Description
Image imported from Derpibooru.
Original: https://derpibooru.org/images/2963445
Metadata at the time of import:
Original: https://derpibooru.org/images/2963445
Metadata at the time of import:
| Field | Data |
|---|---|
| Favorites | 25 |
| Upvotes | 46 |
| Downvotes | 5 |
| Score | 41 |
| Comments | 11 |
| Uploader | Anonymous |
Original description below this line
Fluttershy has some fun in mud.
No matter what I did, I couldn’t get Astralite Heart’s pony-diffusion-v1 model to generate a pony completely covered in mud. Phrases like “covered in mud” or “muddy” &c would only generate mud around where the pony’s body contacted the mud puddle, even though his model must have seen pictures of ponies totally drenched in mud. So I eventually broke out textual inversion: I show ~10 pictures to the AI and ask it, “What’s the common theme in these pictures that you know about, even if you don’t have a human readable name for it?” “It’s [6ab8].” “Cool. Generate a picture of [6ab8] Fluttershy…”
Asking the AI what it thought this concept was (and then naming it “amuddy1” for future reference) resulted in the above pic. Best generation out of 10, then went through img2img maybe five times to clean up the eyes, which were way lopsided in the first generation.
I tried 1, 2, and 4. This image was made with an embedding of a single vector, while 2 and 4 destroyed the ability to make faces.
How do I download the model?
See the first link in Background Pony #B396’s comment below.
That would explain why they tend to clean off so well in the show, too.
thanks
The model is https://huggingface.co/AstraliteHeart/pony-diffusion and the program I used for inference was https://github.com/AUTOMATIC1111/stable-diffusion-webui. I uploaded the 4 kilobyte concept embedding file where I interrogated the AI what it thought muddy was here: https://mega.nz/file/BaonyBBT#YSuiWthxETX_5i3tFmjfkTvgGpWomqxJOcDmVhgdcco . Put it in your embedding/ directory and you should be able to directly use the word “amuddy1” to refer to this concept. (Note that embedding file will only work with the pony-diffusion model. Waifu Diffusion or one of the furry models “know” different things so you’d have to repeat the process of showing them pictures to find what they call the concept…or maybe they understand “covered in mud” in the first place!)