二次元復興時代來臨

所以我要玩你的Lora~

@mr_Jack hahaha~~~

Have you tried doing what the error told you to try? worked for me.

have you tried the sdxl vae fix?, usually give that error with the standard sdxl vae

@elguachiiii Yeah thx, i ask the question on huggingface it's the cause, i change it to sdxl.vae.fp16 it's work now

this is avoidable.
it's the effect on the aesthetic 3rd training to draw the average style of the model. most of the pic would have this as common trait.
but this flaw always answer a prompt and that might not be the word you look for. for example "nose reflection" or "nose shine" in negative prompt won't do much cause it's not in the image description of the training data at any stage (most of the time).
most user spam the word (masterpiece, highres) without thinking about the consequence it imply.
there is many common train using theses average words, and that include de nose highlight in it. most artist put their best efforts in their masterpiece work and the nose highlight is most common in that "masterpiece".

我在其他的xl模型上遇到过相同的问题，解决办法是删除相互矛盾的提示词，但是我不了解原理，也没有办法复现，所以仅供参考

1girl, young man in suit...

3.0 base is the raw result of the first two step training before quality reinforcement. here is what it says on hugging face :

Base:

Feature Alignment Stage: Utilized 1.2m images to acquaint the model with basic anime concepts. Refining UNet Stage: Employed 2.5k curated datasets to only fine-tune the UNet.

Curated:

Aesthetic Tuning Stage: Employed 3.5k high-quality curated datasets to refine the model's art style.

so i will assume that base 3.0 Animagine skipped the latter stage. why ?
it's to make the model less grounded to it's 3.5k aesthetic pictures making it more flexible and therefore better to train further on without much knowledge loss.

if you plan to train this model, pick the base one as it able to catch new concept better.
if you only generate, you will have a much better experience with the classic one.

In my case, accidentally using "latest, new" as part of the prompt was causing this issue. Changing it to "newest, late" solved it.

@Weltleere I did not use such tags

I had this problem when I used LoRA for feets/hands.
Also some prompt caused this bug also after.

Read the "About this version" in the top right.

I believe it's more for training? At least for what I understood.

Base is something you can build upon, while the other is the model focused on generation rather than training.

3.0 base is the raw result of the first two step training before quality reinforcement. here is what it says on hugging face :
Base:

Feature Alignment Stage: Utilized 1.2m images to acquaint the model with basic anime concepts. Refining UNet Stage: Employed 2.5k curated datasets to only fine-tune the UNet.

Curated:

Aesthetic Tuning Stage: Employed 3.5k high-quality curated datasets to refine the model's art style.

so i will assume that base 3.0 Animagine skipped the latter stage. why ?
it's to make the model less grounded to it's 3.5k aesthetic pictures making it more flexible and therefore better to train further on without much knowledge loss.

now that is a wonderful base model that i would happily use !
v3.0 base for training, (lora/dreambooth and finetune)
V3.0 for inference quality

@Le_Fourbe It's worth noting, also, that LoRAs learn whatever about your training set is different from the model. So if you train an anime character against a realistic model, it'll mostly learn 'anime'. Or... if you have a dataset that's got variable, middling quality images—and an aesthetically tuned model, like the 3.0 non-base one—then it'll learn your character, sure, but also "mediocre quality".

Using the base model instead of the full 3.0 model essentially lets you factor out "quality", and create LoRAs that learn just the character without affecting the output quality of the model they're applied to. You can use LoRAs on models that are closely related to the one it was trained on, and it'll usually work fine- which means if you train a LoRA on 3.0-base and use it on 3.0, you'll typically get higher-quality output than was in the training set!

This is helpful.

I agree, all the anime models of SDXL up to that point were tragic. Animagine is finally a sensible SDXL anime model.

@CagliostroLab mine doesn't load too, I got cuda.OutOfMemoryError .

@khanhdz7945 get a better graphics card or maybe you need to update your environment. this is a SDXL model (not TURBO), which still cost a lot of GPU memory.

Question is, does your setup load any other SDXL (not SD1.5) models?

@CagliostroLab it said failed to load checkpoint, restoring to previous... You know what, I think I can paste a few lines of interest here:

Failed to load checkpoint, restoring previous

Loading weights [5493a0ec49] from C:\AI STABLE\stable-diffusion-webui\models\Stable-diffusion\abyssorangemix3AOM3_aom3a1b.safetensors

Applying cross attention optimization (Doggettx).

changing setting sd_model_checkpoint to animagineXLV3_v30.safetensors [1449e5b0b9]: RuntimeError

Traceback (most recent call last):

File "C:\AI STABLE\stable-diffusion-webui\modules\shared.py", line 516, in set

self.data_labels[key].onchange()

File "C:\AI STABLE\stable-diffusion-webui\modules\call_queue.py", line 15, in f

res = func(*args, **kwargs)

File "C:\AI STABLE\stable-diffusion-webui\webui.py", line 199, in <lambda>

shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights()), call=False)

...

model.load_state_dict(state_dict, strict=False)

File "C:\AI STABLE\stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 2041, in load_state_dict

raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(

RuntimeError: Error(s) in loading state_dict for LatentDiffusion:

size mismatch for model.diffusion_model.input_blocks.4.1.proj_in.weight: copying a param with shape torch.Size([640, 640]) from checkpoint, the shape in current model is torch.Size([640, 640, 1, 1]).

size mismatch for model.diffusion_model.input_blocks.4.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([640, 2048]) from checkpoint, the shape in current model is torch.Size([640, 768]).

size mismatch for model.diffusion_model.input_blocks.4.1.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([640, 2048]) from checkpoint, the shape in current model is torch.Size([640, 768]).

size mismatch for model.diffusion_model.input_blocks.4.1.proj_out.weight: copying a param with shape torch.Size([640, 640]) from checkpoint, the shape in current model is torch.Size([640, 640, 1, 1]).

@CagliostroLab I am not sure why this happened, can you stuff like Kenshi and Abyssorange just fine

@Heartbreak117 Are those SD1.5 or SDXL models? If they are SD1.5, then most likely you don't have enough VRAM to run SDXL.

@Heartbreak117 abyss model and the other you have are made on an older architechture so it load as normal.
Animagine is a more recent base using the newest and bigger and more powerful model SDXL.
the error log is weird and suggest that you run a an older version of the WEBUI :
try to make a a brand new folder to install the webui again and put Animagine. it should work right away if you have at least 12Gb of video memory.

@Le_Fourbe I will test it out later then. Though if it doesn't work then it maybe hardware limitation at this point

it doesnt do landscapes well... pretty much only does girls and the same anime guy all the time

@jaimelugo428 prompt and workflow skill issue.

Yes, I can encounter this as well. Keep in mind that this is what I personally experienced, and they could've been coincidences or random, but usually I've found that:

- Changing the placement of the added word that had caused the noise result, can lead to a normal image.

- That maybe the added word is causing noise-filled images because there is another word that conflicts, for example: if you had both "from front" and "from behind", but then that could've been a coincidence on my end, it's hard to tell; I've had "conflicting" words before and the image generated fine.

- That even though there is no "conflicting" word in regards to the added word that led to the noise-filled image, there is another word that for some reason, when typed in the order it is in, or just its presence in the prompt, leads to a noise-filled image when in the same prompt as the added word.

- That increasing weights of words, the whole (word:multiplier) thing, whether they are in the positive or negative prompt, can also lead to a noise-filled image. I say this because I had increased the weights of some words before, then got a noise-filled image, decreased their weights, and got a normal image. Whether this actually means that weights can affect whether or not you get a noise-filled image, I'm not sure. I only know that it just so happened to be the case for that prompt I had.

- That typos may lead to noise-filled images, they sometimes did for me, and upon correcting the typo, the image came out normal. Additionally, extra spaces can lead to it, but not always. For example: "one apple" vs "one apple". I feel there is some randomness to it.

Basically, my process for when I get noise-filled images, is to:

1.) Check that I didn't have any typos/extra spaces.

2.) If I didn't then I would move the added word around, and generate with different spots for the word until I get a normal image.

3.) Lower word weights if I have any that have been increased from the default 1.0, or if I don't want to do that, I move to step 4.).

4.) I remove words that I maybe don't need, or replace words with different worded equivalents, and generate to see if a normal image comes out,

or

4a.) I try for equivalents of the word I'm trying to add.

5.) If still, I get noise-filled images, then I usually just give up on the whatever I was trying to add.

