Anima is a 2 billion parameter text-to-image model created via a collaboration between CircleStone Labs and Comfy Org. It is focused mainly on anime concepts, characters, and styles, but is also capable of generating a wide variety of other non-photorealistic content. The model is designed for making illustrations and artistic images, and will not work well at realism.
It is trained on several million anime images and about 800k non-anime artistic images. No synthetic data was used for training. The knowledge cut-off date for the anime training data is September 2025.
Versions
Anima-Base
The pretrained, unrefined base model. Maximum flexibility, diversity, and style adherence.
LoRAs should be trained using this version.
Anima-Aesthetic
Fine-tuned for better consistency and a higher quality default art style.
Anima-Turbo
Distilled version for fast generations.
Use at CFG 1 and 8-12 steps.
The distillation process also increases stability and gives the model a strong default style, but reduces diversity.
I recommend starting with Anima-Turbo. On average, it is only slightly worse than Anima-Aesthetic, while being very fast to generate (and much cheaper if you use it on an online platform that scales the cost with step count). This makes it very convenient for quickly iterating on prompts. The increased stability can even make it better than Aesthetic in some cases.
Installing and running
Get the text encoder and VAE from the HuggingFage page.
The model is natively supported in ComfyUI. The model files go in their respective folders inside your model directory:
anima-base-v1.0.safetensors goes in ComfyUI/models/diffusion_models
qwen_3_06b_base.safetensors goes in ComfyUI/models/text_encoders
qwen_image_vae.safetensors goes in ComfyUI/models/vae (this is the Qwen-Image VAE, you might already have it)
Generation settings
Works at resolutions between 512^2 and 1536^2 pixels.
30-50 steps, CFG 4-6.
The Aesthetic version can tolerate lower CFGs such as 3, and often looks better with them.
A variety of samplers work. Some of my favorites:
er_sde: neutral style, flat colors, sharp lines. I use this as a reasonable default.
euler_a: Softer, thinner lines. Can sometimes tend towards a 2.5D look. CFG can be pushed a bit higher than other samplers without burning the image.
dpmpp_2m_sde_gpu: similar in style to er_sde but can produce more variety and be more "creative". Depending on the prompt it can get too wild sometimes.
euler: a basic sampler that is a bit more creative than er_sde. Good with the Turbo and Aesthetic versions, since those are naturally more stable.
If going for a more realistic / painterly look, the beta57 scheduler (ComfyUI RES4LYF custom node pack) can help make better textures, since it puts more emphasis on low-noise timesteps.
Prompting
The model is trained on Danbooru-style tags, natural language captions, and combinations of tags and captions.
Use lowercase for tags, and spaces instead of underscores. Score tags are the only tags that use underscores.
Recommended positive prefix: "masterpiece, best quality, score_7, safe, "
Recommended negative: "worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration"
When using a tag that is different between Danbooru and Gelbooru, prefer the Gelbooru version.
Prompt weighting works, but needs a weight higher than typically used for SDXL. Example: "(chibi:2)"
Aesthetic Version Prompting
Anima-Aesthetic is fine-tuned only on high quality images, with all of the quality tags stripped out from the captions. You don't need to use quality tags in the positive at all, but "masterpiece, best quality, " is safe to leave in. I recommend not using score_* tags in both the positive and negative prompt. It is already high quality enough and the score tags can push it too hard into slop territory.
Tag order
[quality/meta/year/safety tags] [1girl/1boy/1other etc] [character] [series] [artist] [general tags]
Within each tag section, the tags can be in arbitrary order.
Quality tags
Human score based: masterpiece, best quality, good quality, normal quality, low quality, worst quality
PonyV7 aesthetic model based: score_9, score_8, ..., score_1
You can use either the human score quality tags, the aesthetic model tags, both together, or neither. All combinations work.
Time period tags
Specific year: year 2025, year 2024, ...
Period: newest, recent, mid, early, old
Meta tags
highres, absurdres, anime screenshot, jpeg artifacts, official art, etc
Safety tags
safe, sensitive, nsfw, explicit
Artist tags
Prefix artist with @. E.g. "@big chungus". You must put @ in front of the artist. The effect will be very weak if you don't.
Full tag example
year 2025, newest, normal quality, score_5, highres, safe, 1girl, oomuro sakurako, yuru yuri, @nnn yryr, smile, brown hair, hat, solo, fur-trimmed gloves, open mouth, long hair, gift box, fang, skirt, red gloves, blunt bangs, gloves, one eye closed, shirt, brown eyes, santa costume, red hat, skin fang, twitter username, white background, holding bag, fur trim, simple background, brown skirt, bag, gift bag, looking at viewer, santa hat, ;d, red shirt, box, gift, fur-trimmed headwear, holding, red capelet, holding box, capelet
Tag dropout
The model was trained with random tag dropout. You don't need to include every single relevant tag for the image.
Dataset tags
To improve style and content diversity, the model was additionally trained on two non-anime datasets: LAION-POP (specifically the ye-pop version) and DeviantArt. Both were filtered to exclude photos. Because these datasets are qualitatively different from anime datasets, captions from them have been labeled with a "dataset tag". This occurs at the very beginning of a prompt followed by a newline. Optionally, the second line can contain either the image alt-text (ye-pop) or the title of the work (DeviantArt). Examples:
ye-pop
For Sale: Others by Arun Prem
Abstract, oil painting of three faceless, blue-skinned figures. Left: white, draped figure; center: yellow-shirted, dark-haired figure; right: red-veiled, dark-haired figure carrying another. Bold, textured colors, minimalist style.deviantart
Flame
Digital painting of a fiery dragon with glowing yellow eyes, black horns, and a long, sinuous tail, perched on a glowing, molten rock formation. The background is a gradient of dark purple to orange.Natural language prompting tips
Follow standard English capitalization rules for character and series names.
If using pure natural language, more descriptive is better. Aim for at least 2 sentences. Extremely short prompts can give unexpected results.
You can mix tags and natural language in arbitrary order.
You can put quality / artist tags at the beginning of a natural language prompt.
"masterpiece, best quality, @big chungus. An anime girl with medium-length blonde hair is..."
Name a character, then describe their basic appearance.
"Digital artwork of Fern from Sousou no Frieren, with long purple hair and purple eyes, wearing a black coat over a white dress with puffy sleeves..."
This is extra important when prompting for multiple characters. If you just list off character names with no description of appearance, the model can get confused.
Limitations
The model doesn't do realism well. This is intended. It is an anime / illustration / art focused model.
The model may generate undesired content, especially if the prompt is short or lacking details.
Avoid this by using the appropriate safety tags in the positive and negative prompts, and by writing sufficiently detailed prompts.
The model isn't great at text rendering. It can generally do single words and sometimes short phrases, but lengthy text rendering won't work well.
The base version is a true base model. It hasn't been aesthetic tuned on a curated dataset. The default style is very plain and neutral, which is especially apparent if you don't use artist or quality tags.
Finetuning tips
Don't train the LLM adapter. My own training script, diffusion-pipe, lets you set llm_adapter_lr=0 to completely disable training it, and the example config has this as a default.
Other trainers like sd-scripts have similar options that should be used.
