Papers
Topics
Authors
Recent
Search
2000 character limit reached

Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion

Published 8 Feb 2026 in cs.CV | (2602.07775v1)

Abstract: Recently, autoregressive (AR) video diffusion models has achieved remarkable performance. However, due to their limited training durations, a train-test gap emerges when testing at longer horizons, leading to rapid visual degradations. Following Self Forcing, which studies the train-test gap within the training duration, this work studies the train-test gap beyond the training duration, i.e., the gap between the limited horizons during training and open-ended horizons during testing. Since open-ended testing can extend beyond any finite training window, and long-video training is computationally expensive, we pursue a training-free solution to bridge this gap. To explore a training-free solution, we conduct a systematic analysis of AR cache maintenance. These insights lead to Rolling Sink. Built on Self Forcing (trained on only 5s clips), Rolling Sink effectively scales the AR video synthesis to ultra-long durations (e.g., 5-30 minutes at 16 FPS) at test time, with consistent subjects, stable colors, coherent structures, and smooth motions. As demonstrated by extensive experiments, Rolling Sink achieves superior long-horizon visual fidelity and temporal consistency compared to SOTA baselines. Project page: https://rolling-sink.github.io/

Summary

  • The paper introduces a training-free Rolling Sink approach that mitigates AR drift by dynamically sliding cached semantic content and positional indices.
  • The paper demonstrates that systematic cache management significantly enhances video stability and temporal consistency in long-horizon autoregressive models.
  • The paper validates its method on VBench-Long benchmarks, achieving state-of-the-art performance and stable video generation well beyond the 5-second training scope.

Rolling Sink: Training-Free Mitigation of Long-Horizon Drift in Autoregressive Video Diffusion

Introduction and Motivation

Autoregressive (AR) video diffusion models have recently advanced high-fidelity video generation, leveraging next-frame prediction to accommodate theoretically unbounded sequence lengths. However, a fundamental limitation remains: these models are typically trained on short, fixed-duration video clips (e.g., 5 seconds at 16 FPS). When deployed for test-time generation over substantially longer time horizons, they experience cumulative artifacts denoted as AR drift—manifesting as degraded subject identity, over-saturation, collapsed structures, and loss of temporal consistency. This mismatch between limited-horizon training and open-ended testing is identified as a specific form of exposure bias, which has previously been explored in sequence modeling literature.

To address this critical discrepancy, the paper introduces Rolling Sink, a training-free mechanism for maintaining AR cache consistency at test time—bridging the gap between training and long-horizon testing. The method systematically analyzes the cache behavior and conditions for stability, formulating a sequence of cache management strategies culminating in the Rolling Sink approach, which demonstrably enables stable video rollouts exceeding 30 minutes despite only 5s training windows. Figure 1

Figure 1: Rollin unlocks open-ended AR video generation. Despite a 5s training duration, Rolling Sink effectively scales the AR video synthesis to minutes long during testing.

Methodological Framework: Systematic Cache Analysis

The core insight is that, in AR video diffusion, the principal factor influencing stability over long horizons is the AR cache configuration, including both semantic consistency and positional encoding of the cached prior blocks.

AR Cache Mechanisms and Drift

Baseline approaches (e.g., Self Forcing) use a bounded context window, providing streaming efficiency but demonstrating severe visual instabilities when extrapolated to long horizons. The paper first characterizes the within-duration AR cache behavior—minimally drifted blocks and appropriate “sliding” in both positional indices and semantics—as the gold standard for long-horizon fidelity. Rolling Sink builds on stepwise analytical improvements:

  1. Attention Sink: Pinning a static prefix of early, self-generated blocks (the attention "sink") in the cache, which helps stabilize color but does not resolve temporal instabilities or repetition.
  2. Sliding Indices: Treating time indices of sink blocks as a moving window, the method updates their positional encodings, mitigating flickering artifacts but still leaving residual subject and structure inconsistencies.
  3. Sliding Semantics (“Rolling Sink”): Approximating true sliding semantics by rolling the semantic content of previously generated, minimally drifted within-horizon blocks through the sink at each generation step. This step produces the most stable, drift-free rollouts. Figure 2

    Figure 2: Overview of our analysis and the proposed Rollin. The progression includes Self Forcing’s bounded cache, static attention sink, sliding time indices, and finally sliding (rolled) semantics enabling drift-free extrapolation.

