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

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 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.

Figure A1. MIGA enables temporally consistent, infinite-frame (∞) video generation in a training-free manner. We present three long videos (1000+ frames) generated by MIGA, while the foundation model used by MIGA, Wan2.1-1.3B, supports only 81 frames by default.

① Two-Stage Training-Inference Alignment (TTA)

Train-free frame-level autoregressive frameworks (e.g., FIFO-Diffusion) require the foundation model to denoise latents that span a wide range of noise levels at inference, while the model is trained on latents sharing a single noise level — a mismatch that hampers generation quality. TTA alleviates this gap by explicitly shrinking the noise span seen by the model. In Stage 1, we maintain a zigzag-structured latent queue that changes the noise level only every L_zig latents instead of every single frame, providing a smoother input distribution. In Stage 2, once all latents are denoised to the same noise level, a unified denoising pass is performed, exactly matching the noise condition seen during training.

Figure 1. Inference framework comparison between FIFO-Diffusion and our Two-Stage Training-Inference Alignment (TTA) mechanism. By introducing a zigzag-structured queue and a unified denoising stage, TTA proactively reduces the noise span of latents fed to the foundation model.

② Dual Consistency Enhancement (DCE)

Although TTA closes the training-inference gap, long-term temporal consistency still requires explicit modeling. DCE addresses this from two complementary directions on the maintained long latent queue. The self-reflection approach focuses on the queue's tail: it adaptively detects abrupt drops in self-similarity among newly added high-noise latents, and triggers an expanded local search only at those anomaly points to correct them — avoiding the redundant computation of fixed-step search. The long-range frame guidance approach focuses on the queue's head: late, low-noise latents that already cover broad temporal context are injected into each local denoising iteration, enabling feature interaction between distant frames and steering the generation toward globally consistent content.

Figure A2. Illustration of the Dual Consistency Enhancement (DCE) mechanism. Self-reflection corrects early high-noise frames at the tail of the queue, while long-range frame guidance leverages later low-noise frames at the head of the queue to jointly improve temporal consistency.