The LLM adapter processes the text embeddings before they get to the diffusion model, and therefore has an outsized influence on the generated images. The adapter itself contains a surprising amount of knowledge and is easy to degrade by training it.
Use a low learning rate. For a rank 32 LoRA, start with 2e-5 and adjust up or down from there.
As a base model, there is no aggressive aesthetic tuning or RLHF you need to overcome when finetuning.
The model has an extremely large and diverse amount of visual concepts baked in already. A light touch is all you need.
Example of a style LoRA, with dataset and configs shared.
Online platforms
In addition to CivitAI, the following platforms also officially support Anima for hosted image generation.
License
This model is licensed under the CircleStone Labs Non-Commercial License. The model and derivatives are only usable for non-commercial purposes. Additionally, this model constitutes a "Derivative Model" of Cosmos-Predict2-2B-Text2Image, and therefore is subject to the NVIDIA Open Model License Agreement insofar as it applies to Derivative Models.
If you would like a commercial license, please email [email protected]
Built on NVIDIA Cosmos.
Description
Fully trained base model
FAQ
Comments (842)
Showing latest 310 of 842.
Simple One-Page LoRA Trainer for Anima (Portable, Auto-Captioning, Smart Cropping & Bucketing)
Raw image folder ➔ 2 clicks to Auto-Prep ➔ Hit Start Training.
Link: https://github.com/ThetaCursed/Anima-TrainFlow
I am Commander Shepard and this is my favorite anima lora trainer on this website
(i tried only this one)
Also, i tried first version with JoyCaption and WD14 tagger both. Should i continue to use JoyCaption or just stick to tags?
@Lostcut Thanks, Commander!
You should test both to see what fits your dataset best, though WD14 works great for me. If you plan to use natural language along with tags in your prompts, a hybrid approach is usually the way to go.
turbo lora v0.2 release: strength 0.7, steps 12, cfg 4, er_sde beta
is the WAY, produce the same result without turbo at steps 30 cfg 6, cut gen time half. All quality tags negative prompt stay the same as in non-turbo generation. Using my own loras with turbo, zero artifacts very satisfying.
Regarding Anima, I’ve given up on trying to achieve consistency using @Style tags. Instead, I’ve found that training a custom LoRA to handle the generation yields much more stable results—at least for characters that Anima has already learned.
My setup was minimal: I prepared about 18 images of the character in my preferred style, set the batch size to 2, and ran it for just 150 steps. Even with such short training, the LoRA outputs are surprisingly stable.
Granted, this method relies on two preconditions: Anima's base model must already know the character, and the LoRA should focus strictly on the character's stylistic features rather than teaching it specific outfits. Still, if you have the environment to train your own LoRAs, I highly recommend giving this approach a try.
Does this work on ForgeUI? If so, how? I get an error that says "ValueError: Failed to recognize model type!"
I heard that Forge Neo does work
i'm getting the exact same error, but the weirdest thing is that it worked partially at start showing the generating process, and then got a blank result, ending up by giving me this error message everytime
Working fine for me in forge-neo
I had the same issue when downloading the model through an API. It downloaded a 1.1Go file when the model is nearly 4Go. Try downloading it manually then move it yourself to the correct folder (I don't use ForgeUI but I guess it ld be the "stable diffusion models" folder). It worked for me
Forge-neo works,
@Proteinique Have you used Forge Neo with the same sttings as in this guide? If not try to use them.
Anima Artist Mixer v2 has now been released on GitHub!
The newly added parameters in the "advance" node can significantly alleviate the issue of style drift under different seeds
I am constantly running into "AttributeError: '_CrossAttnWrapper' object has no attribute 'q_proj'" when trying to use this node.
I got back to image gen after 3 years. Some folks I know were talking about this model so I tried it. Then tried the so-called go-to anime model - WAI-Illustrious. Anima is so much more aesthetically pleasing at basic prompts than WAI-Illustrious it is insane. WAI feels exactly like the models from 3 years ago, maybe slightly better. The natural text undestanding in Anima is great too. Only downside compared to WAI is that the generation takes 2x more time. Amazing base already, looking forward to how it evolves.
Wai is a finetune developer, not a style, and he also has an excellent finetune for anima (even though it was made on Preview 3, it's much better than even those that were fine-tuned on the base version). He squeezed everything he could out of Illustrious.
there is turbo lora recently realesed for anima base 1.0,reduce gen time from 50 seconds to 10-12 for me, and you can simply set clip strengt to 0,so it will only decile without changing style
Is there a character list for this model?
@ghosy01 no CSV like previous models?
@poisenbery there is one,actually.
https://github.com/ThetaCursed/Anima-Style-Explorer
@aa110902 looks like that only has styles/artists
On complex prompts I have low diversity on different seeds, compared to Illustrious. Any advices to increase it, besides reducing prompt and using wildcards? I don't use turbo, just in case
If you are using the turbo lora, it says "You can decrease the LoRA strength a bit below 1 for more variety." It just came out with version 0.2 btw.
There is the nuclear option of having an LLM/SLM remix the prompt.
Most likely the problem is in char loras
https://www.reddit.com/r/StableDiffusion/comments/1tobzgq/comment/oo03b0f/
Also tried a few shitmixes, and diversity is significantly reduced even more (char lora + merge - 0 diversity)
Possible solution - split samplers (first without lora (or with low weight), maybe about 33% steps, second with lora), Need to test lora scheduler also
It's just a characteristic of DiT models like Anima, Flux and basically anything moderately modern (as opposed to U-Net models like SDXL). They don't get confused by prompts nearly as much, but that confusion often manifested as perceived creativity. Basically, it's the reward for chasing "prompt adherence", you can't really have both. The idea is that if you want something different, tell it what you want different.
You will likely not see a new non-SDXL derived model that does not do this at least until some completely new architecture comes along (and probably not then either).
@John_KSampler interesting analysis.
I've noticed it happening myself, and wondered why modern models do it.
Maybe using wildcards would help with this issue?
Can we still use wildcards in things like ComfyUI? Idk.
I've never used them, but i'd say now is the best time to do so. We have a genuine use for Wildcards now.
@Dazrock wildcards just require some random script/extension.
One way to give your prompt more variety is to have an SLM/LLM remix your prompt automatically.
Shat out something, need further testing
https://github.com/Murdunbad/lora-scheduled-comfyui/tree/main
@John_KSampler This is a good explanation! 💯 Thank you.
I think it makes an interesting environment for AI art because if you want those rare creative treasures you have to go back to the old models. Although new models are "better" the old ones are still valuable.
I wonder if its possible to partially break the text encoder to make it unstable/creative... or for that matter simply use a different text encoder than than recommended one. Hacking the text encoder might be the next level trick for these modern times?
@CitronLegacy Text encoder seems to have a lot to do with this. As much as I know the CLIP(s) used in SDXL models encodes the prompt text much differently than modern text encoders using llm/t5. Clip-L and Clip-G encode text using byte-pair encoding, which seperately encodes letter combinations into tokens ("Hatsune Miku" would probably be encoded as "Hats" "une" "Miku"). Modern text encoders/tokenizers use something like Sentencepiece to encode with more context. But I don't really know about details.
Someone made CLIP work with anima: https://github.com/Anzhc/Anima-Mod-Guidance-ComfyUI-Node
But I heard the prompt adherence is much worse than using qwen te.
