PaperlayerView exact revision on arXiv

Source-linked AI summary

Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos

X. Feng, J. Zhu, M. Wu, C. Chen, F. Mao, H. Guo, J. Wu, X. Chu, K. Huang

arXiv:2605.18233v1cs.CV

TL;DR

Train-free frame-level autoregressive methods face a training-inference gap and long-term consistency challenge in generating long videos. MIGA combines two-stage alignment and dual consistency enhancement, achieving state-of-the-art results on VBench and NarrLV.

  • Problem

    Train-free frame-level autoregressive generation faces a training-inference gap and lacks explicit long-range frame dependency modeling, leaving long-video quality suboptimal.

    Source: §1 · p. 2

  • Method

    MIGA combines two-stage noise-span alignment with self-reflection and long-range frame guidance to mitigate the training-inference gap and improve long-term consistency.

    Source: §1 · p. 2

  • Results

    MIGA achieves state-of-the-art results on VBench and NarrLV compared with existing train-free methods.

    Source: §5 · p. 9

  • Takeaways & Limitations

    MIGA provides a train-free approach for extending the generated length of existing foundation models.

    Source: C. Limitations and Future Work · p. 26

    Longer video duration increases the risk of unintended model behaviors, including a generated cat’s head and tail suddenly switching places.

    Source: C. Limitations and Future Work · p. 26

Abstract

from arXiv · show

Without incurring significant computational overhead, train-free long video generation aims to enable foundation video generation models to produce longer videos. Frame-level autoregressive frameworks, e.g., FIFO-diffusion, offer the advantage of generating infinitely long videos with constant memory consumption. However, the mismatch between training and inference, coupled with the challenge of maintaining long-term consistency, limits the effective utilization of foundation models. To mitigate these concerns, we propose \textbf{MIGA}, a novel infinite-frame long video generation method. Firstly, we propose an effective two-stage alignment mechanism that mitigates the training-inference gap by reducing the excessive noise span fed to the model. We then introduce an innovative dual consistency enhancement mechanism, where the self-reflection approach corrects early high-noise frames and the long-range frame guidance approach leverages later low-noise frames with broad coverage to steer generation, jointly improving temporal consistency. Extensive experiments on VBench and NarrLV demonstrate the state-of-the-art performance of MIGA. Our project page is available at https://xiaokunfeng.github.io/miga_homepage/.

1. Introduction

MIGA is a training-free method for infinite-frame video generation that addresses the training–inference gap and long-term consistency. It combines two-stage noise-span alignment with self-reflection and long-range frame guidance.

Introduction

  • Training-free extension of foundation video models offers a practical alternative to building long-video models from scratch, while prior window-based methods can require memory proportional to video length.FIFO-Diffusion instead supports infinite-frame generation with fixed memory consumption.Source: §1 · p. 1
  • Frame-level autoregressive models face a training–inference mismatch that can cause content drift and visual artifacts, while FIFO-Diffusion lacks explicit long-range frame dependencies (Kim et al., 2024).FIFO-Diffusion facilitates feature interaction between neighboring chunks through lookahead denoising.Source: §1 · p. 2
  • MIGA narrows the training–inference noise-span gap with a zigzag latent queue followed by unified denoising once all latents reach the same noise level.The unified denoising stage matches the training-phase noise span.Source: §1 · p. 2
  • MIGA improves VBench subject consistency by 4.7% and background consistency by 2.0% over FIFO-Diffusion, and achieves state-of-the-art performance on VBench and NarrLV.On NarrLV, MIGA also demonstrates exceptional capability in generating rich narrative content.Source: §1 · pp. 2–3
  • MIGA promotes long-term consistency by using self-similarity to correct early high-noise latents and incorporating later low-noise latents as long-range frame guidance during denoising.The self-reflection approach does not rely on external evaluators.Source: §1 · p. 2

2. Related Works

Video generation has advanced from short-video models toward longer outputs, but many long-video approaches require costly specialized training. Train-free methods extend foundation generators more efficiently, and FIFO-Diffusion supports infinite frames with fixed memory; MIGA builds on it to address training-inference gaps and consistency limitations.

Text-to-Video Generation

  • Video generation moved from 2D spatial/1D temporal designs toward 3D full-attention and foundation models, while diffusion models generally remain trained on short, fixed-length videos.Examples include VideoCrafter and Stable Video Diffusion, followed by Video Diffusion Models and CogVideoX; HunyuanVideo and Wan are cited as newer foundation models.Source: §2 · p. 3

Long Video Generation

  • Train-free long-video methods avoid large-scale specialized training, using sliding windows, noise or feature integration, and temporal-position refinement to extend outputs.Gen-L-Video merges overlapping subsequences; FreeNoise, FreeLong, and FreePCA integrate local and global features; RIFLEx reduces periodic repetition.Source: §2 · p. 3
  • FIFO-Diffusion enables infinite-frame generation with fixed memory through frame-level autoregression, and MIGA builds on it to address training-inference gaps and consistency limitations.FIFO-Diffusion uses a noise-space design for frame-level autoregressive generation.Source: §2 · p. 3

3. Methods

MIGA generates long videos through frame-level autoregression, reduces the training–inference noise mismatch with two-stage alignment, and enhances long-term consistency with self-reflection and long-range frame guidance.

3.1. Preliminaries: Train-Free Frame-Level Autoregressive Generation.

  • Frame-level autoregressive generation maintains a queue of latents at progressively increasing noise levels, repeatedly denoises the queue, and shifts its first latent to produce the next frame.FIFO-Diffusion uses sliding-window inference when the queue is longer than the foundation model's supported frame count.Source: §3.1 · pp. 3–4

3.2. Two-Stage Training-Inference Alignment (TTA).

  • TTA slows noise changes across queued latents in Stage 1, then denoises them at a unified noise level in Stage 2 to better align inference inputs with training.Stage 1 processes latents in groups of Lzig and only partially denoises them; Stage 2 uses sliding-window denoising to complete the frames without memory growing with video length.Source: §3.2 · pp. 4–5

3.3. Dual Consistency Enhancement (DCE).

  • DCE combines tail-focused self-reflection to evaluate and correct new latents with head-focused guidance from long-range, low-noise latents during local denoising.The two mechanisms target complementary parts of the maintained queue to support long-term consistency.Source: §3.3 · p. 5
  • Self-reflection uses latent cosine similarity as a consistency metric and exploits its correlation between early high-noise and final clean latents to avoid extra denoising for evaluation.The metric can identify consistency disruptions, and its fluctuation patterns remain correlated with those of clean latents even at high noise intensity.Source: §3.3 · pp. 5–6
  • At the queue tail, self-reflection compares adjacent latent chunks and triggers expanded search when their consistency-score decrease exceeds δadju, using preceding evaluated latents as guidance for correction.The method evaluates early high-noise latents and searches from randomly sampled Gaussian starting points, correcting the latents after the judgment position.Source: §3.3 · p. 6
  • Long-range frame guidance sparsely samples earlier, relatively clean latents and adds them to the local-window input to guide denoising of current frames.Guidance is not applied to the earliest queue positions when there are too few preceding latents; those latents have already received propagated guidance.Source: §3.3 · p. 7

4. Experiments

Across VBench and NarrLV, MIGA improves long-video quality and narrative expressiveness on VideoCrafter2 and Wan2.1. Ablations support the complementary roles of its two-stage alignment and dual consistency mechanisms.

4.1. Implementation Details.

  • Experiments apply MIGA to VideoCrafter2 and Wan2.1 and evaluate video quality on VBench-Long and narrative expressiveness on NarrLV.VBench-Long metrics include subject and background consistency, motion smoothness, temporal flicker, and their mean Overall Score; NarrLV reports scene attributes, target attributes, and target actions across TNA counts 2, 3, and 4.Source: §4.1 · p. 7

4.2. Comparison with Baselines.

  • MIGA achieves state-of-the-art VBench performance across all metrics on both foundation models, improving consistency over FreeLong and all metrics over FIFO-Diffusion.VideoCrafter2 comparisons use 128-frame videos; Wan2.1 comparisons use 161-frame videos.Source: Table 1 · p. 7, §4.2 · p. 7
  • MIGA adds narrative-expressiveness improvements over FIFO-Diffusion on VideoCrafter2 methods and shows notable narrative advantages with Wan2.1.The NarrLV evaluation uses changing prompts for latents at different temporal positions; prompts span TNA counts of 2, 3, and 4 and assess scene attributes, target attributes, and target actions.Source: §4.1 · p. 7, Table 2 · p. 8, §4.2 · p. 8
  • Three Wan2.1-based MIGA videos of 1k frames closely follow their text prompts while maintaining strong subject and background consistency.Additional visualizations are provided in the appendix.Source: §4.2 · p. 8

4.3. Ablation Study.

  • TTA combines zigzag iterative denoising with unified-noise-level denoising; the first stage reduces baseline anomalies, while the second suppresses video noise.As the first-stage zigzag width increases and the model processes frames across a smaller noise span, Overall Score improves and then saturates; using only stage two sharply reduces performance because it omits autoregressive information transfer.Source: §4.3 · pp. 8–9, Figure 6 · p. 8, Table 6 · p. 9
  • On VBench, TTA and DCE individually improve Overall Score by 2.03% and 1.73%, respectively, and together provide complementary gains.FIFO-Diffusion is the ablation baseline without the proposed mechanisms.Source: §4.3 · p. 8, Table 3 · p. 9
  • DCE combines self-reflection and long-range frame guidance; smaller adjustment thresholds broaden searches and increase corrections, while gains eventually converge.Further lowering the threshold increases evaluated steps without increasing corrected steps, lowering Rsucc; the default δadju is 0.01 to balance performance and computational cost, and the tested guidance-frame count performs best at mguid = 6.Source: Figure 5 · p. 8, §4.3 · p. 9, Table 5 · p. 9

5. Conclusion · Appendix · A. Further Details on Our Method

MIGA builds on train-free frame-level autoregressive generation to address the training-inference gap and long-term consistency in infinite-video generation. Its two-stage alignment and dual consistency mechanisms support state-of-the-art results on VBench and NarrLV among existing train-free methods.

5. Conclusion

  • MIGA achieves state-of-the-art results on VBench and NarrLV compared to existing train-free methods.The conclusion describes its generated long videos as having strong temporal consistency and rich narratives.Source: §5 · p. 9
  • FIFO-Diffusion and similar train-free frame-level autoregressive frameworks can generate infinitely long videos with constant memory cost.MIGA builds on these capabilities while addressing their training-inference gap and long-term consistency limitations.Source: §5 · p. 9
  • MIGA introduces a two-stage training-inference alignment mechanism that mitigates the gap by optimizing the noise span.Source: §5 · p. 9
  • MIGA proposes a dual consistency enhancement mechanism to improve long-term consistency through self-reflection and long-range frame guidance.Source: §5 · p. 9
  • The conclusion characterizes MIGA's long videos as having strong temporal consistency.Source: §5 · p. 9
  • The conclusion also characterizes MIGA's long videos as having rich narratives.Source: §5 · p. 9

A.1. Pseudocode Implementation · A.1.1. PRELIMINARIES: TRAIN-FREE FRAME-LEVEL AUTOREGRESSIVE GENERATION.

The pseudocode section details the procedures for train-free frame-level autoregressive generation, Two-Stage Training-Inference Alignment, and Dual Consistency Enhancement. Its preliminary subsection uses FIFO-Diffusion to explain noise-queue initialization and inference.

A.1. Pseudocode Implementation

  • The pseudocode covers frame-level autoregressive generation, Two-Stage Training-Inference Alignment (TTA), and Dual Consistency Enhancement (DCE).The section introduces FIFO-Diffusion as an example, then points to the TTA and DCE mechanisms described in Secs. 3.2 and 3.3.Source: §A.1 · p. 13

A.1.1. PRELIMINARIES: TRAIN-FREE FRAME-LEVEL AUTOREGRESSIVE GENERATION.

