After a couple of painful weeks of trying to get Lumina 2 LoRA training using tools such as Kohya, Diffusion Pipe, and AI Tookit, I finally present you the first fully working concept LoRA for Lumina 2 (and the first LoRA for Neta Lumina).
This LoRA teaches the cat loaf sitting techniques to cats. Without the LoRA, cats will struggle to loaf in most positions
Trained Models:
Lumina Image 2: neta-art/NetaAniLumina_Alpha Round NNNN EP7 S160000 @ 1024
Developed by: Mr Pops Alot
Model type: LoRA
License: Apache 2.0
Features
Turn cats into cute loafs
Consistent results across different angles, aspect ratios, and art styles
How to Use
The key trigger phrase is: cat loaf
Bias, Risks, and Limitations
This LoRA was made as a test lora to test out training on Lumina 2. While intended to be used as a final product, the LoRA's limited image count may cause issues on out of scope use cases.
Training Details
64 images have been used in the dataset, with another 60 images for regularization. Images used are a mix of own sourced photorealisic pictures and drawn art scraped from Danbooru.
No synthetic data has been used for the model. Tagging (Danbooru tag style) is all manually done.
Training Procedure
Lumina 2
Program: Diffusion Pipe
Optimizer: AdamW
LoRA type: LoRA
Steps: 140
Precision: Full BF16
Hardware: 1x RTX 3090 24GB
FAQ
Will this model be ported to X? This model will be trained on various models when I feel like training on them.
License
This LoRA model is provided under the Apache License 2.0 license.
Special Thanks
Ly, and others in the NoobAI discord to for helping me train a LoRA on Lumina and for emotional support.