Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation

What Changed

On-policy distillation (OPD) has become a cornerstone for adapting diffusion models to specific downstream tasks or accelerating inference by distilling a heavy teacher model into a more efficient student. While these methods have successfully reduced the computational burden of diffusion models, a fundamental issue has persisted regarding how they handle Classifier-Free Guidance (CFG). CFG is the standard mechanism for controlling diffusion generation, typically involving a combination of a conditional prediction and an unconditional (negative) prediction.

Recent research has revealed that existing OPD methods, which rely on naive velocity matching for CFG-composed predictions, suffer from an under-identification problem. Specifically, the objective function used to align the student with the teacher is blind to the individual contributions of the positive and negative branches of the CFG. This leads to a failure mode termed Negative Branch Asymmetry (NBA), where the student model inadvertently optimizes the positive branch at the expense of the negative branch. This discovery shifts the focus of distillation research from simple velocity matching toward branch-aware supervision, providing a more robust framework for training models that rely heavily on guidance scales.

Technical Details

To understand the failure mode, one must analyze the standard CFG-composed prediction. In a typical diffusion model, the guided velocity is calculated as a linear combination of the conditional and unconditional predictions. Existing OPD methods treat this guided velocity as a single target for the student to match. The researchers demonstrate that this approach is under-identified because the positive-branch error and the negative-branch error can compensate for one another within the guided prediction.

In scenarios where the teacher's negative branch contains privileged information—information that the student model cannot access or represent—the optimization process encounters a conflict. When the student attempts to minimize the total guided velocity error, the model may reduce the error in the positive branch while simultaneously increasing the error in the negative branch. This phenomenon, Negative Branch Asymmetry (NBA), occurs because the objective function does not distinguish between the two branches, allowing the negative branch to degrade as long as the overall guided velocity appears to converge.

To mitigate this, the researchers introduced Positive-Direction Matching (PDM). Unlike traditional methods that treat the guided velocity as a monolithic target, PDM decomposes the objective into two distinct constraints: one for the positive prediction and one for the CFG conditional direction. By separately constraining these components, PDM prevents the antagonistic dynamics that characterize NBA. This ensures that the student model learns the correct conditional signal without sacrificing the integrity of the negative branch, which is essential for maintaining the model's ability to steer generation effectively.

Developer Implications

The implications for developers working with diffusion models are significant, particularly for those involved in video generation and complex control tasks. First, the findings suggest that current distillation pipelines that ignore branch-level dynamics may be producing models that are fragile when subjected to varying inference guidance scales. If a model is trained using naive guided matching, it may perform well at a specific guidance scale but degrade rapidly as that scale is adjusted, as the underlying branch errors are not properly balanced.

Second, the introduction of PDM provides a concrete, actionable path for improving model robustness. Developers should consider moving away from aggregate velocity matching in favor of branch-aware objectives, especially when the teacher model relies on complex negative conditioning. This is particularly relevant for dense-to-sparse video control, where the model must maintain temporal consistency while adhering to sparse control signals. In these environments, the sensitivity to guidance scales is high, and the use of PDM allows for more stable knowledge transfer from the teacher to the student.

Finally, this research underscores the necessity of auditing distillation objectives. As diffusion models become more complex, the assumptions underlying our training objectives—such as the assumption that matching the final output is sufficient—must be re-evaluated. Developers should look for ways to decompose their distillation targets to ensure that every component of the model's prediction is being supervised correctly, rather than relying on the model to resolve internal conflicts between competing branches.

Bottom Line

The research into Negative Branch Asymmetry highlights a subtle but critical flaw in how we currently distill diffusion models. By treating the guided velocity as a single target, we have been inadvertently allowing models to trade off performance between positive and negative branches, leading to instability at varying guidance scales. Positive-Direction Matching offers a mathematically sound alternative that enforces branch-aware supervision. For teams deploying diffusion models in production, adopting PDM-like strategies is a necessary step toward building more reliable, steerable, and robust generative systems, particularly in high-stakes applications like video control where guidance sensitivity is a primary concern.