Empirical ablations using the VBench-Long benchmark (covering multiple diagnostic dimensions) show strict, monotonic improvements across this analytical sequence, supporting the necessity of each cache update mechanism. Figure 3

Figure 3

Figure 3: Evaluation results during the systematic analysis, on both 1-minute (left) and 5-minute (right) AR video synthesis. The video quality score consistently improves through each analysis step, with Rolling Sink yielding the highest performance.

Figure 4

Figure 4: Visual comparisons across various sink sizes. Larger sink sizes help, but AR drift still persists until sliding indices and rolling semantics are introduced.

Empirical Evaluation

The evaluation leverages VBench-Long for both 1-minute and 5-minute rollouts, as well as qualitative analysis against established SOTA AR models, specifically Self Forcing and LongLive.

Qualitative Comparisons

Rolling Sink produces videos with stable appearance, coherent structural dynamics, and smooth motion, markedly suppressing AR drift compared to baselines. It preserves subject identity and scene geometry significantly longer than methods trained on similar or even extended durations. Figure 5

Figure 5

Figure 5: Qualitative comparisons of Rollin with SOTA AR video synthesis baselines reveal substantial reduction in AR drift—Rolling Sink maintains subject and scene consistency where baselines degrade.

Quantitative Performance

Rolling Sink achieves the best average rank across nearly all VBench-Long dimensions, including subject/background consistency, color, motion smoothness, and flicker rate, at both 1-minute and 5-minute settings. Notably, it maintains or exceeds the scores of systems trained on much longer durations, such as LongLive (with 1min LoRA adaptation), despite never being exposed to such long sequences during training. Figure 6

Figure 6

Figure 6: Radar-charts of quantitative comparisons on 1-minute and 5-minute AR video synthesis. Rolling Sink yields SOTA visual and temporal metrics across all primary axes.

Ultra-Long Video Generation

Rolling Sink extends AR rollouts to ultra-long durations (e.g., 30 minutes), maintaining coherent and stable video generation throughout—demonstrating generalization far beyond the training regime.

Discussion and Implications

Rolling Sink substantiates the claim that cache stabilization at test time—not extension of the training window nor architectural overhaul—is the critical factor for drift-free, open-ended AR video synthesis. The analysis indicates that simply increasing training horizon is neither efficient nor necessary for overcoming exposure bias-induced drift. Instead, cache management strategies—principally, dynamically sliding both cached indices and semantic content—yield significant practical advantages.

Practical Implications

  • Training-Free Generalization: Rolling Sink obviates the need for retraining or adaptation when scaling to long-horizon deployment, providing immediate utility atop existing AR models.
  • Streaming-Efficient Implementation: The method preserves bounded cache and low computational overhead, making it practical for real-time video synthesis applications.
  • Modularity: Rolling Sink can be integrated as a post-processing or inference-time modification for any Self Forcing-like AR video diffusion model.

Theoretical Implications and Future Research

  • Role of Cache Semantics: The findings imply that AR model generalization critically hinges on test-time cache semantics rather than model capacity or explicit long-horizon training.
  • Beyond Single-Shot Generation: While Rolling Sink is demonstrated for single-prompt, continuous long-form video, future research directions could involve multi-shot synthesis, prompt transitions, or adaptive cache manipulation for conditional, interactive, or story-driven content generation.
  • Towards Complete Exposure Bias Mitigation: The approach does not entirely close the gap created by limited-horizon training but provides a strong baseline; further refinements of cache schemes may absorb remaining instabilities.

