Skip to content

README.md

Latest commit

 

History

History
161 lines (119 loc) · 5.8 KB

README.md

File metadata and controls

161 lines (119 loc) · 5.8 KB
Raw

FLUX.2 Klein LoRA Training and Image Generation

Train your own AI image models with just 20-50 photos, then generate unlimited new images using Black Forest Labs' fastest model.

How it works

These AWS Deadline Cloud job bundles use diffusers and peft to fine-tune FLUX.2 Klein with LoRA (Low-Rank Adaptation), a technique that creates small, efficient model adapters from your images.

Workflow:

  1. Prepare Training Data - Collect images of your subject in a directory (optionally with caption files)
  2. Train LoRA - Submit training job using the lora_training bundle
  3. Download Training Output - Download the trained LoRA weights
  4. Generate Images - Submit generation job using the image_generation bundle with your trained LoRA

Prerequisites

  • AWS Deadline Cloud farm with a GPU-enabled queue (Linux fleet with NVIDIA GPU)
  • Deadline Cloud CLI installed

Job bundle index

This table covers every immediate job bundle in flux2_klein_lora/.

Sample What it demonstrates Start here when
LoRA training Fine-tuning FLUX.2 Klein from an image-and-caption dataset You need to create a reusable adapter for a subject or style
Image generation Parallel inference with a trained LoRA adapter You have LoRA weights and want to generate a set of images

Bundle details

1. lora_training

Train custom LoRA adapters for FLUX.2 Klein models using your own image datasets.

Fleet requirements:

  • GPU with 13GB+ VRAM
  • 64 GiB+ system memory (Recommended)
  • Linux OS

Key Parameters:

  • Model Version: flux.2-klein-base-4b or flux.2-klein-4b
  • Dataset Path: Local directory containing training images (.jpg, .png, .jpeg, .webp)
  • Instance Prompt: Text describing your training images (e.g., "a photo of ohwx dog")
  • Resolution: Training resolution (default: 512, use 512 for 24GB VRAM)
  • Network Dim: LoRA rank (default: 16)
  • Network Alpha: LoRA alpha scaling (default: 16)
  • Max Training Steps: Number of training iterations (default: 1500)
  • Output Directory: Where to save trained LoRA weights

Example: Use the job bundle GUI submitter to select parameter values:

deadline bundle gui-submit ./lora_training

Or, use the CLI submitter (specify --queue-id for GPU queue):

deadline bundle submit ./lora_training \
  --queue-id <gpu-queue-id> \
  --parameter DatasetPath=~/training_images \
  --parameter InstancePrompt="a photo of ohwx dog" \
  --parameter OutputDir=/tmp/lora_output \
  --parameter MaxTrainSteps=1500 \
  --parameter Resolution=512

Output: LoRA weights saved as flux2_klein_lora.safetensors with embedded metadata, plus checkpoints every 300 steps.

Download Output: After training completes, download the LoRA weights to use in generation:

deadline job download-output --job-id <training-job-id> --queue-id <gpu-queue-id>

2. image_generation

Generate images using FLUX.2 Klein with your trained LoRA adapter.

Fleet requirements:

  • GPU with 13GB+ VRAM
  • 64 GiB+ system memory (Recommended)
  • Linux OS
  • Trained LoRA adapter from lora_training job

Key Parameters:

  • LoRA Path: Path to trained LoRA .safetensors file
  • Prompt: Text description of image to generate (include your trigger word, e.g., "ohwx")
  • Number of Images: Total images to generate (parallelized across workers)
  • Width/Height: Output dimensions (default: 1024x1024)
  • Inference Steps: Denoising steps (default: 50)
  • Guidance Scale: CFG scale (default: 4.0)

Example: Use the job bundle GUI submitter to select parameter values:

deadline bundle gui-submit ./image_generation

Or, use the CLI submitter:

deadline bundle submit ./image_generation \
  --queue-id <gpu-queue-id> \
  --parameter LoRAPath=/tmp/lora_output/flux2_klein_lora.safetensors \
  --parameter Prompt="a photo of ohwx dog wearing a tuxedo" \
  --parameter OutputDir=/tmp/generated_images \
  --parameter NumImages=4

Output: PNG images saved as image_0001.png, image_0002.png, etc.

Download Output:

deadline job download-output --job-id <job-id> --queue-id <gpu-queue-id>

Model variants

Model Parameters Best For
flux.2-klein-base-4b 4B Fine-tuning, commercial use
flux.2-klein-4b 4B Fast inference (4 steps)

Captions

Each training image needs a matching .txt caption file. You have two options:

Option 1: Auto-generated (default) If no caption files exist, the script creates them using your Instance Prompt. Every image gets the same caption (e.g., "a photo of ohwx dog"). Simple but limited.

Option 2: Custom captions (recommended for quality) Provide your own .txt files alongside images:

training_images/
├── IMG_001.jpeg
├── IMG_001.txt    # "a photo of ohwx dog sitting on grass"
├── IMG_002.jpeg
├── IMG_002.txt    # "a photo of ohwx dog running on beach"
└── ...

Custom captions teach the model more precise associations. Include your trigger word (e.g., "ohwx") in each caption.

Training tips

  1. Dataset size: 20-50 high-quality images work well
  2. Resolution: Use 512 for 24GB VRAM GPUs; higher resolutions require more memory
  3. Steps: Start with 1500 steps, increase if underfitting
  4. Network dim: 16 is a good default; increase to 32 for complex concepts
  5. Learning rate: 1e-4 works well for most cases

References