VideoMLA applies multi-head latent attention with 3D-RoPE decoupling to autoregressive video diffusion, delivering 92.7% KV memory reduction while matching short-horizon baselines and leading long-horizon VBench scores.
hub Canonical reference
Self-Forcing++: Towards Minute-Scale High-Quality Video Generation
Canonical reference. 83% of citing Pith papers cite this work as background.
60
Pith papers citing it
Background
83% of classified citations
abstract
Diffusion models have revolutionized image and video generation, achieving unprecedented visual quality. However, their reliance on transformer architectures incurs prohibitively high computational costs, particularly when extending generation to long videos. Recent work has explored autoregressive formulations for long video generation, typically by distilling from short-horizon bidirectional teachers. Nevertheless, given that teacher models cannot synthesize long videos, the extrapolation of student models beyond their training horizon often leads to pronounced quality degradation, arising from the compounding of errors within the continuous latent space. In this paper, we propose a simple yet effective approach to mitigate quality degradation in long-horizon video generation without requiring supervision from long-video teachers or retraining on long video datasets. Our approach centers on exploiting the rich knowledge of teacher models to provide guidance for the student model through sampled segments drawn from self-generated long videos. Our method maintains temporal consistency while scaling video length by up to 20x beyond teacher's capability, avoiding common issues such as over-exposure and error-accumulation without recomputing overlapping frames like previous methods. When scaling up the computation, our method shows the capability of generating videos up to 4 minutes and 15 seconds, equivalent to 99.9% of the maximum span supported by our base model's position embedding and more than 50x longer than that of our baseline model. Experiments on standard benchmarks and our proposed improved benchmark demonstrate that our approach substantially outperforms baseline methods in both fidelity and consistency. Our long-horizon videos demo can be found at https://self-forcing-plus-plus.github.io/
hub tools
citation-role summary
background 11
method 1
citation-polarity summary
representative citing papers
MV-Forcing composes temporal and view-sequential autoregression in a single diffusion model, using a recurrent 3D reconstruction model as a geometric bridge to generate arbitrarily long, multi-view consistent videos.
ISPA reduces KV cache size by up to 50% in AR video models by transitioning layers to local attention and applying instance-specific least-squares weight modulation to compensate for lost history.
MemLearner introduces a learning-based adaptive context query method using query tokens in video world models to improve long-term scene consistency over rule-based retrieval.
TempAct introduces a planner-executor RL framework with hierarchical group exploration and rewards to improve temporal consistency in autoregressive video diffusion models.
FLAT maps compressed video diffusion latents to explicit triangle splats via ray-centered rotation parameterization and a product window function, reporting better geometric accuracy than 3D Gaussian baselines under identical training.
TetherCache organizes KV-cache into sink, memory, and recent regions and applies gated recall with attention-diversity balancing plus trusted memory editing to stabilize long-horizon autoregressive video diffusion.
LongLive-RAG formulates long video generation as retrieval-augmented generation by treating self-generated latents as a dynamic searchable history and adding a Window Temporal Delta Loss for better retrieval.
LongLive-2.0 delivers an NVFP4 parallel infrastructure that enables direct training of long multi-shot autoregressive diffusion video models and achieves up to 2.15x training and 1.84x inference speedups on Blackwell and other GPUs.
Echo-Forcing decouples stable anchors, compressed history, and recent dynamics in video diffusion KV caches using hierarchical memory, scene recall frames, and difference-aware decay to support interactive long video generation under bounded cache.
CausalCine enables real-time causal autoregressive multi-shot video generation via multi-shot training, content-aware memory routing for coherence, and distillation to few-step inference.
FreeSpec uses SVD-based spectral reconstruction to fuse global low-rank and local high-rank features, reducing content drift and preserving temporal dynamics in long video generation.
Stream-R1 improves distillation of autoregressive streaming video diffusion models by adaptively weighting supervision with a reward model at both rollout and per-pixel levels.
A survey that groups efficient video diffusion methods into four paradigms—step distillation, efficient attention, model compression, and cache/trajectory optimization—and outlines open challenges for practical use.
Grounded Forcing introduces dual memory caching, reference-based positional embeddings, and proximity-weighted recaching to bridge stable semantics with local dynamics, improving long-range consistency in autoregressive video synthesis.
DreamDojo is a foundation world model pretrained on the largest human video dataset to date that uses continuous latent actions to transfer interaction knowledge and achieves controllable physics simulation after robot post-training.
Causal Forcing distills few-step autoregressive video generators from an autoregressive diffusion teacher rather than a bidirectional one, avoiding conditional-expectation blur and beating Self Forcing on motion and quality metrics.