Conclusion

Rolling Sink delivers a principled, training-free solution for bridging the discrepancy between limited-horizon training and open-ended testing in autoregressive video diffusion. The approach relies on systematic cache maintenance, specifically the introduction of dynamically sliding semantic and positional cache segments, to maintain stability and fidelity over ultra-long horizons—achieving SOTA performance across qualitative and quantitative metrics without additional computational or annotation cost. This methodological advance sets a new standard for practical, stable deployment of AR video diffusion models in settings requiring long-duration coherent generation (2602.07775).

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Sign Up to Generate

Explain it Like I'm 14

Overview

This paper is about making AI models better at creating long, smooth videos. Many video AIs are trained on short clips (like 5 seconds), but people want them to keep going for minutes. When these models try to go much longer than they were trained for, the video quality often falls apart: colors get weird, people or objects change shape, and the motion becomes jerky or repetitive. The authors introduce a simple, training‑free method called Rollin (also referred to as Rolling Sink) that helps these models keep their videos stable and consistent for many minutes.

Key Objectives

The paper asks a straightforward question: How can we help a video model that was trained on short clips keep generating good‑looking video for a long time without retraining it?

More specifically:

  • Why do long videos drift in quality after a while?
  • Can we fix this by managing the model’s “memory” during generation?
  • Can we do it without extra training, so it’s fast and practical?

Methods and Approach (in simple terms)

Think of an autoregressive video diffusion model as an artist drawing a video one small chunk at a time. For each new chunk, it looks back at what it drew before. That look‑back is its “memory,” often called a cache. The cache can only hold a limited number of recent chunks (like carrying a small backpack of reference photos).

The core problem:

  • During training, the model only ever practices making about 5 seconds of video.
  • At test time, we ask it to make minutes. That mismatch is called a train‑test gap (also known as exposure bias). It’s like training for a 2 km run and then being asked to run a marathon—the longer you go, the more likely you drift off pace.

The authors study how to keep that small “memory backpack” in good shape so the model doesn’t drift. They build on a strong baseline called Self Forcing (trained on 5‑second clips) and propose three simple, practical ideas:

  • Attention Sink (anchor blocks): Pin a few of the earliest, cleanly generated chunks in the cache as permanent anchors. These stable references help keep colors and overall look from drifting. Think of them as a couple of reliable reference photos always in your backpack.
  • Sliding Indices (move the timestamps): The model internally marks chunks with time positions. Instead of leaving those anchors locked to the start, slide their time positions forward as the video rolls. This better matches the idea that the whole video is moving through time and reduces flickering.
  • Sliding Semantics (roll the content slice): Not just the timestamps—the content of the anchors should also “move” so the model always sees a fresh, consistent slice of earlier, stable material. The authors approximate this by rolling through the set of clean short‑clip chunks in a forward–backward pattern, feeding the model a moving window of trustworthy history.

Together, these steps form Rollin. Importantly:

  • It requires no extra training.
  • It keeps the cache size small and fixed (streaming‑friendly).
  • It uses the same fast sampling steps as the base model.

Main Findings and Why They’re Important

What changed with Rollin:

  • Much longer, stable videos: Even though the model was trained on just 5 seconds, Rollin keeps videos consistent for 5–30 minutes at 16 frames per second.
  • Fewer visual problems: It reduces color oversaturation, warped faces or objects, flicker, and repetitive motion.
  • Better scores: On a long‑video benchmark called VBench‑Long, Rollin scores higher than strong baselines across many categories (like subject consistency, motion smoothness, and overall visual quality) for both 1‑minute and 5‑minute tests. It achieves the best average rank across most dimensions.

Why it matters:

  • It shows you can push short‑trained models to perform well far beyond their training horizon.
  • It’s training‑free and efficient, making it practical to deploy.
  • It highlights that smart memory management during generation can fix long‑video drift.

Implications and Potential Impact

Rollin is a simple, effective way to stabilize long video generation without retraining. This can help creators, studios, and apps that want continuous, real‑time video generation—like long scenes, livestream storytelling, or extended camera moves.

