TL;DR
MIGA proposes train-free infinite-frame video generation with two-stage training-inference alignment and dual consistency enhancement. Achieving 4.7% and 2.0% gains in subject and background consistency on VBench, it delivers state-of-the-art results while maintaining fixed memory.
MIGA proposes train-free infinite-frame video generation with two-stage training-inference alignment and dual consistency enhancement. Achieving 4.7% and 2.0% gains in subject and background consistency on VBench, it delivers state-of-the-art results while maintaining fixed memory.
论文原始摘要
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/.
Paper Collector 中文速览
MIGA实现无训练长视频一致性生成
方法概述
MIGA采用两阶段对齐减少模型输入噪声跨度，解决训练-推理不匹配；通过自反思修正早期高噪帧，利用后期低噪帧的长程帧引导，双重机制协同提升长视频生成的一致性。
核心贡献
提出两阶段对齐机制减少训练-推理差距，引入双重一致性增强机制（自反思+长程帧引导）提升长视频时序一致性。
原始来源与核验范围
核验仅覆盖论文正文中的主张、证据片段和参考文献关系。NGJOO 未独立运行作者代码、重做实验或验证真实部署效果；页面中的可复现性分数是文档完整度评估，不是复现实验结果。
49
已证实
6
证据不足
3
无法验证
N/A
可复现性
核心问题
How can train-free methods generate infinitely long, coherent videos with fixed memory while addressing the training-inference gap and maintaining long-term consistency?
核心方法
MIGA builds on FIFO-Diffusion's frame-level autoregressive framework with two mechanisms: Two-Stage Training-Inference Alignment (TTA) uses zigzag iterative denoising followed by unified noise level denoising to reduce noise span discrepancy. Dual Consistency Enhancement (DCE) combines self-reflection for high-noise latents and long-range frame guidance for low-noise latents to maintain temporal consistency.
方法组件
- Diffusion-based video generation models consist of a conditional encoder, VAE, and noise prediction network for denoising Gaussian noise to recover video latents.
- Frame-level autoregressive generation maintains a queue of T latents with progressively increasing noise levels.
- Each inference step denoises all latents in the queue, with the first latent becoming fully denoised and being dequeued while a new Gaussian latent is appended.
- FIFO-Diffusion realizes a diagonal denoising paradigm through sliding window processing with fixed window size and stride.
- The queue length T is typically greater than the number of frames f0 the model can process, requiring multiple executions of the standard sampler.
论点验证
已证实 (85%)
we propose MIGA, a novel train-free method for infinite-frame video generation
已证实 (85%)
we propose an intuitive and effective two-stage training-inference alignment mechanism to mitigate the inherent training-inference gap in existing train-free autoregressive frameworks
The two-stage TTA mechanism is fully specified in Section 3.2 with mathematical formulations (Eq. 3-4) and algorithms. Its effectiveness is quantitatively demonstrated through ablation studies showing +2.03% overall score improvement.
The two-stage TTA mechanism is fully specified in Section 3.2 with mathematical formulations (Eq. 3-4) and algorithms. Its effectiveness is quantitatively demonstrated through ablation studies showing +2.03% overall score improvement.
已证实 (85%)
we present an innovative dual consistency enhancement mechanism to promote long-term consistency
The DCE mechanism is fully specified in Section 3.3 with self-reflection and long-range frame guidance components. Effectiveness is demonstrated through ablation showing +1.73% overall score improvement.
The DCE mechanism is fully specified in Section 3.3 with self-reflection and long-range frame guidance components. Effectiveness is demonstrated through ablation showing +1.73% overall score improvement.
已证实 (80%)
To inherit the merits of train-free frame-level autoregressive frameworks while alleviating their limitations in training-inference gap and long-term consistency modeling, we propose a novel infinite-frame generation method, MIGA
This is a restatement of the main contribution. MIGA builds on FIFO-Diffusion's autoregressive framework while addressing training-inference gap (via TTA) and long-term consistency (via DCE), both demonstrated through experiments.
This is a restatement of the main contribution. MIGA builds on FIFO-Diffusion's autoregressive framework while addressing training-inference gap (via TTA) and long-term consistency (via DCE), both demonstrated through experiments.
已证实 (85%)
We design an effective two-stage training-inference alignment mechanism that proactively mitigates the training-inference gap by optimizing the noise span
The two-stage TTA mechanism is fully specified with Stage 1 (zigzag iterative denoising) and Stage 2 (unified noise level denoising). The noise span optimization is mathematically defined and its effectiveness is quantitatively demonstrated.
The two-stage TTA mechanism is fully specified with Stage 1 (zigzag iterative denoising) and Stage 2 (unified noise level denoising). The noise span optimization is mathematically defined and its effectiveness is quantitatively demonstrated.
已证实 (85%)
we introduce an innovative dual consistency enhancement mechanism that promotes long-term consistency through self-reflection and long-range frame guidance
The DCE mechanism with self-reflection and long-range frame guidance is fully specified in Section 3.3. Both components are described with mathematical formulations and their effectiveness is demonstrated through ablation studies.
The DCE mechanism with self-reflection and long-range frame guidance is fully specified in Section 3.3. Both components are described with mathematical formulations and their effectiveness is demonstrated through ablation studies.
已证实 (90%)
The first stage maintains a zigzag-structured latent queue to proactively narrow the noise span of input latents
The zigzag-structured latent queue is clearly specified in Section 3.2 with Equation 3 defining the queue structure where adjacent L_zig latents share the same noise level, illustrated in Figure 2(b).
The zigzag-structured latent queue is clearly specified in Section 3.2 with Equation 3 defining the queue structure where adjacent L_zig latents share the same noise level, illustrated in Figure 2(b).
已证实 (85%)
In the second stage, once all latents are denoised to the same noise level, a unified denoising process is conducted, achieving a noise span that matches that of the training phase
Stage 2 is clearly specified in Section 3.2 with Equation 4 showing all latents at unified noise level τ_e-1. The alignment with training conditions is explained and visual results are referenced in Figure 4.
Stage 2 is clearly specified in Section 3.2 with Equation 4 showing all latents at unified noise level τ_e-1. The alignment with training conditions is explained and visual results are referenced in Figure 4.
已证实 (85%)
For early high-noise latents, we design a self-reflection approach that efficiently evaluates and promptly corrects them, thereby ensuring consistency in the subsequently generated video
The self-reflection approach is fully specified in Section 3.3 with Algorithm 6 providing pseudocode. It evaluates consistency at early high-noise latents using C_score and triggers correction when anomalies are detected.
The self-reflection approach is fully specified in Section 3.3 with Algorithm 6 providing pseudocode. It evaluates consistency at early high-noise latents using C_score and triggers correction when anomalies are detected.
已证实 (85%)
our approach achieves this solely through self-similarity analysis among early latents
The paper clearly specifies using cosine similarity between latents (Equations 5-7) as the consistency metric, avoiding external evaluation models like DINO that prior work (ScalingNoise) requires.
The paper clearly specifies using cosine similarity between latents (Equations 5-7) as the consistency metric, avoiding external evaluation models like DINO that prior work (ScalingNoise) requires.
已证实 (85%)
For the later low-noise latents, we introduce a long-range frame guidance approach that incorporates them into each denoising iteration, facilitating feature interactions between distant frames
The long-range frame guidance is fully specified in Section 3.3 with Equation 8 showing how m_guid latents from earlier positions are incorporated into local denoising. Effectiveness is shown in Table 5.
The long-range frame guidance is fully specified in Section 3.3 with Equation 8 showing how m_guid latents from earlier positions are incorporated into local denoising. Effectiveness is shown in Table 5.
已证实 (90%)
MIGA achieves significant gains of 4.7% and 2.0% in subject and background consistency on VBench, respectively, compared to FIFO-Diffusion with a similar framework
Specific quantitative results are stated in paragraph 6: 'MIGA achieves significant gains of 4.7% and 2.0% in subject and background consistency on VBench, respectively, compared to FIFO-Diffusion.'
Specific quantitative results are stated in paragraph 6: 'MIGA achieves significant gains of 4.7% and 2.0% in subject and background consistency on VBench, respectively, compared to FIFO-Diffusion.'
证据不足 (50%)
evaluations on NarrLV demonstrate that MIGA exhibits exceptional capability in generating rich narrative content
已证实 (90%)
Unlike existing methods that change the noise level with every single latent frame, our queue alters it every L_zig latents
The zigzag structure is clearly specified in Section 3.2 with Equation 3 showing that noise level changes every L_zig latents instead of every single latent.