  • FIFO-Diffusion (Kim et al., 2024) illustrates a noise queue Q whose latents have progressively increasing noise levels.The queue length is the denoising step count T.Source: §A.1.1 · p. 13
  • The queue is initialized with clean initial latents in its last f0 positions, and the sampler adds noise according to scheduler-specified strengths (Ho et al., 2020; Lipman et al., 2022).The queue is longer than the initial latent count, so those clean latents fill only its last f0 positions.Source: §A.1.1 · p. 13
  • The remaining (T −f) latents are initialized by injecting varying noise levels into the first latent.Source: §A.1.1 · p. 13
  • At each inference step, all queue latents lose one noise level, the resulting clean first latent is dequeued and saved, and Gaussian noise is enqueued.This repeats to continue frame generation.Source: §A.1.1 · p. 13
  • Because the model processes only f0 latents while the queue contains T latents, FIFO-Diffusion predicts noise through sliding windows of size f0 and stride lstride.One sampler step over the queue therefore requires multiple foundation-model noise predictions.Source: §A.1.1 · p. 13

A.1.2. TWO-STAGE TRAINING-INFERENCE ALIGNMENT MECHANISM. · A.1.3. DUAL CONSISTENCY ENHANCEMENT MECHANISM.

MIGA’s two-stage training-inference alignment reduces the inference noise span, while its dual consistency enhancement combines self-reflection and long-range frame guidance to improve consistency. The sections detail how the latent queue is initialized and denoised, and where each consistency approach operates.

A.1.2. TWO-STAGE TRAINING-INFERENCE ALIGNMENT MECHANISM.

  • TTA mitigates the training-inference gap by reducing the noise span of latents fed to the model during inference.Source: §A.1.2 · p. 15
  • Initialization assigns adjacent Lzig latents the same noise level and progressively guides noisy latents from clean queue-tail latents instead of duplicating semantically identical frames.Source: §A.1.2 · p. 15
  • TTA first dequeues partially denoised latents while enqueuing Gaussian noise latents, then applies remaining denoising steps at identical noise levels.Stage 1 proceeds until N latents have been dequeued; Stage 2 denoises those latents.Source: §A.1.2 · p. 15

A.1.3. DUAL CONSISTENCY ENHANCEMENT MECHANISM.

  • DCE comprises Long-Range Frame Guidance and Self-Reflection, with guidance processing sparse long-range and dense local latents together during sliding local segments.Source: §A.1.3 · p. 15, Figure A1 · p. 17
  • Self-Reflection evaluates early high-noise latents for consistency anomalies and corrects them before each queue inference; expanded tail-latent search can address detected anomalies.The figure illustrates an anomaly in which a car changes from black to white.Source: §A.1.3 · p. 17, Figure A1 · p. 17
  • Long-Range Frame Guidance targets low-noise latents at the queue head to facilitate interactions among distant latents and is integrated into each queue inference.Source: §A.1.3 · p. 17

A.2. Analysis of Framework Unification · A.3. Multi-prompt Conditional Generation

MIGA’s TTA and DCE mechanisms are interdependent parts of one frame-level autoregressive framework: TTA operates at the queue level, while DCE guides denoising within each queue. The framework also supports multi-prompt control by assigning different text conditions to latents at different temporal positions.

A.2. Analysis of Framework Unification

  • The autoregressive process maintains a queue whose frame noise increases along time, reduces all queued noise levels at each step, and pops the clean frame from the head.The procedure repeats iteratively to generate frames autoregressively.Source: §A.2 · p. 17
  • Because foundation models process only a limited number of frames at once, a sliding window sequentially refines the queue’s noisy frames during each inference operation.Source: §A.2 · p. 17
  • TTA divides queue-level inference into two stages with different noise distributions, reducing the noise span fed to the model and transferring foundation models’ short-term generative capabilities to long-video generation.The passage characterizes this as mitigating the training-inference gap.Source: §A.2 · pp. 17, 20
  • DCE enhances consistency by guiding denoising with low-noise frames over long ranges and timely self-reflection correction of early high-noise frames.These operations focus on denoising within each queue iteration.Source: §A.2 · p. 20
  • TTA and DCE form a unified framework, with TTA operating along queue-level inference and DCE along frame-level denoising to support consistent, high-quality long-video generation.The paper describes the mechanisms as interdependent rather than independent modules.Source: §A.2 · pp. 17, 20