Limitations and future directions:

  • The paper focuses on single‑prompt, single‑shot videos (one scene that keeps going). Real movies use multiple shots and prompts that change over time. A next step is to adapt these ideas to multi‑shot generation, ensuring smooth transitions while keeping consistency over very long timelines.

In short, the paper’s big idea is: if you keep the model’s small memory backpack well organized—using stable anchors and sliding them through time and content—you can turn short‑trained video models into long‑distance runners without extra training.

View Paper Prompt  View All Prompts 

Knowledge Gaps

Below is a single, focused list of the paper’s unresolved knowledge gaps, limitations, and open questions. Each point is stated concretely to guide future research.

  • The method is evaluated and implemented only on Self Forcing; generalization to other AR video diffusion architectures (e.g., DF/TF training regimes, different DiT variants, non-Wan VAEs, state-space memory models) is not tested.
  • The choice of bounded cache capacity K=6 and sink ratio S/K=83% is heuristic; a principled or adaptive strategy for selecting K and S per prompt, scene type, motion complexity, or rollout length is missing.
  • The “Sliding Semantics” rolling schedule (alternating forward/reverse segments) is ad hoc; no analysis or learning-based approach to derive optimal content rotation, ordering, or periodicity is provided.
  • Theoretical understanding of why sliding indices and rolling semantics mitigate exposure bias is absent; formal drift models and error propagation analyses across AR steps (and across cache states) are needed.
  • RoPE time index extrapolation to very large indices (minutes/hours) is assumed but not studied; the behavior of positional embeddings under extreme extrapolation (e.g., saturation/aliasing in RoPE) and alternatives (ALiBi, learned PE, hybrid PE) remain unexplored.
  • Residual flicker and consistency artifacts are acknowledged but not systematically characterized; a taxonomy of failure modes and targeted mitigations (e.g., per-layer cache weighting, temporal smoothing modules) are missing.
  • The approach repeats a finite training window via rolling, which risks periodicity/looping; quantitative metrics for novelty, non-redundancy, and long-horizon “loopiness” are not reported.
  • Multi-shot, prompt-changing long-video generation is explicitly out of scope; mechanisms for cache re-initialization, semantic refresh, transition smoothing, and prompt-conditioned memory persistence remain open.
  • The method assumes a fixed prompt and constant prompt embedding; strategies for gradual prompt evolution, conditioning drift control, or content “aging” without collapse are absent.
  • Layer-wise cache design is uniform (concatenate prior K/V at all DiT layers); the impact of layer-specific cache policies, attention head gating, decay weighting, or learned cache mixing is not investigated.
  • Cache corruption detection and repair are not addressed; adaptive filtering, compression, or confidence-weighted cache updates could help, but are unstudied.
  • Interaction with denoising schedules is fixed (4-step sampler); the relationship between sampling steps/noise schedule and long-horizon drift is not analyzed, nor are adaptive samplers explored.
  • Generalization to different frame rates, resolutions, aspect ratios, and block sizes is not evaluated; how Rollin scales or recalibrates to these settings is unclear.
  • Performance beyond 5 minutes is claimed qualitatively, but quantitative evaluation (metrics) is limited to 1 and 5 minutes; rigorous assessment at 10–30 minutes or longer is missing.
  • Robustness across prompt categories (e.g., high-speed camera motion, heavy occlusion, complex multi-object interactions, large illumination changes) is not systematically tested.
  • The evaluation suite samples only 10 prompts per VBench-Long dimension; statistical significance, variance across seeds, and sensitivity to prompt sampling are not reported.
  • Comparisons exclude several relevant training-free or memory-augmented baselines (e.g., MemFlow, VideoSSM, Deep/Hybrid Sink, Backwards Aggregation); a broader, head-to-head benchmarking is lacking.
  • Effects of training duration on Rollin’s gains are unexplored; how benefits scale when the base model is trained on 10s/60s/100s clips or minute-scale datasets is an open question.
  • No mechanism is provided to inject new semantics over time without destabilizing the cache; strategies for controlled content introduction (semantic gating/curriculum) are open.
  • Identity and structure consistency improvements are shown, but dedicated identity metrics (beyond VBench-Long proxies) and human studies for perceptual continuity over minutes are missing.
  • Compute and memory trade-offs are not deeply profiled; latency, throughput, and GPU memory overhead of sliding indices/semantics versus standard SF at different K/S are not reported.
  • Safety and content governance for ultra-long rollouts (e.g., preventing unwanted repetition, drift to unsafe content) are not discussed; monitoring and intervention hooks are an open area.
View Paper Prompt  View All Prompts 