Step Forcing lets a 4-step video world model generate 30-second closed-loop rollouts, and a VLM judge scores them to reproduce the real RoboArena policy ranking at r=0.989.
Aligning temporal granularity, action subspaces, and train-test conditioning yields SOTA long-horizon mobile and fine-grained manipulation success for a unified world-action model.
InteractiveAvatar is a real-time infinite-streaming avatar video generation system using autoregressive distillation, Long-Short Visual Memory for consistency, and a Reasoning-Reaction Module for intent-aware interactions.
SurgVista mitigates spatial interaction incoherence and temporal fidelity collapse in surgical world models through trajectory-based contrastive regularization and drift-perturbed training, outperforming prior methods on a new long-horizon benchmark.
UniTemp enables arbitrary temporal order video generation in autoregressive diffusion models via bidirectional distillation and blockwise anchor latents.
AnchorEdit is the first autoregressive diffusion framework for causal multi-turn image editing, achieving claimed SOTA consistency over 10+ rounds via three-stage training and a memory mechanism.
DisCo uses discrete action primitives for camera control in video world models to achieve more reliable action following than continuous trajectories.
citing papers explorer
-
VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion
VideoMLA applies multi-head latent attention with 3D-RoPE decoupling to autoregressive video diffusion, delivering 92.7% KV memory reduction while matching short-horizon baselines and leading long-horizon VBench scores.
-
MV-Forcing: Long Multi-View Video Generation via 4D-Grounded Spatio-Temporal Self-Forcing
MV-Forcing composes temporal and view-sequential autoregression in a single diffusion model, using a recurrent 3D reconstruction model as a geometric bridge to generate arbitrarily long, multi-view consistent videos.
-
Towards Memory-Efficient Autoregressive Video Generation via Instance-Specific Parametric Absorption
ISPA reduces KV cache size by up to 50% in AR video models by transitioning layers to local attention and applying instance-specific least-squares weight modulation to compensate for lost history.
-
MemLearner: Learning to Query Context memory for Video World Models
MemLearner introduces a learning-based adaptive context query method using query tokens in video world models to improve long-term scene consistency over rule-based retrieval.
-
TempAct: Advancing Temporal Plausibility in Autoregressive Video Generation via Planner-Executor RL
TempAct introduces a planner-executor RL framework with hierarchical group exploration and rewards to improve temporal consistency in autoregressive video diffusion models.
-
FLAT: Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation
FLAT maps compressed video diffusion latents to explicit triangle splats via ray-centered rotation parameterization and a product window function, reporting better geometric accuracy than 3D Gaussian baselines under identical training.
-
TetherCache: Stabilizing Autoregressive Long-Form Video Generation with Gated Recall and Trusted Alignment
TetherCache organizes KV-cache into sink, memory, and recent regions and applies gated recall with attention-diversity balancing plus trusted memory editing to stabilize long-horizon autoregressive video diffusion.
-
LongLive-RAG: A General Retrieval-Augmented Framework for Long Video Generation
LongLive-RAG formulates long video generation as retrieval-augmented generation by treating self-generated latents as a dynamic searchable history and adding a Window Temporal Delta Loss for better retrieval.
-
LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation
LongLive-2.0 delivers an NVFP4 parallel infrastructure that enables direct training of long multi-shot autoregressive diffusion video models and achieves up to 2.15x training and 1.84x inference speedups on Blackwell and other GPUs.
-
Echo-Forcing: A Scene Memory Framework for Interactive Long Video Generation
Echo-Forcing decouples stable anchors, compressed history, and recent dynamics in video diffusion KV caches using hierarchical memory, scene recall frames, and difference-aware decay to support interactive long video generation under bounded cache.
-
CausalCine: Real-Time Autoregressive Generation for Multi-Shot Video Narratives
CausalCine enables real-time causal autoregressive multi-shot video generation via multi-shot training, content-aware memory routing for coherence, and distillation to few-step inference.
-
FreeSpec: Training-Free Long Video Generation via Singular-Spectrum Reconstruction
FreeSpec uses SVD-based spectral reconstruction to fuse global low-rank and local high-rank features, reducing content drift and preserving temporal dynamics in long video generation.
-
Stream-R1: Reliability-Perplexity Aware Reward Distillation for Streaming Video Generation
Stream-R1 improves distillation of autoregressive streaming video diffusion models by adaptively weighting supervision with a reward model at both rollout and per-pixel levels.
-
Efficient Video Diffusion Models: Advancements and Challenges
A survey that groups efficient video diffusion methods into four paradigms—step distillation, efficient attention, model compression, and cache/trajectory optimization—and outlines open challenges for practical use.
