Kasucast #16 - Stable Diffusion: Create a PNGtuber with Character LoRA

#stablediffusion #characterdesign #pngtuber #virtualyoutuber #vtuber #conceptart #digitalart #machinelearning #lora #lycoris #locon #loha #characterLORA #style #dreambooth #digitalillustration #aiart #style #artstyle #automatic1111 #kohya_ss

Resources:

Find the art pieces in this video here:
Instagram:

🎉 Social Media:

Images may be fabricated and therefore not real. I am unaware of any illegal activities. Documentation will not be taken as admission of guilt.

Timestamps:
00:00 Intro
00:43 Tools you'll need
01:09 Dataset introduction
01:41 Reasoning for using a character LoHa
02:30 Captioning the character
04:25 Character LoHa training settings
06:35 Initial test of character LoHa
08:33 Choosing the best epoch
09:34 Verifying the chosen epoch LoHa integrity
10:55 Mini veadotube mini overview
12:03 Choosing the style LoRA
12:43 Check the raw style LoRA
13:24 Prompting with both character LoHa and style LoRA
15:32 Technique to vary seed within a radius
16:31 Compositing bust shots together
17:17 Overview of base full body image
18:24 Outpainting the full body
20:27 Outpainting the bust shot
20:59 Upscale the body
21:47 Upscale the bust
22:28 Separate head from background
22:51 Preparing the body for bust photobash
23:41 Rasterizing a smart object
24:08 Liquifying for better seams
24:38 Refining the entire character
27:40 Preparing head for smart object (simplifies character expression replacement)
28:56 Creating the open mouth expression
30:31 Creating the closed mouth expression
31:07 Creating the closed eyes expression
31:36 Making the closed eyes a global toggle
32:00 Removing the background from the entire character
33:08 Exporting in the correct format for veadotube mini
33:34 Setting up pngtuber in veadotube mini
34:42 Using veadotube mini with OBS
35:45 Closing thoughts

[dascodraws]

I appreciate how you explain everything!

[Sheevlord]

Great video, as always!

Here's a fun fact that might be helpful when generating characters with specific facial expressions. You can use emojis and even ASCII smileys in the prompt. Not just for faces either - various hand gesture emojis work too. I sometimes use them when inpainting hands.

[UsaraDark]

Great tutorial, covering multiple concepts to achieve a goal. I haven't yet messed around with inpainting/outpainting, but it looks more powerful than I thought.

Also, what did you do with the full body? In the end, you only used the square crop of the upper body.

[cyril]

Hey thank you for your tutorials, they are all great, although ive been watching all of them cause I want to create a Loha (like here) based on only one picture dataset. Also it is not based on an anime illustration but a photorealistic one. Been scouring the entire youtube for Loha/Locon to do this and nothing. Could you maybe do one if you get any good result with it, because im sure it will have great success cause a lot of ppl are trying to create their AI twins (not my case).

Im trying to create an AI twin of an AI girl model (from only one image) to generate her in many different settings. Do you think it is feasible ?

[dascodraws]

so loha is best for character and lycoris locon for style

[Faris-]

I love your detailed tutorials. There's nothing like it on the internet, I really appreciate your great effort here. May I ask where did you get anything v4.5 inpainting model?

[dungeonmaster]

When someone says AI art takes no skill send them this video.

[SuperDao]

I'm confuse... you're using loha for training but when generating you're using a lora. Is it normal ?

[ItsDaBunz]

due to how copyright works around AI generated images, you'd have no copyrights over the character, meaning anyone could use it/become your character, not a good look. In addition, most people worth their salt in this community don't associate with Ai images and those who use them, since it's art theft and people tend to value ACTUAL art. Good luck with that though.