@Atarachi 
thank you for the advice! I have just started using the SDXL model, and it seems to respond to each word much more accurately than the SD1.5 generation. By using sentence-based instructions instead of word-based instructions and eliminating competing prompts, the noise was almost gone.

Hey man, do you have some kind of list of artists?

@Haiigdso3 There are no artists. This process exploits the understanding of the model to fit your needs. There are many methods to do this, I have chosen to use it for artists.
To find "artists", you must consider the behavior of concept mixing for the model you choose. I determine this by mixing two concept. Example: "group" and "tall" may become "lloatrup". If this produces a result of "group" and "tall", then the model is not good for mixing. If this produces an unrelated result, then it is good for mixing. This process is different for each model, so you can find how their perception affects the mixing.

@reptilekiller Thank you for explaining, im kinda starting to understand what you mean. I need more time lmao

@CagliostroLab hi, no problem, well I kind of understood how to use the model better. It is very sensible to everthing in the prompt (which may be good for complex images, but add chaos), so if you prompt something like "girl with giant muscles" it may sometimes do a muscular girl, but other times it will add some muscular dude aswell somewhere. So it may be better to prompt with adjectives like "extreme muscularity" or something. Or just "bodybuilder girl" will sometimes create a girl (muscular or not) and a male bodybuilder.

do you have any prompt tips suren? Would greatly appreciate the help

Make sure to use 1024 x 1024 resolution or higher! A lot of people don't know yet that SDXL doesn't quite work like SD1.5

Don't use any VAE's. The checkpoint already has an integrated VAE. A lot of the other VAE's ruin the images, and if it's a SD 1.5 VAE the results are even worse.

Are there any good alternatives for anime?

Skill issue.

@ZeroFLN dood, model should work from get go without magic prompts ok ? This is how most models work and this abomination is so messed up that even author needs extra prompts to force it to create decent pic, its not how it should be + someone told me not sure if truye that authors of this mod go around looking if someone merged this model with anything and they strike it down which is lunatic behaviour considering they stole most training images !

@CagliostroLab Excuse me, I can load this checkpoint, why?

Don't use VAE's. I've tried using some of the ones that we have for SDXL and for some reason the results are horrible (I haven't tried them on other checkpoints but for Animagine it seems to mess up the image, at least for me). Just load up this checkpoint and start generating without a VAE

@GogetaSSGSS3 Got it thank you!

@CagliostroLab got it! Thanks!

Do you read introduction carefully?This is my method.(I am not good at English xd)
Prompt：(masterpiece,bestquality),style(such as illustration、impasto),
1girl/1boy/2girls/1girl and 1boy,character,source,Feature1,Feature2,angle of view
background,other Quality words
Negative Prompt：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,artist name,(Maybe you can use negativeXL_D? a emb)
sample and step：Euler a ,28
Hire.fix：SwinIR_4x or latent(nearest),15-20 ,0.55,1.5
resolution：768x1280（→1152x1920 by hire.fix）

@Mostima Yes I read it! I've still been having trouble though, especially when there is more than one character in the image. They blend together.

Thank you very much for the suggestions, I will try these out!

And Im facing the same problem as you=(

我也遇到了和你一样的问题。你解决了吗

I also encountered the same problem as you. Have you solved it

much like NAI before it as Cagliostro said.
however you can counter balance it by calling it later on the prompt using square brackets. negative are like that too. You can lower the weight.
also relying on theses words will make the output less creative/random. this is no issue

If you want an easy to use UI without the high Vram usage of the automatc111 and the complexity of comfy ui, then try fooocus, it has an inpaint for sdxl built in(which is actually good)

if you want to learn comfy ui then there are several tutorials on youtube, comfy ui is very powerful interface.

@newtrialtime101704 i like how flexible you are with comfyui but it takes a bit to get into it. i've tried already a bit around with inpainting but without getting the result i was hoping for. that's why i was asking. i guess i will keep trying out stuff. thanks for your suggestion tho

@laborchef1698 then you could try out fooocus ui, it has in painting , outpaining all built in, and is very easy to use, as long as you dont care about regional prompting, animate diff, SVD, or other complex thing, and just want a good ui for generating good quality images out of the box, then fooocus ui is really good. here is the git hub link if you need it. https://github.com/lllyasviel/Fooocus its a simple installation.