Practical Applications

Immediate Applications

Below are practical use cases that can be deployed now using Rollin (Rolling Sink), a training-free cache-maintenance strategy that enables ultra-long, stable autoregressive (AR) video diffusion from short-horizon-trained models.

  • Stable single-shot long-take generation for film and media production (Media/Entertainment)
    • Tools/Workflows: Plug-in for existing AR video generators (e.g., Self Forcing-based pipelines) that exposes sink ratio S/K, sliding indices, and rolling semantics as user-tunable controls to produce minute-long continuous shots.
    • Assumptions/Dependencies: Single-prompt, single-shot content; bounded cache (e.g., K=6, S≈83%); relies on RoPE time indices and a short-horizon-trained AR model; no audio; resolution/framerate constraints from the base model.
  • Digital signage and brand loops (Advertising/Retail)
    • Tools/Workflows: Rollin-enabled “loop generator” to create long, stable background motion graphics for storefronts, event stages, and trade-show displays.
    • Assumptions/Dependencies: Static or slowly evolving scene semantics; quality tied to base model’s learned priors and VAE decoding.
  • VJing and live event visuals (Performing Arts/Creative Tech)
    • Tools/Workflows: Real-time AR video visuals engine with Rollin cache scheduling for LED walls and stage projections; parameters exposed via MIDI/OSC for live control.
    • Assumptions/Dependencies: Real-time inference on GPUs; consistent subject/structure stability prioritized over frequent semantic changes.
  • Streamer overlays and ambient backgrounds (Consumer Software/Content Creation)
    • Tools/Workflows: Desktop/mobile apps that generate long, coherent background videos for livestreams, video calls, and social media posts.
    • Assumptions/Dependencies: Single prompt; safe-content filters; on-device or cloud inference.
  • Previsualization and animatics for long shots (Film/Animation/Pre-production)
    • Tools/Workflows: Storyboard-to-long-take preview tool that reads a static shot description and produces multi-minute previsualizations with consistent IDs/colors/structures.
    • Assumptions/Dependencies: Scene and subject identities remain stable; directors can iterate via prompt refinement rather than multi-shot scripting.
  • Synthetic data generation for long-horizon computer vision tasks (Academia/Software)
    • Tools/Workflows: Dataset pipelines generating long sequences with stable identity, structure, and motion to benchmark tracking, segmentation, and temporal consistency algorithms.
    • Assumptions/Dependencies: Synthetic physics and realism limited by the base model; labels may require auxiliary annotation tools.
  • Long-horizon quality evaluation harness (Academia/Benchmarking)
    • Tools/Workflows: Adoption of VBench-Long with Rollin-enabled rollouts to stress-test AR models beyond training windows; standardize metrics for flicker, consistency, motion smoothness.
    • Assumptions/Dependencies: Access to evaluation models and fixed prompt suites; reproducible inference settings; compute availability.
  • Cost-effective long-form generation without retraining (Software/Cloud)
    • Tools/Workflows: “Training-free long-form mode” in existing video-gen APIs; cloud microservice that wraps AR models with Rollin cache logic to extend horizons.
    • Assumptions/Dependencies: Compatible base AR model; throughput depends on streaming sampler steps and bounded cache; GPU scheduling.
  • Stability-first mode in creative suites (Creative Software)
    • Tools/Workflows: Toggle in video-gen UIs to switch to Rollin for minutes-long shots; presets that manage sink size and rolling cadence for different content types.
    • Assumptions/Dependencies: Single-shot workflows; integration with existing DiT-based backends.
  • Policy-aligned watermarking and provenance for long synthetic videos (Policy/Platform Trust & Safety)
    • Tools/Workflows: Pipeline hooks to embed persistent watermarks and metadata across extended outputs; moderation systems tuned for long-horizon artifacts (repetition, identity drift).
    • Assumptions/Dependencies: Platform cooperation; watermark survives compression; content provenance standards.

