AAM - AnyLoRA Anime Mix - Anime Screencap Style Model - v1

NSFW

AnyLoRA Anime Mix (AAM)

Anime Screencap Style Model

Do you like what I do? Consider supporting me on Patreon 🅿️ or feel free to buy me a coffee ☕.
A ❤️, a kind comment or a review is greatly appreciated.

Please check out AAM XL too

Purpose of this model

Make amazing anime style artworks on its own.
Train character loras where the dataset is mostly made of anime screencaps, allowing less anime style transfer and less overfitting.
Use anime styles if the lora is "weak" and doesn't give enough flat coloring.
It's also pretty good at generating hentai stuff.

Suggested settings

Set the ETA Noise Seed Delta (ENSD) to 31337
Set CLIP Skip to 2
DISABLE face restore. It's terrible, never use it
Use negative prompts and embeddings that don't ruin the style
Use AnimeVideo or Foolhardy as upscalers. If you know what to do, Latent makes an interesting effect (see the first example and the Miku one)
(optional to reproduce some results) Set "Do not make DPM++ SDE deterministic across different batch sizes." Only my old images use that.
(this setting is not actually good, so I'm gonna disable it, but might be the reason some results are different)

LCM version
Use it with 5-15 steps, ~2 cfg. LCM VERSION WORKS ONLY WITH LCM SAMPLER (to get it on Auto1111 you currently need a plugin).

Being a distilled model it has lower quality compared to the base one. However it's MUCH faster and perfect for video and real time applications.

Brief history

Chatting on Discord, I noticed my Mushoku Tensei style LoRA, while trained on AnyLoRA, was performing a bit better on CyberAlchemist's AnimeMix base model. I read on his HF that the model was a supermerge of AnythingElse, his AnimeMix lora and various other style loras. Since AnimeMix was already including many of my old anime style loras, I wanted to experiment and try to make a model that could maximize the result of my Mushoku Tensei style, going beyond what AnyLoRA and AnimeMix could do.

After testing many different versions, adjusting the settings and receiving feedback from the Discord channel, I ended up with a version that works very well on anime loras trained on AnyLoRA (base). Especially my Mushoku Tensei one which felt very weak on AnyLoRA.

I don't advise using this for anime style training (limit to generation after training it on AnyLoRA or NAI), but it might actually be fairly good for character loras where your dataset is made mostly of anime screencaps. It should technically reduce the anime style impact making the character lora less overfitted. I'll definitely experiment with it in the following weeks.

Description

FAQ

Comments (48)

KytraJun 5, 2023· 4 reactions

CivitAI

First! 😉

12user34kn276Jun 5, 2023· 2 reactions

CivitAI

Very cool concept! I'll definitely try training a Character LoRA on this, maybe even compare a model trained on AnyLora with both fan art and screencaps vs trained on this, but only screencaps

Lykon

Author

Jun 5, 2023· 2 reactions

yes please do. I wanted to have some ready before release but didn't have enough time :)

wlmsgJun 6, 2023

CivitAI

Can I think that it would use a key word to present an anime style?Or just batter on anime style lora?

AnyLora is so good ,so I want to know what the difference is.

Lykon

Author

Jun 6, 2023

this defaults to anime screencap style, meaning that used as a base model will increase the flat coloring. During lora training it should reduce the style impact of anime screencaps.

fulforget85886Jun 6, 2023

Flat colors and thick line work. Could probably convert to vector really easily given the right postproduction software.

longkin186947Jun 6, 2023· 1 reaction

CivitAI

this is very useful, thank you

MCBXYOJun 6, 2023· 4 reactions

CivitAI

So good for making anime TV-esque style renders.

Nazuna_VampiJun 6, 2023

CivitAI

What's the best VAE for this model? The same two that you mention on your AnyLoRA base; WD Vae and FT MSE? Is this also a good general anime model like the base?

also

Thank you for your models.

Lykon

Author

Jun 6, 2023· 3 reactions

vae is baked in, so just select Auto.

Nazuna_VampiJun 6, 2023

@Lykon Oh, I see. Thanks.

AnnawnJun 7, 2023

CivitAI

so what did you use to train this? just anime screencaps?

Lykon

Author

Jun 7, 2023

yep.

hfpanJun 10, 2023

CivitAI

So let me get this straight:

If you are using this for training, this will make anime character LORAs be more flexible in various checkpoints.

If you are generating with this LORA, this can help draw out 'anime' style and colors.
Would it perhaps turn realistic character LORAs into better looking anime characters? Interesting

Lykon

Author

Jun 10, 2023· 1 reaction

everything mostly correct, but depends on some smaller factors.

btdoge007Jun 12, 2023· 1 reaction

CivitAI

great work makes me crying,thx!

Nazuna_VampiJun 15, 2023

CivitAI

I'm training a loRA to produce a more consistent cel shading (flat coloring / anime coloring) style on naked bodies. Should I trained it on your model or would be better to use the full leaked NAI model, the 7Gb one? My dataset consists of 80% screencaps and 20% high quality cel shaded ilustrations for a total of 250 images.

Lykon

Author

Jun 15, 2023

you should train on an anime model that does not do flat coloring by default, such as anylora. If you train on a model that does normally flat coloring, your lora will learn nothing. If you read the description of this model, I have already said what it should and should not be used for.

JB2023Jun 16, 2023

CivitAI

Hello, I am getting occasionally black image/output while using this checkpoint model. Is there anyway I could fix this issue?

Lykon

Author

Jun 16, 2023

I doubt it's from the model. Are you using diffusers from python? You sure it's not the NSFW filter?

JB2023Jun 17, 2023

@Lykon Is there a nfsw filter? I did some research and found out the issue might be related to VAE. I used suggested command line argument- --precision full --no-half and now I haven't been getting any black screen output. However, the generation speed was obviously slower than half precision. Before I was getting random black screen output when using hire fix/ upscaler. Any help will be appreciated.

Lykon

Author

Jun 17, 2023

@JB2023 black screens happen from time to time on any model. It's just a bug of the system and has little to do with the checkpoint

ElimBecilJun 23, 2023

Me too

andinmaro146Jul 3, 2023· 1 reaction

CivitAI

I'm getting great results when training and generating using this model. Great job!

Lykon

Author

Jul 4, 2023

Show me!

andinmaro146Jul 4, 2023· 1 reaction

@Lykon Added some images in the review

MonkeyDLuffyJul 4, 2023· 1 reaction

CivitAI

hey i was reading the suggestions, how do I get the AnimeVideo or Foolhardy upscalers?

A1322Jul 16, 2023· 3 reactions

CivitAI

All setting ENSD to 31337 does is let you replicate images by other people who also have their ENSD set to 31337. It has no impact on image quality. ~~It's just a dumb meme that people forgot was a meme.~~ (It's not a meme. It's even worse. See comments. The truth will shock you. It's still not an important setting though.)

Lykon

Author

Jul 21, 2023· 1 reaction

it's not a meme. It was used back in the day to replicate NAI pictures. It just got left in the settings for many users of that period.

A1322Jul 22, 2023· 1 reaction

@Lykon Yeah, I learned that more recently, and it means the actual issue is that someone at NAI still thinks 90s chatroom jokes are funny, which is honestly even worse.

I do think it just confuses new users by having it listed alongside clip skip and VAEs in model descriptions, implying an equivalence in importance. Auto1111 does emphasize that it doesn't affect quality, but modern ENSD use cases are so niche that I'm not sure it's worth mentioning at all. (I know it doesn't really matter. It's just a pet peeve of mine and I wish it'd go away.)

Also it's funny to use "back in the day" to refer to stuff from like a year ago, but time really does move fast as hell here.

pseudokawaiiJul 26, 2023· 1 reaction

@A1322 so you're A1111 + A211

A1322Jul 27, 2023

@pseudokawaii Wow, that took me a moment. I guess I'll have to change my name to ComfyUI211 soon.

ktiseos_nyxDec 9, 2023

I realie this is 5 months old as a comment thread, but y'know - NAI also thinks that copyrighting an open source model is entirely. funny, so y'know that stacks XD

A1322Dec 11, 2023

@duskfallcrew That's the essence of Big Tech, though, isn't it? Taking stuff that already exists and claiming they own it.

(Although now that I say it, that's just what Disney has been doing this for decades. They should copyright it!)

PettankoPaizuriJul 30, 2023· 15 reactions

CivitAI

Was this updated? Civit says it's new but then the date is the same as before and the version seems to be the same as well

feathericeJul 31, 2023

CivitAI

I wanted to generate some semi-realistic 90s anime images and this model has best eye details to get some cool highlights right. Sadly, i can't say the same about hair, seems like dataset had very limited amount of decently drawn retro hair. Any recommendations how to improve it? Tried some prompt tricks and loras, but no luck.

hoshino_HAug 1, 2023

CivitAI

Why do I generate all celluloid-style pictures, help

Adam_LolAug 30, 2023

I’m having the same issue. Did you find a solution to this?

PK02Aug 19, 2023· 12 reactions

CivitAI

Absolutely amazing model, my friend. The outputs are so clean I could lick them.

brandiipvtltd304Aug 12, 2024

wtf

savvascharue5273Aug 22, 2023· 10 reactions

CivitAI

It works phenomenally, dont let the low Download Count deceive you.

Tip: Pair it with Detail Tweaker Lora and this model rocks for Anime Style Art

BaYueCNAug 26, 2023· 1 reaction

CivitAI

What is the "ETA", I am a novice small white, so there are many don't understand the knowledge, please understand

SentenceJan 30, 2024

I'm not sure if you still need it, but here you are:
This option adds that a number to your seed number when you generate images. People set it to 31337 to replicate NovelAI generations. It's important to keep it the same as the image you're trying to re-generate, but it's irrelevant if you're just using random seeds every time. If you still want to set it and use AUTOMATIC's Stable Diffusion, just ctrl+f the "noise seed delta" option in settings.

DomWooferMay 16, 2024

estimated time of arrival

crawlingratOct 31, 2023· 5 reactions

CivitAI

Why does this work so well for character LoRA training when compared to other checkpoints? The results are so far better then AnyLora and NAI!

Lykon

Author

Dec 6, 2023· 2 reactions

if your character datasets are obtained from anime screencaps, this will lower the style learned by the lora, thus making a cleaner character concept.

1622279Jan 5, 2024· 1 reaction

@Lykon So I can use it to train lora's character right?

crawlingratJan 5, 2024· 1 reaction

@malksama Yeah it works well.

Checkpoint

SD 1.5

by Lykon

Download (Beta) View on CivitAI