Blog
Trying to understand things better by writing about them
- June 20, 2026
MeanFlow One-Step Generation
Flow matching models are a type of generative model that shift one data distribution to another by learning the velocity field between the two. They are trained to learn the instantaneous velocity of data samples at a timestep between (the initial data distribution) and (the final data distribution). Generating data from a trained flow matching model involves taking a sample from the initial distribution (like a Gaussian), integrating the velocity field step-by-step from to , and shifting the sample towards the target distribution through the integrated displacement path. However, flow matching can be computationally expensive because the quality of the final sample generally improves with more integration steps. MeanFlow: what is it and why is it important?
- May 4, 2026
Diffusion Models
Before diffusion models, image generation was largely dominated by GANs. This was despite their inherently poor training stability caused by their two competing networks. Diffusion models emerged as a more stable alternative that produce highly realistic images, and now, are the most popular model for image generation tasks. But what are diffusion models, and how do they work? What are diffusion models?