train-character-lora

Train a Character LoRA (local, Flux.1-dev)

Overview

The trainer runs ostris ai-toolkit's run.py inside a headless GPU Docker container, driven through the three train_* MCP tools. You (the LLM) are the UI. Each takes an action: train_prepare_dataset owns the datasets, train_start owns the jobs, and train_doctor owns the trainer itself. You generate the dataset, launch the job, watch progress, and the finished LoRA lands in ComfyUI models/loras/ and the LoRA catalog without further steps.

Base model: FLUX.1-dev (the best proven character consistency; needs ~24GB VRAM with quantization, RTX 4090 class).
Phase-1 scope: character LoRAs only. Style/slider/edit and other bases come later.

The flow (tool sequence)

train_doctor {action:"doctor"}. Preflight once per session. Checks docker daemon, --gpus all GPU passthrough, trainer image, HF_TOKEN. If image:false, run train_doctor {action:"build_image"} (one-time, several minutes, since it builds CUDA plus torch plus ai-toolkit). If hfTokenSet:false, warn the user: the first run downloads FLUX.1-dev (gated HF repo) and needs HF_TOKEN in the MCP server env.
train_prepare_dataset {action:"prepare"}. Stage the images. See "Dataset" below.
train_start {action:"start"}. Launch. Returns a job id at once; training runs detached.
train_start {action:"status", id}. Poll progress (progress.step/totalSteps/loss, recent samples, log tail). Poll on a slow cadence (every few minutes). A 2000-step run is roughly an hour on a 4090. Don't block on it.
Done. status:"completed" means the .safetensors was copied to models/loras/<name>.safetensors and upserted into the LoRA catalog (result has the paths and catalog id). Verify by loading it in a Flux workflow (LoraLoaderModelOnly, strength 1.0) with the trigger word in the prompt.

Dataset guidance

Call train_prepare_dataset {action:"prepare"} with name, items: [{path, caption?}, ...] and a defaultCaption.

10 to 30 varied images of the subject: different angles, expressions, lighting, backgrounds, distances (close-up, half-body, full-body). Variety beats count.
Trigger word: pick something rare and stable (e.g. ohwx, zxc_person), NOT a real word. Use it as defaultCaption and pass it as trigger to train_start.
Captions: describe what changes between images (pose, setting, clothing, expression); the model learns the constant identity from the images themselves. Start each caption with the trigger word, e.g. ohwx person sitting in a cafe, laughing, natural light. Keep them short and factual. When in doubt, the trigger word alone (defaultCaption) is a workable baseline.
Images are copied and renamed img_00001.<ext> etc. Source files are never modified.

Params (sane defaults — override sparingly)

Param	Default	When to change
steps	2000	200 for a smoke test; 1500–3000 real runs. More ≠ better (overbake = plasticky).
lr	1e-4	5e-5 for a tighter/subtler identity.
rank	16	32 for very detailed characters.
resolution	[512,768,1024]	[512] if VRAM-constrained.
quantize	true	Keep true on 24GB.
saveEvery / sampleEvery	250	Lower (100) to watch early progress.

Monitoring & judgement

train_start {action:"status"}'s progress.samples are host paths. Look at them. (ai-toolkit prints no saved-sample lines, so they populate at finalize from the output dir; mid-run you can look directly in the job's output/<name>/samples/ folder.) Identity should be recognizable by ~1/3 of the run; if samples stay generic past halfway, the run will likely underfit. Cancel (train_start {action:"cancel", id}) and check captions and trigger.
Loss should trend down and stabilize (~0.1 to 0.3); wild spikes usually mean lr too high.
Checkpoints save every saveEvery steps under the job's output/ dir, so a cancelled run isn't a total loss.

Failure modes

no_docker / no_image from train_start {action:"start"}: run train_doctor {action:"doctor"}, follow its hints.
OOM / CUDA errors in the log tail: drop resolution to [512], keep quantize:true, batch stays 1.
handoff failed in job error: training itself finished; the LoRA is still under the job's output/<name>/ dir. Copy it into models/loras/ manually and upsert the catalog.
First run is slow before step 1. FLUX.1-dev download (~24GB) plus latent caching. As long as the log tail moves, it's fine. The HF cache persists across runs.

Sources

Official: none found.
Empirical: sampler values, wiring, and prompt notes from working graphs in packs/ and observed renders; not a vendor prompting guide.