@RisingV That makes sense! "Modern text encoders/tokenizers use something like Sentencepiece to encode with more context." matches what I've heard other people say.
I always thought CLIP was kind of the same thing as text encoding? If so then this is exactly the kind of "Hacking the text encoder" that I was imaging!
@CitronLegacy Yeah, so I don't really know what I am talking about here :D
CLIP (or rather some sorts of CLIP) is/are the text encoder(s) used for SD1.5 and SDXL. As much as I understand CLIP is more like encoding every tag or rather part of the tag seperately in one token, while more modern text encoders encode longer text sequences at once. I guess this is part of the reason why anima has better prompt adherence/can handle more context, but illustrious is more "creative". But I guess we have to ask someone who knows more than me about this, like n_Arno or some CLIP expert like Felldude. Also there has already been some debate about aplying other versions of qwen llm as text encoder on the anima huggingface page, but since they are not trained with anima it seems they don't give better output.
will it still be possible to train anima in 2026 dataset for the next version if there is one? because the character info trained in anima is kind of lacking
I guess there won't be a "next version" by circlestonelabs with updated dataset. You have to look out for finetuned versions for that.
Hi, I use Forge, where should I put the downloaded files???
neoforge
The main model goes in the "Stable-diffusion" folder. For the required files, the Text encoder (gwen_3_06b) goes into "text_encoder" and the VAE (qwen_image_vae) goes into "VAE"
Is there a way to make tags work for specific characters? Like how Pony had the BREAK tag to help with separating chracteristics.
It works with describing a character in natural language. The text encoder of anima is much better understanding that than the one of SDXL models (Pony, illustrious).
If you use Stability Matrix and forge-neo,put qwen_3_06b_base.safetensors in Models\TextEncoders
Can i learn LORA for Anima on Osiris AI?
For some reason its saving images to temp instead of output folder in comfy. And I dont have this issue with any other model.
If you are using a custom wokflow make sure it has the Save Image node. I saw some wokflows with Preview Image node, which only shows the image without saving it. Comfy saves all images to temp, but they may get deleted after you close the app.
@NNAI That was it, the official worfklow example on huggingface with the anime girl didnt include the save node, just the preview node. lol
Will the next version possibly have an updated dataset?
Yup they still does not have furry and they need more to fix the pixel images
@Gabsnsu I don't see any announcement that they are continuing with training?
V1 Base is our foundation model.
I believe it's up to the community to improve it from here on out now.
Like we did with SDXL.
@Dazrock are you dumb
@Gabsnsu I think Dazrock is right. There hasn't been any change in dataset from preview1 version other than including some more images for additional styles. Currently tdrussel is working on a turbo version of anima and it does not seem like there will be any further training with an expanded dataset. Better quality can probably achieved using the turbo version, but for other things you have to rely on a community finetune. Also insulting someone will not get you anywhere.
@RisingV What I'm trying to say is that it's impossible for them not to keep updating this there are so many things it's missing Stop this stupid drama, blah blah, too much talk.
@Gabsnsu Chill. Go use any other models such as pony, IL etc if you want that furry, anthro bullshit.
@chrisss1 are you a dumbass too
@Gabsnsu I can totally understand you want more (furry) characters or capabilities. But I guess that's just not what circlestonelabs is gonna do. It was the same with illustriousXL. They trained version 0.1 and then refined it with 1.0 and 2.0, but they did not add much additional data. It was only that someone else used version 0.1 to train NoobaiXL with more anthro/furry capabilities by adding dataset from e621. Anima base is a true base model not designed for specific use cases. You just have to wait until someone picks it up and makes a finetune that fits your use case. Though I don't know if anyone will do a large finetune of anima given the less permissive license compared to illustrious. So maybe use some LoRAs made by the community for this?
@Gabsnsu Shut up furry. This model wasn't specifically trained on either furry or real-person content, but you can absolutely go train and use LoRAs to fix that, instead of shitting on other community members here. Don't tell me you're too lazy to lift a finger. Use it if you want it, don't use it if you don't, then fuck off
@Gabsnsu pay up another 50k dollar for him to train full e621 or artstation if you want, if not then fuck off. Even just captioning a full danbooru/e621 with a half decent vlm captioner can cost 10-20k of budget. Anima is a finished product and anything beyond it is community incentive and support. Or you can hope another DiT with mixed dan/e621 will be released soon (there is).
Hands struggle a lot around objects a lot. Hand detailers are not helping too much.
Does this model not work with Forge/Pinokio? It keeps giving me 'model cannot be recognized' errors anytime I try to generate
Use Forge NEO, there is even a preset for Anima
can someone explain what is "shift" in Forge NEO - Anima preset?
I can't tell what it does
In comfyui, higher shift value can shift more denoising to early stage, meaning the model spends more time working on outlines and composition instead of texture and fine details.
@necrophagism777 so you mean lower the shift make the textures better and higher shift made better dynamic motions and composition?
what is the sweet spot? the default is 4 I guess
手指和脚趾的数量控制很差，很难做到正确的5根
Any plans of updating VAE so it doesn't create that half-tone texture?
That only seem to happen with Euler/Euler A. "updating" VAE would mean retraining the entire model.
@deitychaser Some samplers are better, yeah, but it's still there, just have to look in the right spots for it. Or maybe I'm doing it wrong. I'm not sure what retraining entails but someone released a different VAE for Anima that reduces that texture a lot, just not completely, so there's hope for a proper fix.
@arnvc The model is not exactly great with details. Alot of it is due to not being trained properly on highres. Finetunes with better training nailing better details will probably fix this.
Can I just ask that you not make the Flux Klein error of reducing the steps down the the lowest possible for the distilled model, like 12ish steps would be good, right now ~40 steps is too long, but 12 steps is reasonable, that is if you can't just speed up the latency of the steps if you can speed on the speed of each step by ~4x then you can stick with 40 steps or ya know a combination, I'm just saying around 4x the speed would be just right, you don't need to make it as fast as possible people can chill for a minute for a high quality image.
CLIPTextEncode
ERROR: clip input is invalid: None
Use qwen text encoder: https://civitai.red/api/download/models/2945208?fileId=2824387
Does anyone know why adding lighting prompts causes black squares to appear and makes characters occupy only a tiny fraction of the frame?
Try placing "border" in negative.
apparently er sde sampler with simple scheduler fix it. this happen not just lighting prompts but also other prompts too with some have high or low chance to get it. training it seems to make that worse or reduce depending. its very hard to tell. because i fine tune it. so er sde sampler with simple scheduler and turn that shift thing to 3 as default. that would be hopefully fix it. i don't know what that thing is. but it is heavily baked into anima base itself.
For people struggling with hands, I highly recommend Linear Quadratic scheduler (found in WebUI Forge - Neo, not sure about other stuff). From my testing it does hands and fine details much better, especially at higher steps.
Good call, but also use dpmpp_2m_sde_heun_GPU with it. This will give better quality at the expense of time. CFG on 6 will also help. And try learning ComfyUI, it will help you push your generations even further
When using artist styles I use "watermark, signature, username," in the negative prompt, and it does help but sometimes I still get signatures and text etc. on the images. Any tips to make them go away 100% ??
Flux Klein can remove it.
