This is a LoRA model for MINA from “ROBOT ON THE ROAD.”
It was trained using 125 images (images from the artbook and variations created using nanobanana based on those images).Number of steps: 1,500
This is my first LoRA. I tried to avoid overfitting, but if you have it generate content involving sex, the composition might sometimes be off. Also, it may not work well if you only use Danbooru tags. It seems to work better when using natural language.
Since this is one of my favorite characters, I plan to keep updating it in the future.
--------------------Ver 1.1 Update----------------------------------------------------------
I reviewed the parameters and retrained the model. It appears that the issue with prompts not working well, which was present in V1.0, has been largely resolved.
I canceled training for the Text Encoder and switched to U-Net Only.
Unet lr : 0.0001
Network Rank Dim: 16
Network alpha: 8
I ran the model for 20 epochs (5,000 steps) using a dataset of 250 images.
Description
To prevent overfitting, I adjusted the parameters.
Network Rank Dim: 16
Network alpha: 8
I canceled the Text Encoder training and switched to “U-Net Only.”
I ran the model for 20 epochs (5,000 steps) using a dataset of 250 images.
There were some issues with prompt compliance, but they seem to have been largely resolved. While V1.0 only accepted natural language, the model now works properly even when prompts are written using only combinations of Danbooru tags.