meredy/メルディ v2 (fairy tail) (sdxl (pony)/sd1.5)

I would prefer if you guys don't use this model for Commercial Use.

You can check out my pixiv to see what model I am working on and what ai art I create.

https://www.pixiv.net/en/users/34413372

meredy pony sdxl v2.5

update: I uploaded a 16 epoch because the other lora are not very flexible. One problem is the compatibility with other lora and checkpoints models.

I used the same dataset that I used on sd1.5 and pony sdxl v2. The difference is that I have added a new tags and trained the lora different this time. Use the tag crenel in prompt to get consistent lower part of her dress. The improvement with this lora comes from that it is more better at making the lower part of ther clothes and in general makes better quality images. The model is also smaller than pony sdxl v2, I made it smaller because it was easier to train it. The v2 I trained with A100 gpu (google colab) with way more steps. I was not satisfied withe result and made a new one.

I suggest you use 648x976, 768x768, 976x648, 648x1152.

use epoch 25 for more flexibility and 53 if you didn't like the other lora. It should be better at clothes.

prompt: meredy76, ponytail, ear muffs, yellow waist sash, detached sleeves, black stockings, cleavage, long hair, ponytail, pink hair, crenel, bare shoulders, red dress,

SDXL (pony) v2:

sdxl: pony version: I learned about the pony model and like it a lot. I wanted to update my meredy lora to sdxl version. I would suggest using this lora on pony model and can't say it will be best at other non pony models. the sweet spot strength for this lora is 0.8. I would recommend adding furry, pony, horse in negative. the pony model will sometime turn the characters into furry or give tail to the characters. using the keyword pony tail will trigger this. pretty bad for this model because the character in the lora have pony tail.

I am using the same dataset that I used on SD1.5 lora, but cropped at 1024x1024

SD1.5

I made 2 different sd1.5 lora you can choose. If you don't like the first one you can try the other one.

In this lora model I have trained more images than the last v1 lora. The character meredy don't have many artwork or good enough screenshots online so most of the images are screenshots that I took from the anime. because she appears very rarely, I was able to take 149 images.

The model is also larger and flexible. You can use it in other models that are not anime models. You can use it to create meredy clothes and other types of clothes. v1 somewhat struggle with it. I trained this lora on anylora-anime mix.

I tried to make lora that can create meredy clothes 100%, but it still struggle with it.

I recommend 0.5-0.8 lora, 7-12 cfg scale, 512x512, 512-768, 768-512.

meredy in anime use a blue cape a lot of the time so the model is bias and will by default include cape in images. to remove them, have cape in negative prompt and if you still struggle with it, you can include blue also in negative prompt.

if you want to create meredy from anime: meredy76, red dress, short dress, yellow ribbon, ears muffs, pony tails, pink hair, detached sleeves, off shoulder, aqua eyes, medieval, fantasy, black stockings, browns boots, cleavage,