MIGA, from Alibaba's AMAP research team, has been accepted to ICML 2026. The paper tackles a concrete but critical engineering problem: how to make off-the-shelf short-video diffusion models generate long videos without retraining or extra memory.
The answer runs along two parallel paths: first align the noise worlds of training and inference, then let the model self-reflect while looking into the distance.
---
Root Cause: Trained on "Neat Queues", Facing a "Motley Crew" at Inference
Off-the-shelf video diffusion models (VideoCrafter2, Wan2.1, etc.) share a tacit assumption at training time: all frame latents fed to the model sit at the same noise level — e.g., a batch of 16 frames all at timestep τ=25. The model learns to denoise a batch at one specific noise level.
But training-free infinite-frame frameworks like FIFO-Diffusion maintain a noise queue Q to autoregressively advance: the first frame is nearly clean (τ≈0) while the last is pure noise (τ=T). Each denoising step takes f_0 frames from the sliding window — frames spanning a huge noise range.
The model never saw such inputs during training. This is the essence of the training-inference gap (Figure 2(a) shows the diagonal noise queue). Consequences: content drift, visual artifacts, deforming subjects.
---
Path 1: Two-Stage Alignment (TTA) — Reorganizing the Motley Crew
Core intuition: compress the noise span in the queue before the model actually processes it.
Stage 1: Zigzag Iterative Denoising — Turn the Diagonal into a Zigzag
Instead of every frame latent having a distinct noise level, MIGA changes the noise level only every L_zig frames: {τ_e, τ_e, τ_e, τ_e, τ_{e+1}, ...}. This "proactively narrows the noise span" within each window. Ablations (Table 4) show L_zig=4 is the sweet spot.
Stage 2: Unified Noise-Level Denoising — March Together Once Everyone Arrives
After n zigzag iterations, the queue holds n*L_zig frames that are only locally consistent. Stage 2 waits until all frames drop to the same level τ_{e-1}, then denoises them together — matching training conditions exactly.
Note: skipping Stage 1 and directly unifying all latents causes a sharp performance drop — Stage 1's autoregression builds implicit inter-frame information transfer that Stage 2 depends on. Ablations (Table 6): baseline 95.02 → +Stage 1 96.78 → +Stage 2 97.05.
---
Path 2: Dual Consistency Enhancement (DCE) — Self-Reflection + Seeing Far
TTA fixes input quality, but long videos also demand long-range consistency; sliding windows are "near-sighted."
Self-Reflection: Predicting Final Consistency at High Noise
Prior methods (e.g., ScalingNoise) use external models (DINO) requiring pixel decoding — expensive and rigid. MIGA's insight from an unexpected finding: consistency-score fluctuation patterns of early high-noise latents strongly correlate with final clean latents. Even at noise level 40 (max 50), C_score curves correlate highly with clean latents (Figure 3(c,d)) — so failures can be predicted before generation completes.
Mechanism: 1. Define a judgment node f_judg at the queue tail 2. Compute C_score via cosine similarity of adjacent frames (entirely in latent space, no decoding) 3. Trigger correction when adjacent-chunk C_score drops by more than δ_adju = 0.01 4. Generate n_samp candidate continuations guided by verified frames; keep the most consistent
This is Test-Time Scaling (TTS) realized in video generation. Smaller thresholds mean more searching: better quality, more compute. δ=0.01 is the balance point.
Long-Range Frame Guidance: A Telescope for the Sliding Window
At position l, the input is no longer just [z_l, ..., z_{l+f_0-1}] but includes m_guid sparsely sampled distant low-noise frames: [z_1,...,z_m_guid, z_l,...]. Ablations (Table 5): m_guid=6 optimal — 0 gives O.S. 95.80, 6 jumps to 96.87.
---
Experiments: Numbers Plus 1000+ Frames
Tested on VideoCrafter2 and Wan2.1-1.3B.
VBench (VideoCrafter2-based, 128 frames):
- FIFO-Diffusion: S.C. 92.92, B.C. 95.01, O.S. 95.02
- MIGA: S.C. 97.66, B.C. 96.99, O.S. 97.82
- Subject consistency +4.7%, background consistency +2.0%
- Only text conditioning verified; action/camera-motion conditioning remains open
- Slow degradation can still appear at extreme lengths
- Multi-prompt narrative coherence has room to improve
- Feng et al. (2026). Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos. arXiv:2605.18233. ICML 2026.
NarrLV: Wan2.1-based MIGA leads across TNA=2/3/4; notably Wan2.1 is a photorealistic model where consistency is harder to maintain, yet MIGA still works.
Figure 1 shows 1000+ frame generation, while Wan2.1-1.3B natively supports only 81 frames.
Memory: inherits FIFO-Diffusion's fixed-memory advantage — no growth with video length.
---
Why It Matters
Video generation is splitting into two routes:
1. Retraining route: SkyReels-V2, MAGI-1, etc., training long-video models from scratch with massive data and compute 2. Training-free route: building frameworks atop existing short-video models
MIGA belongs to the latter but rivals or exceeds some training-based methods (Appendix B.4). For resource-constrained settings (ordinary users, real-time apps, edge devices), training-free + fixed memory is highly attractive.
More deeply, MIGA successfully transfers Test-Time Scaling from LLMs to diffusion video generation — Self-Reflection is "try multiple times, pick the best," the same coin as o1/R1-style inference-time scaling.
---
Limitations
From Appendix C:
---
One-Line Summary
MIGA smooths the noise span with Zigzag scheduling, aligns training and inference via two stages, predicts consistency failures early with Self-Reflection, and fixes sliding-window myopia with distant-frame guidance — together letting off-the-shelf models generate stable infinite-length videos with no extra memory.
The paper didn't retrain the model — it redesigned the environment the model works in.
---
Reference: