For about a year now, I’ve been creating LoRAs using the trainer that CivitAI itself provides to users, allowing us to bring some of our favorite characters into the world of AI. For a long time, I had wanted to create my own LoRAs because I often couldn’t find models for the characters I wanted to generate. So I started dedicating myself to completing CivitAI’s daily quests, and I even found myself buying Buzz just to keep training models instead of spending it all on image generation.
Over time, I realized I was wasting the potential of my own hardware, since I have a high-end PC. That made me start researching how to train my own models locally. I began with OneTrainer, and the learning curve felt very steep. It took real effort just to get my first model running. In fact, I ended up abandoning the project for about two months because at the time I didn’t understand that I needed to load a base model myself. But once I got past that point, it turned into a nonstop self-taught journey.
Not only did I realize that I could train better and higher-quality models locally, but I also discovered that I could train models with surprisingly good quality and size efficiency and even reduce their file size afterward so they wouldn’t be so heavy. A big shoutout and thanks to the person who made this article, which helped me a lot:
https://civitai.com/articles/10659/guide-lora-resizing
Because of my stubborn (and even perfectionist) personality, I ended up sticking with OneTrainer, mainly because Kohya_ss felt even more complex to me. Since then, I’ve read many tutorials and asked for advice from ChatGPT and other platforms in order to keep improving. And thanks to the results of a poll in my Discord server, people asked me to focus more on creating character LoRAs rather than just generating images.
The truth is that this is an art form that is difficult to master, because training a good LoRA is divided into very important stages:
Gathering source material in the highest quality possible, then collecting and preparing cropped images. (Easy difficulty)
Searching for and tagging every image carefully, making sure not to include tags that might cause the checkpoint to learn things you do not want it to learn. (Medium difficulty)
Masking the cropped images in order to prevent too much background from leaking into the training. (Medium to high difficulty — at least until very recently, as of the time of this post)
Training the LoRA and testing it afterward, with the goal of finding not only the right epoch but also a version that feels faithful to the character while still being flexible for the public to use. (Easy difficulty, but very time-consuming)
Each one of these steps can take me anywhere from 1 to 4 hours depending on the size of the dataset, but the full process itself takes days. Because of that, from now on, please expect all of my models to receive proper attention to quality and detail, so I can avoid bringing you LoRAs that suffer from overfitting, underfitting, or unwanted artifacts, objects, or extra characters appearing in the scene. I may even remake some of my older models so they can be updated to my current standards.
Of course, my models will never be perfect, because perfect models do not exist. But I do this altruistically and as a hobby, using the tools and resources available to me. I plan to share the datasets on my Patreon page in the future, and I will continue collecting and making more characters over time, and as AI keeps advancing every single day, I know they will be very useful, since better ways of training models will almost certainly keep appearing over time.
So if you enjoy my work, I would truly appreciate it if you supported, shared, and helped spread my models and creations whenever possible. Thank you very much, my dear followers.
TL;DR Expect to see even better, accurate and flexible models :3