-
Grounded Forcing: Bridging Time-Independent Semantics and Proximal Dynamics in Autoregressive Video Synthesis
Grounded Forcing introduces dual memory caching, reference-based positional embeddings, and proximity-weighted recaching to bridge stable semantics with local dynamics, improving long-range consistency in autoregressive video synthesis.
-
DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos
DreamDojo is a foundation world model pretrained on the largest human video dataset to date that uses continuous latent actions to transfer interaction knowledge and achieves controllable physics simulation after robot post-training.
-
Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation
Causal Forcing distills few-step autoregressive video generators from an autoregressive diffusion teacher rather than a bidirectional one, avoiding conditional-expectation blur and beating Self Forcing on motion and quality metrics.
-
RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation
Step Forcing lets a 4-step video world model generate 30-second closed-loop rollouts, and a VLM judge scores them to reproduce the real RoboArena policy ranking at r=0.989.
-
ABot-M0.5: Unified Mobility-and-Manipulation World Action Model
Aligning temporal granularity, action subspaces, and train-test conditioning yields SOTA long-horizon mobile and fine-grained manipulation success for a unified world-action model.
-
InteractiveAvatar: Real-Time Streaming Video Generation for Consistent and Intent-Aware Avatars
InteractiveAvatar is a real-time infinite-streaming avatar video generation system using autoregressive distillation, Long-Short Visual Memory for consistency, and a Reasoning-Reaction Module for intent-aware interactions.
-
SurgVista: Long-Horizon Surgical World Modeling with Plausible Instrument-Tissue Dynamics
SurgVista mitigates spatial interaction incoherence and temporal fidelity collapse in surgical world models through trajectory-based contrastive regularization and drift-perturbed training, outperforming prior methods on a new long-horizon benchmark.
-
UniTemp: Unlocking Video Generation in Any Temporal Order via Bidirectional Distillation
UniTemp enables arbitrary temporal order video generation in autoregressive diffusion models via bidirectional distillation and blockwise anchor latents.
-
AnchorEdit: Maintaining Temporal Consistency in Multi-turn Image Editing via Causal Memory
AnchorEdit is the first autoregressive diffusion framework for causal multi-turn image editing, achieving claimed SOTA consistency over 10+ rounds via three-stage training and a memory mechanism.
-
DisCo: World Models with Discrete Camera Motion Control
DisCo uses discrete action primitives for camera control in video world models to achieve more reliable action following than continuous trajectories.
-
Echo-Infinity: Learning Evolving Memory for Real-Time Infinite Video Generation
Echo-Infinity replaces handcrafted KV-cache schedules with end-to-end optimized Memory Queries and a Unified Relative RoPE recipe to support real-time infinite video generation in diffusion transformers.
-
AAD-1: Asymmetric Adversarial Distillation for One-Step Autoregressive Video Generation
AAD-1 uses a causal generator with a bidirectional holistic discriminator plus phased distribution matching before adversarial training to reach state-of-the-art one-step autoregressive video generation on VBench.
-
LiveBand: Live Accompaniment Generation in the Audio Domain
LiveBand generates high-fidelity music accompaniments to live audio in real time via a causal transformer in audio latent space trained with adversarial sequence-level supervision.
-
AR Forcing: Towards Long-Horizon Robot Navigation World Model
AR Forcing trains diffusion world models by integrating standard noise prediction loss into an autoregressive loop that uses self-generated predictions as context, reducing train-inference mismatch for improved long-horizon image consistency and trajectory accuracy on navigation datasets.
-
Robust Dreamer: Deviation-Aware Latent Gaussian Memory for Action-Controlled AR Video Generation
Robust Dreamer uses Latent Gaussian Memory anchored to diffusion latents and Deviation Learning with a Dynamic Deviation Archive to reduce drift in long-horizon action-controlled image-to-video generation, reporting SOTA results on ScanNet, DL3DV, and OmniWorldGame.
-
OmniMem: Scalable and Adaptive Memory Retrieval for Long Video Generation
OmniMem enables scalable long video generation via adaptive sparse KV retrieval that addresses local bias and union explosion while preserving explicit historical access.
-
StreamChar: Long-Horizon Streaming Character Audio-Video Generation with Decoupled Orchestration
StreamChar decouples LLM-based orchestration from DiT denoising to achieve real-time long-horizon streaming character audio-video generation with reduced drift and misalignment.
