Alibaba's AMAP research team has landed a paper at ICML 2026: MIGA, addressing a specific 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.
The Root Problem: Training Saw a "Neat Queue", Inference Faces a "Ragged Mob"
Off-the-shelf video diffusion models (VideoCrafter2, Wan2.1, etc.) share a tacit assumption during training: all input frame latents sit at the same noise level. A batch of 16 frames is all at timestep τ=25. The model learns to denoise a uniformly noisy batch.
But training-free infinite-frame frameworks like FIFO-Diffusion maintain a noise queue Q for autoregressive rollout. The first frame is nearly clean (τ≈0), the last is pure noise (τ=T). When the model denoises a sliding window of f_0 frames, those frames span a huge noise range — inputs the model never saw in training. This is the essence of the training-inference gap (Figure 2(a): FIFO-Diffusion's queue is a diagonal line, noise growing smoothly from 0 to T).
Consequence: content drift, visual artifacts, subjects deforming over time.
Path One: Two-Stage noise Alignment (TTA)
Core intuition: compress the noise span inside the queue before the model processes it.
Stage 1: Zigzag Iterative Denoising
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}, τ_{e+1}, ...} — one step per 4 frames. This "proactively narrows the noise span" within each sliding window. Ablation (Table 4) shows L_zig=4 is the sweet spot.
Stage 2: Unified Noise-Level Denoising
After n rounds, the queue holds n*L_zig frames that are consistent only within groups of 4. Stage 2 waits until all frames drop to the same level τ_{e-1}, then denoises uniformly — matching training conditions exactly.
Note: skipping Stage 1 and doing Stage 2 directly collapses performance. Stage 1's autoregressive process builds implicit inter-frame information flow that Stage 2 depends on.
Ablation (Table 6): baseline 95.02 → +Stage 1 96.78 → +Stage 2 97.05.
Path Two: Dual Consistency Enhancement (DCE)
TTA fixes input quality, but long-video generation also needs long-range consistency — sliding windows are "nearsighted."
Self-Reflection
Prior methods (e.g., ScalingNoise) assess consistency with external models like DINO, requiring pixel decoding — expensive and inflexible. MIGA's surprising finding: the fluctuation pattern of consistency scores on early high-noise latents correlates strongly with the final clean latents (Figure 3(c,d): even at noise level 40 of 50, correlation remains high).
So you can predict failures without finishing the video: 1. Define a judgment node f_judg at the queue tail 2. Compute a consistency score C_score via cosine similarity of adjacent frames — fully in latent space, no decoding 3. When adjacent chunks' C_score drops by more than δ_adju=0.01, trigger correction 4. Generate n_samp candidate continuations, guided by already-verified frames, and keep the most consistent one
This is Test-Time Scaling (TTS) realized in video generation. Smaller thresholds mean more search, better performance, more compute; δ=0.01 is the sweet spot (Figure 5 shows correction rate R_corr and success rate R_succ vs. threshold).
Long-Range Frame Guidance
MIGA's fix is minimal: sparsely sample m_guid low-noise frames from the queue's front and concatenate them into the current window. Input becomes [z_1,...,z_m_guid, z_l,...,z_{l+f_0-m_guid-1}] instead of just [z_l,...,z_{l+f_0-1}] — a "telescope" providing global context. Ablation (Table 5): m_guid=6 is optimal, lifting O.S. from 95.80 (m_guid=0) to 96.87.
Experiments: Not Just Numbers, But 1000+ Frames
Tested on both 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 unexplored
- Slow degradation possible at extreme lengths even for 1000+ frames
- 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 (narrative content): Wan2.1-based MIGA leads across TNA=2/3/4 — notable since Wan2.1's photorealistic style makes consistency harder than VideoCrafter2's animated style.
Figure 1 shows 1000+ frame generation, while the base Wan2.1-1.3B only supports 81 frames.
Memory: inherits FIFO-Diffusion's fixed memory footprint — no growth with video length.
Why It Matters
Video generation is splitting into two routes:
1. Retraining route: SkyReels-V2, MAGI-1, etc. — massive data and compute 2. Training-free route: frameworks on top of existing short-video models
MIGA belongs to the latter yet matches or beats some trained methods (Appendix B.4). For resource-constrained settings (ordinary users, real-time apps, edge devices), training-free + fixed memory is highly attractive.
Deeper still: MIGA successfully transfers the Test-Time Scaling idea from LLMs to diffusion video generation. Self-Reflection is essentially "try more at generation time, pick the best" — the same coin as o1/R1's inference-time scaling.
Limitations
From the paper's Appendix C:
One-Sentence Summary
MIGA smooths the noise span with zigzag scheduling, aligns training and inference conditions in two stages, predicts consistency anomalies at high noise via Self-Reflection, and cures sliding-window myopia with long-range guidance — letting off-the-shelf short-video models generate arbitrarily long videos stably, with no extra memory.
The paper doesn't retrain the model. It redesigns the environment the model works in.
Reference: