TL;DR
AGAVE (Adaptive Guidance-Aware Velocity Ensembling) transforms classifier-free guidance from a single, rigid scalar into a dynamic field that varies across time, space, and video frames. By mathematically "baking" this guidance schedule directly into the training target of a one-step model, the authors achieve the quality of complex, multi-step adaptive sampling in a single function evaluation with zero extra inference cost.
What Problem Does This Paper Address?
Diffusion and flow-matching models are the current gold standard for video generation, but they are notoriously slow, requiring dozens of network evaluations to produce a single clip. "One-step" generators like MeanFlow aim to solve this by jumping directly from noise to data in a single step.
However, these fast models have a "guidance problem." Classifier-Free Guidance (CFG)—the "knob" used to trade off diversity for prompt alignment—is usually treated as a single constant scalar. But research shows that guidance shouldn't be constant: it should be low at the start of denoising (to avoid errors), peak in the middle, and vary spatially based on where the action is.
Multi-step models can easily change the guidance scale at every step. One-step models, by definition, have no "steps" to schedule over, forcing them to use a crude, fixed value that often leads to over-saturated colors or "collapsed" motion in videos.
Methods
The authors resolve this tension with a clever mathematical observation. MeanFlow models don't just predict a single point; they learn an average velocity over a noise interval.
The Interval-Integrated CFG (I²-CFG) Identity
The core contribution is Theorem 1. The authors prove that because the MeanFlow target is an integral, you can place a time-varying guidance function inside that integral. This allows a single network evaluation to "compile" the effect of an entire multi-step guidance schedule.
AGAVE: Three Axes of Guidance
The researchers instantiate this theory as AGAVE, which modulates guidance across three dimensions:
- Noise-time: Uses a learnable Beta-shaped schedule. This allows the model to "rediscover" the optimal strategy of peaking guidance in the middle of the process.
- Space (Motion-Saliency): Automatically amplifies guidance in regions with coherent motion (foreground) while keeping it lower for static backgrounds.
- Frame (Position Asymmetry): Learns a profile that treats interior frames differently than boundary frames to improve temporal consistency.
To learn these parameters, they use adaptive-guidance distillation, where a multi-step "teacher" model provides a reference for what a high-quality, guided sample should look like.
Datasets
The authors evaluated AGAVE on several standard video generation benchmarks:
- VBench: A comprehensive suite for video quality, including motion, consistency, and aesthetics.
- UCF-101: Used for measuring FVD (Fréchet Video Distance).
- MSR-VTT: Another standard dataset for text-to-video evaluation.
- ImageNet 256²: Used for a proof-of-concept image generation task.
Key Insights and Ideas
- Guidance is a Field: The most important shift is moving from a scalar $\omega$ to a field $\omega(z_\tau, \tau, f)$. This acknowledges that different parts of a video need different amounts of "push" from the prompt.
- Amortization of Control: Instead of paying the computational price of a schedule at inference time, AGAVE pays it once during training.
- The "No Discretization Error" Advantage: Surprisingly, a one-step model using I²-CFG can actually outperform a multi-step model. Multi-step models only approximate a guidance schedule through discrete steps (incurring error), whereas the one-step model integrates the schedule exactly.
Key Results
- State-of-the-Art: AGAVE set a new SOTA for one-step video generation on VBench, surpassing previous leaders like iMF and rCM.
- Recovering Motion: It specifically improved "Dynamic Degree" (motion) by +4.7 points on the 14B parameter model, fixing the "static video" problem common in distilled models.
- Efficiency: It matches the quality of a 50-step teacher model in just 1 step, with zero extra latency and fewer than ten additional learnable scalars.
- Learned Behavior: Without manual tuning, the model learned a bell-shaped temporal schedule that naturally vanishes at high noise—validating existing "zero-init" heuristics used in multi-step sampling.
Limitations and Future Work
- Distillation Dependency: While the schedule parameters are few, AGAVE currently requires a multi-step teacher to provide the distillation signal.
- Background Consistency: The authors note that while AGAVE improves motion, some baselines score slightly higher on background consistency simply because they produce nearly static videos—a trade-off AGAVE intentionally breaks to provide more dynamic content.
- Broader Control: The "Interval-Integrated" framework could potentially be used for other types of inference-time controls beyond just CFG, such as structural guidance or style transfer.
Summary updated