The zigzag structure is clearly specified in Section 3.2 with Equation 3 showing that noise level changes every L_zig latents instead of every single latent.
已证实 (80%)
This zigzag structure provides the model with a smoother noise span across inputs, contributing to mitigating the training-inference gap
The design rationale is explained in Section 3.2, and the effectiveness is demonstrated through ablation studies (Table 6) showing performance improvements from the zigzag structure.
The design rationale is explained in Section 3.2, and the effectiveness is demonstrated through ablation studies (Table 6) showing performance improvements from the zigzag structure.
已证实 (85%)
Since all latent frames share the same noise level, the model processes latents with identical intensity at each denoising operation. This setup aligns well with the conditions seen during training
The design rationale is clearly explained in Section 3.2. During training, the model sees latents at unified noise levels, and Stage 2 replicates this condition by processing all latents at the same noise level τ_e-1.
The design rationale is clearly explained in Section 3.2. During training, the model sees latents at unified noise levels, and Stage 2 replicates this condition by processing all latents at the same noise level τ_e-1.
证据不足 (50%)
As the foundation model handles a fixed latent length per pass, memory usage does not grow with longer videos
已证实 (85%)
the self-reflection approach focuses on latents at the tail of the queue, efficiently evaluating and correcting newly added latents
The self-reflection approach is described in Section 3.3, focusing on the judgment index f_judg at the tail of the queue for evaluating and correcting newly added high-noise latents.
The self-reflection approach is described in Section 3.3, focusing on the judgment index f_judg at the tail of the queue for evaluating and correcting newly added high-noise latents.
已证实 (85%)
the long-range frame guidance approach targets latents at the head of the queue, incorporating long-range, low-noise latents into each local denoising process
The long-range frame guidance is described in Section 3.3, targeting latents at the head of the queue (earlier, lower-noise positions) to guide local denoising.
The long-range frame guidance is described in Section 3.3, targeting latents at the head of the queue (earlier, lower-noise positions) to guide local denoising.
已证实 (80%)
our self-reflection approach integrates TTS with the characteristics of frame-level autoregressive generation for long videos
The paper explains in Section 3.3 how self-reflection integrates test-time scaling (TTS) techniques with the frame-level autoregressive generation framework for long videos.
The paper explains in Section 3.3 how self-reflection integrates test-time scaling (TTS) techniques with the frame-level autoregressive generation framework for long videos.
... 共 58 个论点
可复现性评估
较低可复现性 (0%)
缺失的复现细节
- No code available - implementation cannot be verified or reused
- No data available - evaluation datasets and prompts not accessible
- Random seeds not specified - results cannot be exactly reproduced
- Hardware/environment specifications not provided (GPU type, memory requirements, software versions)
- Computational cost and inference time not detailed in main text
- Appendix details referenced (App. A.1.1, A.3, A.4, B.1) but not accessible for complete implementation
- User study details incomplete - prompt selection criteria, annotator instructions, and statistical analysis methods not specified
- Baseline configurations for FIFO-Diffusion, Scaling-Noise, FreePCA, FreeLong not detailed
- Exact evaluation metrics implementation for VBench and NarrLV not specified
- Preprocessing steps and initialization procedures only referenced in appendix
局限与证据边界
- We observe that this frame-level autoregressive generation framework is difficult to apply to certain foundation models based on the MMDiT architecture. The main reason is that these models concatenate text and video features, and jointly interact with the noise timestep condition
- While longer video duration offers greater space for content creation, it also increases the risk of unintended model behaviors
- the beginning of the generated video follows the text prompt well, with a cat walking from left to right. However, after some time, the cat's head and tail suddenly switch places. This phenomenon can be regarded as a hallucination of the video generation model, or as evidence of the lack of underlying physical knowledge
- In future work, we aim to incorporate additional conditioning signals beyond text instructions to enable the generation of more realistic long videos
本分析由 PDF 阅读助手 自动生成，仅供参考，不构成学术评审意见。验证结论和可复现性评估基于论文文本自动分析，可能存在偏差。原始论文请参阅 arXiv。
分析时间：2026-05-24T13:21:11+00:00 · 数据来源：Paper Collector