A Kohya_SS inspired LoRA trainer because everything else is dead or broken.
The AI space is full of abandoned projects and "solutions" that break if you look at them wrong. It's a nightmare to find an SD trainer that actually works with modern workflows and enough friendly to not scare people away.
So I built Onika. It works, heavily inspired by Kohya_SS, but actually flexible for the future.
Early access. Expect chaos as I don't have $$$ Video Card to test every function one by one
It works, but it's not perfect yet. You'll find bugs. You'll have ideas. Tell me about them. I'm building this for us, so your feedback actually matters. I might even fix it.
Onika is a powerhouse for training LoRA, Lycoris, and LoHa models.
- The Good Stuff: Supports SD1.5, SDXL, SD3.0, SD3.5, Flux, and Flux2.
- Runs on anything: Whether you have a 24GB beast or a potato 4GB card, Onika adapts. Even supports CPU training if you hate yourself.
- Smart Optimization: It has dedicated button and detects your hardware and optimizes memory/speed automatically.
- Dynamic Presets: It auto-generates presets for your dataset. You can also manage your own. Manual config is for professionals also.
- Auto-Updates: Includes an update checker so you don't have to manually git pull. It updates itself.
Your mess, your problem.
I provide the tool; you provide the data. You are responsible for your datasets and what you generate. Don't do illegal stuff. I'm just the dev.
If this tool saves you from a mental breakdown, consider supporting the project.
Maintenance is hard work and I'm doing this for free. Seeing people use it keeps me motivated to not let this end up in the "abandoned" pile.
.
