Train a Personal LoRA with fal.ai and Flux
Introduction
AI personalization is now within everyone’s reach. Using LoRA (Low-Rank Adaptation), you can customize powerful AI models with just a few personal images and modest computing resources. Unlike traditional fine-tuning, LoRA efficiently trains a small set of parameters while keeping the base model frozen — perfect for individual creators without specialized hardware.
I’m big fan of Superhero movies and always wanted to see myself as one. I’m addicted to the text to image models since the start and an option to create images with myself is them is a blessing.
Thanks for reading Shmulik’s Substack!
In this guide, I’ll walk you through training a personalized LoRA using the fal platform and Flux model to transform yourself into superhero personas. By the end, you’ll be creating AI-generated content featuring your likeness with impressive accuracy.
As of the time of writing this, there is a free promo code of 50$ for fal.ai that you can use for training a few Lora Models for Free!
Go to aiengineerpack.com to redeem it.
Let’s transform you into your own AI superhero!
Shmulik, The unbreakable. My favorite result
What You’ll need
10–20 photos of yourself: Different angles and expressions for best results
fal.ai account : Sign up at fal.ai
LLM access (like Claude or ChatGPT): For generating creative superhero descriptions
$5–20 budget : For LoRA training costs (of multiple models), or use the free $50 credit from aiengineerpack.com (if you can)
That’s it! The fal.ai platform handles all the technical aspects, from computing resources to model integration. Just upload your images, set up training (15–30 minutes), wait for completion (15–30 minutes), and start creating!
Understanding the Components (Optional Section)
If you’re curious about the technology behind your AI superhero transformation, here’s a brief explainer of the key components:
Flux Model : A state-of-the-art image generation model similar to Stable Diffusion, optimized for high-quality, detailed outputs. It excels at creating realistic human features, making it perfect for superhero transformations.
LoRA (Low-Rank Adaptation) : A fine-tuning technique that allows you to customize large AI models efficiently. Rather than retraining the entire model (which would require massive computing resources), LoRA trains a small set of parameters that “steer” the base model toward your specific appearance.
fal Platform : A cloud service that simplifies AI deployment, offering pre-configured environments for training and running models without dealing with complex setup processes. It provides the necessary GPU resources on demand, making high-end AI accessible to everyone.
You don’t need to understand these technical details to successfully create your AI superhero — the fal interface guides you through the process step by step — but knowing the basics helps you make better choices during setup and prompt engineering.
Step 1: Gathering Your Training Data
Collect 10–20 clear and high quality photos of yourself with varied angles, expressions, and lighting (Google Photos could be great tool for helping with it).
Avoid group shots, heavy filters, hats or sunglasses. Here are some examples for good and bad photos (of myself, of course):
Up — Good photo(high res, alone, clear background), Down— Bad Photo (low res, with my lovely wife, with hat and noisy background)
Step 2: Training Your Personal LoRA
There are multiple train style to train LoRA with flux, I chose the next three with the same photos and the trigger word Shmulik:
flux-pro-trainer, flux-lora-fast-training, flux-lora-portrait-trainer
The pricing is listed on each train style and is based on the steps param. For the first model I suggest to just use the default parameters, if you are not satisfied, you will be able to try again with different params and photos.
With the default steps param. the training should take about 15–30 minutes for each train style with price of 12$. During the process, you can follow the training progress with the logs of your run or take a break until it finishes.
Step 3: Check your LoRA Model
After the train finished on any of the train styles, you will have the model on your Training history. It’s possible to take the generated LoRA model and use it locally or on another service, but in this tutorial we will stay in fal.ai.
After Clicking on Run Inference, you will just to a flux-lora with your own trained LoRA.
Try a new prompts with your trigger word and see what you get (Don’t be afraid if they aren’t the best yet, the detailed prompt will help a lot).
I started with the simple prompt Shmulik Looking great (like always) and got the photo:
Shmulik Looking great (with suit)
It showed me that the LoRA worked but it’s not the best yet.
After some more failed attempts, I tried to create myself as captain america, with the prompt Envision a fully body portrait of Shmulik as captain america and got the photo:
Envision a fully body portrait of Shmulik as captain america
which showed me that the model itself it great, but there are some bad photos in my dataset (these with the hat). I cleaned up some photos from the dataset, added more good photos and trained again.
Step 4: Creating Your Own Superhero Images
For generating the superhero prompots with the trigger word, I used the ChatGPT, you can use it or any other strong LLM Claude or Gemini.
I used very simple prompt —
“Please write a detailed superhero prompt for me about Shmulik for the image to text model flux”
You should use your own trigger word and you can use this promps or can more complex one
With this simple prompt (after I told him to remove the mask), I got:
Superhero Image Prompt — Shmulik, The Unbreakable
A powerful superhero named Shmulik stands atop a futuristic skyscraper, overlooking a cyberpunk cityscape filled with neon lights and towering structures. His face is fully visible — strong and determined, with piercing blue eyes that radiate confidence and wisdom. His short, dark hair is slightly tousled by the wind, and a faint scar runs across his cheek, hinting at past battles.Shmulik wears a high-tech armored suit in deep blue and silver, crafted from an advanced extraterrestrial alloy that shimmers under the city lights. The suit is sleek yet powerful, with glowing energy lines running along his arms and chest, pulsating as he channels his abilities. His broad shoulders and strong stance exude an aura of unstoppable resilience.
A flowing midnight-blue cape, lined with golden circuitry, billows behind him, reacting dynamically to the wind. His right hand is clenched into a fist, crackling with energy, while his left is extended, forming a shimmering, translucent force field. His superhuman abilities allow him to generate unbreakable barriers and manipulate energy with precision.
Above him, the stormy sky is alive with lightning, casting dramatic highlights on his face. The neon glow of the city reflects off his armor, adding a futuristic edge to the scene. The towering buildings around him display digital billboards, some showing cryptic messages and others featuring the emblem of Shmulik — a golden shield with an infinity symbol at its core.
The entire composition captures a sense of heroism, mystery, and raw power, making Shmulik an awe-inspiring figure of justice and strength.
That generated for me the first image —
Shmulik, The unbreakable.
After you have one, you can generate many more by just telling him to make another one with the same trigger word. For each prompt I used the LoRAs of all the train styles and saved the best. The Inference price is much lower, compared to the training price.
With This Approach I got the next images:
Shmulik, Guardian of the Infinite Scroll
Shmulik, The Clockwork Sentinel
Shmulik, The Thunderheart
That’s it! You’ve now created a customized AI model that can transform you into any superhero character you can imagine.
Summary
In this tutorial, we learned how to train a personal LoRA for flux with fal.ai. I hope you will try it yourself and send me some of your best creations, superhero or whatever else you want to do with it.
Enjoy!
Thanks for reading Shmulik’s Substack!
Originally published on AI Superhero
Top comments (0)