Perhaps the rating system on Pony could be removed? I think this type of rating system, where the impact is unclear, actually puts users in a difficult position.
I agree. I don't want to see anything related to old models such as pony etc. It kinda weird to have score 7, score 8 etc in a model like Anima.
I do agree, but it's probably also worth not including the quality tag to begin with. That, and Anima exposes again the fundamental problem of UNETs struggling to understand how quality and detail correlate. The answer tends to be in the art style, which is why we're seeing users rapidly turn to Style LoRAs.
Even the Turbo LoRA drastically improves the detail, if at the expense of prompt adherence.
Idk why but version v1.0 has a harder time generating characters that look more mature than just lolis. It's becoming like wai-illustrious.
And doesn't really know mixing two artists art styles than illustrious-based models
illustrious is old you can't do these things With modern technologies
For me, the biggest problem with Anima is that it can no longer draw the “fat mons”. It's a pity. I wonder if future versions will add data on this aspect, or if I should wait for some kind soul to create a relevant LoRA.
I can't find any form or consistency to this model. Even with sample prompts, quality and style vary wildly, going from anime to flat color, to cartoon, to default AI plastic gloss, to straight abominations, even if you define the style.
It's slower than other models and implodes if you try to generate anything in high quality, and the prompt adherence is terrible, even with a high CFG, causing the model to hallucinate.
I was hoping to find a newly released model to rival SDXL, but this isn't it.
Well, it's a true base model and given that I think you can produce very good results with it. Style may vary if you do not use a style lora, artist style tag or a style block, yeah, but that's because there is no strong default style.
Maybe you are not using the right settings?
I haven't had much problems using it and prompt adherence is much better than with illustriousXL most of the time.
For reference I am using Sampler: ER_SDE, Scheduler: Beta, steps: 30, shift: 3, CFG: 5-6 most of the time.
@RisingV : Sorry for the silly question, but what do you mean by “shift 3”? I use almost the same settings and get really good results of surprisingly high quality.
Even in some finetunes/merges, I was making anime figures, some seeds the model goes a little too hard and adds doll joints which is fine but some seeds the model just shits out some giga slop hyper glossy skin images.
@tilleul97455 I don't really know much about it myself, but Shift is a parameter that is used by DIT models (as much as I know).
Here an explanation copied from a comment of another user:
"In comfyui, higher shift value can shift more denoising to early stage, meaning the model spends more time working on outlines and composition instead of texture and fine details."
If you get good results without setting a shift value, I would not bother.
In Forge Neo you can toggle on the shift slider in the settings under Presets->Anima.
I think the default value is 3...
In other inference tools (like comfy) I don't know how to set shift.
@tilleul97455 The other reply is better written, but I'll add a simplistic description of Shift. It's basically a 'creativity' limiter. Lower numbers mean more variation/hallucination.
What it's doing behind the scenes is changing how much noise is present at each step on a curve determined by the value you set.
It gets used in a lot of WAN video workflows. I'm a noob here, but it seems to work well as a way to set how much movement you want.
@RisingV You can set Anima's shift in Comfy by using the ModelSamplingAuraFlow node. Personally I haven't really had any images consistently come out better on higher shift (over 3), but maybe that's because I mostly use style Loras.
I recomend to use dpmpp_2m_sde_heun_GPU and linear quadratic. Heun is a second order generation, meaning it stabilizes the image better for more coherent detail, this way you can generate up to 3072x2304 before the image becomes just noise. I do recomend to use 2048x1536 since that is the point with my testing where the quality of the image is at its strongest. You can ofcourse, use lower resolution if you want, but the model will have less of a canvas to work with. The style is just your workflow being bad. I don't have those issues in the slightest, I can't help you without any details. But try adding modelsamplingauraflow node. The style and quality is probably because you have added SDXL Lora's on a different completely different architecture.
Has anyone tried using the Anima model to train some less prominent painting styles? I used the same set of parameters to train several Anima LoKRs. For the more prominent styles, they could basically match those trained with Illustrious. However, for the one less prominent style, its quality is far worse than those trained with Illustrious. After multiple attempts adjusting parameters and data with no success, I’m starting to suspect that the issue lies in the lack of prominence of the art style
：（
It's a bit of a 25/75 situation for me. I like crazy illustrations, but anima doesn't like that, as I assume that's a base model issue. In other cases, like my oil painting one, it executed it better.
Anima—not cut out for realism? Are you sure? With a photorealistic Lora and one for enhancing details (especially skin), we’ll be almost there.
I really like this model; it’s not perfect, but it’s only the first iteration.
Personally, I think it adheres very well to the prompt—even excellently. Compared to SDXL, it’s like night and day.
As for me, it’s “Goodbye, SDXL.”
For the image I posted this morning, I only needed 4 generations with ComfyUI, and that's really awesome.
i get just noise and im using all the correct vae, encoder, and recommended settings.
I recomend to use dpmpp_2m_sde_heun_GPU and linear quadratic. Heun is a second order generation, meaning it stabilizes the image better for more coherent detail, this way you can generate up to 3072x2304 before the image becomes just noise. I do recomend to use 2048x1536 since that is the point with my testing where the quality of the image is at its strongest. You can ofcourse, use lower resolution if you want, but the model will have less of a canvas to work with
It can do generic tv anime screencaps well, but for Illustrations like those on Pixiv, unfortunately no. I will be waiting for Illustrious V2 instead of this.
Instantly fell in love with this special one ♥
THANK YOU ⌒ω⌒
just love the possibilities.... every seed ~ like a little sparkling world waiting to be discovered. Just love it ⌒ . ⌒
a pure joy to explore ♥
Q: will there be an controlnet for Anima someday?
It already exists: https://huggingface.co/kohya-ss/Anima-LLLite/tree/main
Is experimental tho, but at least the depth CNet works nicely.
@Silvicultor These LLLite controlnets are okay but we really need an actual inpainting and anytest CN model
@Silvicultor Do you know how to make it work on Forge NEO?
@_Jarvis_ Just add them to the controlnet models folder and use the default settings (leave it on Any and leave the weighting the same). You can set your denoise fairly high around 0.5-0.6 and then just inpaint as you would normally. Definitely experiment with mask blur though, the results vary a ton and also be ready to retry a bunch since it rarely gets it right first try (esp. if you're using soft inpainting)
@_Jarvis_ If your Neo is up-to-date it should work out of the box with the Controlnet Integrated. At least the depth CNet worked perfectly fine for me.
@SilvicultorUnderstood. Thanks for the detailed explanation. I'll try again, because the first time I had a lot of errors in the Forge command line.
@Silvicultor did not know, I find Huggingface a pain to seach on XD
But I guess LLLite model is not quite what I think about, Im used to that an controllnet is as big as the model itself, so is this more like an ControllNet lora?
can someone give me the workflow
click into the preview image, on the bottom right, you can copy the nodes and paste them directly into comyfui
@izoraa I'm sorry, I can't find what you mentioned. Could you please explain it more clearly? I really don't understand this interface.
@jun60516750 Click on the model's homepage, which is also Anima's homepage. You'll see a red-haired girl with a blue butterfly on her finger, right? Click on that image, and on the right side, you'll see various meta data. Find "comfy: 10 Nodes," click to copy, then go back to comfyui, press Ctrl+V to paste, and you'll see the workflow.
