FLUX.2 [klein] 4B Style LoRA Trainer: Training

Produce style-focused LoRA adapters on top of the FLUX.2 [klein] 4B architecture. Upload a ZIP dataset with training images and optional captions. FLUX.2 [klein] 4B Style LoRA Training is a training workflow for producing style-focused LoRA adapters on top of the FLUX.2 Klein 4B architecture. Upload a ZIP dataset containing images and optional text captions, specify a trigger word and step count, and receive downloadable safetensors model output with instant platform AIR deployment.

Interactive model controls

Produce style-focused LoRA adapters on top of the FLUX.2 [klein] 4B architecture. Upload a ZIP dataset with training images and optional captions.

📝

Overview

About this model

FLUX.2 [klein] 4B Style LoRA Training is a training workflow for producing style-focused LoRA adapters on top of the FLUX.2 Klein 4B architecture. Upload a ZIP dataset containing images and optional text captions, specify a trigger word and step count, and receive downloadable safetensors model output with instant platform AIR deployment.

1Produce style-focused LoRA adapters on top of FLUX.2 Klein 4B

2Zipped dataset training with downloadable safetensors output and instant AIR deployment

💰

Pricing & Value

Cost analysis

Provider	Cost	Notes
MuAPI	$4.50 per 1,000 steps ($0.0045 / step)	Best price & high performance

MuAPI$4.50 per 1,000 steps ($0.0045 / step)

Best price & high performance

** Competitor pricing is estimated based on similar model architectures and usage tiers.

⚙️

Technical Details

Configuration schema

Parameter	Type	Description	Default
Dataset Zip URL	string	URL to ZIP archive containing training images and optional .txt caption files.	`undefined`
Trigger Word	string	Word or phrase used to activate the trained concept at inference time.	`my_style`
Training Steps	int	Total number of optimization steps to run during training.	`1000`
Learning Rate	float	Step size applied at each training update.	`0.00005`

Dataset Zip URLstring

URL to ZIP archive containing training images and optional .txt caption files.

Default Valueundefined

Trigger Wordstring

Word or phrase used to activate the trained concept at inference time.

Default Valuemy_style

Training Stepsint

Total number of optimization steps to run during training.

Default Value1000

Learning Ratefloat

Step size applied at each training update.

Default Value0.00005

📖

Implementation Guide

Developer documentation

How to Use FLUX.2 [klein] 4B Style LoRA Trainer

Prepare Your Dataset
- Collect 10-50 images representative of the style or subject you want to train, and zip them (optionally with matching .txt caption files).
Host the Dataset
- Upload the .zip to a publicly reachable URL and set it as dataset.
Choose a Trigger Word
- Pick a short, unusual word or phrase as trigger_word to activate the trained concept at inference time.
Set Training Parameters
- training_steps: default 1,000, adjustable up to 4,000. More steps generally improve fidelity but cost more.
- learning_rate: default 0.0005; lower values train more conservatively.
Submit and Wait
- Training is asynchronous. Poll the status endpoint until complete.
Use the Trained LoRA
- Pass the returned .safetensors URL to a FLUX.2 [klein] 4B-compatible LoRA inference endpoint.

❓

Common Questions

Frequently asked

How many training images do I need?

10-50 images work well. Prioritise variety in pose/lighting for a subject, and consistency in aesthetic for a style.

Do I need caption files?

No, they're optional. If included, name each `.txt` file identically to its image.

How many training steps should I use?

1,000 steps is a reasonable starting point; up to 4,000 for higher fidelity at higher cost, priced at $4.50 per 1,000 steps.

How does this differ from the other FLUX.2 [klein] trainer size?

This is the 4B architecture tier. The other size trades off quality against training/inference cost — check the sibling FLUX.2 [klein] trainer for the alternative.

What is the trigger word for?

It's the token you include in inference prompts to activate the trained style or character. Choose something short and unlikely to collide with normal vocabulary.