Prompts:
The model is dreamboothed on tagged suisei no majo images; some prompts that work are
- suletta mercury
- miorine rembran
- gundam aerial
Training details:
Trained with kanewallmann Dreambooth repository using tags as captions
- Trained for 10000 steps probably at the default learning ratet lr=1e-6
- Dataset: around 500 tagged images of suise no majo + thousands of customized reg images
Problems:
As the model is trained only on tagged images, it is more flexible but it is only harder to prompt. Some detailed description may be needed to get the character right, especially when trying to prompt suletta and miorine in the same image.
Example Generations:
