train-character-lora

1# Train a Character LoRA (local, Flux.1-dev)
2 
3## Overview
4 
5The trainer runs ostris ai-toolkit's `run.py` inside a headless GPU Docker container,
6driven through the three `train_*` MCP tools. You (the LLM) are the UI. Each takes
7an `action`: `train_prepare_dataset` owns the datasets, `train_start` owns the jobs, and
8`train_doctor` owns the trainer itself. You generate the dataset, launch the job, watch
9progress, and the finished LoRA lands in ComfyUI `models/loras/` and the LoRA catalog
10without further steps.
11 
12- Base model: FLUX.1-dev (the best proven character consistency; needs ~24GB VRAM with
13  quantization, RTX 4090 class).
14- Phase-1 scope: character LoRAs only. Style/slider/edit and other bases come later.
15 
16## The flow (tool sequence)
17 
181. `train_doctor {action:"doctor"}`. Preflight once per session. Checks docker daemon,
19   `--gpus all` GPU passthrough, trainer image, HF_TOKEN. If `image:false`, run
20   `train_doctor {action:"build_image"}` (one-time, several minutes, since it builds CUDA
21   plus torch plus ai-toolkit). If `hfTokenSet:false`, warn the user: the first run
22   downloads FLUX.1-dev (gated HF repo) and needs `HF_TOKEN` in the MCP server env.
232. `train_prepare_dataset {action:"prepare"}`. Stage the images. See "Dataset" below.
243. `train_start {action:"start"}`. Launch. Returns a job id at once; training runs
25   detached.
264. `train_start {action:"status", id}`. Poll progress (`progress.step/totalSteps/loss`,
27…