@izoraa OMG this is a game changer! Thank you so much dude!!
Doesn't work on Automatic1111
A1111 is already abandoned, use Forge Neo for gradio webui: https://github.com/Haoming02/sd-webui-forge-classic/tree/neo
I did a bit of testing with the model, and you can generate images with a resolution of 2048x1536 natively. This is aprox 3mp. It is a bit dependant on the sampler and scheduler, which I recomend to use dpmpp_2m_sde_heun_GPU and linear quadratic. Higher resolutions are posible, at 3072x2304 was the highest I tested that worked, but the quality dropped significantly at this point. 2560x1920 might be a good middle ground of resolution and quality.
Anima is Illustrious V2 but with no furry or mix styles i hope they fix it In the future,
If I have certain unique preferences and I know I’m in the minority, I’ll try to research and create the images I want myself, rather than asking the majority to accommodate my preferences. I hope the creator won’t feel pressured by certain comments to include elements that aren’t helpful to the model or might even negatively impact most users.
遭遇了一个提示词相关的问题，请求大佬解答：我想设计一个半透明连衣裙下的露乳头胸罩的服饰，但人物的衣着为sundress与nipple cutout bra的时候，无论怎么调整权重和注释，模型总是会混淆。连衣裙上
出现不该有的开口，或者露出的乳头部分没有被连衣裙覆盖的效果。我该如何解决这一问题？
I translated this question with AI:
I encountered a promptrelated issue and would love to get some help from the experts: I want to design an outfit featuring a nipple cutout bra worn under a see-through sundress. However, when the character's clothing is set to a 'see-through sundress' and a 'nipple cutout bra', the model always gets confused no matter how I adjust the weights or annotations. Either unwanted openings appear on the sundress itself, or the exposed nipples are not covered by the sundress as they should be. How can I resolve this issue?
顺带一提，自然语言的方式我也试过，但效果依旧不甚理想。By the way, I have also tried using natural language prompts, but the results were still far from ideal.
我试了一下 不知道是不是你想要的效果
I’ve tested a variety of methods, but Anima’s upscaling consistently introduces dense black speckle artifacts like those shown in the image when the resolution goes beyond a certain point. https://civitai.red/images/133149340
I have also experimented with different VAEs, but none of them resolved the issue. I hope Anima Base will continue to receive updates and improvements to address this problem in future releases.
not good at inpaint either, have to use SD XL 😿 or will just become noisy and blurry
CLIP loading error. Anyone know how to fix this?
which clip node u using gguf node or safetensor node
@himanshu991123744 已經解決，謝謝
My Clip also has this problem. How did you solve yours?
@naa451448247 "I updated and restarted it a few times, and it suddenly started working. I think it’s pretty magical too."
i should start- spent hours having a blast with this base once i added my secret sauce : ) lol
-amazing work, its funny, cute, clever
--- question
To improve style and content diversity, the model was additionally trained on two non-anime datasets: LAION-POP (specifically the ye-pop version) and DeviantArt.
-are every single image from DeviantArt hand picked and approved/allowed by artist? -you should be above to give access to this, if so
- as well as this data set LAION-POP (specifically the ye-pop version) --i assume you went thru it and cleaned it
- i will have a modified version of your model that is amazing -- i will be in top ten with it : ) so i want to have as much info about this model because it has alot of potential
- i will have a presentation of myself as an artist for hire/work with/ etc... will post link when done
- every dataset i have seen -- i can spot hundreds of future issues --if you did not clean your dataset --with a trained artistic eye from decades of experience-- that needs to be done --now-- and release a new base -- anyway -- maybe more talk after i have my stuff up : )
Reading remarks abou this on the anima huggingface page may help: https://huggingface.co/circlestone-labs/Anima#dataset-tags
did a quick search for DeviantArtt::
https://huggingface.co/circlestone-labs/Anima/discussions/34
-- no reply from source
-- every major model maker does not place correct value to the full equation of importance::
code for system 50% important + a perfected dataset 50% important= we have not seen it
yet --- this is a Large selling point of a model if it was done, because manually going thru 100,000 image --1 by 1 and its txt file --- is a ton of work --you would let everyone know you did this : )
-- will eventually layout my findings
-are every single image from DeviantArt hand picked and approved/allowed by artist? -you should be above to give access to this, if so
shut up Meg
It's absolutely addictive, I'm completely captivated. It absolutely deserves the $5 sponsorship. Looking forward to more models :)
I wonder how much money they made from civitai
thanks
I cannot get this to produce good images. I have the ext encoder and VAE, but for some reason the images turn out too dark or not right at all. Any help is appreciated. I've tried settings/prompts that other people have used and it does not work for me. I'm using swarm UI if that helps
if you're using the basic generator-try DPM++ 2M SDE as the sampler and use "simple" for the scheduler. ER-SDE-Solver and simple also works. the workflow on the preview images should work fine if you're using the comfy workflow page
@listlessness I'll try it. Thank you!
I’ve been trying to get good images, but this model and its finetunes are so inconsistent for me. With the Turbo LoRA, I expected choppy results, but even the regular model doesn’t produce good results. The text gets completely massacred somehow, and there are multiple anomalies when the model tries to generate hands or legs.
I feel like I’m cursed or something, but Anima just feels like it doesn’t work for me. The workflows here aren’t that great to follow either; most of them don’t turn out the way they should for some reason. I think Illustrious is still king for me
its trained on relatively small res arts, best is follow up with trained res and then upscale then the results are stunning
try forge neo, euler a + sgm uniform works good for me.
I don't know what you're doing wrong, but I'm getting great results!
The best model for producing classic artistic images with a combination of low and high CFG. Images with a dance of brilliant colors and contrasts. Thank you.
I'm pretty dumb, so maybe someone can help me. There's no text encoders folder in the stable-diffusion-webui folder (and I've never used them before), so where should I install it? THX
I'm assuming you're using either A1111 or Forge. Anima does not work with either of those since those are older webui's that are no longer being updated to work with newer models. You have options, download Forge NEO (This one is the closest to what you're familiar with), SwarmUI, or Comfy. It's self explanatory from there.
Try using COMFYUI and move forward. Don't stay.
@Big_Soda Thank you!
Love it, for how long will the 5Buzz charge stay? its eating my stockpile
Forever, or maybe soon they will make it more then 5
Why no gay images here or males models made with this
Be the change you want to see
You clearly haven’t been watching the Gallery very much.
Because the statement is wrong.
@nogo maybe
my new favorite model i can't believe how good this. i mainly use it for photo real stuff. but it is awesome. color saturation, creativity, love, love, love it. thankyou for your effort. thankyou for diffusion-pipe.
Should we use the commas in a grammatical way when using natural language prompts? And should we ever use periods?
Very good question.
I dont think I saw much if any difference between using commas or punctuations at ends of phrases/sentences
Yes, as stated in the model card on huggingface.
Write like you write to any other LLM. Natural language!
Does anybody else genuinely not feel ANY difference at all from anima preview 3? i cant be the only one, in all my testing results are very similar in terms of quality and prompt adherence
Most of the training done after preview3 was for generation of higher resolution images. I guess the model already had solidified on most of the concepts/styles in preview3 version already.
