TRAINING GUIDE
In previous article we learn how to train your own model on Flux.
This time I'll share with you my experience creating a Illustrious model on-site training feature. After many models trained I finally found the best configuration settings that worked for me and I hope it works for you as well.
First thing first
Tensor.art Account
Set up an account on Tensor.art and ensure it has a minimum of 700 credits.
Dataset Guidelines
For optimal results, it is advisable to include between 10 - 30 images in your dataset. The majority of these should be close-up photographs, complemented by a few mid-range shots and at least one full-body image. Additionally, incorporating side profile images alongside front-facing ones will enhance the dataset's diversity.
Here is my dataset that I've used for this LoRa.
Guidelines for Captioning Photos
Tensor.art automatically put captions on your images but consider adding:
Unique Trigger Word: Begin each caption with a distinctive trigger word that identifies the dataset.
Also consider this on your captions.
Descriptive Tags: Include tags for all elements you wish to vary in the generated images. For instance, if you intend to modify attributes like hair color and length, ensure these are detailed in every photo caption.
Consistent Elements: Any aspect not explicitly described and that remains consistent across all photos in the dataset will automatically be associated with the trigger word
Parameter Settings and Likeness Capture
Utilizing a collection of approximately 10 portrait photographs, a highly effective and adaptable likeness model can be developed.
Parameter Guidelines
• Higher rates = stronger character features but potential loss in image quality
• Lower rates = better image quality but weaker character features
• Most character models work well with UNET around 0.0005 and TE around 0.00003
• 5-10 repeats, 10 epochs
In my case I want my model looks like picture but with a little variation in the style and face.
Knowing this I used the follow configuration settings for my model:
AdaFactor
Scheduler
Cosine
UNET Rate: 0.0005
Text Encoder (TE) Rate: 0.00003
DR: 64 AR: 32
So Here is my Parameter Settings on Tensor Art in-site training.
First Upload your Images
Note: Tensor Art Automatically tag your images
Next select Profesional Mode in the settings
Select the base model Illustrious
Next click on NetworkModule menu and select Dora (better than LoRa)
Write the trigger word
And then change the parameters to this
Next fill the positive and negative prompt of the image you want to generate as preview of your model while is training, and the image sise.
Finally, click on Start Training Now
When the training is done, here comes the hard part because you have to choose which one to publish on your profile or download.
Usually is recommend the last one. So click on publish and click on create a project
Choose a A Project Name, Channel and Tags related to your model
In the next window change the Version of your model
Scroll down and change Showcases and Permission Settings if you want to.
Finally click on Publish.
And that's it.
You can try it in your profile but you have to wait a few minutes until it’s deployed.
This are the results for my model
Here si the link if you want to try it!
https://tensor.art/models/803009896502202836/Merry-Illustrious-or-Christmas-Walkthrough-Final
I Hope everything goes fine and you enjoy the results.
Good luck and THANK YOU for reading this article!
