// the find
kijai/ComfyUI-FluxTrainer
A ComfyUI custom node that wraps kohya-ss's sd-scripts to let you train Flux LoRAs, and experimentally full finetunes and LyCORIS, without leaving ComfyUI. It's for people who already have a ComfyUI inference setup and want to train on the same models/python env instead of standing up a separate training pipeline.
Lets you reuse the exact model files and python environment you already have for inference, so there's no second toolchain to maintain or version-match. Builds on sd-scripts, LyCORIS and prodigy-plus-schedule-free rather than reimplementing training from scratch, so the underlying math is coming from actively-used implementations. The node-graph format means you can wire up multiple training configs side by side and compare runs visually instead of juggling config files.
The author flags upfront that they have little training experience and that the shipped defaults are just whatever happened to work on their own dataset, not tuned values - and explicitly says not to ask for help, so you're on your own if something breaks. Full finetune support is marked untested, only the LoRA path has actually been exercised. There's a hf_token.json file sitting in the repo tree, which is the kind of thing that shouldn't be committed at all and is worth checking before you install. No tests and no CI beyond a publish workflow, so nothing catches drift when upstream kohya/LyCORIS code changes underneath the vendored copies here - and the last push was April 2025, so it's been stale for a while in a space that moves fast.