Generate an image of 1536x1536 size and you will see the difference.
hii ! anyone can help me ? i download stable diffusion on A1111 and there is no text_encoders folders so i have grey generated pics.... someone know how install it ?
You downloaded a different version of Stable Diffusion. You need Stable Diffusion Forge Neo
You need to download either Comfy_UI or Stable Diffusion Forge Neo
I recommend Comfyui, use claude to help you through it, it gets things wrong at times but it is very helpful overall
doesn't seem to be able to do double penetration :(
either that or my prompt isn't good enough
I can do triple and even quadruple penetration. Set weights strength of matching piece of the your prompt by selecting text and using ctrl + arrow up/down set weight.
example (triple penetration):
https://ibb.co/JwRdW3W8
Check a double penetration post on gelbooru (which is what the model is trained on) and see which tags they use.
First anime model I've ever seen where "very dark skin" can actually work on the woman instead of the man when both are present honestly quite incredible
The concept is separated by two booru tags:
"dark-skinned female" and "dark-skinned male". Both work with the "very" prefix.
Anima can't generated cute chibi like Illustrious it seems. Hopefully next version will improved on that one.
skill issue.
skill issue
https://ibb.co/4wqbwSmh
@enjidifussion20820 what the fuck does that mean around your shit
you must be tarded, cuz I literally just put "chibi" with weight of 2 and got it perfectly.
@Meowzilla Most of the time you would get is deformed, not really chibi
After testing and comparing them, I can safely say that Anima falls way short of ILL. It's much easier to get higher-quality results with ILL, and its artistic styles are better—not to mention that training a character LoRA is way easier on ILL.
And yeah, there's a lot that Anima just can't do compared to ILL.
Use Chibi only or Chibikemo, Chibi is a general tag, that's why. But yeah, on tag knowledge Chenkin/NAI is better and the illustrious models you know, have Chenkin or NAI merged in, pure Illustrious is meh.
Can I not use this with forge classic?
no, you need forge neo. i installed it using this app mgr. https://github.com/LykosAI/StabilityMatrix
only for forge NEO use
18 credits for 1 image is crazy
Run it locally.
Quite possibly the best model to ever do it
Great detail lvl, outperforming Illustrious/Pony/ classic SDXL models.
It is heavier on the GPU side and longer generation time, but nothing even 10gb vram can't handle and tweak.
Prompting requires way more careful building and adjusting habits than older anime models, this one actually handles "natural" tagging like flux very well, not just booru tagging, it's safe to say it's more reliant and skill dependent on prompting than Illustrious/Pony/SD1.5 ever.
Read more info on the author hugging face, great must do prompting rules for effectiveness.
This model holds very strong potential and fair to assume many will be defaulting to it.
Cool stuff!
Anyone knows how to use this model properly on comfyui locally? I can't make any useful nodes work because it works as a diffusion model instead of a regular checkpoint so I can't use adetailer, inpaint or some useful nodes :(
try swarmui, its pretty good
I can help you figure it out.
First make sure your version of ComfyUI is up to date, it won't work otherwise as you will be missing the nodes you need. Then you can just use the built in Anima template in Comfy UI, or drag and drop my image into comfy and it should work... https://civitai.com/posts/29511171
I only use 1 custom node, for wildcards, but you can just disconnect that and then type in your own prompt like you normally would.
Was this model made from scratch or is based from another base model like SDXL?
Cosmos 2
This model still needs to be reworked, I find it unstable and the results are sometimes strange.
I guess they are waiting Cosmos 3 Edge 4b to use it as a foundation for Anima 2.0
@Lynx2025 While exciting for a cosmos upgrade, that would mean that all character/style/etc loras would need to be retrained. Worth it, but hope they tell us soon if this is the case.
@TTTTT55 there are already problems with finetuning Anima into a new version of AnimaYume
Too bad it's very complicated to get this model working with Comfyui. I can't even generate images although I've downloaded everything that I need.
Yeah i have the same problem.working with comfyui is very very complex and difficult to create a workflow without any errors.i almost downloaded over 150gb of safetensors model including checkpoints diffusion models vea loras and other stuff.but still struggling to create a good image/anime art i wish there was another alternative with a friendly user interface to work with it
i also thought it was complicated, but its actually very easy, just watch a tutorial
@neisangtr499 Forge Neo is supporting anima. It's actually very easy to use there I think.
Try using SwarmUI and utilize the Comfy interface (or just use the simple generating tab) it has in it.
Just try Invoke AI: it has built-in support for Anima and even can download and install it for you.
I'd like to ask if Anima will still be updated this month.
No lol Unfortunately, this is the final version But they will continue to update it. They didn't say that, but I'm sure of it because it still needs some things and fix too
Yeah It's kinda radio silence from Anima creator. Need public monthy roadmap from him like in game industry. And with ComfyUI as a partner he'd make a better nodes like Ideogram director for Anima
I just hope they keep upgrading the quality, especially in terms of artist accuracy, copyright anime styles, because most fine-tuned models people make just make it worse.
Out of memory error on my 12GB of VRAM. Anyone else?
I advise you to check if these 12 GB of video memory are free. In general, Grok is well versed in helping with ComfyUI and working with other interfaces.
Lower resolution, reduce steps and cfg, 4.5 + 24 steps is good low.
Upscaling is much more straining on this that previously, so play about your initial output ratio and tweak upscaling to be lighter.
I don't think this advice has anything to do with reality. He mentioned that he has 12 GB of video memory, which is more than enough for Anima and for models that are much more demanding than it. He probably has some kind of incorrect setting within the system itself or the interface.
check your batch count
I run it on 8Gigs, no problems. I use Invoke AI, not Comfy, but it should work in Comfy as well.
my 6gb GTX1660 runs it quite fast, so you should be fine
@sten4950 @IdoJiro Could you share your workflow? (If using ComfyUI). If you share any image here on Civitai, I can copy it. I'd be very grateful 🙏
Im praying for more updates in the future
I'm glad that the basic CLIP has been replaced by a proper text encoder within an SD-like model. I hope that in the future, the vast family of Pony and Illustrious (and others) will also get a proper text encoder, which is much more convenient and smarter than the basic CLIP—which can't even properly grasp contextual relationships.
If they keep updating this will they be able to remove deformed and blurry stuff finally ? I know Ai images is Random But is this possible
Greetings. The checkpoint is awesome, but I have a few questions: first, what is the safe weight limit since the checkpoint works with an increased weight? And secondly, the model produces decent quality on the Eyler A, but it does not always follow prompt. And on DPM2++ SDE, in principle, follows, but it often gives out a fierce mess. Is there any way to fix this with settings?
Do you have a working IPAdapter model for Anima?
After switching to "anima," my generation speed increased about threefold—is that just how it works? Or is there a trick to it? If anyone knows, please let me know!
so far i can tell the same thing not sure why might be the way Anima is trained less stress and space on your hardware
Try with TurboLora, can increase your speed again, partially with I think more than x2.
Depends on the pipeline. I had this happen when I simply layered it into my old Illustrious workflow. FacialDetailer for example is literally just a hamper, you don't need it. But additionally things like that aren't necessarily tuned for Anima, resulting in longer generation times.
Anima also prefers in general to have less. Less CFG, Less steps. Bit different from Illustrious.
@__VL why? Detailer is pretty important, just don't use it together with starting generation, refine later
@Aratoum I think it works good without. I don't use it.
@Aratoum Detailer isn't quite needed on Anima, Anima is more well trained. Additionally it's the architecture of Anima compared to Illustrious, Pony, etc. Detailer is simply built different, it doesn't necessarily understand Anima as well, leading to longer generation in my experience. Meanwhile Pony, Illustrious, etc all have loosely similar architecture.
It is important, just for checkpoints that aren't Anima.
Is there an issue with the Qwen-Image VAE?
I'm generating at 920×1536 and using a standard hires fix workflow:
Generate the initial image
Upscale with an external upscaler (ESRGAN/UltraSharp/AnimeSharp, etc.)
Encode back to latent with the VAE
Run a second sampling pass
Everything works fine when the upscale factor is 1.5x. However, once I increase the upscale factor to 1.75x or higher, the image develops a large amount of circular/grainy noise after VAE encoding and the second sampling pass.
I've tested multiple external upscalers and the issue appears regardless of which upscaler I use, so it doesn't seem to be specific to a particular ESRGAN model.
Has anyone else experienced this? Could it be a limitation or issue with the Qwen-Image VAE?
It doesn't like hires fix, when you go start with anything beyond 1408. You need tiled for that.
I noticed the same thing, but only when using a realistic Sam Anima variant, If you are using a turbo lora I found that turning down the number of steps helps to get rid of the artefacts. You can check out my img2img workflow to see my steps: https://civitai.com/articles/31980/sam-anima-anime-to-realistic-image-transformer-with-auto-captioning-and-pure-high-res-fix
Its so great.
Are there any plans for controlnet models in the future? The natural lang processing is amazing compared to the open source illustrious models we have access to. However, having more control via inpainting and scribble or canny models would be AMAZING.
Kohya-ss has controlnet model for Anima, check it out
Kohya-ss control net works fine to me: https://huggingface.co/kohya-ss/Anima-LLLite
@FinisherStrike @RavirKun Ah shucks, I'm terrible with nodes. I personally use Invoke which supports Anima but not Kohya-ss via their canvas and nodes, at least not yet. Worst case I'll have to learn how to use comfy to work with canny, linear, and softedge models.
@Anominalmoose use this workflow, work fine: https://civitai.red/models/2576647/animasimple-t2i-workflow-with-upscale-detailers-and-controlnet
Why, when I use this workflow, do male penises or female vaginas always turn into a solid white patch? Where is the setting wrong?
Share prompt and settings first to answer on your question.
here is my prompt:
i use 2 anima lora one is crotch blowjob and anime character.
this is my prompt:
Expressions of contempt and anger; nipples ring; cum in mouth; (nipples: 1.3); nude; completely nude;
looking_at_viewer, blush, 1boy, hetero, penis, solo_focus, pov, oral, fellatio, pov_crotch.
black blazer, blue plaid skirt, red stripe bowtie, grey hair, light brown hair, green hair bow, bangs, one side up, long hair, brown eyes,lv_kotori.
Negative:
worst quality, low quality, score_1, score_2, score_3, old, jpeg artifacts, logo, watermark, bad anatomy, bad hands, missing finger, head out of frame, wide eyed, red light, green light.
the out is everything good, but the penis is fill white.
i try to add penis,vagina,uncensored, no censorship, detailed pussy, detailed penis.
Negative add censored, white patch, white blur, solid white, mosaic, censorship.
sometimes is working (3/10) images. almost the penis or vagina fill white.
my workflow is Anima (Preview) Workflow v6 detailer. and just enable "Use SAMLoader and Save Image" ,1024*1536 ,step 30,cfg 4.0,
i try to "Enable NSFW ADetailer" but still has same problem.
i try to use some user create image like 81189, use the same lora + checkpoint. but still has this promlem.
@acer1204
Some Ideas:
1. Don't write with underlines like "solo_focus", do "solo focus" the underlines only with score-tags
2. Try use the true promptstructure:
- quality/meta/year/safety tags
- 1girl/1boy/1other
- character
- series
- artist
- general tags
3. I'm not sure that the ADetailer works with Anima, try without it some times.
4. Try alternate Checkpoints, I like Silvermoon
5. If nothing helps, you can change the workflow
It's a fantastic, top-notch model!!
Quick question, mate. How can I use this in SwarmUI? Is there a way to integrate this?
literally just follow the installation instructions they give you(SwarmUI/models/diffusion models......etc etc) don't worry about manually choosing the encoder and vae in the app, swarm will do it automatically. choose it from the model dropdown like any other illustrious checkpoint. S/S: euler ancestral and normal. others will work but you have to figure that out on your own. if you're using the comfy workflow just drag and drop any preview image into the app and it will place the nodes for you
Amazing model but I have an issue I haven’t faced when using illustrious, sometimes it makes a style more dominant above another one like when using a style lora and using an artist tag it prioritizes the artist or even with only artist tags. The concept of style merging I think can be better
Why does it add "@photograph" at the beginning of my prompt?
probably some lora you are using that adds it
@nogo I was using the same LoRA with a different model (Anima based) and it didn't add anything. Also the one I'm using is the official Turbo LoRA, which shouldn't be a problem.
Update: I tried removing the LoRA, but the @ is still added at the beginning of my prompt.
Could be workflow related, if you don't use the one provided.
In case anyone has the same issue:
I found the problem. It was the model-keyword extension; apparently it detects the anima model and adds a style tag at the beginning. I just disabled it and went back to normal.
@Fatbuns I'm using Forge Neo, but thanks for the advice
Awesome base!
Edit model would be PEAK
Can you use controlnet depth and controlnet open pose with this model?
In my view, this is a far superior model to ILL. It outperforms ILL significantly in complex scenes and multi-character interactions. Anima may not perform well in certain scenarios (such as with specific tags or upscale), but as it is a base model, you can easily use fine-tuned versions created by others or train your own Lora to compensate for some of its shortcomings.
Creating a Lora model isn’t too difficult either; apart from the tag order and a few training parameters, it’s identical to the ILL model that uses Danbooru tags.
The speed and size are also excellent; the 2B size ensures it runs smoothly even on graphics cards like the 4060/5060, which only have 8G VRAM. As it uses the DiT architecture, it’s normal for it to be slower than the U-NET-based SDXL model.
If possible, I hope that future versions will change the LLM -> T5 text encoding architecture to a single LLM architecture, as this would take the model’s text comprehension to the next level (the token limit is 512; the model crashes when processing text exceeding this limit).
As for upscale, you can use Impack's Make Tile SEGS+Tag Generator + Detailer (SEGS) to do this; in my view, 1284x1824 -> 2568x3684 works very well.
@ThreadZen sorry to bother but i am training a lora first time i dont have enough gpu ,so i am running on free google colab tier,may i ask you which model illustrious or anima is faster to train on t4 gpu google providees
@LastDelivery4801226 My GPU is Arc B580. At a resolution of 1535^2 and with a batch_size = 1, the speed when using sd-scripts is 3.86 s/it. use torch compile, VRAM usage is 12 G. If you use a CUDA device, the VRAM usage should be lower.
