Train LoRA with PixAI
With PixAI's new LoRA training feature, you can simply upload the images of your favorite characters or art styles to customize your own LoRAs. Then, you will be able to teach the AI to generate more amazing images with the same essense!
LoRAs (Low-Rank Adaptation) are like add-on modifiers for AI models.
You can add loras while you generate to teach the AI to do a specific thing, such as a character, an art style, a pose, or an aesthetic.
Have you ever been frustrated when the AI doesn't seem to understand a character name or word you put in the prompts? That is because the AI model don't have that word in its vocabulary.
This is where LoRAs come in handy.
LoRAs will apply small changes on where the image and the prompt meet, which means that LoRA will help the AI understand specific prompt better during generation.
Similar to the process to training an AI model, to make an effective LoRA, we need to feed it a set of images to teach the LoRA about your desired character, style, or pose. This progress is called Training.
Through the images you provide, the LoRA will be able to 'recognize' a new word you enter in to the prompts and draw it accordingly.
Usually, training a LoRA is quite a inconvenient, multi-step process. But have made it a lot easier here on PixAI!
You may found this button on the side bar
or this button on top bar, generated tasks, Model Market, LoRA/Model selection pop window in generation panel↓
Once you click on them, they will bring you to the LoRA training page.
First, you need to choose your base model for the LoRA training.
Base model will affect the AI's understanding and the LoRAs' style.
After training this LoRA, what model will you use with it for image generation later? If you see your go-to model on the list, go ahead and choose that one.
The base model you choose will make your LoRA more compatible with it.
eg. LoRA trained with Moonbeam as a base model will work well when I use it with Moonbeam; but it may not work as well with other models.
If you are unsure, the more fundimental/original models (Like Anything V5) will make a more versatile LoRA.
We highly recommend choosing Anything V5 as the base model for anime style LoRA, and RealisticVisionV51 for realistic style LoRA
Give your LoRA an awesome name!
Is your LoRA about an anime character, a realistic character, a scenery, or a style?
Choose the option that best catagorize your LoRA as we will apply slightly different processing techniques depending on its purpose!
The trigger words you write down will be added in the prompts to activate this LoRA later on during image generation.
Remember that LoRAs will help an the AI learn new words or character names? You can think of trigger words as the vocabulary you wish the AI to learn.
We recommend writing down the name of your character or style, plus some basic traits or descriptions.
eg. ganyu, blue hair, horns
To train a LoRA, you will need to feed the AI a set of images to use as reference.
Astonishing reference images nearly always yields astonishing training results and the final result highly depends on the reference images you choose. Try spend a bit more time preparing them.
The AI will learn the common characteristics or the 'essence' in your images and associate it with the trigger word you put in.
Upload at least 10 images.
Your images shall only include one subject/character in a clearly and in-focus (not blurred) manner.
Try to include different images that includes your subject in different contexts, angles and lighting conditions.
We suggest using images with a short-side dimension of at least 1024 pixels.
Caution: Each training will cost 25,000 credits except special discount. Also, you won't be able to change training parameters (including reference images) after you submit the training task.
example page
You can check your LoRA training status in your Personal Profile Page. LoRAs you trained will show up in your Models/LoRAs tab. It will tell you if the LoRA training is completed or not.
If LoRA training is completed, it will show as 100%.
Otherwise, it will show as progressing percentage.
Use your newly trained LoRA for a task to test how your LoRA works.
Choose a model similar to your LoRA's base model:
For example, if you chose an anime-styled base model for your LoRA, please choose an anime-style model for your image generations also.
Furthermore, same with normal image genaration, the more specific prompts you describe the character or style trait in your LoRA, the better test result you may have with the LoRA.
You can choose to publish trained LoRA to PixAI community. Add a cover image for your LoRA and you're good to publish!
What is a LoRA
LoRAs (Low-Rank Adaptation) are like add-on modifiers for AI models.
You can add loras while you generate to teach the AI to do a specific thing, such as a character, an art style, a pose, or an aesthetic.
Have you ever been frustrated when the AI doesn't seem to understand a character name or word you put in the prompts? That is because the AI model don't have that word in its vocabulary.
This is where LoRAs come in handy.
LoRAs will apply small changes on where the image and the prompt meet, which means that LoRA will help the AI understand specific prompt better during generation.
Training a LoRA
What "Train" means
Similar to the process to training an AI model, to make an effective LoRA, we need to feed it a set of images to teach the LoRA about your desired character, style, or pose. This progress is called Training.
Through the images you provide, the LoRA will be able to 'recognize' a new word you enter in to the prompts and draw it accordingly.
Usually, training a LoRA is quite a inconvenient, multi-step process. But have made it a lot easier here on PixAI!
Where to Train a LoRA in PixAI
You may found this button on the side bar
or this button on top bar, generated tasks, Model Market, LoRA/Model selection pop window in generation panel↓
Once you click on them, they will bring you to the LoRA training page.
LoRA Training Page
1. Base Model
First, you need to choose your base model for the LoRA training.
What is a Base Model?
Base model will affect the AI's understanding and the LoRAs' style.
After training this LoRA, what model will you use with it for image generation later? If you see your go-to model on the list, go ahead and choose that one.
The base model you choose will make your LoRA more compatible with it.
eg. LoRA trained with Moonbeam as a base model will work well when I use it with Moonbeam; but it may not work as well with other models.
If you are unsure, the more fundimental/original models (Like Anything V5) will make a more versatile LoRA.
We highly recommend choosing Anything V5 as the base model for anime style LoRA, and RealisticVisionV51 for realistic style LoRA
2. Basic Settings
Name
Give your LoRA an awesome name!
Purpose
Is your LoRA about an anime character, a realistic character, a scenery, or a style?
Choose the option that best catagorize your LoRA as we will apply slightly different processing techniques depending on its purpose!
Trigger Words
The trigger words you write down will be added in the prompts to activate this LoRA later on during image generation.
Remember that LoRAs will help an the AI learn new words or character names? You can think of trigger words as the vocabulary you wish the AI to learn.
We recommend writing down the name of your character or style, plus some basic traits or descriptions.
eg. ganyu, blue hair, horns
3. Reference Images
To train a LoRA, you will need to feed the AI a set of images to use as reference.
Astonishing reference images nearly always yields astonishing training results and the final result highly depends on the reference images you choose. Try spend a bit more time preparing them.
The AI will learn the common characteristics or the 'essence' in your images and associate it with the trigger word you put in.
What kind of images shall I upload?
Upload at least 10 images.
Your images shall only include one subject/character in a clearly and in-focus (not blurred) manner.
Try to include different images that includes your subject in different contexts, angles and lighting conditions.
We suggest using images with a short-side dimension of at least 1024 pixels.
Caution: Each training will cost 25,000 credits except special discount. Also, you won't be able to change training parameters (including reference images) after you submit the training task.
Check Trained LoRA
example page
You can check your LoRA training status in your Personal Profile Page. LoRAs you trained will show up in your Models/LoRAs tab. It will tell you if the LoRA training is completed or not.
If LoRA training is completed, it will show as 100%.
Otherwise, it will show as progressing percentage.
Test Your Result
Use your newly trained LoRA for a task to test how your LoRA works.
Choose a model similar to your LoRA's base model:
For example, if you chose an anime-styled base model for your LoRA, please choose an anime-style model for your image generations also.
Furthermore, same with normal image genaration, the more specific prompts you describe the character or style trait in your LoRA, the better test result you may have with the LoRA.
Publish Trained LoRA
You can choose to publish trained LoRA to PixAI community. Add a cover image for your LoRA and you're good to publish!
Updated on: 14/03/2025
Thank you!