Fooocus ui is great inpainting and outpainting, and has good inpainting models for fixing faces. I use it to generate 5120x1440 wallpapers. YMMV if you use the Fooocus V2 style though, I have had mixed results with how clean the art comes out. Also make sure you change to Euler A in advanced, advanced, dev options if you want to do as the author recommends since Fooocus defaults to dpm 2 sde

Which LoRA?

@Erocle anyone But when I use other checkpoints it works

@rdfssqa Just making sure, you are using SDXL LoRAs correct? SD1.5 LoRA will have no effect on SDXL checkpoint.

@Erocle thanks that help me solve the problem

Hello, I am not an expert in AI Art generation by any means, but I think you can use this model in conjunction with a LoRA dedicated to the specifc character to get the type of generations you are specifying.

Here's are some LoRAds trained on Ayanami Rei: https://civitai.com/search/models?sortBy=models_v5&query=Ayanami%20Rei

Read the LoRA page to learn how to trigger the LoRA on generation, and refer to a guide or video if you're not sure how to use LoRA's in the first place.

@BigChungus696969 Thanks, I am very well aware of how to use a LoRA. I personally had trained and published a bunch.

The problem is, this is a SDXL checkpoint and it DOESN'T work with any of the SD1.5 character LoRAs out there. 99% of the LoRAs on this website are not compatible with this checkpoint.

As Animagine XL says it was trained on anime pictures and had characters baked in, I found it surprising that it is not compatible with EVA characters. In the SD1.5 world there is even a joke that an anime based checkpoint must pass the Asuka test - it should be able to draw Asuka Langley out of the box. I am just a bit disappointed that Animagine XL, the best anime SDXL checkpoint, isn't able to draw Ayanami.

@phageoussurgery439 It's because animagine is trained on a subsection of danbooru, about a sixth of the potential images. Asuka / NGE was not a part of the categories chosen. 1.5 benefits from a corporate leak that means that it has seen every major anime character, so to get similar performance, animagine would need at least another $5-10k to see the rest of danbooru. SDXL is expensive to train and requires fixing the bugs in the available trainers for best performance. It's not that surprising that a SDXL model that has seen all of danbooru doesn't exist yet.

You might not know this, but the dataset of Animagine is open-source. I saved a copy of the image URLs and tags list and searched for "Neon Genesis Evangelion". It shows 474 results, which is probably why it cannot generate EVA characters properly.

In contrast, the term "Gawr Gura", which is 1 character, returns with 9679 results.

@kayfahaarukku Thank you all, that makes sense.

you may wanna try animeillustdiffusion if ur looking specifically for a good model to generate eva characters with no lora, v7.1 has very high weighted character tags that work almost as good as a lora

@nanamine Thanks for the suggestion! I just tested Anime Illust Diffusion XL. Yes, it can do Ayanami Rei out of box. Not as good as the LoRAs but pretty decent.

it will never supported, sdxl architecture are already different, you will still need seperated lora

People using SD1.5 LoRAs in SDXL and saying it works is placebo; as x1101 said, they are completely different

bra, transparent shirt, see through . maybe

@x1101 it just cutoff the chest part and show bra :-(

i got it to work with crop top overhang sometimes but no luck with anything else

@nise888 just use lora if not working

People will complain about anything even when it's given for free lol. This model is by far the best anime model we have and people are still complaining. Pony has better prompt understanding, but animagine has way better style/aesthetics in my opinion. Both are good in their own ways. Both are the best we have right now.

How much RAM does your GPU have?

@spicy_deluxe 6GB its a 1660 ti, since i downloaded the model the entire SD broke xD, I tried to install everything again but it hasnt worked

@adht100701219 The model alone is 6.46GB. Maybe take the model out of the folder so SD doesn't try to load it on startup? I've already gotten "out of memory" errors on some XL projects with my 24GB 4090.

maybe try to extend your pc's virtual memory can deal with your problem?

@spicy_deluxe thanks for answering, I couldnt make it work but after the third reinstall is working

@adht100701219 Try to use WebUI Forge instead of Automatic1111

I haven't tried it with this model, but with PonyXL, the best I've used is SwinIR. The ESRGAN models just didn't do so well for me and led to a lot of mottling and moire patterning unless I kept the denoise value very, very low. Strangely, even R-ESRGAN was giving me better results at one point.