Long-Term Applications

The following use cases require further research, scaling, or development, including multi-shot prompting, semantic transition design, audio, and higher-fidelity/physics-aware generation.

  • Multi-shot, script-level long video generation with coherent transitions (Media/Entertainment)
    • Tools/Workflows: Shot scheduler that extends Rollin principles to multi-shot prompts; semantic memory across scenes; transition-aware cache management.
    • Assumptions/Dependencies: New cache policies to handle changing semantics; continuity tools (character/location tracking); multi-prompt conditioning.
  • Virtual production integration for LED volumes and camera tracking (Film/Production Tech)
    • Tools/Workflows: Rollin-powered generative backgrounds that remain consistent over long takes; integration with camera metadata and lighting pipelines.
    • Assumptions/Dependencies: High-fidelity rendering, real-time constraints, physics-aware motion; safety rails for artifact detection.
  • Real-time interactive long video with dynamic user inputs (XR/Interactive Media)
    • Tools/Workflows: Multi-prompt Rollin that can update semantics mid-rollout without destabilizing identity/motion; control via gestures or controllers.
    • Assumptions/Dependencies: Robust cache update strategies; latency-optimized inference; user-guided semantic management.
  • Extended world simulators for planning and training (Robotics/Autonomy/AI Research)
    • Tools/Workflows: Long-horizon simulated visuals for pretraining temporal reasoning, tracking, and planning models; stress-testing autonomy pipelines.
    • Assumptions/Dependencies: Physics consistency and domain realism; coupling with dynamics engines; validation against real-world data.
  • Generative cinematography assistants (Media Tools)
    • Tools/Workflows: Tools that suggest camera moves and scene evolution while maintaining long-horizon coherence; AI “long-take director” assistants.
    • Assumptions/Dependencies: Higher-level narrative modeling; editable caches; integration with editing timelines.
  • Long-form educational content generation (Education)
    • Tools/Workflows: Generation of continuous demonstrations (e.g., lab simulations, process animations) with stable concepts over minutes/hours.
    • Assumptions/Dependencies: Domain accuracy; expert-in-the-loop validation; content safety.
  • Compliance frameworks for extended synthetic media (Policy/Standards)
    • Tools/Workflows: Standards for watermark persistence across long sequences; provenance tracking; disclosure policies for generative broadcast content.
    • Assumptions/Dependencies: Cross-platform adoption; robust detection/watermarking; regulatory alignment.
  • On-device long video generation for AR glasses and mobile (Hardware/Edge)
    • Tools/Workflows: Lightweight Rollin variants optimized for mobile NPUs/GPUs; energy-aware cache scheduling.
    • Assumptions/Dependencies: Model compression; thermal/power budgets; UI constraints.
  • Cross-modal long-horizon generation (Audio+Video, Narration) (Creative Tech)
    • Tools/Workflows: Unified cache policies across video and audio models to maintain temporal coherence in soundtrack and visuals.
    • Assumptions/Dependencies: Multi-modal AR architectures; synchronization strategies; evaluation metrics for cross-modal consistency.
  • Production-scale “Rollin-as-a-Service” and memory scheduling libraries (Software/Platforms)
    • Tools/Workflows: Managed services that provide standardized cache management (sliding indices/semantics) across generative backends; SDKs for developers.
    • Assumptions/Dependencies: Vendor-neutral APIs; observability and QA hooks; autoscaling for long-form workloads.

