Autoregressive Video Generation

TL;DR

Autoregressive video DiTs generate long-form video by repeatedly conditioning each new clip on past frames, in contrast to bidirectional/full-sequence T2V models that synthesize a fixed-length window jointly. The filed papers split the long-horizon problem into three complementary subproblems — what to remember (PFP’s pretrained memory compression), how to stay on the data manifold (Context Forcing’s long-context teacher + error-recycling distillation), and how to control multi-minute generation (LongVie 2’s multi-stage curriculum of multimodal guidance + degradation-aware training + history-context guidance). All three attack the same forgetting-vs-drifting dilemma that the Self-Forcing / LongLive / Self-Forcing++ / Rolling Forcing / FramePack-F1 streaming-tuning lineage struggles with past ~10s.

Key claims

• The fundamental failure mode of streaming-tuning approaches is a student–teacher mismatch — the student rolls out for 20–60s but the teacher only sees 5s windows, capping learnable context length (Context Forcing: Consistent Autoregressive Video Generation with Long Context §1, Fig. 2(b)).
• This manifests as a forgetting–drifting dilemma: short memory → minimal error accumulation but identity/scene loss; long memory → identity preserved but drift away from the real-video manifold (Context Forcing: Consistent Autoregressive Video Generation with Long Context §1, Fig. 1).
• The long-horizon objective can be addressed by an explicit memory-compression encoder pretrained with a “preserve any random frame perfectly” objective — random-index frame reconstruction at 4×4×2 compression yields PSNR 17.41 / SSIM 0.596 / LPIPS 0.171, vs. FramePack-style large-patchifier at PSNR 12.93 / LPIPS 0.365 (Pretraining Frame Preservation in Autoregressive Video Memory Compression §3.1, Table 1).
• …or by partitioning the KV cache into functional regions — Attention Sink + Slow Memory (surprisal-promoted) + Fast Memory (FIFO) + bounded RoPE — extending effective context past 20s vs. 1.5–9.2s for prior SOTA (Context Forcing: Consistent Autoregressive Video Generation with Long Context §3.3, Table 2).
• …or by a multi-stage training curriculum that separates controllability, degradation-robustness, and cross-clip consistency: LongVie 2 sequences multimodal guidance → degradation-aware input-frame training → history-context guidance and claims 5-minute continuous generation (LongVie 2: Multimodal Controllable Ultra-Long Video World Model project page).
• …or by a structured stream buffer with named functional slots — [reference | long-term queue (≤3) | short-term | denoising stream (3×3)] — turning the cache into an explicit data structure rather than a windowed KV (FlowAct-R1: Towards Interactive Humanoid Video Generation §2.1).
• …or by a recurrent memory vector — a single fixed-size representation built by recurrent attention over the hidden states of prior segments, cross-attended at every Transformer block during denoising; relative positional encoding inside the memory-attention layer is the headline length-extrapolation mechanism (MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation §3, Fig. 2A).
• …or by an inference-time alignment scheme — WMReward treats a frozen AR base model (MAGI-1 24B) as a sample-able support and uses a latent-world-model (VJEPA-2) surprise reward + best-of-N / ∇-guidance to steer toward physically plausible trajectories; PhysicsIQ V2V 55.22 → 62.00 (+6.78), 1st place at ICCV 2025 (Inference-time Physics Alignment of Video Generative Models with Latent World Models §3.1, Table 1).
• Drift correction during training requires the teacher / supervisor / input distribution to look drifted: Context Forcing achieves this via Error-Recycling Fine-Tuning that injects Bernoulli past-residual noise into the teacher’s context (Context Forcing: Consistent Autoregressive Video Generation with Long Context §3.4); LongVie 2 calls the equivalent recipe “degradation-aware training on the input frame” (LongVie 2: Multimodal Controllable Ultra-Long Video World Model project page); FlowAct-R1 implements the same idea by probabilistically swapping ground-truth memory latents for “generated-GT” latents re-noised+denoised by an intermediate model (FlowAct-R1: Towards Interactive Humanoid Video Generation §2.2); MALT Diffusion is the originating instance — it injects Gaussian noise directly into the memory vector built from ground-truth latents during training so the model learns to denoise under imperfect memory (MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation §Training, Fig. 5B).
• Pretraining a memory encoder on a clean, isolated objective transfers to the AR fine-tune — pretrained vs. random-init at the same architecture / 100k steps differs by ~+9 cloth-consistency and +4 identity points (Pretraining Frame Preservation in Autoregressive Video Memory Compression Table 2).
• Bounded Positional Encoding alone is responsible for ~9 VBench Total points (82.45 → 73.52 when removed) and is essentially free to add (Context Forcing: Consistent Autoregressive Video Generation with Long Context Table 3).
• Surprisal-driven cache consolidation — promote a Fast-Memory token to Slow Memory when its key-vector cosine distance from its predecessor exceeds threshold τ — outperforms uniform-interval slow-memory sampling (VBench Total 82.45 vs. uniform-interval-1 80.82 / interval-2 81.11) (Context Forcing: Consistent Autoregressive Video Generation with Long Context §3.3, Table 3).
• Dense-shot-shift training data effectively hides error accumulation for short-form content; long-single-shot continuous generation is the harder regime (Pretraining Frame Preservation in Autoregressive Video Memory Compression §4.2).
• AR backbones already place physically plausible trajectories somewhere in their support; the inference-time challenge is searching that support efficiently — VJEPA-2 surprise is a stronger search signal than pixel-MSE (VideoMAE) or VLM-judge rewards (Inference-time Physics Alignment of Video Generative Models with Latent World Models §3.1, Table 1).
• Inference-time drift correction need not be guidance-based: a periodic re-noise-and-denoise pass on the most error-prone memory component (short-term memory in FlowAct-R1) using the more-stable components (reference + long-term) as guidance is sufficient to keep streaming generation on-manifold over multi-minute durations (FlowAct-R1: Towards Interactive Humanoid Video Generation §2.5).
• An asymmetric “fake-causal” attention mask — denoising stream attends to memory + reference + itself, memory + reference do not attend to the denoising stream — keeps fully denoised tokens from being corrupted by the still-in-progress denoising stream as it propagates through deep DiT layers, while also cutting attention compute (FlowAct-R1: Towards Interactive Humanoid Video Generation §2.2).
• Streaming-AR video can be pushed to a 3-NFE per-chunk inference budget at 25 fps / 480p via a 3-stage distillation pipeline (CFG-elimination → naive step distillation into 3 segments → chunk-aware DMD), supported by FP8 quantization, frame-level hybrid parallelism, fused kernels, and async VAE/DiT pipelining (FlowAct-R1: Towards Interactive Humanoid Video Generation §2.4).
• The asymmetric-distillation pipeline behind CausVid / Self Forcing has two gaps to close, not one: the sampling-step gap (multi → few-step) is well-understood; the architectural gap (bidirectional full attention → causal attention) was not previously addressed theoretically (Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation §2.2).
• ODE distillation is a flow-map-matching procedure and requires frame-level injectivity — each noisy frame must map to a unique clean frame under the teacher’s PF-ODE; distilling an AR student from a bidirectional teacher violates this and collapses to a conditional-expectation solution rather than recovering the teacher’s flow map (Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation §3.2).
• The fix is to fine-tune the bidirectional base into an AR diffusion teacher first (via teacher forcing, not diffusion forcing) and use that AR teacher for causal ODE distillation; the resulting Causal Forcing pipeline beats Self Forcing by 19.3% Dynamic Degree, 8.7% VisionReward, 16.7% Instruction Following at matched training/inference budget (Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation §3.3, §4).
• Diffusion forcing (CausVid-style noisy-prefix AR training) is suboptimal for AR diffusion training — but bidirectional diffusion forcing for video continuation remains valid; the specific refuted practice is conditioning on a noisy prefix during AR training (Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation §3.2).
• Causal ODE distillation and causal consistency distillation target the same object (the AR-conditional flow map) but differ in supervision shape: ODE needs full PF-ODE trajectories precomputed and stored; CD takes a single online teacher step between adjacent timesteps on real videos, eliminating the trajectory-storage bottleneck (Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation §3).
• Causal CD as the initialization stage (Causal Forcing++) under a frame-wise 2-step setting surpasses 4-step chunk-wise Causal Forcing by +0.1 VBench Total / +0.3 VBench Quality / +0.335 VisionReward, while cutting first-frame latency 50% and Stage-2 training cost ~4× (Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation Abstract, §4).
• DMD (reverse KL, mode-seeking) gives sharp first-few-frame quality but amplifies exposure bias over long AR rollouts; CD (forward KL, mode-covering) is the right objective for the initialization stage that feeds into long rollout — the two should be used at different stages of the same pipeline, not interchanged (Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation §4, Fig. 5(b)).

