Introducing Calixyne,
My first Character Lora.
She's 100% AI generated, starting with QWEN, and then creating different poses, to generate the dataset. Trained locally with diffusion pipe.
----
Added v2 for QWEN, added more images to dataset, and retrained. Generally feels like a better identity lock without over-fitting.
If you're doing a single pass, then a LoRA strength of 1 works nicely, if you're doing a latent upscale/second pass, I'd suggest 0.75 on the first run and 0.25 on the second.
----
Added QWEN. - All of the example images were generated both with the QWEN Image base model, and the Hyphoria QWEN checkpoint to show the variation with the underlying Checkpoint.
Relatively happy with the results, although you probably want to use a weight of 0.6-0.7.
-----
Added WAN2.2. - Note that most of the identity is in the LOW LoRA. You can use both, but definitely need the LOW!
-----
Added Illustrious/SDXL. - Created a handful of portraits with different checkpoints to show the identity across differing models.
Description
Retrained with some additional images, generally find this to have a better character lock.
Strength of 1.0 when using a single pass, if doing latent upscale I would recommend 0.75 on first pass and 0.25 on the second to avoid over fitting.