Drifting generative models speed up with multihead scale approach
Convergence rates for generative drifting flows: fixed-scale obstructions and multihead acceleration
Machine Learning
Summary
Generating realistic new data can be slow because models often struggle to capture details at every scale. The authors identify that using a fixed level of detail causes slow learning, especially for fine features. By combining information from multiple scales simultaneously with a multihead method, their approach learns much faster and gets exponentially closer to the target data's distribution. This could enable generative models that create new samples in just one step more efficiently.
What this means in practice
- •For machine learning engineers: Develop faster one-step generative models for image or audio synthesis by integrating multihead scale normalization to improve training efficiency.
- •For computer vision developers: Enhance detail recovery in generative models producing high-resolution images by overcoming fixed-scale bottlenecks with multi-resolution aggregation.
A theory result. No direct application yet.
Authors
Arthur Stéphanovitch, Eddie Aamari
Abstract
Drifting models offer a promising route to faster generative AI: they perform gradual transport during training, while generating new samples in a single step. This paper asks whether the underlying drifting process can converge rapidly to a target distribution under ideal conditions, before finite-data or optimization effects are introduced. We show that its convergence rate depends critically on how it handles spatial scale. With a single fixed resolution, fine-scale features of the target can become nearly invisible, leading to extremely slow convergence. We introduce a multihead approach that combines scale-normalized information across a continuum of resolutions. We prove that this multihead approach restores exponential convergence near standard reference distributions. These results identify fixed resolution as a key bottleneck and provide a simple route to faster one-step generative models.