diffusion-pipe for production training
author of the fork, 2026
This fork converts diffusion-pipe, a tool for training image and video generation models, from a local script into an installable package that takes Hugging Face model and dataset identifiers directly.
The diffusion-pipe software trains image and video generation models, and upstream runs as a script on a single machine using local files. This fork is an installable package that accepts Hugging Face model and dataset identifiers directly, uses the ComfyUI LTS package in place of a bundled ComfyUI git submodule, and adds what multi-node training requires: Kubernetes launchers, experiment tracking, a shared dataset cache, node-local weight staging, and network pinning for training over Thunderbolt.