Blog
Notes on generative models, geometry, and making models run fast on accelerators.
Topics
All posts
- Diffusion-Series 3 - One Step: Consistency Models, Flow Maps, MeanFlow, and Distillation How to generate in a single network evaluation: the flow map that every 1-NFE method learns, the from-scratch line from consistency models to MeanFlow and drifting, the three families of distillation, and where the ImageNet numbers stand. 17 min read
- Diffusion-Series 2 - Flow Matching: Diffusion as a Velocity Field From DDIM's deterministic sampler to flow matching and rectified flow: what the network learns, why the loss works, how it relates to epsilon-prediction, and why straight training paths still give curved sampling paths. 8 min read
- Advanced Unified-Models Recording the reading progress of recent advanced unified-models and diffusion-LLMs 4 min read
- Diffusion-Series 1 - DDIM: Making Diffusion not Diffusioning This is a simple explanation of DDIM, a fast sampling method for diffusion models. It is based on the original DDPM, but it is not a diffusion model anymore. 4 min read