# I trained a tattoo AI model with Replicate, share the experience!
## Causes
I've been working on AI-assisted design lately, and it just so happens that I'm particularly interested in tattoo design. After a period of fiddling around, a pretty good tattoo design model was trained using the Replicate platform. Currently it has been integrated into the platform.
In fact, at first I was also torn between training my own model, after all, there are quite a few ready-made models on the Internet. But after using it for a period of time, I found that the generic model is still a little less interesting to understand the tattoo design, especially in the processing of some details. So just thought, instead of using other people's models to make do, why don't you train one yourself.
## Why train your own models?
Honestly, it's mainly for these reasons:
1. **More understanding of tattoos**: Ordinary models may not quite understand the special requirements of tattoos, such as the smoothness of the lines, the layering of shading
2. **Good Adjustment**: change your own model as you wish, without being limited by others' settings
3. **Better results**: after specialized training, the tattoo design area is indeed much better than the generic model
4. **It's actually pretty cheap**: one training session cost less than $2, which is a pretty good deal!
## Hands-on tutorials
If you'd like to try it too, I've organized the steps:
### First step: preparation
Need to prepare these:
- Replicate account
- 20-30 high quality images of tattoos
- Around $2 for training
### Step 2: Prepare the picture
1. **Collection of photographs**:
- Pick your favorite tattoo style
- Pictures should be clear
- It's better to have a unified style.
2. **Organization of documentation**:
```bash
# Create a folder for pictures
mkdir training_data
# Pack it up
zip -r training_data/*
```
### Step three: start training ###
You can use web pages or code to train:
```python
import replicate
# Build a new model
model = (
owner="Your username",
name="tattoo-style-lora",
visibility="public",
description="Tattoo design model"
)
# Start training
training = (
version="ostris/flux-dev-lora-trainer:4ffd32160efd92e956d39c5338a9b8fbafca58e03f791f6d8011f3e20e8ea6fa",
input={
"input_images": open("", "rb"),
"steps": 1000,
"trigger_word": "TATTOO_STYLE" # Trigger word, random name
},
destination=f"{}/{}"
)
```
### Step 4: Test the results
Once you're trained, try a cue word like this:
```
A TATTOO_STYLE design of a dragon, black and grey style, detailed linework
```
## Some tips and tricks ##
1. **Regarding the selection of maps**
- Try to pick one that's as close in style as possible
- Picture quality should be good
- Get a few more samples of angles
2. **Training parameters**
- Just use the default learning rate.
- 1,000 steps is almost enough for training.
- All the other parameters are fine with the defaults.
## Actual results ##
If you want to see the effect, you can just go ahead and try it. We've put this model up, type in the style of tattoo you want and you'll see the results.
[Hot TattooI](/)
## Written at the end ##
To be honest, AI-assisted design is really quite convenient nowadays, but it still depends on the designer to control the final result. This model is also an auxiliary tool to help provide some inspiration and reference.
If you are interested, you can try the free [AI Tattoo Generator](/)
, those who want to train on their own can also try following this article. Feel free to share any problems you encounter!
Fine-tune FLUX.1 with your own images - Replicate blog