skills.rest
artokun
artokunCOMMUNITY
Explore all skills by this author →
Website𝕏@artokunGitHub

train-character-lora

Trains character LoRA models on FLUX.1-dev using local GPU Docker containers.

715|111|Updated Feb 15, 2026
Search Tags:#character#docker#training#GPU#ComfyUI#FLUX#Lora
One-click install
npx skills add https://github.com/artokun/comfyui-mcp --skill train-character-lora
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: train-character-lora
Source: https://github.com/artokun/comfyui-mcp/tree/main/plugin/skills/train-character-lora
Command: npx skills add https://github.com/artokun/comfyui-mcp --skill train-character-lora

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Training a custom character LoRA normally requires deep knowledge of training frameworks, dataset preparation, and GPU configuration. This Skill automates the entire pipeline through three MCP tools, letting you train an identity LoRA from your own photos on a local GPU without manual setup.

Core Features & Use Cases

  • Dataset Preparation: Stage 10-30 subject images with captions and a trigger word via train_prepare_dataset, with automatic renaming and no modification of source files.
  • Managed Training Jobs: Launch detached training runs with train_start, monitor progress, loss, and sample images, and cancel underperforming runs.
  • Environment Preflight: Verify Docker, GPU passthrough, trainer image, and HF_TOKEN with train_doctor before starting.
  • Use Case: A user provides 20 photos of a person, and the Skill trains a FLUX.1-dev LoRA that lands directly in ComfyUI's models/loras/ folder, ready to load with LoraLoaderModelOnly in any Flux workflow.

Quick Start

Train a character LoRA from the photos in my dataset folder using the trigger word ohwx and monitor the training progress until it completes.

Frequently Asked Questions about train-character-lora

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I train a character LoRA locally on my own GPU?▼

Use the train_prepare_dataset tool to stage 10-30 varied images with a trigger word, then launch training with train_start. The job runs detached in a Docker container and the finished LoRA is copied into ComfyUI's models/loras folder automatically.

How many images do I need to train a character LoRA?▼

Use 10 to 30 varied images showing different angles, expressions, lighting, backgrounds, and distances. Variety matters more than count, and each caption should start with a rare trigger word like ohwx.

What GPU do I need for FLUX.1-dev LoRA training?▼

An RTX 4090-class GPU with about 24GB VRAM is required when quantization is enabled. If you hit OOM errors, drop the resolution to [512] and keep quantize set to true.

Why does my LoRA training fail with no_docker or no_image errors?▼

These errors mean the Docker daemon or trainer image is missing. Run train_doctor with the doctor action to diagnose, then use the build_image action to build the CUDA, torch, and ai-toolkit image once.

How do I know if my LoRA training run is going well?▼

Poll train_start with the status action and inspect the sample images. Identity should be recognizable by about one third of the run, and loss should trend down and stabilize around 0.1 to 0.3.

Can I train style LoRAs or use base models other than FLUX.1-dev?▼

No, this Skill is scoped to character LoRAs on FLUX.1-dev only. Style, slider, and edit LoRAs or other bases like WAN and Z-Image are handled separately through the ai-toolkit-trainer skill.

Related Skills

Explore More Design & Creative Skills→
raydocsraydocs
1

photo-standardization

Standardize photo capture with lighting, angle, and distance guidance using Swift and Vision.

Community
Basic
MCKRUZMCKRUZ
86

comfyui-lora-training

Automate dataset preparation, training configuration, and evaluation for character LoRAs in ComfyUI.

Community
Intermediate
Mathews-TomMathews-Tom
310

gpu-optimizer

Optimize PyTorch and CUDA training on NVIDIA GPUs with 8–24GB VRAM.

Community
Advanced
inference-shinference-sh
688

flux-image

Generate images with FLUX models via the inference.sh CLI.

Official
Intermediate