My first LoRA training with OneTrainer
Stable Diffusion models may lead to “standarized” results. Images may be generated with the same face for different prompts. So, to obtain personalized results (a face with specific features, a specific dress/clothing, more detailed eyes, mouth, hands/feet, and so on) Stable Diffusion models (checkpoints) may need a LoRA (or more). This post is about my first LoRA training.
What is a LoRA
Very briefly (as I published a more detailed article about LoRA’s here): a LoRA is a small fine-tuning SD model that will help the base-model (or checkpoint) in generating specific features, like a specific character face (it could be an actor, yourself, an anime/cartoon hero), a specific style (like a photographic style, an anime style or a cartoon style, a painter style) or a specific kind of detail for the image (clothing style for example). It’s small (compared to the 7Gb size of a SDXL checkpoint), easy to train and very easy to apply to any SD base model you will use for your image generations.
OneTrainer
I used Nerogar’s OneTrainer (as a package of Stability Matrix) to train my first LoRA. The subject I wanted to train on was “Kiara Aigen”, a fictional AI generated character I “tested” on a couple of Socials (Instagram and Threads) for a few months (with excellent results by the way). At the time I generated most of the images I used with the FaceSwap technique, but I soon understood it was a terrible solution. The images generated had all the same identical expression, same face, there was no diversity.
I was an absolute beginner in the AI-generated image world. Reading online I found that the best solution would have been to train a specific LoRA that would allow me to generate consistent images of the same “person”. So, I checked the hundreds images I created for “Kiara Aigen” and chose the best 24 pictures I had. These were the base of my LoRA training. I will write a detailed guide of how to use OneTrainer very soon (the first draft is already written, I should publish it by the next week hopefully ), OneTrainer is probably the best tool available today for LoRA training. It’s easy, fast and has a great UI.
The results
I did a few test to get the best LoRA I wanted, and maybe I could still improve it a little. But the results are good and I am satisfied with what my LoRA is now. Maybe I will work for a v2.0 later on, but for now I am happy. I tested different Epochs values, different data-type values, different Batch-sizes, Learning rates and Ranks.
Below you can view three images I created using this LoRA.
The LoRA was trained on “Kiara Aigen” face. As I said before, Kiara is a fictional, AI generated, character I created in january 2024. She is around 22 years old with short red hair. In the images above I set the location of the images in New York, at sunset, during summer.
These are just a few examples, I am still testing the LoRA with different checkpoint trying to reach the most realistic results. But for the moment I manage to obtain different expression, face positioning and look.
I will publish more pictures of her later on in my Portfolio, in the OneTrainer gallery (not yet ready at the moment).
There is more to come… so stay tuned!
Leave a Reply