38. 2025Pack and Force Your Memory: Long-form and Consistent Video Generationcited by 26

50 most cited of 569 All 569 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2026In-Distribution Forcing for Long Video Generation at Test Time
  2. 2025Playing with Transformer at 30+ FPS via Next-Frame Diffusion
  3. 2026Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation
  4. 2026Custom Forcing: Training-Free Subject Customization for Autoregressive Video Generation
  5. 2026Memory Forcing: Attendable Mid-Horizon History for Streaming Video Generation
  6. 2026HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoising
  7. 2026SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching
  8. 2026Future Forcing: Future-aware Training-free KV Cache Policy for Autoregressive Video Generation