The year/Independent research

Paper 2605.18233

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

Published
May 2026
Research lab
Independent
Citations
1
GitHub
Not linked

01 In brief

Summary

The paper introduces MIGA, a training-free method for infinite-frame long video generation that builds on frame-level autoregressive frameworks like FIFO-Diffusion.

MIGA addresses two key limitations: the training-inference gap and long-term consistency.

It proposes a Two-Stage Training-Inference Alignment (TTA) mechanism that reduces the noise span of input latents via zigzag iterative denoising (Stage 1) and unified noise-level denoising (Stage 2).

Additionally, a Dual Consistency Enhancement (DCE) mechanism uses self-reflection to correct early high-noise latents and long-range frame guidance to incorporate low-noise latents for improved temporal consistency.

Experiments on VBench and NarrLV show MIGA achieves state-of-the-art performance, with significant gains in subject and background consistency over FIFO-Diffusion.

The method supports multi-prompt control and generates videos of 1000+ frames with constant memory consumption.

02 From the paper

Abstract

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/.