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

TL;DR

MIGA is a train-free recipe for turning a pretrained short-clip video diffusion model into an infinite-frame long-video generator on top of the FIFO-Diffusion frame-level autoregressive framework. It targets two failure modes of that framework: (a) the train/inference mismatch where the model is asked at inference to denoise a queue of latents spanning a wide range of noise levels even though it was trained at a single noise level, and (b) long-term temporal drift. A two-stage alignment mechanism narrows the noise-span seen by the model, and a dual consistency enhancement combines self-reflection on newly added high-noise frames with long-range frame guidance from already-denoised low-noise frames. The authors report SOTA on VBench and NarrLV.

Key claims

Frame-level autoregressive long-video frameworks like FIFO-Diffusion let you generate infinitely long videos with constant memory but suffer a training–inference mismatch: the model was trained on latents at a single noise level and at inference is asked to denoise a queue spanning a wide range of noise levels [project page, Introduction].
The “Two-stage Training-inference Alignment” (TTA) mechanism explicitly shrinks the noise span fed to the model. Stage 1 maintains a zigzag-structured latent queue where the noise level changes only every L_zig latents rather than every frame, smoothing the input distribution [project page, TTA section].
The Dual Consistency Enhancement (DCE) mechanism operates on two ends of the long latent queue: self-reflection targets the queue’s tail (newly added high-noise latents), adaptively detecting abrupt drops in self-similarity and triggering an expanded local search only at those anomaly points — avoiding the redundant computation of a fixed-step search [project page, DCE — self-reflection].
The long-range frame guidance arm targets 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 generation toward globally consistent content [project page, DCE — long-range frame guidance].
MIGA reports state-of-the-art performance on VBench and NarrLV [project page, Abstract].

Method

MIGA sits as a sampler-level wrapper around a frozen text-to-video diffusion model. The base substrate is FIFO-Diffusion’s diagonal denoising idea: a long queue of latent frames where each successive frame is at a slightly higher noise level than the previous; on every step the queue is denoised once and the cleanest latent is dequeued as an output frame while a new pure-noise latent enters at the tail. MIGA makes two changes to that loop:

Two-stage Training-inference Alignment (TTA) changes the shape of the noise schedule across the queue. Instead of incrementing the noise level at every position, Stage 1 holds the noise level constant across blocks of size L_zig before stepping — a zigzag schedule. This is a pure inference-time intervention that shrinks the noise span any single forward pass has to handle, narrowing the gap to the single-noise-level regime the model was trained for.
Dual Consistency Enhancement (DCE) runs over the maintained long latent queue. Self-reflection monitors the queue’s high-noise tail for abrupt drops in self-similarity between newly added latents; only at those anomalies does it spend extra compute on an expanded local search, rather than paying the cost of a fixed search interval everywhere. Long-range frame guidance takes already-denoised low-noise frames at the head of the queue and injects them as additional conditioning into the per-step denoising, so that distant frames can interact via attention.

The two changes are complementary: TTA closes the training–inference gap per step, and DCE closes the cross-time-consistency gap across steps. Both are train-free — no weights are updated.

Results

The project page reports state-of-the-art performance on VBench and NarrLV. Quantitative tables are not on the landing page; the actual VBench / NarrLV numbers would live in the (currently unposted) full paper. NarrLV (LongVie 2: Multimodal Controllable Ultra-Long Video World Model cites it; the original NarrLV paper is Feng et al., 15 Jul 2025) decomposes prompts into Temporal Narrative Atoms and uses an MLLM-based progressive QA evaluator over up to n=6 narrative units, so a MIGA win on NarrLV would mean wins on Narrative Element Fidelity, Narrative Unit Coverage, and Narrative Unit Coherence — directly probing the long-range consistency that DCE targets.

Why it’s interesting

MIGA is the rare train-free entry in the long-video Pareto. The four filed operating points in Context Length / Quality Trade-off in Video Generation — explicit compression (PFP), functional cache partitioning (Context Forcing), curriculum-based consistency (LongVie 2), learned mid-buffer compression with top-k retrieval (PackForcing) — all require training. MIGA achieves long-video extension by changing the noise schedule and adding cross-time attention at inference time, on a frozen base model. That makes it directly comparable to other train-free FIFO-Diffusion descendants but also a clean baseline for the question: how much of the long-video win in trained recipes comes from explicit memory machinery vs. from closing the FIFO-Diffusion train/inference mismatch? It also contrasts with Rolling Forcing: Autoregressive Long Video Diffusion in Real Time and LongLive: Real-time Interactive Long Video Generation, both of which fine-tune for streaming AR but report effective context windows of only 3–6s — MIGA’s claim is that with the right inference-time scheduling, the frame-level AR substrate can be pushed much further without any training.