As for the ESRGAN models, though, generally speaking, I don't like the "Sharp" models for anime, even 4x-AnimeSharp. If you want sharp, crisp upscaled images, and especially if the image doesn't have too much intricate detail, try 4x_NMKD-Superscale-Artisoftject_210000_G. If it does have a lot of fine detail, try 4x_NMKD-Siax_200k instead.

If you try them, please let me know how it goes with this model.

to com a mesma duvida

Had the same doubt, after searching a while I found this:
"Animagine XL 3.0 Base is the foundational version of the sophisticated anime text-to-image model, Animagine XL 3.0. This base version encompasses the initial two stages of the model's development, focusing on establishing core functionalities and refining key aspects. It lays the groundwork for the full capabilities realized in Animagine XL 3.0. As part of the broader Animagine XL 3.0 project, it employs a two-stage development process rooted in transfer learning. This approach effectively addresses problems in UNet after the first stage of training is finished, such as broken anatomy.

However, this model is not recommended for inference. It is advised to use this model as a foundation to build upon. For inference purposes, please use Animagine XL 3.0."

@Ailight Thank you. So for simple images generation it is better to use version 3.0 instead of 3.0 Base, if I understood everything correctly.

https://docs.qq.com/doc/DWGdwZ3lxUHVIQVRa?_t=1708191839528 try it

@rmllxx You are a hero

Can nai3 generate illustrations with special perspective based on prompt words? Like fisheye perspective, or the perspective of someone who's being trampled underfoot。

we don't train inpaint model

Can you also share which ones you found out?

@shadebedlam577820 some of my favorite artist like horosuke,eufoniuz,niliu chahui, torino aqua. you can also try some artists you like, maybe it will give decent results. To try it you can use tag like this: by artist name, for exampe, by torino aqua

@pecorine Thanks. I tried like 6 of my favourites but I dont feel like it recognized them but I will try more and some of your suggestions

@pecorine Also I wanted to ask do you just put the artist name in the prompt or do you do like :"art by ...", "by ..." or "artist: ..." what works best?

@shadebedlam577820 i didn't try it but I always use "by ...", it works well with me.

@shadebedlam577820 where do you come from

idk, maybe next month

it's already built in, but you can use this VAE if you unsure
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

It's shared publicly but due to it contain sensitive information we don't announce it everywhere

True. It knows nikke and blue archive pretty well. Surprisingly

I agree

sd3 would probably be too expensive to fine tune, sdxl is 3.3B parameters, and finetuning it is expensive in terms of hardware requirement. sd3 is 8B parameters, unless we are talking about sd3 800m parameter one, not sure if the quality would be better than sd1.5 tho, considering sd1.5 was 866 million parameters.

@newtrialtime101704 but you can hope that when its released, it will be optimized and maybe the architecture its a little different? why would they release something people cant use?

@torkothey do not want people to fine tune it. They want people to use their services as it is.

it comes back ੭ ᐕ)੭*⁾⁾

https://github.com/cagliostrolab/cagliostro-webui/blob/main/wildcard/character.txt

I think the training is based on booru tags. Still some natural language prompts work for me, but not complicated ones.

Heres some stuff that should help no matter the model https://github.com/comfyanonymous/ComfyUI_examples/tree/master/area_composition

Does other SDXL work for you? Have you checked if you have the latest updates and versions of torch, python, and if you have xformers?
I'm no pro at this in any way but if you treat it like working down a checklist I hope you'll find a solution.

there many solution, for start what kind of image if you send me one and send me your comfyui image i can help you,

@Hati_Hrodvitnisson i have everything up to date, and other models works fine, this is the only one that never worked for me so I eventually removed it. I dont know how to check if i have xformers... im still learning but I tried every troubleshooting I've found online and just doesn't work. I'm using Automatic1111

@cherodesu i was trying to use the Rabbit Hole Lycoris or even just random generated images and all I get is a garbled image with random colors, i dont use comfyui but Automatic1111 

@moon_sorrow I have counter the same issue before and solve it by removing my Lora.

