FLUX.1 [dev] Style LoRA Trainer: Training

Produce style-focused LoRA adapters on top of the FLUX.1 [dev] architecture. Upload a ZIP dataset with training images and optional captions. FLUX.1 [dev] Style LoRA Training is a training workflow for producing style-focused LoRA adapters on top of the FLUX.1 [dev] architecture. Upload a ZIP dataset containing images and optional text captions, specify a trigger word and step count, and receive downloadable safetensors model output you can plug straight into flux-1-dev-style-lora-inference.

Interactive model controls

Produce style-focused LoRA adapters on top of the FLUX.1 [dev] architecture. Upload a ZIP dataset with training images and optional captions.

📝

Overview

About this model

FLUX.1 [dev] Style LoRA Training is a training workflow for producing style-focused LoRA adapters on top of the FLUX.1 [dev] architecture. Upload a ZIP dataset containing images and optional text captions, specify a trigger word and step count, and receive downloadable safetensors model output you can plug straight into flux-1-dev-style-lora-inference.

1Produce style-focused LoRA adapters on top of FLUX.1 [dev]

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

💰

Pricing & Value

Cost analysis

Provider	Cost	Notes
	$2.50 per 1,000 steps ($0.0025 / step)	Best price & high performance

$2.50 per 1,000 steps ($0.0025 / 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.	`0.0005`

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 Value0.0005

📖

Implementation Guide

Developer documentation

❓

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 $2.50 per 1,000 steps.

Where do I use the trained LoRA?

Pass the returned `.safetensors` URL to flux-1-dev-style-lora-inference, or any other FLUX.1 [dev]-compatible LoRA inference endpoint.

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.