A.3. Multi-prompt Conditional Generation

  • The frame-level autoregressive framework enables multi-text control by applying different text conditions to latents at different temporal positions during queue inference.With one text condition, the condition remains constant; with multiple prompts, the sliding-window process guides latents at different positions with different conditions.Source: §A.3 · p. 20
  • With nprom prompts, each prompt sequentially controls Nprom frames, producing N = npromNprom clean latents for the final video; a Wan2.1-based MIGA example illustrates three prompts.The input text condition is specified for the model’s current queue position and the number of dequeued clean latents.Source: §A.3 · p. 20, Figure A2 · p. 20

A.4. Implementation on Different Foundation Models · B. Further Details on Experimental Analysis

MIGA adapts frame-level autoregressive generation to VideoCrafter2 and Wan2.1 by modifying their noise-prediction and denoising processes, but the framework is difficult to apply to some MMDiT models.

A.4. Implementation on Different Foundation Models

  • MIGA equips VideoCrafter2 with frame-level autoregressive generation by modifying its noise-prediction model and sampler.VideoCrafter2 is described as an early, representative text-to-video foundation model used by existing train-free long-video methods.Source: §A.4 · pp. 20–21
  • At each inference step, VideoCrafter2’s noise-prediction model receives latents with different noise levels determined by frame index.Each frame’s latents interact with their respective noise-level conditions.Source: §A.4 · p. 21
  • VideoCrafter2’s sampler denoises each frame’s latents separately, conditioning on their respective noise levels.Source: §A.4 · p. 21
  • For Wan2.1, MIGA modifies the noise-prediction model and sampler, and its default UniPC sampler requires both to retain and use information from previous steps.Source: §A.4 · p. 21
  • MIGA can migrate to VideoCrafter2 and Wan2.1 because cross-attention lets latents use frame-specific noise timesteps while treating all latents together for text conditioning.Source: §A.4 · p. 21
  • The framework is difficult to apply to some MMDiT models because concatenated text and video features jointly interact with noise timesteps that cannot be distinguished at the frame level.Source: §A.4 · p. 21
  • Injecting an intermediate timestep into CogVideoX-5B text features to enable migration produced abnormal video outputs.Figure A3 illustrates this bad case.Source: §A.4 · p. 21, Figure A3 · p. 21

B.1. Implementation Details of Ablation Studies · B.2. Qualitative Comparison between VideoCrafter2-based MIGA and Wan2.1-based MIGA

The appendix specifies the settings and conditions used to ablate TTA and DCE, and compares how VideoCrafter2-based and Wan2.1-based MIGA differ in video style and evaluation strengths.

B.1. Implementation Details of Ablation Studies

  • For VideoCrafter2-based VBench ablations, FIFO-Diffusion is the baseline; TTA-only adds two denoising stages, DCE-only adds self-reflection and long-range guidance, and the final method combines both.The TTA setting uses Lzig = 4 and τe = 10; DCE uses δadju = 0.01 and mguid = 4.Source: §B.1 · p. 21
  • The TTA ablation adds Stage 1 zigzag denoising with Lzig = 4 to FIFO-Diffusion, then adds Stage 2 unified-noise-level denoising with τe = 10.The baseline matches the setting used in the core-mechanism ablation.Source: §B.1 · p. 21
  • TTA hyperparameter studies use a random 50% prompt subset; running Stage 2 alone for all 64 steps removes autoregressive generation, unlike configurations retaining Stage 1.Stage 1’s iterative generation connects latents so cleaner latents can guide subsequent high-noise latents; synchronous denoising of independent latents weakens their correlation and consistency.Source: §B.1 · p. 22
  • Stage 1 builds latent connections through autoregressive early denoising, while Stage 2 completes denoising at the training-matched noise span of 1 and suppresses visual artifacts.The paper attributes semantic and spatial consistency to Stage 1 and improved visual details to Stage 2.Source: §B.1 · p. 22
  • DCE ablations vary the self-reflection threshold from a no-search baseline of δadju = 0.07 and progressively increase the number of long-range guiding frames mguid.The threshold study tracks overall score and provides individual-metric results separately; the guidance-frame study uses the same baseline.Source: §B.1 · pp. 22–23