Cross-cutting assumptions and dependencies

  • Works best for single-shot, single-prompt scenarios; multi-shot requires new cache policies to handle semantic transitions.
  • Depends on an AR video diffusion backbone (e.g., DiT-based models) with RoPE or compatible temporal indexing, a bounded cache, and short-horizon training (e.g., ~5s clips).
  • Quality and realism are constrained by the base model’s training data, resolution, framerate, and VAE decoding.
  • Content safety, copyright, and provenance must be addressed for commercial deployment (e.g., watermarking, moderation, dataset licensing).
  • Compute considerations: real-time or near-real-time inference requires capable GPUs; streaming efficiency hinges on few-step samplers and cache size management.
View Paper Prompt  View All Prompts 

Glossary

  • AR cache: The conditioning memory of previously generated frames/blocks that the autoregressive model uses to predict the next block. "we conduct a systematic analysis of AR cache maintenance."
  • AR drift: Progressive degradation and instability during long autoregressive generation caused by accumulated errors or distribution mismatch. "Such AR drift is commonly attributed to error accumulation."
  • Attention Sink: A technique that pins a static prefix of early tokens/blocks in the attention cache to stabilize generation over long horizons. "We first apply Attention Sink (i.e., pinning the first SS blocks as sink blocks where both the time indices and semantics are static), and analyze the effect of different sink ratios (SK\frac{S}{K})."
  • Autoregressive (AR) video diffusion models: Video generative models that predict frames sequentially by conditioning on previously generated frames within a diffusion process. "Recently, autoregressive (AR) video diffusion models has achieved remarkable performance."
  • Bidirectional attentions: Attention mechanisms that allow information flow in both directions across time or tokens, often used in non-autoregressive diffusion transformers. "they usually rely on bidirectional attentions in DiTs~\cite{peebles2023scalable} and denoise all frames simultaneously, making them incompatible with such ``open-ended'' setting."
  • Denoising diffusion model: A generative model that iteratively removes noise from an input across a schedule of timesteps to produce clean samples. "each AR generation step (i.e., conditional distribution) is modeled by a denoising diffusion model GθG_\theta~\cite{lipman2022flow,liu2022flow}."
  • Denoising timesteps: The discrete time indices in a diffusion process that control noise levels during iterative denoising. "We term the set of denoising timesteps as {t0,t1,…,tT}\{t_0, t_1, \dots, t_T\}, where t0=0t_0 = 0 and tT=1000t_T = 1000."
  • Diffusion forcing (DF): A training strategy where the model conditions on ground-truth frames that have been corrupted with noise at randomly sampled levels. "During training, major techniques for building the AR cache are teacher forcing (TF)~\cite{gao2024ca2,hu2024acdit,jin2024pyramidal,zhang2025test}, diffusion forcing (DF)~\cite{chen2024diffusion,yin2025slow,chen2025skyreels,gu2025long,teng2025magi,song2025history,po2025bagger}, and self forcing (SF)~\cite{huang2025self,yang2025longlive,cui2025self,hong2025relic,lu2025reward,yin2024improved,yin2024one,yi2025deep}."
  • DiTs (Diffusion Transformers): Transformer-based architectures tailored for diffusion modeling in vision/video, often employing attention over tokens/frames. "they usually rely on bidirectional attentions in DiTs~\cite{peebles2023scalable} and denoise all frames simultaneously"
  • Exposure bias: The mismatch between training and testing conditions in autoregressive models that causes errors to compound at test time. "we further interpret it through the lens of ``exposure bias''~\cite{bengio2015scheduled,ranzato2015sequence,lamb2016professor,zhang2019bridging,schmidt2019generalization,li2023alleviating,ning2023elucidating}"
  • Latent frames: Compressed representations of video frames in a learned latent space used by the model prior to decoding to pixel space. "Specifically, it's trained on sequences of 21 latent frames (corresponding to 81 frames after VAE~\cite{wan2025} decoding)."
  • Long-horizon drift: Visual degradation and inconsistency that emerges when generating far beyond the training duration. "We characterize the long-horizon drift in AR video diffusion as the exposure bias from a train-test horizon mismatch"
