Unifying Distributional Training for One-Step Visual Generation

Imagine you're developing a system that generates images from text descriptions. The challenge lies in ensuring that the generated images not only look realistic but also accurately reflect the input descriptions. Current methods often struggle with this, leading to issues like mode collapse, where the model fails to capture the diversity of possible outputs, resulting in repetitive or low-quality images. This is what's called mode collapse, and it can severely limit the effectiveness of generative models in real-world applications.

To address these challenges, the authors propose a new framework that separates the tasks of distribution modeling and matching discrepancies in generated features. This framework, called MGFlow, uses Gaussian mixtures to model feature distributions at different levels of granularity, allowing for more flexible and accurate image generation. By coupling mass-constrained sample assignments with updates to the model components, MGFlow effectively mitigates mode collapse, leading to significant improvements in performance on benchmarks like ImageNet. Compared to previous methods, MGFlow not only achieves state-of-the-art results but also enhances the capabilities of existing models, such as transforming a four-step text-to-image generator into a more efficient one-step process.

Introduces a unified framework for visual generation that connects distribution modeling and feature matching.

Demonstrates substantial improvements over existing methods with clear metrics on standard datasets.