B.2. Qualitative Comparison between VideoCrafter2-based MIGA and Wan2.1-based MIGA

  • VideoCrafter2-based MIGA scores higher on subject and background consistency, while Wan2.1-based MIGA performs better on motion smoothness and temporal flicker and shows stronger narrative generation on NarrLV.The comparison attributes the consistency difference partly to VideoCrafter2’s animation-like outputs versus Wan2.1’s realistic videos with richer texture details.Source: Figure A4 · p. 23, §B.2 · p. 23

B.3. Computational Efficiency Analysis · B.4. Comparison with Training-Based Methods · B.5. Human Evaluation Experiments

MIGA is reported as comparable to training-based methods and stronger than FIFO-Diffusion in human comparisons, while its efficiency analysis describes the costs and quality gains of its alignment and consistency mechanisms.

B.3. Computational Efficiency Analysis

  • With TTA alone, MIGA matches FIFO-Diffusion’s computational cost because both use T denoising steps per latent, while its overall score increases by 2.03%.Source: §B.3 · p. 23
  • Stage 2 processes a queue that grows with video length, but sliding-window denoising keeps each foundation-model window constant-sized, so memory overhead does not rise with frame count.Measured peak memory rises moderately with generated length because intermediate variables are stored, while the additional overhead remains minimal and acceptable.Source: §B.3 · p. 24, Table A3 · p. 24
  • Long-range frame guidance adds only a small number of noise-prediction steps per queue inference, making its extra computational overhead negligible despite considerable performance gains.The added steps are determined by niter = ⌈(T − f0 + mguid)/lstride⌉ + 1, compared with ⌈(T − f0)/lstride⌉ + 1 without guidance.Source: §B.3 · p. 24
  • Self-reflection uses extended sampling as test-time scaling, and its degree is controlled by the adjustment threshold δ_adju and sample count n_samp; performance improves as computational cost increases.Each noise prediction is a computational unit, and each triggered sample requires ⌈(L − f_judg + 1)/L_zig⌉ additional noise-prediction steps.Source: §B.3 · p. 24
  • Without DCE, MIGA adds only +0.05 s per frame over FIFO-Diffusion while improving VBench Overall Score by +1.73 O.S.; DCE further improves performance at higher computational cost.Efficiency is measured as average time to generate a single frame.Source: Table A4 · p. 24, §B.3 · p. 24

B.4. Comparison with Training-Based Methods

  • Despite no large-scale training, MIGA achieves comparable performance across all metrics against train-based long-video methods under the same evaluation settings.The compared methods use frame-level autoregressive inference but train-specific strategies and architectures.Source: Table A5 · p. 25, §B.4 · p. 25

B.5. Human Evaluation Experiments

  • Across 48 prompt-generated video pairs judged by eight annotators, MIGA outperforms FIFO-Diffusion on subject and background consistency, motion smoothness, and temporal flicker.Annotators made pairwise comparisons on each dimension and could select a tie.Source: §B.5 · p. 25

B.6. More Qualitative Results

MIGA produces approximately one-minute videos with both Wan2.1-1.3B and VideoCrafter2. The appendix shows additional generation cases from both models.

B.6. More Qualitative Results

  • Wan2.1-based MIGA generates 1,000-frame videos, while VideoCrafter2-based MIGA generates 600-frame videos, each approximately one minute long.These counts follow the models’ default fps settings.Source: §B.6 · p. 25
  • Additional MIGA generation cases using Wan2.1-1.3B and VideoCrafter2 appear in Figures A6 and A7.Figure 1 also presents three long-video cases generated with Wan2.1-1.3B.Source: §B.6 · p. 25

C. Limitations and Future Work

MIGA extends foundation models to longer videos, but longer duration can expose unintended behaviors such as hallucinations or failures of physical knowledge; future work will explore conditioning beyond text.

C. Limitations and Future Work

Related papers

Loading 2605.18233v1…