This model has been commissioned by https://civarchive.com/user/Morenoged. He said "please". And gave me buzz. Pay him a visit.

Nota : Pony being based on SDXL, my pony / SDXL models can work on SDXL / Pony checkpoints... or not, and in sometimes surprising ways. You have to try.
Sometimes, a Pony/SDXL lora can even work on SD 1.5, but of course isn't made for good results like that.

Here's my version of Dixie Catillac. Dixie is a character from an old cartoon called Heathcliff & the Catillac Cats, in the years 1980's.

Trigger : The trigger is dixiecatillac.

Roxie is a color swap (yes, that's that sort of cartoon). Try with brown body fur, Black hair / tail / eyes. The bangs are opposite to Dixie, but SD generally do bangs in the same colors than hair, and turned in the same direction than head / body.
"COLOR belly" is generally understood by checkpoints (Roxie is beige, I think).

As the choice of training images is poor, you will have better results by adding more prompts used by the training.

Change the word COLOR for the color you want, you silly goose ! Or just don't give color at all. Looks at my examples !
Honestly, look at the examples. Do you know the time it took me to do them on all those checkpoints ? I don't have a 600$ graphic card !

the word color is meant to be replaced by the color you want. Throw as much paint on the poor girl as you want to !

What version to use :

SD 1.5 for "normal" SD 1.5 checkpoints.

SDXL for SDXL checkpoints.

Pony for Pony checkpoints.

SDXL frequently works on Pony, and Pony on SDXL (you have to try). It's more rarer for SD 1.5 but can happen.

Version note : I'm still not very happy with Pony version, but it's the better I can do for the moment with my actual knowledge of training. Problems seems to come from the artifacts of the AI generated images, based of 8 very low / blurry base images (Dixie is a basically unknown character). Some checkpoints cant's repair the hands, and I think I see from where came the strange "points" you sometimes found under her mouth line.

The very low (when not blurry) 8 images of the starting dataset are a killer.

I'm not looking for an excuse. Or just a bit...

Prompting :

The points here are captioned (used in the training), meaning have strong effect when you use them in your prompt :

COLOR hair : The color of the hair. By default, Dixie has blonde hair

COLOR body fur : She's supposed to be white.

COLOR eyes : By default, they are generaly light blue, but can be green.

COLOR eyeliner : By default, generaly purple.

COLOR shirt : Her classic sport uniform. By default, pink. You will have, maybe, to weight it a little. Dixie seems to love pink.

COLOR shorts : Her classic sport uniform. By default, pink.

COLOR belt : Her classic sport uniform. By default, magenta.

COLOR scarf : Her classic sport uniform. By default, magenta.

cat tail is captioned. Sometimes you will need to add this for the poor cleo to have a tail, but can add another tail for... reasons.

Apparently, even if I captioned as such, "blonde" is not a color for a tail. For generating images, use yellow for the color of her tail

Other prompt :

All images have also been captioned with anthro and furry. Can help on Anime checkpoints, or having no effect at all, maybe also on furries, you try. Particularly true on Pony.

That can also help.

Note for SDXL :

anthro, furry in the prompt can be necessary on some checkpoints. Or have no effect at all. It's very curious.

Tips:

Basically, the more detail you add in your prompt, the better Dixie will be.

This is because the few number of images I could find for the training dataset.

clip skip can be 1 or 2. You should try.

Lora weight : between 0.7 and 1. depends of the checkpoint.

Trigger : The trigger is dixiecatillac. You can put no weight on it, or give it 1.1 or 1.2, particularly if you put a lower weight to the lora.

Also :

Generation with DPM++ 2M / SDE and Automatic / Karras works well

20 - 25 sampling steps should be largely enough.

Upscaling upgrade the image a lot. I use R-ESRGAN 4x+ Anime6B, hires steps 0, Denoising strength 0.35, upscale by 2, and no refiner.

Older versions :

Pony V 0.5

First version on Pony. While trained with the same dataset than SDXL, the result was far less good in my opinion.

First Try - not Final !

First try at the character Dixie, from the cartoon Catillac cats. It's going to need more training because my first dataset is nearly only cartoon shots, and as we know all those drawings were not very good.

Trigger : dixiecatillac.

Other prompts : eyes, hair, body fur, cat tail, eyeliner, scarf, belt (some checkpoints don't like give her a tail, for some reason).

Beware, the LoRa love to add stands, lamps and colors on the ground (because the Dataset).

As this LoRa have been trained on the same moment the server had his problem on training, apparently it was trained on SDXL when I asked on Pony. Metadatas from the model say it's a base SDXL model.
And it seem to works a little better on SDXL than on Pony (you can see the difference in the examples). Ratatoskr give particularly neat results.

I'm going to work for a second training with new images generated with this version. That's sort of a challenge for me, now. I'm not going to be beaten by a cat.

I know it's bad, but really Dixie is a really unknown character, there is nearly no correct images from fans, and images from the show are, well, 80's child cartoon show level.
Basically, for this first training, I have nearly only around 8 bad images, and all with the same very simplified cartoon style... The horror.

Also, maybe a little over-trained. 20 epochs and 40 repeats...

If you want to help me in this work, remember :

• When you use the Add Post button to post an image, CivitAI give me some buzz (75). It WILL help a lot, because training take a lot of buzz and time, particularly on these rare characters. I don't know if there is a limit (probably)

• A like on a model you like will help it to be more know. More people loving my work, more chance for them to post images, more buzz.

• One like on one of my images you like, CivitAI give me 3 buzz if the image is less than one month old.

Support confirmed we don't get the big Buzz for images posted in "crossposting" (probably the ones posted by the Create > Post images menu), but only with the "Add Post Button" on the page of the LoRa.