Recent contributions

• [2026-07-20] Happy Oyster — Real-Time World Model for Interactive Creation: Happy Oyster (Alibaba Token Hub) — closed real-time interactive WFM landing page (Directing + Wandering modes, up to 3 min at 480p/720p). Almost certainly an autoregressive/streaming rollout given the “keeps listening and responding throughout generation” framing and multi-minute duration, but no architecture is disclosed. Filed as a product-surface marker in the closed real-time interactive-video race alongside Genie 3 / GWM-1 / HY-World 1.5 / LingBot-World.
• [2026-07-17] HDR: Hierarchical Denoising for Multi-Step Visual Reasoning: HDR adds a hierarchical planning axis to streaming AR video: latents form a temporal-resolution tree and generation proceeds coarse-to-fine, with a Sparse Hierarchical Attention Pattern (local + parent + first-frame) that overlaps with Rolling Forcing’s attention-sink + local-window recipe but adds a vertical plan → detail dimension via a shared KV cache. Reports +26 absolute overall success on a six-task long-horizon reasoning benchmark vs. a streaming AR-diffusion baseline while keeping 0.70s-per-latent streaming — an alternative to the memory-compression / KV-partition / degradation-training axes for long-horizon coherence.
• [2026-07-16] VideoRAE: Taming Video Foundation Models for Generative Modeling via Representation Autoencoders: VideoRAE feeds AR video generators with a discrete multi-codebook high-dimensional quantization of frozen VFM (V-JEPA 2 / VideoMAEv2) features, and reports UCF-101 class-to-video gFVD 40 with an AR generator on this latent — a data point that AR video generators can be pushed further by inheriting a video-foundation-model latent rather than training a bespoke VQ tokenizer from scratch.
• [2026-07-14] Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model: Xiaomi-Robotics-U0’s FlashAR+ inference (diagonal-parallel decoding + vLLM paged KV cache + acceleration adaptation) delivers up to 82.9× generation speedup and drops 1024² single-sample latency from 450.77s to 5.44s, putting AR image/video serving back on the Pareto frontier against diffusion.
• [2026-07-10] LingBot-VA 2.0: Native Video-Action Pretraining for Generalizable Robot Control: LingBot-VA 2.0 extends the causal-video-DiT recipe by pretraining it from scratch for embodied control (rather than retrofitting a bidirectional generator) with a sparse MoE-13B-A1.9B video stream, Multi-Chunk Prediction (imported from Next Forcing) supervising 3 future chunks via lightweight auxiliary heads at intermediate layers to break the “next chunk looks the same as this chunk” appearance-copying failure mode, and Foresight Reasoning async inference with learned-forward-dynamics re-grounding for real-time closed-loop use. Reported 225 Hz peak inference frequency.
• [2026-07-09] LingBot-World 2.0 / LingBot-World-Infinity — Infinite Worlds with Versatile Interactions: LingBot-World 2.0 moves the LingBot recipe from “block-causal distillation only” (v1) to causal pretraining as the base training objective, then re-distills a real-time student at 720p / 60 fps. Inference exposes explicit attention-sink + local-window CLI flags (--sink_size 6 --local_attn_size 18), matching the sink+bounded-RoPE machinery from Context Forcing. Claims unbounded interaction horizon vs v1’s ~10-minute ceiling; no VBench numbers disclosed at filing.
• [2026-07-04] RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation: RoboWorld introduces Step Forcing — denoise from a one-step self-forwarded prior under a fixed few-step schedule (matching train and inference), with probabilistic data-grounded anchor steps to preserve action controllability; 15.3 FPS at 4 steps with frame-causal attention + KV-cache.
• [2026-06-29] PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation: PhysisForcing applies its physics-aware auxiliary losses on top of Wan2.2-I2V-A14B and Cosmos3-Nano (both autoregressive/I2V video DiTs) and shows the gains transfer across very different backbones, suggesting the recipe is architecture-agnostic.
• [2026-06-24] Next Forcing: Causal World Modeling with Multi-Chunk Prediction: Next Forcing introduces a new supervision-shape axis orthogonal to memory/drift fixes: lightweight auxiliary multi-chunk-prediction heads (next¹/²/³, chained across depths) port LLM-style MTP into causal-AR video. Reports 2.3× faster convergence vs. LingBot-VA at 50 fps, new SOTA on RoboTwin (94.1/93.5), >50% FVD reduction on general video pretraining, and an optional 2× inference speedup when heads are retained for parallel next-chunk prediction.
• [2026-06-23] Jun Gao thread (4/5) on Cosmos2.5-2B finetuned with cleaned data + kinematic skeleton conditioning: Single-GPU finetune of Cosmos2.5-2B with skeleton conditioning suggests small Cosmos backbones can compete on action-following character generation with much heavier Wan-based pipelines.
• [2026-06-23] Causal World Modeling for Robot Control (LingBot-VA): LingBot-VA extends causal-AR + chunkwise flow matching from pure-generative video to closed-loop control: the same KV-cache + causal-attention discipline used in Context Forcing / Causal Forcing, but the AR sequence interleaves video latents with action tokens, and fresh real observations replace generated tokens at each step, removing the train-test distribution mismatch teacher forcing usually suffers. Variable chunk-size training (1/4/8/16) lets one model trade per-step compute for reactivity at deployment.
• [2026-06-09] Latent Spatial Memory for Video World Models: Mirage adds a sixth operating point to the forgetting–drifting Pareto: an external 3D-indexed latent cache (depth-guided back-projection of diffusion latent tokens; novel views via direct latent warping). Direct successor to SPMem (Video World Models with Long-term Spatial Memory) — same external-cache architecture, but storing latents rather than RGB, eliminating the per-query VAE encode/render round-trip. Reports 10.57× speedup and 55× memory reduction over explicit-3D baselines.
• [2026-06-08] Scaling Video Training with Parallelism: Identifies a load-balancing failure mode specific to AR teacher-forced video training: the [clean history ; noisy target] sequence layout means generic Ulysses-style SP leaves some ranks with mostly clean context and others target-heavy. Balanced SP (used in LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation) fixes this by making the SP unit a temporal chunk that holds matched clean + noisy latents on the same rank, with VAE encoding pulled into the SP partition via a left halo.
• [2026-06-06] Direct Video-Action Models — Causal Video Models Are Data-Efficient Robot Policy Learners: Rhoda AI’s DVA introduces Context Amortization — a from-scratch causal-video training recipe where at every position along a long sequence the model sees noise-free history and predicts the noisy future, amortizing many supervision points over one context encode while strictly matching the inference mask. Positioned as an explicit response to Diffusion Forcing’s per-frame-independent-noise training distribution rarely producing noise-free-history → noisy-future arrangements, and as an alternative to the explicit-KV-cache partitioning used by Context Forcing / PackForcing / Rolling Forcing. First filed long-context training recipe motivated by closed-loop robot control rather than long-video synthesis.
• [2026-05-27] Planning at Inference: MCTS Test-Time Scaling for Long Video Generation: Attacks the chunked-extension semantic-drift problem with inference-time MCTS over candidate continuations on Cosmos-Predict2, no retraining of the base AR video DiT. Search-based test-time scaling beats Best-of-N / Greedy / Beam on object permanence, temporal coherence, and text–video alignment; produces coherent >20s videos.
• [2026-05-26] MIGA: Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos: MIGA — train-free FIFO-Diffusion descendant. Keeps the frame-level diagonal-denoising substrate (constant memory, infinite output length) but fixes its training–inference mismatch with a zigzag noise schedule and adds cross-time attention via anomaly-triggered self-reflection at the queue tail and long-range frame guidance from the queue head. No fine-tuning of the base T2V model.
• [2026-05-24] MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation: MALT Diffusion (Yu et al., Google + KAIST + Luma AI, Feb 2025; surfaced via Google patent US20250363171A1 May 2025) — the originating method in this lineage. Recurrent memory vector built by attention over prior segments’ hidden states, cross-attended at every diffusion-Transformer block during denoising, with relative positional encoding for length extrapolation, noise augmentation on the memory vector during training to close the train/inference gap, and stop-grad through prior segments so training memory does not grow with video length. Trained on ≤5s clips, claims >120s and ≥1024-frame generation at 8 FPS. PFP, Context Forcing, PackForcing, FlowAct-R1, and LongVie 2 can each be read as evolutions of one or more of MALT’s three core ingredients.
• [2026-05-23] Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation: Causal Forcing++ — replaces the causal-ODE-distillation initialization of Causal Forcing with causal consistency distillation (single online teacher step on real videos, no offline trajectory storage). Frame-wise 2-step beats chunk-wise 4-step Causal Forcing by +0.1 VBench Total, +0.3 VBench Quality, +0.335 VisionReward, with 50% lower first-frame latency and ~4× lower Stage-2 training cost. First open release of frame-wise 1/2-step AR video models. Pipeline extends to action-conditioned world models (minWM).
• [2026-05-23] Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation: Causal Forcing — identifies a frame-level-injectivity violation in Self-Forcing/CausVid’s ODE-init stage (bidirectional teacher → AR student cannot recover the teacher’s flow map). Fix: three-stage pipeline (teacher-forced AR diffusion training → causal ODE distillation from the AR teacher → asymmetric DMD). Beats Self Forcing by 19.3% Dynamic Degree, 8.7% VisionReward, 16.7% Instruction Following at matched budget. Sits upstream of the rollout-drift fixes (Context Forcing, PFP, FlowAct-R1, LongVie 2) at the asymmetric-distillation stage they all build on.
• [2026-05-22] FlowAct-R1: Towards Interactive Humanoid Video Generation: A real-time (25 fps / 480p, TTFF ~1.5 s) interactive humanoid streaming AR-video system on Seedance MMDiT — chunkwise diffusion forcing + Self-Forcing++-style memory-error injection + 4-slot structured memory bank + fake-causal attention + inference-time Memory Refinement + 3-NFE multi-stage distillation + MLLM action planner. Application-specialized cousin of the Self-Forcing / Context Forcing lineage; user-study evaluation only, no shared-benchmark numbers.
• [2026-05-22] Inference-time Physics Alignment of Video Generative Models with Latent World Models: Inference-time alignment of MAGI-1 24B via WMReward (VJEPA-2 surprise) with best-of-N and ∇-guidance — no weight updates, monotone scaling with particle count, 1st place at ICCV 2025 PhysicsIQ Challenge. Complements the three training-time approaches below; works on a frozen AR base model.
• [2026-05-22] LongVie 2: Multimodal Controllable Ultra-Long Video World Model: Multi-stage training curriculum (multimodal guidance → degradation-aware training → history-context guidance) pushing continuous AR video generation to 5 minutes; introduces LongVGenBench (100 × 1-minute high-resolution videos). Project-page-only; no numbers or code yet.
• [2026-05-22] Context Forcing: Consistent Autoregressive Video Generation with Long Context: Names the forgetting–drifting dilemma; addresses it with a Context Teacher (long-context, error-recycling-finetuned) + Slow-Fast Memory KV cache + Bounded Positional Encoding. 60s VBench subject-consistency 94.88 on Wan2.1-T2V-1.3B at 17 FPS.
• [2026-05-22] Pretraining Frame Preservation in Autoregressive Video Memory Compression: Pretrains a memory encoder with a random-index frame-preservation objective so it can drop into Wan/HunyuanVideo as a LoRA-fine-tuned AR memory module; 20s history at ~5k context on consumer 12GB GPUs.

Open questions

• Are the filed approaches stackable? PFP keeps context cheap (compression), Context Forcing keeps context useful (distillation + cache management), LongVie 2 schedules training stages, FlowAct-R1 productizes a slot-structured memory bank, Causal Forcing/++ fix the initialization of the few-step AR student, MALT is the recurrent-memory-vector ancestor of PFP/Context-Forcing/FlowAct-R1 — nothing structurally conflicts, but no filed paper has tried more than one axis. Causal Forcing in particular is upstream of the others (it fixes Stage 0 of their pipelines) and should compose cleanly.
• MALT’s noise-augmented memory, Context Forcing’s Error-Recycling Fine-Tuning, LongVie 2’s degradation-aware input training, and FlowAct-R1’s generated-GT memory sampling are four different injection sites for the same idea — perturb the memory during training to look like rollout. Which injection site dominates, and do they stack?
• Does PFP’s memory encoder benefit from Context Forcing’s Bounded Positional Encoding? PFP says nothing about RoPE behaviour over long sequences; Context Forcing reports BPE is responsible for ~9 VBench points. MALT uses relative positional encoding inside memory attention as its own length-extrapolation mechanism — does BPE or MALT-style relative PE win on the same backbone?
• Can the WMReward inference-time recipe be composed with the training-time fixes — e.g. does Context Forcing’s drift-suppressed model still benefit from VJEPA-2-steered BoN, or are the gains substitutes rather than complements?
• FlowAct-R1’s Memory Refinement (periodic re-noise+re-denoise of short-term memory using reference + long-term as guidance) is a third class of drift fix that sits at inference but doesn’t use a reward model. Does this trick generalize beyond humanoid generation? And does it stack with training-time fixes or substitute for them?
• The training-time papers are built on Wan-family backbones (Wan2.1-T2V-1.3B, Wan 2.2 5B/14B); WMReward uses MAGI-1; FlowAct-R1 uses Seedance; Causal Forcing/++ use Wan2.1; MALT uses W.A.L.T. How much of the headline numbers transfer across backbones?
• LongVGenBench (LongVie 2), the MovieGenBench-derived prompts (Context Forcing), the user-study ELO setup (PFP), PhysicsIQ (WMReward), and FlowAct-R1’s GSB user study are five different evaluation protocols. None of the five papers compares on the others’ yardstick. MALT reports Kinetics-600 long-video prediction — a sixth.
• FlowAct-R1’s “fake-causal” attention mask is a clean statement of an asymmetry that other AR-video systems handle ad hoc. Is this a free win for any chunkwise-rollout DiT, or does it specifically need the slot-structured memory bank to work?
• Causal Forcing++‘s DMD-vs-CD analysis (CD for initialization, DMD for final polish) is a sharper statement than the field had — does the same split apply to other AR-video distillation pipelines, and where does it break down? In particular, FlowAct-R1’s 3-NFE pipeline uses chunk-aware DMD throughout; does swapping the first stage to causal CD give the same wins reported in Causal Forcing++?

Papers

2026-05

• AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation — AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation (published 2026-05-23)
• TokenTrim: Inference-Time Token Pruning for Autoregressive Long Video Generation — TokenTrim: Inference-Time Token Pruning for Autoregressive Long Video Generation (published 2026-05-23)
• LongVie 2: Multimodal Controllable Ultra-Long Video World Model — LongVie 2: Multimodal Controllable Ultra-Long Video World Model (published 2026-05-22)
• LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation — LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation (published 2026-05-18)
• Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation — Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation (published 2026-05-14)

2026-04

• Introducing Odyssey-2 Max: Scaled World Simulation — Introducing Odyssey-2 Max: Scaled World Simulation (published 2026-04-21)
• VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization — VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization (published 2026-04-14)
• Envisioning the Future, One Step at a Time (Myriad) — Envisioning the Future, One Step at a Time (Myriad) (published 2026-04-10)
• LPM 1.0: Video-based Character Performance Model — LPM 1.0: Video-based Character Performance Model (published 2026-04-09)

2026-03

• Matrix-Game 3.0: Real-Time and Streaming Interactive World Model with Long-Horizon Memory — Matrix-Game 3.0: Real-Time and Streaming Interactive World Model with Long-Horizon Memory (published 2026-03-27)
• PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference — PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference (published 2026-03-26)
• Wan 2.7 upcoming features preview (video editing, V2V, real-time) — Wan 2.7 upcoming features preview (video editing, V2V, real-time) (published 2026-03-22)
• Into the Dreamverse: Vibe Directing in FastVideo — Into the Dreamverse: Vibe Directing in FastVideo (published 2026-03-15)
• OmniForcing: Unleashing Real-time Joint Audio-Visual Generation — OmniForcing: Unleashing Real-time Joint Audio-Visual Generation (published 2026-03-12)
• Streaming Autoregressive Video Generation via Diagonal Distillation — Streaming Autoregressive Video Generation via Diagonal Distillation (published 2026-03-10)
• MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines — MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines (published 2026-03-06)
• Helios: Real Real-Time Long Video Generation Model — Helios: Real Real-Time Long Video Generation Model (published 2026-03-04)
• CausalWan2.2-I2V-A14B-Preview — FastVideo causal I2V model on Wan2.2-A14B — CausalWan2.2-I2V-A14B-Preview — FastVideo causal I2V model on Wan2.2-A14B (published 2026-03-02)

