I've been messing around trying to figure out LoRA training recently and decided to train a style LoRA on the original Marsey Telegram stickers.
Should more flexible than MarseyGen since this is trained on a newer uncensored model, but it's booru tag-based so it's not as easy to prompt.
Recommended Settings:
Requires a UI that supports SDXL V-Pred models such as reForge/Comfy UI
Model: NoobAI-XL V-Pred-1.0 (may work ok with other NoobAI based models)
LoRA strength: 1.0
Sampler: Euler A
Scheduler: Beta
Steps:
30CFG Scale: 5
CFG Rescale: 0.5-0.7
Zero SNR: On
Use the following tags to generate images of Marsey:
domestic cat, white fur, orange fur, black eyes, striped tail
Full example prompt:
masterpiece, best quality, marseystyle,
solo, domestic cat, white fur, orange fur, black eyes, striped tail, sitting, :D, blush, black background
<lora:marseystyle:1>
Negative prompt: worst quality, low quality
Some example images (with metadata from reForge UI):
Jump in the discussion.
No email address required.
your examples for neg prompts are shit. learn 2 prompt.
apart from that, good job!
did you use a guide for training the lora? can you link it?
Jump in the discussion.
No email address required.
I skimmed through a few guides and this was probably the best/most up-to-date one I found:
https://civitai.com/articles/3522/valstrixs-crash-course-guide-to-lora-and-lycoris-training
Looking at the metadata of loras other people trained was also useful to get an idea of what settings work for a specific model:
https://xypher7.github.io/lora-metadata-viewer/
Still a decent amount of trial and error though, adjusting the settings/captions and such and re-running the training.
Jump in the discussion.
No email address required.
More options
Context
More options
Context