Guidance Is a Field, Not a Scalar: Interval-Integrated Classifier-Free Guidance for One-Step Video Generation
summary
The gist
One-step generators built on MeanFlow synthesize a video in a single function evaluation, yet inherit a rigid treatment of classifier-free guidance (CFG)…
In short
One-step video generators treat classifier-free guidance (CFG) as a fixed number, which is insufficient because optimal guidance changes based on noise, space, and frame position. The authors propose AGAVE, a method that treats guidance as a spatially and temporally varying field. This field is integrated into the model's structure to achieve superior performance compared to fixed scalar conditioning.
Key concepts
- Classifier-Free Guidance (CFG)
- CFG is a technique used in AI generation where the model learns to generate content by comparing predictions made with and without specific class information. In this paper, the authors argue that CFG should not be a single fixed value but a complex field that adapts to different parts of the video frame.
- Interval-Integrated CFG Identity
- This is the mathematical foundation proving that an arbitrary guidance field can be successfully incorporated into one evaluation step. By treating guidance as a function integrated over a noise interval, the authors show it can be absorbed into the model's target structure without introducing discretization errors.
- AGAVE (Adaptive Guidance-Aware Velocity Ensembling)
- AGAVE is the proposed method that realizes guidance as a field. It is parameterized by three axes: noise-time, space, and frame position. This allows the guidance to adapt dynamically across different aspects of the video generation process.
- Field-Guided Average Velocity
- This is the resulting concept where the model calculates a velocity based on an average guided by a spatially and temporally varying field. Because this construction inherits none of the errors from finite schedulers, it can provably outperform multi-step methods.
Terminology used across episodes
This episode discusses
- Guidance Is a Field, Not a Scalar: Interval-Integrated Classifier-Free Guidance for One-Step Video Generation · Paper Radio
- Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets
- CFG-Zero*: Improved Classifier-Free Guidance for Flow Matching Models
- Mean Flows for One-step Generative Modeling
- Improved Mean Flows: On the Challenges of Fastforward Generative Models
- Stage-wise Dynamics of Classifier-Free Guidance in Diffusion Models
- HunyuanVideo: A Systematic Framework For Large Video Generative Models
- OmniDirector: General Multi-Shot Camera Cloning without Cross-Paired Data
- TARS: Timestep-Aware Data Scaling for 3D-Free Video Re-Shooting
- Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models
- Classifier-free Guidance with Adaptive Scaling
- Decoupling Semantics from Distortions: Multi-Scale Two-Stream Vision-Language Alignment for AI-Generated Image Quality Assessment
- ParaScale: Scale-Calibrated Camera-Motion Transfer via a Gauge-Invariant Parallax Number
- ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation
- OmniDrive: An LLM-Choreographed Multi-Agent World Model with Unified Latent Co-Compression for Multi-View Driving Video Generation
- Dynamic Classifier-Free Diffusion Guidance via Online Feedback
- Classifier-Free Guidance: From High-Dimensional Analysis to Generalized Guidance Forms
- UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild
- Towards Accurate Generative Models of Video: A New Metric & Challenges
- Wan: Open and Advanced Large-Scale Video Generative Models
- LiveAct: Towards Real-Time Multi-Agent World Models with Heterogeneous Action Interfaces · Paper Radio
The paper
Guidance Is a Field, Not a Scalar: Interval-Integrated Classifier-Free Guidance for One-Step Video Generation · Read on arXiv
Peking University
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Guidance Is a Field, Not a Scalar".
