A Dark Race Becomes Magnificent!👑
Follow me on X/Twitter for exclusive content: https://x.com/Citron_Legacy 😁
I've made over 200 Art styles. Checkout my collection here: https://civarchive.com/collections/10766871
This is a style Lora not a character Lora: it knows some characters those are just bonus content.
Please support the artist by checking out their page: https://twitter.com/HARATAKE
Description
<lora:HaradaTakehitoIXL_v3:1.0>,
FAQ
Comments (14)
Im happy you are making illustrious versions of everything! To be honest, i kinda did not like the uhh..."pony era". Illustrious is the most fun i ever had with ai! On pony, some tags didn't even really work and had no replacement :x
Thanks! Yeah I really enjoy Illustrious! Let me know what things I should upgrade to Illustrious.
I'm just randomly upgrading things so I'd like to hear ideas.
Illustrious feels like the same magic when I first used AI. Its unbelievable how powerful it is. Its so much cleaner than Pony. I'm still going to make Pony things but I prefer Illustrious.
Happy to see this, as someone that really enjoys Harada's work. It's overall great fun.
Been using it quite a bit, though my silly attempts at trying to make an Artina/ Vulcanus or Sapphire haven't panned out too well.
Still, great fun!
Glad you like it. For at least a decade this was my favorite style because I grew up playing Disgaea/Nippon Ichi games.
I was thinking about making separate Loras for the characters that this can't create. I can Artina and Sapphire on my idea board.
@CitronLegacy it'd be great! Take your time in future endeavors.
Most I've attempted besides the main three that are perfected are: Raspberyl, Seraphina, Rozalin, Fuuka and Pleinair.
If it serves anything, this model does get part of Artina's hair right (the front hair locks), but forgets the large braid and her eye color a lot. Meanwhile it's difficult to keep Sapphire's skin tone. Haven't attempted to make the D7 gals yet. Side note, I've loved this model specifically for that green haired angel I'm posting a lot.
@Resshu Vulcanus is ready: https://civitai.com/models/1062511
Her ponytail braid works with my Lora :)
I'm surprised Sapphire works without a Lora because there is not as much fanart of her.
Illustrious is trained on Danbooru2023 Dataset so it should have at least some knowledge of D7.
Glad you like it! Your green hair angel is very cute! I really like that hair style.
just type harada takehito in prompt, this lora is not necessary.
My Lora is trained on Disgaea RPG art so the style is focused on the modern style whereas "harada takehito in prompt" produces a different era of Disgaea style.
Yes you dont need my Lora to make a Disgaea style but you do need it to make this style.
Pease see this comparison:
https://civitai.com/images/46254135
It's a really great LORA! I also made PONYLORA. I was successful with PONYLORA, but not with IllustriousLORA. IllustriousLORA cannot be made with your quality. Are the settings significantly different? LORA is created with kohya-ss
Thanks glad you like it!
Training Pony vs Illustrious is a little bit different. I like using more Steps for Illustrious training.
The Illustrious base model is very strong so it can handle extra steps without being overcooked.
.
Here is an example:
For my Diives Style I use the same dataset of about 100 images but Illustrious training had more than 3 times as many steps.
Diives Style Pony: 774 Steps (100 images, 10 Epochs, Batch size 5, Repeats 4)
Diives Style Illustrious: 2475 steps (99 images, 10 Epochs, Batch size 4, Repeats 10)
.
This Harada Takehito Illustrious style had 3800 steps.
There might be other tricks to training on Illustrious but this is what works for me.
@CitronLegacy Thank you for providing such useful information!
We use over 700 types of materials at all times.
It may have failed because I just set the repeat to 1 and increased the epochs. By the way, do you want to change the values of Learning rate and Text Encoder learning rate?
@miyasuke358858 I never change the Learning rate or Text Encoder Learning.
Usually the only settings I touch are Epochs, Repeats, and Batch.
Occasionally I change the Dim and Alpha when I need to better control a style. (All of my Flux loras have adjusted Dim and Alpha: I dont use CivitAI's defaults for Flux Training).
Sometimes in Pony I change the resolution: for example if the dataset has really old images that are blurry I don't use 1024 for XL because sometimes the automatic upscale in trainers still makes blurry images. This very experimental so I don't recommend changing resolution unless you are really trying to explore training techniques.
@CitronLegacy Thank you for the great information! Thanks to you it worked! It took me 30 hours to make 130 pieces, but I got everything I wanted!
@miyasuke358858 Happy to help! Glad it worked!
Let me know if you post your creation I'd be interested to see your success :)
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.