
This is my first try on LoRa, it was trained for around 12 to 15 Epochs on an RTX 3090 for about 12 hours based on 95 images taken from the best poses over all Episodes. I used the BluRay remaster as a source.
The main checkpoint I used for this training was Albedobase. I noticed you can start to get interesting creative results around 0.5 to 0.75, going over 0.75 makes the LoRa sometimes too close to the original. This is my first try so it's not perfect, also the source was hand painted so the character lines weren't always very consistent, but with some tries you can get really cool results with a clean retro-aesthetic.
I recommend AlbedoBase XL with SDE Karras, but you can try anything you want.
PS : I used descriptions during training like : iria, anime girl, brown hair, brown cape, futuristic female armor, blue tshirt
描述:
First release
训练词语: iria
名称: iria_zeiram.safetensors
大小 (KB): 1781930
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success