🌈Read Me Please on How to Use 🦄

Pony FaeTality is a versatile Pony Model that has a natural 2.5D look but can also do many other styles without having to utilize style LoRAs. 🥰

Version 1.1 Update (Chat GPT Captioned)

I made a small update focused on improving the lighting, I trained a lora on 120 images that were specifically captioned with GPT-4 to test how well GPT handles captions. Although the improvements in lighting aren't dramatically significant, there is a slight enhancement. I'll have to experiment again (whenever I have time 😿) using a much larger dataset to see more substantial results probably. 🦄

Using descriptions like this helps a lot with the effect: bright colors, high contrast, dark background, vivid lighting

Example of an image in the GPT-4 dataset along with the caption given

heart-shaped candle, vibrant melting wax, dynamic splashes, candle flame, bright colors, high contrast, dark background, vivid lighting, celebration, romance, abstract, love concept, glowing effect, sparkling particles, multicolored drips, wax flow, illuminated, artistic, creative photography, energy, melting in motion, high resolution, centerpiece, decorative, festive mood

Tips and Tricks

• Always use score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up

• Use Clip Skip 2

• This model was trained using both blip and WD captions and tags. Utilize danbooru tags when prompting.

• Certain keywords will have a bigger effect on the outputs, they are: realistic, freckles, large lips, beautiful eyes

• When prompting less is better I've found, don't go crazy adding a bunch of descriptions

• Go easy on the negatives, 3D in the negatives has one of the biggest effects, and this will tone down the trained style a lot if you don't like the look. If you're going to use negatives make sure they're targeted ones.

• Because it's a Pony base, you can utilize the data selection tags of: 'source_pony', 'source_furry', 'source_cartoon' and 'source_anime' and ratings of 'rating_safe', 'rating_questionable' and 'rating_explicit'.

• Recommended aspect ratios of 1024x1024, 832x1216, 1216x832

What I'm Good At

• 2.5D People, Animals, Furries

• Adds pretty details and backgrounds a lot of the time naturally

• NSFW

• Cute Things

What I'm Bad At

• Non-NSFW Complex Scenes or Objects

• Sometimes I make people have giraffe necks, you can fix this by adding 'long neck' in the negatives. 🦒

• Photorealism

• Sometimes I like to make random things into animals 🤷‍♀️

How I Was Made

I started this project over a month ago by trying to finetune Faetastic SDXL to have a more 2.5D style. The entirety of the dataset is generated images. 95% of the images were my generated images of a prompt style that gave an awesome 2.5D look with FaeTasticSDXL...and I wanted a checkpoint that always had that style.

One of my images from the dataset

I was successful and the finetune is really great, but I decided I wanted it to be good at NSFW, which turned out to be pretty difficult and I had some pretty hilarious outputs.

Around this time Pony started to get really popular, because I'm busy so much with Civitai I just decided to take a shortcut and swap to Pony as the base model to achieve my goals.

So I finetuned then trained 4 different LoRAs and smashed those babies in. LoRAs were to nail down the style more + I also added my FaeTastic Details dataset into them, which seems to have given it that detailed cute vibe that I enjoy.

What I'm Going to Try to Improve (When I Can 😅)

• Better Darkness/Shadows

• Try to get rid of oddities

• A better understanding of "normal things"