Jane: Here's the gist of the paper. One-step generators built on MeanFlow synthesize a video in a single function evaluation, yet inherit a rigid treatment of classifier-free guidance (CFG): the guidance scale is frozen during training or, at best, exposed as a single scalar This conflicts with mounting evidence that the optimal guidance is not a constant but a field that varies with the noise level, with spatial content, and, for video, with frame position The core problem addressed The central tension explored is that existing one-step models treat classifier-free guidance (CFG) as a fixed scalar rather than a spatially and temporally varying field This rigidity conflicts with empirical findings showing that optimal guidance should be time-varying, dependent on noise level, spatial content, and frame position While multi-step samplers can re-schedule guidance along a trajectory, one-step models lack such a trajectory The authors resolve this by treating guidance as a field that can be absorbed into the target integral structure of MeanFlow The mathematical foundation of the solution The authors formalize this concept through the interval-integrated CFG identity, proving that an arbitrary guidance field can be compiled into one evaluation Specifically, because the MeanFlow target is an integral over a noise interval, it can absorb an integrand that varies arbitrarily along the noise path This yields the field-guided average velocity, which satisfies a MeanFlow-type identity The key result is that this construction inherits none of the discretization error that every finite scheduler must pay, meaning it can provably surpass any finite multi-step sampler The structure of the proposed method The paper introduces AGAVE (Adaptive Guidance-Aware Velocity Ensembling) as the instantiation of this field concept This field is parameterized by three axes: noise-time, space, and frame position Noise-time modulation is achieved via a learnable Beta schedule over noise level τ Spatial modulation uses a motion-saliency map M over space Frame modulation employs a frame-position asymmetry profile ρ(f) The training and empirical results AGAVE is trained by distilling a multi-step adaptively guided teacher to learn the schedule parameters ϕ, which are fewer than ten extra scalars The training objective involves fitting the network to match the field-guided average velocity target Empirical results show that AGAVE surpasses all one-step baselines on VBench, UCF-101, and MSR-VTT while matching multi-step guided sampling Table 3 demonstrates that AGAVE at one step matches or beats adaptive guidance run on the 50-step teacher The significance of the findings AGAVE is the first one-step generator to rival multi-step adaptive guidance, setting a new state of the art on VBench, UCF-101, and MSR-VTT The gains are broad rather than localized, with AGAVE attaining the best VBench total among one-step methods on both backbones Furthermore, sweeping the guidance scale ωmax shows that the AGAVE field strictly dominates the iMF scalar over the entire FVD–CLIP-T curve This confirms that a single AGAVE checkpoint is strictly more controllable than scalar conditioning The paper concludes that this construction is unavailable to single-time models such as consistency models, whose prediction f(zt, t) exposes no interval
r, t: over which a within-trajectory schedule could be integrated References
Paper summary: Tom: So to recap what we just heard, this paper is about overcoming the limitation in one-step video generation models where classifier-free guidance is treated as a fixed scalar.
Jane: The main idea they push is that optimal guidance isn't fixed; it's actually a field that depends on things like noise level and where you are in the video, or frame position.
Lu: They tackle this by showing that the integral structure of MeanFlow lets you absorb an arbitrary guidance field directly into the target itself during one evaluation.
Tom: This is formally done through what they call the interval-integrated CFG identity, which proves that this compiled guidance doesn't introduce any schedule discretization error.
Meng: The key result here is that this method can match or exceed whatever performance you get from a finite multi-step sampler by treating the guidance as a field.
Jane: They then define AGAVE as the specific system that implements this field, using noise-time, space, and frame position as its three modulation axes.
Lu: The paper also shows how these axes are parameterized: a learnable beta schedule for noise level tau, a motion saliency map M for space, and a frame-position profile rho of f.
Tom: And they show that the training is done by distilling knowledge from a multi-step teacher to learn these parameters.
Jane: So, it’s about learning these schedules rather than just fixing them beforehand. This distillation process only requires fewer than ten extra scalars to be learned.
Conclusion: Tom: The title itself, "Guidance Is a Field, Not a Scalar," really captures the essence of what they are doing here—moving away from that rigid scalar thinking.
Jane: They show that this field concept allows one-step models to achieve a level of control over guidance that previously only existed in multi-step samplers.
Lu: The implication is significant because it suggests the computational substrate of MeanFlow isn't just for efficiency, but for handling these continuous guidance fields.
Meng: For those of us shipping models, this means we can have much finer control over the output quality without needing to run a lengthy multi-step process every time.
Tom: It’s about moving from asking what the single best number is to asking what the best function is along the generative path.
Jane: This work sets a new benchmark for one-step video generation on benchmarks like VBench, UCF-one hundred one and MSR-VTT because it matches or beats multi-step adaptive guidance results.
More episodes
- 2603.28113-The multiplicative Lipschitz bound as an inductive bias
- 2603.28595-Optimistic Actor-Critic with Parametric Policies for Linear Markov Decision Processes
- 2603.29765-Training-Free Dynamic Upcycling of Expert Language Models
- 2605.00778-Observable Performance Does Not Fully Reflect Adaptive System Organization: A Multi-Level Analysis of Gait Dynamics Under Occlusal Constraint
- 2605.06906-TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond
- 2605.09165-Sparse Layers are Critical to Scaling Looped Language Models
- 2605.09439-Inverse Design for Conditional Distribution Matching
- 2605.10313-Signature Approach for Contextual Bandits with Nonlinear and Path-dependent Rewards: Regret Bound and Statistical Inference
- 2605.10878-Minimum Neural Weight Norm = Kolmogorov Complexity
- 2605.11617-MIST: Reliable Streaming Decision Trees for Online Class-Incremental Learning via McDiarmid Bound