-
StreamEdit: Training-Free Video Editing via Few-Step Streaming Video Generation
StreamEdit enables high-quality training-free video editing by adapting streaming video generation models with dual-branch fast sampling, self-attention bridge, cross-attention grounding, source-oriented guidance, and visual prompting, outperforming prior methods in few-step regimes.
-
DySink: Dynamic Frame Sinks for Autoregressive Long Video Generation
DySink replaces static early-frame anchors in autoregressive video generation with dynamically retrieved, visually relevant history, filtered by an inter-head consensus gate, improving long-video temporal scores.
-
FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching
FlowLong generates videos several times longer than native model windows by blending adjacent predictions with Tweedie matching to enforce manifold and temporal consistency while using stochastic noise injection early and deterministic sampling later.
-
RAVEN: Real-time Autoregressive Video Extrapolation with Consistency-model GRPO
RAVEN aligns training and inference for causal autoregressive video diffusion via interleaved rollout repacking and introduces CM-GRPO for direct RL on consistency-model kernels, claiming better quality than recent baselines.
-
Head Forcing: Long Autoregressive Video Generation via Head Heterogeneity
Head Forcing assigns tailored KV cache strategies to local, anchor, and memory attention heads plus head-wise RoPE re-encoding to extend autoregressive video generation from seconds to minutes without training.
-
HorizonDrive: Self-Corrective Autoregressive World Model for Long-horizon Driving Simulation
HorizonDrive is a new anti-drifting autoregressive training and distillation method that enables minute-scale stable driving video rollouts by making the teacher model rollout-capable via scheduled rollout recovery and teacher rollout DMD.
-
AsymTalker: Identity-Consistent Long-Term Talking Head Generation via Asymmetric Distillation
AsymTalker uses temporal reference encoding and asymmetric knowledge distillation to produce identity-consistent talking head videos up to 600 seconds long at 66 FPS.
-
Mutual Forcing: Dual-Mode Self-Evolution for Fast Autoregressive Audio-Video Character Generation
Mutual Forcing trains a single native autoregressive audio-video model with mutually reinforcing few-step and multi-step modes via self-distillation to match 50-step baselines at 4-8 steps.
-
Memorize When Needed: Decoupled Memory Control for Spatially Consistent Long-Horizon Video Generation
A decoupled memory branch with hybrid cues, cross-attention, and gating improves spatial consistency and data efficiency in long-horizon camera-trajectory video generation.
-
Repurposing 3D Generative Model for Autoregressive Layout Generation
LaviGen turns 3D generative models into an autoregressive layout generator that models geometric and physical constraints, delivering 19% higher physical plausibility and 65% faster inference on the LayoutVLM benchmark.
-
From Synchrony to Sequence: Exo-to-Ego Generation via Interpolation
Interpolating only the video frames between synchronized exo and ego clips already turns discontinuous cross-view generation into continuous sequence modeling and measurably improves diffusion-based Exo2Ego synthesis.
-
Long-Horizon Streaming Video Generation via Hybrid Attention with Decoupled Distillation
Hybrid Forcing combines linear temporal attention for long-range retention, block-sparse attention for efficiency, and decoupled distillation to achieve real-time unbounded 832x480 streaming video generation at 29.5 FPS.
-
GeoWorld: Geometric World Models
GeoWorld applies hyperbolic geometry to JEPA world models and introduces geometric reinforcement learning, reporting modest success-rate gains of ~3% and ~2% on 3- and 4-step planning tasks versus V-JEPA 2.
-
Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion
Rolling Sink is a training-free cache adjustment technique that maintains visual consistency in autoregressive video diffusion models for ultra-long open-ended generation beyond training horizons.
-
WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling
A real-time video diffusion world model that uses dual action control, reframed position encodings, and context-aligned distillation to keep generated environments consistent over hundreds of frames.
-
Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillation
Reward Forcing combines EMA-Sink tokens and Rewarded Distribution Matching Distillation to deliver state-of-the-art streaming video generation at 23.1 FPS without copying initial frames.
-
Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length
Live Avatar reports real-time streamable generation from a 14B audio-driven diffusion model at ~20 FPS on 5 H800s with stable identity over 10,000 seconds.
-
Directing the World: Fast Autoregressive Video Generation with Compositional Human-Camera Control
A decoupled-control autoregressive video model using Fast-Slow Memory training, dynamic projection, and staged camera control to produce stable long-horizon outputs with human and viewpoint guidance.
-
Ultra Flash: Scaling Real-Time Streaming Video Generation to High Resolutions
Ultra Flash introduces a cascaded streaming super-resolution framework with specialized training, upsampling, and optimization to enable real-time high-resolution video generation from low-res diffusion models.