Use around 1.5 strength for better character consistency, especially if stacking other LoRAs such as SNOFS for NSFW.

Trying out a character lora workflow driven by agentware. The agents work through various procedures for creating multiple looks, camera angles, backgrounds, lighting, while keeping a consistent character, then caption them for me.

This group was trained on 36 images created via Grok image edit based originally on a single reference. 72 images were submitted (short caption + longer caption versions) for better prompting. 2200 steps, 11 epochs. Likeness was best at full training (11 epochs).

Optional keyword blondwavesgf may or may not be useful.

Publishing as a public LoRA on Civitai so I can easily access the generation tools as I have lots of buzz. :)

Feel free to use her in your creative projects.

This was the original source:

Dataset looked like this: