All the above images are generated using my Teacache implementation with an L1 threshold of 6, and adopt the Euler A sampler and my Smooth scheduler.
muse dash style for Neta Lumina. Run it with the trigger word <style>muse dash</style>.
It is generated with Euler A and my Smooth under the Netayume model.
You can use my Smooth scheduler and Teacache implementation by adding the GitHub repository to your custom_nodes folder.
Description
Comments (5)
Is there a tutorial somewhere on how to train Loras on Lumina based models? I'd love to train some of my own, nice model BTW!
I trained the lora by cloning the original kohya's repo in the branch sd3 which supports lumina.
@spawner6 What was the vram requirement for it? if you don't mind me asking. Appreciate the LoRA btw.
@vesola3327205 I trained it with a 4090d 24g.
For those who can read Chinese, here’s a detailed tutorial on training common types of LoRA (not LyCORIS). Additionally, I’ve submitted a PR to the dev branch of the LyCORIS repo. When using this PR to train LyCORIS LoRAs, please note that you need to follow the installation method outlined in the dev branch’s README to set up Lumina’s LyCORIS support — specifically, run:
git clone https://github.com/KohakuBlueleaf/LyCORIS.git -b dev
After that, you can start your LyCORIS LoRA training as usual.
Important: Before you begin, make sure to run pip uninstall lycoris-lora to remove any previously installed version of the lycoris-lora library.