yes , u will see on top of the showoff images ,it has "animagineXL3.1" and on the side has the other versions

@GoskK Thank you! What I was referring to was if it was possible to reuse version 3.0 in civitai where I usually make images, my computer is not good enough so I use the site to do this

@sofiadamasko257 oh ,in this case I dont think its possible in civitAI ,but I know that in Tensor art you can choose the checkpoints versions

Do you have a link to the sliders you mentioned?

@Aruhshura Certainly. I realize my wording was unclear: you can use "cinematic lighting" to improve the overall lighting. The other is indeed a slider, and it's the "great lighting" slider: https://civitai.com/models/228974/great-lighting-sliders-ntcaixyz

@einar_rainhart no worries, thanks for the link. I'll try the term in my prompts as well

I think the same, the bodies look weird, I liked version 3.0 better and the bad thing now is that I can't use version 3.0 in civitai where I usually make images, my computer is not good enough to do this

Images look worse in 3.1 cause this model has learned the concept of "aesthetic" during its latest training process, so it tends to produce 'unaesthetic' content when there is no aesthetic modifier in prompt. In my own experiments, just follow the official prompt guideline will produce decent images. For example, add (masterpiece), (best quality), (very aesthetic), (intricate details) at the end of prompt, and put 1girl, solo at the beginning.

is your venv file corrupted? try deleting it and try again. also pay attention to parameters modifying the float 32 and fp16 variables

要是愿意接受质量下降的话，3.0的lora勉强可以用到3.1上，不过也只是勉强

use LORA

我·想·我·能解决

Unfortunately pretty much every anime model everywhere struggles with short-haired women.

it is baked in. try with/without a vae and it makes no difference

@temp10317164399 Thank u for responding, I've done that and you're right. Idk why this time the SDXL VAE is suggested

Check out the lists in https://huggingface.co/spaces/cagliostrolab/animagine-xl-3.1/tree/main/wildcard

@_rob332 thank you so much!

Just use the lightning Lora

I would check the "Clip Skip" option and set it to 2 if you don't already have it.

You can also check the resolution and the number of steps, some models are adapted for certain resolutions and perhaps the one you chose is not suitable and you could leave it at 40 steps.

Or, try adding the following to the negative prompt: (worst quality, low quality:1.2)

If you can't get the image you want, use Text to Image, fill in the prompt with more details so that the AI understands in more detail what it is seeing, for example:

+

(masterpiece), 1girl, long hair, white hair, hair between eyes, heterochromia, white dress, sleeveless, black thighhighs, white boots, dress lift, best quality, expressive eyes, perfect face, full body, white background,

-

(worst quality, low quality:1.2), nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark , username, blurry

1111

i agree

your choice would be pony model, autism mix confetti.

This is SDXL model, so this is quite normal.

Hey, I've been meaning to ask this too since I've seen how radically different certain artists can do for any character. So I've been on a binge for days worth of testing between 4 of my favorite characters from different series (Hanayo, Tohru, Hamakaze and Nessa) with various clothing and poses to showdown between artist tags.

You can find my findings here: https://www.reddit.com/r/WaifuDiffusion/search/?q=author%3Aacoolrocket&include_over_18=on&restrict_sr=on&t=all&sort=new

In short, its unfortunately a tie at times for certain characters and a good amount of times they have their own desirable specific artist for that anime series.

Although if I were to draw 3 of the best, it would be between terufuu, negom and sekina.

this should be pinned.

Where do I put the artist's style tag below?

「(quality tag), (1girl or 1boy), (character name), (work name), (any other prompt)」

@mumumu1295  I put it at the very beginning, (by [artist tag]:1.2) setting the weight from 1 to 1.3. But that's not a guideline from me.

My advice, always experiment with tags and placement. It's always kind of a lottery, where generations can turn out to have interesting results.

@Hannibal_Lector Thank you

hero!

Actually artist tags would have a better effect when strung together rather than used separately, maybe checkout my brief article?

https://civitai.com/articles/4476/for-nai3-and-animaginev3-artists-name-could-be-the-key-to-control-the-style-of-images

@Shallowpuddle I read the article. Interesting. Based on my experience and experiments, the use of a large number of artist tags harms the end result, often adding anatomical distortions, unnecessary objects and artifacts. I use no more than 3-4 artist tags at a time. But usually 2-3. This has become kind of my rule.

I tend to think that the weight and order of the artists' tags give a greater impact than their number.

I came up with three ways to use tags. Nothing special, but it works fine for me:

The first option is the usual one. Just the tags in a row. And here either all weights are the same (usually I put (1.15-1.25)), or with different weights.

If different, then the first tag is the main tag that defines the style, the secondary tags add details to the main style.

Usually the weight of the first tag is in the range (1.2-1.3). For the secondary ones in the range (0.85-1.15).

EXAMPLE 1: "(by sylvia_\(huajiuhuajiu\):1.25), (by kaguura_\(kagu\):1.25), (by ask_\(askzy\):1.25),"

EXAMPLE 2: "(by sylvia_\(huajiuhuajiu\):1.3), (by kaguura_\(kagu\):1.15), (by ask_\(askzy\):1.1),"

The second option is to use Automatic1111's Prompt Editing feature. [From:To:When] [ARTIST_TAG1:ARTIST_TAG2:STEP TO SWITCH]

We generate some kind of artist tag up to a certain step, after which we switch to other tags. In my opinion, it is a very underestimated feature that is rarely used.

EXAMPLE 1: "[(by kaguura_\(kagu\):1.3):,(by sylvia_\(huajiuhuajiu\):1.3),(by ask_\(askzy\):1.1):0.25],"

In this example, we use the Kaguura tag for up to 25% of the generation, after which we switch to Sylvia+Ask. It is very convenient to control the composition and style. I often look for a seed with a certain artist first (or even without artist at all), after which somewhere from 25-50% (0.25-0.5) I set the style that I want to see in the final version. Very similar to Refiner.

EXAMPLE 2: "[:(by sylvia_\(huajiuhuajiu\):1.3),(by kaguura_\(kagu\):1.15),(by ask_\(askzy\):1.1):0.33],"

The third option is to use Automatic1111's Alternating-Words feature.

[ARTIST_TAG1|ARTIST_TAG2]. Where the first tag alternates with the second after each step. You can have more than two tags, but I usually use alternating two.

This option is very... random and often the result is /interesting/, but it can be very harmful to the composition, details, and quality. And well, I use it in combination with the second option (EXAMPLE 2).

EXAMPLE 1: "[(by kaguura_\(kagu\):1.3)|(by sylvia_\(huajiuhuajiu\):1.3),(by ask_\(askzy\):1.1)],"

EXAMPLE 2: "[:[(by kaguura_\(kagu\):1.3)|(by sylvia_\(huajiuhuajiu\):1.3),(by ask_\(askzy\):1.1)]:0.33],"

And again, this is my way and experience. Maybe I'm doing something wrong...
Anyway, I think experimentation is great.
I apologize for the wall of text, it somehow turned out by itself :D

@Hannibal_Lector unique insight, my stereotype stems from the convention of NAI3 community, people there usually compose long string of artists tag(more than five) to achieve particular style, but invest little time in the weight and order of them.
From my experience, large number of artist names do sometimes undermine the overall style, but select tags carefully could resolve this problem, for example impasto style(ask, wlop, modare and much more) is an ideal basis for accommodating other styles, so this kind of artist tag is universal in almost all NAI3 images.
Anyway I appreciate your valuable experimentation, later I'll try it out myself.

@Hannibal_Lector The alternating prompt is also extremely useful for colors. It tends to be easier and more reliable to mix basic colors than to try to prompt a particular tone/saturation directly.

HERO!

Character list here: https://huggingface.co/spaces/cagliostrolab/animagine-xl-3.1/blob/main/wildcard/character31.txt

@Aruhshura Thank you =D

@Aruhshura To simplify, I have converted the tag list into Markdown and filtered it. If you like, you can integrate the list into the model. It should simplify the overview.
https://tagindex.lenz-host.de/en/Models/Animaginexl31

@Aruhshura where's nine the phantom from blazblue, it's too many to find and difficult to find.

@J_gumbainia69 bro just ctrl + f and search

@Aruhshura hi, do you know how we can add our own characters in the model?

i didnt seen yuno gasai in the list... where is my book...

It's an inherent issue with AI generated images. You can use Adetailer and/or LoRas like badhands to help mitigate it.

@smurfme is any lora or emmbedding are useful for fixing hand in sdxl model? cause i have only have easynegative for sd15

sadly SDXL models lack of hands Lora

I had the same problem the only thing that worked for me slightly better than Adetailer was just doing inpainting on the hands with very simple prompts and doing it iteratively. Do high batch count, select one which looks the best and then inpaint on that one and continue improving until you find something that looks good.

What comes closest to a solutions is LoRA's that fix hands, but they aren't foolproof: https://civitai.com/models/411088/detailed-perfection-style-xl-hands-feet-face-body-all-in-one-sd15 does do a pretty okay job though!

Like other comments you can surely try inpaint or add embeddings but there's another (1/5) solution

only by working on negative prompts:

https://civitai.com/posts/2706757

If all else fails, you can do a quick and crude fix in photoshop (or any image editor) and then inpaint over that.

use prompts like "Waving" or "holding" or "Gloves"
it's bad at hands if you don't specify how the hand should look like, but does well with hands when it knows how to make the hand

yep, they have it

@happyhen where can i find it?

@hikaw42805487 https://huggingface.co/spaces/cagliostrolab/animagine-xl-3.1/commit/cd54846a61ef14270a2146dce034ac1cb0fe2633#d2h-194052

@happyhen thank you

you got the wrong model💀

Try pony.This model is for general anime characters😅

if you want animal, go for pony. animagine is designed solely for anime.

@Madafada1991 but there are animals in anime

Same :c

Better use 4xAnimeSharp or R-ESRGAN 4x+ Anime6B, Euler A is not bad but images generated with this sampler are often more "soft" looking, other sampler like DPM++ 2M Karras or SDE are more defined.

@norim I used DM++2Sa or other sampling method and it showed the same problem.

U dont use sharper model like DPM Karras on enhancer with this model, u use softer model like Euler, always use either euler a or euler. cfg 6-7 doesnt matter. resolution of 768x1536, 1536 x 768, 1024 x 1024, 768 x 1024, is recommended.

有試過加入NFSW 提示詞嗎

或者你可以試試融合了本模型的Anything XL模型

768x1536 full body.
1024x1024 upper body view
1024x1024 squatting/sitting

先画小的图，再用upscale放大

U can use Counterfeit

I just noticed it was missing, too. It definitely limits what the model can do if we can't check what it knows. 🙁

this is the reason why I ALWAYS save useful information locally.

uploaded to google drive:
https://drive.google.com/drive/folders/1X-EobkAeXPRAN2DkKydSVxSV7OR8EM59?usp=sharing

there may be even more information here than was on the site (I didn’t use it myself) because I parsed training dataset by myself and divided it according to sfw ratings. (maybe it’s not as convenient as it was on the site, but what can you do)

it is here for all 4000+ characters: wildcard/characterfull.txt · cagliostrolab/animagine-xl-3.1 at main (huggingface.co)

pony is pony, this one follows sdxl line more.

@Madafada1991 thank you

use animagine as your training base checkpoint to get the best outcome

its actually quite hard, because the prompt dont listen to human-like prompt, it needs a "booru-ish" tag a.k.a robotic like tags.

try using a LoRA. i find that animagine sucks with everything involving feet and hands.

@Hiroshi_Ookami right now i am trying to switch to pony diffusion but still got some problems with lora training

I use "none". The image is generated at 720 x 1280 and then scaled 1.5x to get a full HD image.

autismmix is good for cartoon gens but it doesn't do that well when it comes to anime. No matter what lora you use, it still has that western cartoon look to it that all pony mixes have.

Animagine far exceeds autism

so many pony haters 😂

@Hapse That's what happens when a model is trained on a low quality jpeg dataset + extreme learning rate + no dropout + etc etc (I'm referring to Pony here)

Animagine does not suffer from any issues Pony has, since it is trained using actual settings but the dataset just isn't as big.

i do not know whether civit implements this feature; but many of the underlying inference engines / text transformers allow for exactly what you are asking (separate prompts for separate objects within an image). you'd do that by separating (effectively independant) prompts with the keyword "BREAK".

example: "woman, smiling

BREAK

male, serious"

i advise experimentation and would appreciate feedback if it results in anything

if youre doing it locally google "regional prompts"

Skill issue

@Ayat70 bro is not gonna sugarcoat it

idk bro, u could try using loras for nsfw, or using style loras for pony, they are very useful

Not skill issue, it's actually bad for nsfw. Try Obsession (Illustrious - XL) it's the much better anime model for nsfw.

@DarkenDark there is clearly a gap between bad and so damn bad. Also, animagine is not bad on nsfw. People who said it's bad are just comparing it with the so nsfw-specialized model pony

@Emon_Dante Eh he's probably using hyperbole. And I mean, it is bad for nsfw. Sometimes it will produce okay, and rarely good results; but more often than not it makes very simple mistakes, inaccurate characters, flawy shapes for sexual parts, artifacts... And the reason why it is bad does not matter when the end result is just... bad. This model -while can be considered good generally- is simply not a functional model for nsfw works without an unnecessary amount of work.

@DarkenDark tested the model you suggested, thank you. Thanks to you, I literally discovered the best model of all time. If I was already getting amazing results with animagine refined, illustrious is like Lv. 9999999 boss. I achieved peak generation illustration with my proper prompt, you changed lives.

How about Pony AuraFlow?

@TomLucidor, It hasn't been released yet.

@CuauhtemocI5MAL A man can hope for the future

if you want a new character, you should go to UrangDiffusionXL page. one of CagliostroLab staff updates new character there so you can ask him to put new characters there.

Try adding undesirable things to the negative prompt. Also, put "nsfw" as the first thing in the negative prompt!

add more negatives

It is worth trying “aged up” at the prompt.

@Hackgets tried Mature, Adult, Grown and even at 20/30s and it destroyed character face. Gonna try Aged Up. Thanks

https://huggingface.co/spaces/cagliostrolab/animagine-xl-3.1/tree/9cdf32cb0983ca5b59cb82a4528bdc9e4ec34427/wildcard

@ChrisArgenteo tqsm

yes then choose it from the ui you use, what vae does is normally fixing colors and adding details

if you are using comfyui, everything is pretty easy

give it more steps to process complex posturesm this model is well trained on hands and stuff but it need more steps to fully process those details

I suggest switching over to Draw Things if you are on iOS. I used DiffusionBee in the past but prefer Draw Things nowadays.

@AonekoChannel yeah its the best way to do some final tweaks maybe the frame is great but is missing a finger i really dont mind doin more post work to get a proper frame goin on some times, most times is pretty complete and accurate tbh it just a bigger database and deffo gotta be more specific

((side lighting:1.2)) as a starter point as you up on the weight youll gt different angles on the position of the light, ex. ((sidelighting:1.5)) will get you that 5pm bloom goldenm and ((side lighting:0.4)), will only give you a suttle hair reflection,

((ambient lighting:1.2))

((spotlight:1.2))

((volumetric lighting:1.4))

((single point lighting:1.2))

((backlight:1.2))

mix and match this sget me good results, you can also trigger lighting with other words its just you experimenting a bit, theres also specific words for lighting, like colors and temperatures.

in a sense weight :1 is top lighting and as you go on further weights the angle changes, :1.3 - 1.5 ive even gone as far as 1.6-1.7 to get some over the top lighting effects

bcs no body, betting on it anymore

I can, through ComfyUI and Python combined. Using it in DreamIO game and generating pictures just for fun. 3D models are great if you put some effort to positive and negative prompts.

its works smooth for me with any tbh, you just need to figure out what you doing, SDXL give it more steps to process it like more steps, Grapefruit you can get cool results with 25+ steps and it will also handle lora stacks and it also likes more steps to process complex prompts wich Grapefruit is very good at handling complex prompts