2026-02

• Solaris: Building a Multiplayer Video World Model in Minecraft — Solaris: Building a Multiplayer Video World Model in Minecraft (published 2026-02-23)
• MonarchRT: Efficient Attention for Real-Time Video Generation — MonarchRT: Efficient Attention for Real-Time Video Generation (published 2026-02-12)
• Xmax X1 — Real-Time Interactive Video Model (Product Announcement) — Xmax X1 — Real-Time Interactive Video Model (Product Announcement) (published 2026-02-08)
• Context Forcing: Consistent Autoregressive Video Generation with Long Context — Context Forcing: Consistent Autoregressive Video Generation with Long Context (published 2026-02-05)
• Quant VideoGen: Auto-Regressive Long Video Generation via 2-Bit KV-Cache Quantization — Quant VideoGen: Auto-Regressive Long Video Generation via 2-Bit KV-Cache Quantization (published 2026-02-03)
• Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation — Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation (published 2026-02-02)

2026-01

• Advancing Open-source World Models (LingBot-World) — Advancing Open-source World Models (LingBot-World) (published 2026-01-29)
• Project Genie: Experimenting with infinite, interactive worlds — Project Genie: Experimenting with infinite, interactive worlds (published 2026-01-29)
• FlowAct-R1: Towards Interactive Humanoid Video Generation — FlowAct-R1: Towards Interactive Humanoid Video Generation (published 2026-01-15)
• Inference-time Physics Alignment of Video Generative Models with Latent World Models — Inference-time Physics Alignment of Video Generative Models with Latent World Models (published 2026-01-15)
• Fast Autoregressive Video Diffusion and World Models with Temporal Cache Compression and Sparse Attention — Fast Autoregressive Video Diffusion and World Models with Temporal Cache Compression and Sparse Attention (published 2026-01-01)

2025-12

• Pretraining Frame Preservation in Autoregressive Video Memory Compression — Pretraining Frame Preservation in Autoregressive Video Memory Compression (published 2025-12-29)
• Yume-1.5: A Text-Controlled Interactive World Generation Model — Yume-1.5: A Text-Controlled Interactive World Generation Model (published 2025-12-22)
• End-to-End Training for Autoregressive Video Diffusion via Self-Resampling — End-to-End Training for Autoregressive Video Diffusion via Self-Resampling (published 2025-12-17)
• HY-World 1.5 (WorldPlay): A Systematic Framework for Interactive World Modeling with Real-Time Latency and Geometric Consistency — HY-World 1.5 (WorldPlay): A Systematic Framework for Interactive World Modeling with Real-Time Latency and Geometric Consistency (published 2025-12-17)
• Flowception: Temporally Expansive Flow Matching for Video Generation — Flowception: Temporally Expansive Flow Matching for Video Generation (published 2025-12-12)
• Runway GWM-1 launch broadcast — world model, avatars, audio, multi-shot editing — Runway GWM-1 launch broadcast — world model, avatars, audio, multi-shot editing (published 2025-12-11)
• EgoEdit: Dataset, Real-Time Streaming Model, and Benchmark for Egocentric Video Editing — EgoEdit: Dataset, Real-Time Streaming Model, and Benchmark for Egocentric Video Editing (published 2025-12-05)
• TV2TV: A Unified Framework for Interleaved Language and Video Generation — TV2TV: A Unified Framework for Interleaved Language and Video Generation (published 2025-12-04)
• RELIC: Interactive Video World Models with Long-Horizon Memory — RELIC: Interactive Video World Models with Long-Horizon Memory (published 2025-12-01)

2025-11

• STARFlow-V: End-to-End Video Generative Modeling with Normalizing Flows — STARFlow-V: End-to-End Video Generative Modeling with Normalizing Flows (published 2025-11-25)
• Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout — Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout (published 2025-11-25)
• PAN: A World Model for General, Interactable, and Long-Horizon World Simulation — PAN: A World Model for General, Interactable, and Long-Horizon World Simulation (published 2025-11-12)
• Neodragon: Mobile Video Generation using Diffusion Transformer — Neodragon: Mobile Video Generation using Diffusion Transformer (published 2025-11-08)
• InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation — InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation (published 2025-11-06)
• MotionStream: Real-Time Video Generation with Interactive Motion Controls — MotionStream: Real-Time Video Generation with Interactive Motion Controls (published 2025-11-03)

2025-10

• Generative View Stitching — Generative View Stitching (published 2025-10-28)
• LongCat-Video Technical Report — LongCat-Video Technical Report (published 2025-10-25)
• Krea Realtime 14B: Real-Time, Long-Form AI Video Generation — Krea Realtime 14B: Real-Time, Long-Form AI Video Generation (published 2025-10-15)
• Self-Forcing++: Towards Minute-Scale High-Quality Video Generation — Self-Forcing++: Towards Minute-Scale High-Quality Video Generation (published 2025-10-02)

2025-09

• Radamés Ajna — testing nano-banana with recursive editing as long video gen — Radamés Ajna — testing nano-banana with recursive editing as long video gen (published 2025-09-30)
• SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer — SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer (published 2025-09-29)
• Rolling Forcing: Autoregressive Long Video Diffusion in Real Time — Rolling Forcing: Autoregressive Long Video Diffusion in Real Time (published 2025-09-29)
• LongLive: Real-time Interactive Long Video Generation — LongLive: Real-time Interactive Long Video Generation (published 2025-09-26)
• X-Streamer: Unified Human World Modeling with Audiovisual Interaction — X-Streamer: Unified Human World Modeling with Audiovisual Interaction (published 2025-09-25)

2025-08

• InfinityHuman: Towards Long-Term Audio-Driven Human Animation — InfinityHuman: Towards Long-Term Audio-Driven Human Animation (published 2025-08-27)
• Wan-S2V: Audio-Driven Cinematic Video Generation — Wan-S2V: Audio-Driven Cinematic Video Generation (published 2025-08-26)
• InfiniteTalk: Audio-driven Video Generation for Sparse-Frame Video Dubbing — InfiniteTalk: Audio-driven Video Generation for Sparse-Frame Video Dubbing (published 2025-08-19)
• Yan: Foundational Interactive Video Generation — Yan: Foundational Interactive Video Generation (published 2025-08-12)
• Genie 3: A new frontier for world models — Genie 3: A new frontier for world models (published 2025-08-05)

2025-07

• MirageLSD: The First Live-Stream Diffusion AI Video Model — MirageLSD: The First Live-Stream Diffusion AI Video Model (published 2025-07-17)
• NeuralOS: Towards Simulating Operating Systems via Neural Generative Models — NeuralOS: Towards Simulating Operating Systems via Neural Generative Models (published 2025-07-11)

2025-06

• Qwen VLo: From "Understanding" the World to "Depicting" It — Qwen VLo: From “Understanding” the World to “Depicting” It (published 2025-06-26)
• Hunyuan-GameCraft: High-dynamic Interactive Game Video Generation with Hybrid History Condition — Hunyuan-GameCraft: High-dynamic Interactive Game Video Generation with Hybrid History Condition (published 2025-06-20)
• VideoMAR: Autoregressive Video Generation with Continuous Tokens — VideoMAR: Autoregressive Video Generation with Continuous Tokens (published 2025-06-17)
• Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation — Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation (published 2025-06-11)
• Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion — Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (published 2025-06-09)
• Video World Models with Long-term Spatial Memory — Video World Models with Long-term Spatial Memory (published 2025-06-05)
• TalkingMachines: Real-Time Audio-Driven FaceTime-Style Video via Autoregressive Diffusion Models — TalkingMachines: Real-Time Audio-Driven FaceTime-Style Video via Autoregressive Diffusion Models (published 2025-06-03)

2025-05

• Test-Time Training Done Right (LaCT) — Test-Time Training Done Right (LaCT) (published 2025-05-29)
• Let Them Talk: Audio-Driven Multi-Person Conversational Video Generation (MultiTalk) — Let Them Talk: Audio-Driven Multi-Person Conversational Video Generation (MultiTalk) (published 2025-05-28)

2025-04

• GenDoP: Auto-regressive Camera Trajectory Generation as a Director of Photography — GenDoP: Auto-regressive Camera Trajectory Generation as a Director of Photography (published 2025-04-09)

2025-02

• MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation — MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation (published 2025-02-18)

Unknown date

• Magica: AI UGC game engine (Dynamics Lab) — Magica: AI UGC game engine (Dynamics Lab) (published unknown)
• AI humans from Runway (Runway Characters / GWM Avatars product announcement) — AI humans from Runway (Runway Characters / GWM Avatars product announcement) (published unknown)