  • LoRA: Low-Rank Adaptation; a parameter-efficient fine-tuning technique used to adapt large models with minimal additional parameters. "LongLive's LoRA weights (further trained on 1 minute videos) isn't loaded (i.e., w/o LoRA)."
  • Open-ended testing: Test-time generation where the model continues producing content for arbitrarily long horizons without a predefined end. "Since open-ended testing can extend beyond any finite training window, and long-video training is computationally expensive, we pursue a training-free solution to bridge this gap."
  • Prompt embedding: The vector representation of the input text prompt that conditions the generative model throughout synthesis. "Since the prompt embedding stays fixed throughout AR video synthesis"
  • RoPE (Rotary positional embeddings): A positional encoding technique that encodes relative positions by rotating query/key vectors in attention. "The time indices are embedded using rotary positional embeddings (RoPE)~\cite{su2024roformer}."
  • Rollin: The proposed training-free method that maintains cache behavior via sliding indices and rolling semantics to mitigate long-horizon drift. "we propose Rollin, a training-free approach for bridging the gap between limited-horizon training and open-ended testing."
  • Rolling Sink: The specific instantiation of Rollin that uses a large attention sink and rolling of within-duration content to stabilize ultra-long generation. "Built on Self Forcing (trained on only 5s clips), Rolling Sink effectively scales the AR video synthesis to ultra-long durations (e.g., 5-30 minutes at 16 FPS) at test time"
  • Self Forcing (SF): A training paradigm where the model conditions on its own generated frames rather than ground-truth, reducing train-test cache mismatch. "During training, major techniques for building the AR cache are teacher forcing (TF)~..., diffusion forcing (DF)~..., and self forcing (SF)~..."
  • Sliding Indices: Assigning sink blocks time indices via a fixed-length sliding window aligned with the current step to better match temporal behavior. "Sliding Indices: Treating the time indices as a global axis i∈[0,∞)i\in[0,\infty), at each AR step ii, we shift sink blocks' time indices as a fixed-length (i.e., SS) sliding window on this axis."
  • Sliding Semantics: Updating the semantic content cached in sink blocks by rolling segments from within-duration history to approximate endless, drift-free evolution. "Sliding Semantics: Ideally, the sink blocks' semantic content should also slide along the a drift-free, global video manifold that lasts endlessly."
  • Streaming efficiency: The ability to generate long sequences continuously under bounded memory and compute by limiting cache size and steps. "the total cache capacity KK is strictly bounded for streaming efficiency."
  • Teacher forcing (TF): A training strategy where the model is conditioned on clean ground-truth previous frames during autoregressive learning. "During training, major techniques for building the AR cache are teacher forcing (TF)~\cite{gao2024ca2,hu2024acdit,jin2024pyramidal,zhang2025test}"
  • Train-test gap: The performance mismatch arising when test-time conditions (e.g., longer horizons) differ from those seen during training. "a train-test gap emerges when testing at longer horizons, leading to rapid visual degradations."
  • umT5: A multilingual variant of T5 used to encode prompts for conditioning video generation. "the global prompt embedding (encoded by umT5~\cite{chung2023unimax}) is constant during AR video synthesis"
  • VAE (Variational Autoencoder): A generative model that encodes frames into a latent space and decodes them back, used here to map between latents and pixels. "corresponding to 81 frames after VAE~\cite{wan2025} decoding"
  • VBench-Long: A long-video evaluation benchmark that measures quality across multiple diagnostic dimensions. "quantitative evaluations using VBench-Long \cite{huang2023vbench,huang2025vbench++,zheng2025vbench2} on both 1-minute (Tab.~\ref{tab:1min}) and 5-minute (Tab.~\ref{tab:5min}) AR video synthesis"
  • Video manifold: The conceptual high-dimensional manifold of valid video trajectories whose local slices represent temporally coherent content. "the sink blocks' semantic content should also slide along the a drift-free, global video manifold that lasts endlessly."
View Paper Prompt  View All Prompts 

Open Problems

We found no open problems mentioned in this paper.

Collections

Sign up for free to add this paper to one or more collections.

Sign Up

GitHub

  1. Rolling Sink