Fine-Tune FLUX.2 Klein for Custom Image Generation

Basic idea

This playbook shows how to fine-tune FLUX.2 Klein with multi-concept LoRA (Low-Rank Adaptation) using AI Toolkit for custom image generation on your hardware platform. Unified memory and GPU acceleration let you train and generate with the Diffusion Transformer, Qwen text encoder, and autoencoder.

Multi-concept LoRA teaches FLUX.2 Klein new concepts, characters, and styles. Trained LoRA weights drop into existing ComfyUI workflows for prototyping and experimentation. The same path supports high-resolution training and inference at 1024px and above.

What you'll accomplish

You'll have a fine-tuned FLUX.2 Klein LoRA that generates images with your custom concepts and is ready for ComfyUI workflows on your hardware platform.

• Fine-tune FLUX.2 Klein 4B or 9B with AI Toolkit
• Train on sample concepts (tjtoy toy and sparkgpu GPU) or your own dataset
• Run high-resolution (~1K) diffusion training and inference
• Integrate LoRAs into ComfyUI visual workflows
• Use Docker images for reproducible train and inference environments

What to know before starting

Required:

• Basic Linux command line and Docker container usage
• Familiarity with generative image concepts (prompts, diffusion models, LoRA)
• Hugging Face account and token for gated FLUX.2 Klein access

Optional:

• Prior ComfyUI experience (node graphs and workflow JSON)
• Experience preparing image datasets for concept training

Supported hardware platforms

Use the matrix below to confirm your hardware platform, recommended default local settings, and whether multi-node applies.

Prerequisites

Hardware requirements

• Supported hardware platform — see Supported hardware platforms matrix above
• Sufficient memory for FLUX.2 Klein training and 1024px inference
• Enough free disk for model downloads, the AI Toolkit cache, container images, and LoRA checkpoints
• No other heavy GPU workloads running during train or inference

Software requirements

• NVIDIA Docker / NVIDIA Container Toolkit: docker --version and nvidia-smi inside a GPU container
• Network access to Hugging Face for gated FLUX.2 Klein and text-encoder downloads
• Hugging Face access token (HF_TOKEN) with access granted on the FLUX.2 Klein model card for your system
• Web browser access to port 8188 for ComfyUI

Ancillary files

All required assets can be found in this playbook repository.

• assets/download.sh — Downloads the FLUX.2 Klein model, Qwen text encoder, and VAE into models/
• assets/Dockerfile.train / assets/launch_train.sh — Build and run AI Toolkit LoRA training
• assets/config/flux2_klein_4b.yaml / assets/config/flux2_klein_9b.yaml — Define datasets and training parameters
• assets/Dockerfile.inference / assets/launch_comfyui.sh — Build and run ComfyUI inference
• assets/flux_data/ — Sample multi-concept dataset
• assets/workflows/base_flux2.json — ComfyUI workflow for base FLUX.2 Klein inference
• assets/workflows/finetuned_flux2.json — ComfyUI workflow for LoRA-conditioned inference

Time & risk

• Estimated time: 4 HOURS, including first-time downloads, container builds, training, and inference
• Risk level: Medium
  • Docker permission issues may require a group change and new login session
  • Gated model access and large downloads can fail without a valid Hugging Face token or enough disk
  • Best results need hyperparameter tuning and a high-quality dataset
• Rollback: Stop and remove Docker containers; delete downloaded models, the AI Toolkit cache, and LoRA checkpoints if needed
• Last Updated: 10/21/2026
  • Fine-tune FLUX.2 Klein with AI Toolkit and ComfyUI inference on supported hardware platforms