Version 5

I didn't upload version 4 due to it being really flawed due to errors in my training process as I was getting used to local training on a new GPU.

Version 5 added more images to the dataset and seems to have reinforced the look and feel I was going for with her.

When trying to update the LoRAs for Pony and SDXL I just couldn't get the consistency I was looking for so have left them as they are for now.

Version 3

The same dataset was used but the images were all retagged to a better standard and a Flux model was added.

Version 2.0

I've added another 170 images to the training set to try to improve consistency and reinforce the traits and personality that Sophie was meant to convey.

This improved dataset has been working well for me in the new sample images I've added to her showcase.

Going forward I'll try to use these same prompts and the same seed so that versions can be easily compared so that anyone trying to decide which one is right for them can perform a more meaningful comparison.

Original message:

After a few attempts at creating a LoRA for a character I'm wanting to use for a Sci-Fi story I'm creating I have gotten to a point where I'm happy to share it with the wider world to see what other people with more skill and talent than I do can do with it, so please upload any examples of images that you create using it so that I can learn more.

Sophie is described as a tall 25 year old ginger woman with freckles and small breasts in the prompts used to create the character. The training images were almost entirely created with the Duchaiten-RealNSFW-XL model and a few different LoRAs dropped in to create the character I now have. It hasn't been built with NSFW imagery in mind but a few of the images used to create it were NSFW in nature to help reinforce the character being created.

I find that the model works best when set to a strength of 0.75 and used with realistic checkpoints and LoRAs. Pony XL checkpoints don't seem to work with this model but I've had good outcomes also from Juggernaut XL Rundiffusion.

Any thoughts and feedback are welcome and I'd love to see what can be created with this model so please do upload any examples of images you create so I can see what other adverntures she can be sent out on!

描述:

I added an additional 170 images to the original ~40 images used to train V1. I did use the LoRA in the process of creating these images, but went back to the base prompt used to generate the initial 40 as I thought that would help to weed out any inconsistencies that I'd baked in with a weaker initial dataset.

The result appears to be a more consistent and better detailed face for the majority of the time. I created 23 images using the same prompt and seed using both the old and new LoRA and the new one came out clearly best in about 15 of these, slightly ahead in 2, about equal in another 2 and better in the old model with only 4 cases.

I found 0.75 to be a good srength for the LoRA but have done no testing on higher or lower strengths as what I got to start was perfectly suitable for me.

训练词语: sophie-molgrin