See if following steps help:

a) Switch the model to SD1.5 model, waiting it finish load, and then another SDXL model, and switch it back.

b) Try if the model work with some simple prompt in text to image WITHOUT ANY LORA e.g. solo, 1girl. ← if not work, my solution not works for you.

b) If it works, try image-to-image with simple prompt e.g. solo, 1girl.

c) If that works too, start to add Lora trigger word in image-to-image.

For me it usually works after step (c).

@moon_sorrow Try using Forge UI, it is a more compact A1111 and a complete standalone (you don't have to manually tinker the dependencies) it is very user friendly and reliable. My other tip is, try to use it without LoRA or Lycoris or any embeddings, just use it as it is and see if it works.

I also encountered problems with this model. The resolution must be higher than 1024x1024. You also need to select the required VAE in the checkpoint settings, or use the one that comes with the checkpoint.

You are using the SDXL VAE, correct, and not any other VAE?

Same situation for me, any other checkpoints works well, but this one just get me some garbled illustrations somehow, no matter how I change the parameters.

@kmn199x3 im not using any VAE

@moon_sorrow Sometimes disagreements between the prompt and the negative can produce a black or red picture

@moon_sorrow I believe this needs a SDXL VAE specifically. Does this one work https://civitai.com/api/download/models/290640?type=VAE?

@maisi i'll try soon and let you know if it works, but I dont think it did last time

what are your computer specs? what is the UI you are using? if your VRAM is less than 8 GB then use Forge or Comfy UI, as they are much more optimised.

as for the lightning model, you could run Animagine 3 with lightning 4 / 8 step Lora, but expect some dip in quality, not too much tho.

@newtrialtime101704 oh, thanks for the advice. i didn't know you could run "normal" models with lightning loras.

i use forge already and have a RTX2060 (6GB VRAM)

while generating with animerge i get microfreezes and the sound starts to stutter. it takes way longer than generating with AAM XL, too. but it manages to finish if you wait long enough.

i guess my card simply runs OOM while generating with animerge and starts offloading things to RAM/swap file or something. i dunno how exactly it works. i tought maybe its because animerge is full fp32 and AAM XL only fp16, which should need less VRAM. (at least thats what i've read)

@ElevatedKitten 6gb should be more than enough to run sdxl on forge, try adding --xformers on the command line argument on webui-user.bat , also on forge there is an inbuilt extension called never oom enable it for unet model, also you can put --all-in-fp16 in the webui-user.bat to make all the models run in fp16. also if its really slow, you can disable offload to ram on Nividia control panel. tell me if any of this works. also first try using this model without any loras.

@newtrialtime101704 never OOM + --all-in-fp16 did the trick, thanks!

If your computer has less than 20GB of memory, it is recommended to increase the memory. With 16GB of memory on my computer, running SDXL twice with A1111 will cause a crash. Running SDXL with Forge will directly exit the program. However, after adding an 8GB memory stick to increase the total memory to 24GB, all issues disappeared. I found that the required memory usage can reach up to 19GB.

Translation:

It seems that the characters from the Touhou series are completely absent in this model. I believe there will be improvements in the 3.1 version. In the meantime, you can try this Lora model: https://civitai.com/models/287769/multi-in-one-model-for-touhou-characters

FP8 weight 

(Use FP8 to store Linear/Conv layers' weight. Require pytorch>=2.1.0.).
I know why. I turned on this setting. The rendering speed became very fast. After turning it off, it became slower, but I could see that LoRa was working. Some models can also be used with this turned on, but this model doesn't work.

I think it's not a problem with the model, but the UI.

where do you get 3.1 i only see 3.0 here ?

i have a doc with all the character that 3.0 had is there an updated version with the new added characters in 3.1?

@Phenix60 You can easily visit it at HuggingFace or at SeaArt now.

@Nightcliff cagliostrolab/animagine-xl-3.1 at main (huggingface.co) here it is

https://huggingface.co/cagliostrolab/animagine-xl-3.1/tree/main

@aiRabbit0 Thats not answer for my question, lol.

I guess v3.1 will be available on civitai later. If you need it urgently, you can easily visit it at HuggingFace or at SeaArt now.

@iNcUb From what I've read it's part of a sponsor thing, so they probably can't share the model here yet

here https://huggingface.co/spaces/cagliostrolab/animagine-xl-3.1/tree/main/wildcard

@114514ghs thx a lot mate