Illustrious--------------------

Same dataset as pony, but seem very different, much more closer to the style

Little bit over-trained, using around 0.8

Tested on waiNSFWIllustrious_v130.safetensors

DPM++ 2M SDE Karras

34 step

Hires.fix + Adetailer on the post images

896 * 1152

POS: 1girl,

NEG: anime, 2d, bad quality, worst quality, worst detail, sketch, censor, blurry,

same dataset

new understanding

new model

new training ways

Video gen with wan2.1 :)

V2--------------------

For goodFitPonyv4.1 or realistic models. + lower size

Training on 500+ images from overwatch CGIs Sort image set from v1, delete most of the D.va image due to weird face. the overall face should be cuter than v1.

All faces are From Mercy, Tracer, Widowmaker, Mei, D.va, Kiriko, and Sombra. No male face has been trained. I will make a spread Lora for all CG styles of characters.

New training ways will cost this Lora much more cartoon, than v1 it will be more blurry, on anime or cartoon checkpoints.

-

All Post images are made on GoodFit-PonyV4.1 with

Steps: 35,

Sampler: Euler A SGMUniform,

CFG scale: 7,

Size: 896x1152,

Hires.fix no adetailer. (still needed when far)

v1---------------------

Training on 700+ images from overwatch CGIs. Aiming for the style of the overwatch CGI, not for a special character. Face should be all mixed in with the styles.

All train images are on 1024*1024

using settings other than 1024*1024 may give a funny result

The size has gotten up to close to 900MB, tried to lower the size but seems like it didn't hold the style well. This version should be pretty strong on different kinds of checkpoints.

Works great with the Pony itself on weight 0.9 or lower(pony without tune seems to understand it very well, but the result is just too blurry, I think you should be able to fix it with some extra prompts)

With basic prompts, the best result seems to be on 2.5d Pony v10.
v15 is a little too cartoon. even the same checkpoint with different versions might give huge differences. I will also try to push a merge checkpoint to see if it can give a better result.

The trigger word is still "owstyle", but seems like it isn't changing much on the result, at least on the test I have done.

Should be working well on weight 1, if you start to see broken body parts or blurry image lower it to 0.9.

missing some of the tags in this new version, tagging 700+ image is no fun :( and it takes too long to keep retrain it :(

so you may see plant in the background + some of the character cloth may overfit. if you want you can leave a comment or in discord to let me know what tags may not work well. ones I have enough I may make a new version. :-/

Give it a try, I hope you like it :)

Share some post and join the discord to chat :D

-

-

v0.8

Post images used high res fix + Adetailer, v0.8 is all on checkpoint 2.5D pony

All train images are on 1024*1024

using settings other than 1024*1024 may give a funny result

edit: Find out this version is very d.va faver, fix version is coming.

Tune for the 2.5d pony checkpoint, much more stable, and much close-looking than v0.5

Using this with 2.5D pony will give a much better owstyle result.

It is a really good checkpoint, but sadly it is not big enough for civit to put it in onsite-generate :-/

This is not a character Lora, is it just for the general style in Overwatch or for makeup :-?

-

I love how Lora is able to put the best overwatch character faces into the person with what is best fitting the characteristic.

-

example: when you trying to do a queen. Mercy's face shows up, and her halo will become her crown :)

Or when you trying to do a witch and widowmaker face will somehow show up(don't know is it because her helmet is more fitting, or just the face) xD

and somehow d.va face will show up as some innocent girl :-?

Guess this is how AI sees things, which is kind of understandable is funny to see :D

-

Well, Give it a try, and see if this is better :)

Still a test, not a v1, So feel free to leave a comment here or in Discord :D

If you can gives me some image you have that you think will improve this, pls do so.

I will try my best to improve this Lora.
Join the discord to chat and share some post :D

-

-

v0.5---------------------------------------------------------

when you give enough tries there is just one that works :^

( I don't even know if I can recreate the same one again :')

Train on many different overwatch CGI images, many of which I used to train 1.5 Lora

Able to make very overwatch-style faces.

Still a test, this is my first time making an sdxl Lora, So I don't think this is good enough yet :-?

Feel free to give me some feedback here or in Discord, I will keep improving this Lora

All train images are on 1024*1024

using settings other than 1024*1024 may give a funny result

Post images used high res fix + Adetailer

-

Trained mostly for the face, Style is adding to the model, but somehow didn't change the background enough which makes it kinda look like an oil painting (I think this is even better :3 )

-

-

Spend all that time waiting for SD3, it just takes too long so I gave sdxl a try :)

Share some posts :)

Join Discord to chat :D

Search AI Models & Addons

very few results