Train FLUX.1 Image Models Fast with Your Photos—No Coding Needed!

type

status

date

slug

summary

Key Steps of Training FLUX.1

Prepare training images

at least 10 high-quality facial photos taken from different angles and lighting conditions.

Select Trigger Word: Create a unique "trigger word" that will be used to activate the model later in the prompt text.

Create and train the model: Upload images and trigger words on Replicate for training, which takes about 20 minutes.

Generate images: Use the trained model to generate detailed description text containing trigger words to generate images related to the prompt.

Step 0: Preparation

Before you start fine-tuning your FLUX.1 model, you will need the following:

A Replicate account

A few photos of yourself as training pictures

$2-3 to cover model training costs

Step 1: Collect training images

You will need to collect several photos of yourself for training. It is best to use photos from different angles, lighting conditions, and backgrounds. It is recommended to have at least 10 high-quality facial photos , but at least 2 are required.

Image requirements:

Supported formats: WebP, JPG, PNG

Resolution: 1024×1024 or higher is recommended

File name: You can name it as you like. The file name does not affect the training.

Aspect ratio: Images can be square, landscape, or portrait

It is recommended to have at least 10 pictures. The more pictures, the better the effect.

Choose a variety of pictures, such as different backgrounds, clothing, lighting, angles, etc.

Preparation steps:

Store the images in a folder named, for example data.

Compress the folder into .zip a file named data.zip.

Step 2: Choose a unique trigger word

When fine-tuning the FLUX.1 model, you need to select a unique trigger word that will be used later in the image generation prompt text.

Trigger word requirements:

When choosing trigger words, keep the following in mind:

It should be unique, i.e. MY_UNIQ_TRGGR

think "vanity license plate", but of any length.

It should not be an existing word in any language, such as dog or cyberpunk

Don't use TOKcombine it with other tweaks .

because it will conflict if you want to

Capitalization doesn’t matter, but capital letters can help you differentiate your trigger words from the rest of the text in your prompt.

For my Zeke/Ziki-Flux spin, I chose ZIKI as the trigger word. Short, unique, and easy to remember.

Have you decided on a trigger word? Remember it, you’ll need it in the next step.

For example, the author used "ZIKI" as the trigger word in the example, and you can choose a similar unique character combination according to your preferences.

Step 3: Create and train the model

Next, you will upload the training images on the Replicate platform and start training the model. You can choose to use the web form for training, or automate the process through the API.

Web page training steps:

Visit the Flux tuning form.

Choose where to publish the model: You can choose to publish the fine-tuned model as public or private.

Upload training images: In input_images field, upload the compressed package created previously data.zip.

Enter Trigger Word: In trigger_word field, enter the unique trigger word you selected previously.

Select the number of training steps: The default setting is 1000 steps. Fewer steps may not be able to learn the concepts in the image well, while more than 1000 steps may be a waste of time and cost.

Click Create training to start training.

Step 4: Wait for training to complete

The training process is pretty fast, but still takes a few minutes. If you use ten images and 1000 steps, the whole process should take about 20 minutes. Use this time to get up from your computer, stretch your arms and legs, drink some water, etc.

When you come back, your model should be ready.

Step 5: Generate an image using a web page

Once training is complete, your fine-tuned model is ready to use. Enter the prompt sentence through the web form to generate an image.

Visit the Replicate platform web playground.

Enter the prompt: Include the trigger word you set previously in the prompt. For example:

"photo of ZIKI looking super-cool, riding a segway scooter"

The FLUX model works better with detailed prompts, so provide as much specific as possible.

Step 6: Generate an image using the API

If you don't want to manually enter the prompts on the web page to generate images every time, you can use Replicate's API to automate the generation process.

The Web Playground is a great place to start playing with your new model, but generating an image with every click can quickly become tiresome. Fortunately, your model is also hosted in the cloud and provides an API, so you can run it from code in the programming language of your choice.

When you run the model, you’ll see tabs for different languages, such as Node.js and Python. These tabs contain code snippets that show you how to construct the API call to replicate the exact parameters you just entered into the browser form.

Click the Node.js tab in the web playground to view the API code:

Training FLUX.1: in the web playground to view the API code

API code examples: You can find API code examples in different languages in the web playground, such as Python or Node.js. Here is a simplified Node.js code example:

With this API, you can generate images from your own programs.

Step 7: Generate better prompts using LLM

Sometimes, you may have difficulty coming up with a suitable prompt to generate an image. In such cases, you can use a language model to generate detailed prompts.

Prompt generation example: Use LLM to generate detailed prompts such as the following:

This generates some interesting prompt:

To get started writing your own prompts, check out Meta Llama 3.1 405b , a fast and powerful language model that you can use on the web or through the API like your own models:

Step 8: Iterate and share results

Now that you have a fine-tuned image generation model and a language model to help generate prompt words, it’s time to start playing around and generating interesting images.

If you need inspiration, check out the Flux Tweaks collection on Replicate to see what others have created.

By following the above steps, you can successfully fine-tune the FLUX.1 model and generate personalized images. If you encounter any problems, you can refer to Replicate's support page for help.

🔥

Server Speed, Redefined: Discover the future of RYZEN VPS with RackNerd