Bleach: Thousand Year Blood War Style LoRA
Reviews are very appreciated and show me you enjoy my work, keeping me motivated in releasing more Models in the future! Would love to see the community create images with my work.
v1.0 (Anima Base 1.0)
Trained on the Anima 1.0 Base model for compatibility with other Anima models.
All preview images are created using the MiaoMiao Harem 1.1 Checkpoint.
For more information check the "About this version".
The settings used to create the images were:
Steps: 30
CFG: 4
Sampler: Eular a
Scheduler: Normal
Denoise: 1Best Results are at 1 weight, using the following trigger word is optional but increases influence:
@vslinxbleachtybwv1.0 (Anima 0.1)
Trained on the Anima 0.1 Preview model for compatibility with other Anima models.
All preview images are created using the WAI-Anima 1.0 Base Checkpoint.
For more information check the "About this version".
The settings used to create the images were:
Steps: 30
CFG: 4
Sampler: Eular a
Scheduler: Normal
Denoise: 1Best Results are at 1 weight & using the trigger word:
@vslinxbleachtybwv1.0-v1.1 (NoobAI V-Pred)
Trained on the base NoobAI 1.1 V-Pred Version model for compatibility with all other NoobAI V-Prediction models.
All preview images are created using the Noobai-Cyberfix-V2/Perpendicular-Cyberfix-V2 (VPred) Checkpoint.
The settings used to create the images were:
Steps: 30
CFG: 5
Sampler: Eular
Scheduler: SGM Uniform
Denoise: 1Best Results are at 1 weight & using the trigger word:
vslinxbleachtybwv1.0 (NoobAI EPS)
Trained on the base NoobAI 1.1 EPS Version model for compatibility with all other NoobAI models.
All preview images are created using the NoobAICyberFix 3.0 Checkpoint.
The settings used to create the images were:
Steps: 30
CFG: 7
Sampler: Eular
Scheduler: Normal
Denoise: 1Best Results are at 1 weight & using the trigger word:
vslinxbleachtybwv1.0 (IL)
Trained on the base Illustrious-XL Model for compatibility with all other Illustrious models.
All preview images are created using the NoobAI-XL Epsilon-pred 1.1-Version.
The settings used to create the images were:
Steps: 28
CFG: 5.8
Sampler: DPM 2
Scheduler: Karras
Denoise: 1Best Results are at 1 weight & using the trigger word:
vslinxbleachtybw styleAll images were upscaled using the Remacri-4x Model.
All images were generated using my ComfyUI workflow that you can find here, attached to preview images or in the suggested resources. This includes a Detailer that automatically inpaints body parts like eyes, lips, hands etc.
So if your results do not match mine it's probably because of this. I recommend using the workflow which is also attached to every image.
IMPORTANT
Creating the Characters from the Show is completely dependent on the Base Model or other LoRA's you're using. This was done to ensure the LoRA would only affect the style of the image and not the depiction of characters and their outfits.
The character names were added to the prompts while not removing their characteristic features to ensure the LoRA has an idea of who the character is without overwriting the prompt you're using. Meaning if the base model or another LoRA you're using do NOT know the character, it has to be prompted by their characteristic features and not by their names.
These Characters are included in the training data:
inoue orihime
kurosaki ichigo
sado yasutora
ishida uryu
genryusai yamamoto
young genryusai yamamoto
child nelliel tu odelschwanck
kyoraku shunsui
ukitake juushirou
mallia menoly
loly aivirrne
emilou apacci
franceska mila rose
cyan sung-sun
abarai renji
hitsugaya toushirou
sui feng
bambietta basterbine
sajin komamura
matsumoto rangiku
unohana retsu
unagiya ikumi
kurosaki masaki
kurotsuchi mayuri
asguiaro ebern
ayasegawa yumichika
yhwach
tier harribel
urahara kisuke
quilge opie
ishida ryuuken
haschwalth jugram
kuchiki byakuya
as nodt
kuchiki rukia
zaraki kenpachi
sasakibe choujirou
shinji hirako
aizen sousuke
mcallon meninas
tenjiro kirinji
hikifune kirio
hyosube ichibe
shutara senjumaru
shiba kukaku
kotetsu isane
kuna mashiroDescription
Last iteration of this model for now, shoutout to TinySandwich for help with the v-pred parameters.
Increased token length for more accurate and better prompt adherence
Fixed weighted captions that accidentally got introduced in V-Pred 1.0
Using zero terminal snr now for weight normalization across all epochs and repeats
In rare cases if your checkpoint's style influence is too heavy you can increase the lora strength to 1.4~
FAQ
Comments (7)
if its okay with you can you share the config you used for the lora ? im just curious because the final lora ended up so well and i cant get my config to work like that
heya @ore1,
Of course, i'm a big friend of open source and sharing sources/materials :D
Are you using kohya_ss? Since i don't traing with anything else.
Otherwise you can always pretty easily check the configs people have used to train their lora's by going to https://lorainfo.tools/ and dropping the lora there, it'll give you all the meta information like total steps, epochs, amount of images, words and prompts it's trained on as well as every parameter you can set in the training software :)
Awesome style!
Why isn't taichou haori working with this LoRA? It should have worked... hmm 😢
It was present in the training data but not explicitly tagged to make the LoRA more versatile with other character/clothing lora's.
That basically means that this only injects the style in the image but is not trained on any new concepts like clothing :/ so if the base model you're using can't generate the clothes with this tag - this lora sadly won't do it either.
I have checked some illustrious models and some noobai models and sadly none of them could do it straight out of the box, at least from the ones i tested. Sorry 🥲
@vslinx These checkpoints really need to be updated. Many simple tags aren't applying properly, hence the reason models like Illustrious and NoobAI were created, to reduce the reliance on LoRAs. I can still generate the desired look using 'white haori' + 'black kimono,' so at least that works for now. Even for Unohana Retsu, the 'front braid' tag should work, especially given her braid type. But these checkpoints are still outdated,it's been a while since we've had a better understanding of how checkpoints should work.
@monicalucci yeah 100%, noobAI has a good normalization algorithm where even things with a smaller amount of tags work relatively well - at least for style's.
But some clothing combination is very hard to achieve for nieche tags or series specific clothing/concepts.
There are always workarounds but obviously direct solutions would be better :D
I'm still glad we have illustrious based checkpoints like noobAI since it's a completely different world to 1.5, SDXL or even Pony which were unusable without loras ^^
Details
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.