Anima to me, represents the evolution of general AI image crafting. With the Turbo Lora this thing is a beast. I've always had 30s per image on Illustrious. Anima brings that down to 15s - 20s, with extreme fidelity. FacialDetailer isn't even needed. I cannot praise that enough. I should add though, I genuinely dislike the lack of knowledge in this model.
It acknowledges (loosely) some concepts but not too strictly. Sizing in general it struggles heavily with (if not 100% female anatomy). It always feels like it only understands approx half to roughly 3/4ths of what I prompt. I've even made sure to carefully select Danbooru tags. Honestly, I hope we can continue to see it improve, as I would like to see it simply understand more. There's a nuclear bomb of potential, it's just watered down with a communications barrier.
Use natural language prompts, the Qwen text encoder doesn't seem to respond well to Danbooru tags as one would expect.
@kitekholin natural language has no effect the model fundamentally doesn't understand certain concepts. (Example if making a futanari, try editing the testicles in any way)
You're supposed to use tags in tandem with natural language. Tags first, then natural language.
@__VL I doubt the vanilla text encoder would even respond well to nsfw topics. You might want to grab the heretic-abliterated-uncensored version of the text encoder from huggingface.
@kitekholin the vanilla text encoder responds very well to NSFW prompts. But as I said, not every concept is understood.
I downloaded heretic-abliterated-uncensored and the results are the same. The base model of Anima just lacks fundamental understanding of concepts, which is limiting as most checkpoints built on this.
@kitekholin the text encoder Anima uses is fast but pretty weak at understanding what you write. Danbooru tags are pretty much must, if you want to reliably lead the model
Sometimes Danbooru tags work like natural language, some seeds use a random artist's style, so I never know what's gonna happen, and that's what I love about it; every image and every prompt has to be carefully tweaked, and that makes it special.
@Aratoum I am able to steer my images quite well. If the model doesn't reproduce what I want I often phrase it another way, add more details, look for synonyms, parenthesis the word, and sometimes even add "must have".
For example one time I wanted an image of soldier with todays military kit, added something like soldier with combat armor and it gave me some generic soldier from an anime. So I specified, plate carrier, bags, straps, assault riffle. Got close, then the assault riffle looked weird, so I then specified "AK-47, Kalashnikov riffle" and boom got a soldier with today's military kit and an AK-47 as a weapon.
The same feeling—images generated by Anima are of very high quality, but its semantic understanding in complex scenes is not very good, making it difficult to achieve intricate physical interactions among three people.
Help. I added all the components needed and I get this error pointing to the negative prompt
"RuntimeError: ERROR: clip input is invalid: None"
Any help would be appreciated.
It requires a text encoder / clip model. If you're using ComfyUI this is a required component. If you take an image posted on the page for this, you can drag and drop it into ComfyUI and it'll automatically generate the workflow that was used to create the image.
Has anyone got this running on Forge UI? I get "ValueError: Failed to recognize model type!". Using ComfyUI in the meantime; a shame, since it's so much slower than Forge for some reason.
I've been testing Anima for a while now, and while I can definitely see what it's is trying to achieve, I feel like it gives up too much control in exchange for natural language prompting.
One of the biggest strengths of the SDXL/Illustrious ecosystem is how responsive it is to prompt engineering. Danbooru tags, prompt weights, emphasis, and prompt structure all have a predictable impact on the final image. Small changes often produce small, intentional differences, which makes prompting feel like a skill that can be learned and refined over time.
With Anima, I constantly feel like I'm fighting the text encoder instead of directing the model. Although it still accepts Danbooru-style tags, its Qwen-based encoder is fundamentally designed to interpret prompts more like natural language than discrete concepts. Because of that, tags don't behave with the same precision or responsiveness that CLIP-based models do, and weighting syntax such as (tag:1.2) no longer provides the level of control many experienced users expect. Prompt tuning becomes much more abstract, and debugging why a prompt isn't producing the desired result is significantly harder.
I understand why this approach is attractive for beginners. Being able to describe an image in natural language is far more approachable than learning thousands of Danbooru tags. However, one of CLIP's biggest advantages was exactly the opposite: it treated concepts almost like building blocks. A tag such as red hair consistently activated a learned visual concept, and combining, weighting, or restructuring tags allowed users to fine-tune the generation with remarkable precision. While CLIP wasn't as good at understanding long, contextual descriptions, it was extremely responsive to prompt engineering. Natural language prompting feels much more opaque, making it harder to iteratively refine an image toward a specific artistic goal.
The Qwen encoder is impressive from a language understanding perspective, but I'd love to see better compatibility with traditional prompt engineering techniques instead of replacing them entirely. I think that would make the model appealing to both new and experienced users.
Had zero problems with prompt controll. Use less quality modifiers and good sampler and you will be fine.
I am using anima with tag prompts and tag/nl hybrid prompts and overall the prompt adherence has been better for me with anima than with illustrious. If you are applying weights to tags, you should use higher weights than what you are used with illustrious, e.g. :2, :3, :5.
I can say there are some concepts working better in illustrious, but overall I am amazed at what anima is able to understand (especially in LoRa training).
@RisingV as I mentioned earlier, I've been testing Anima for quite a while now, so naturally I spent some time learning how to prompt it properly. I adapted my old Danbooru-based prompts to fit Anima's recommendations, removing most of the SDXL/Illustrious-specific formatting and restructuring them into a more natural language style.
What I do think Anima excels at is exactly what you mentioned: contextual understanding. It does a great job of grasping the overall scene and intent from just a few lines of text, and that's genuinely impressive.
My issue isn't with the model's ability to understand context, but with fine-grained control. I find it much harder to iteratively refine an image compared to CLIP-based models.
For example, let's say I want a wider framing that shows the character and the background in a very specific proportion. With XL/Illustrious, the model has a strong understanding of concepts like "wide view", and small prompt adjustments usually produce predictable changes, so I know which direction the generation will move.
With Anima, I often feel like I have to keep rewording or expanding the description until the model understands what I'm trying to achieve. It certainly understands the overall intent, but I personally find it much less responsive when I want to make small, precise adjustments to the composition.
That's really the point I was trying to make. I don't think Anima is a bad model, in fact, I think it has a lot of potential. I just miss the level of prompt controllability that CLIP-based models have developed over the years.
Ok, yeah, probably the way you have been using illustrious/sdxl does not work as good with anima. I guess I haven't been using it that way.
CLIP is just very different from any llm text encoder, so it's not a surprise you are getting different results.
I think someone tried to make CLIP work with anima, but I heard from another user that prompt adherence decreases significantly using it. Haven't tried it myself though.
I believe the two AIs have different focuses.
Anima is designed to create raw assets that follow a consistent line of reasoning and focus on context, aiming to solve a pain point SDXL never managed to address: specific details.
I don't think any current SDXL model, including Pony or Illustrious can handle this task as well as Anima does.
On the other hand, SDXL models still have the edge when it comes to more beautiful styles and the specific nuances of unique images.
I highly recommend using Grok to get guidance and learn how to use the tool.
i need anima training in Google Colab or any website to make them for free 🙏🙏
https://civitai.com/articles/28641/anima-lora-trainer-app
I found this bro, you can try it
@hakfull95539 This is the old version i can't
Yes me too
@Ishsbskskhxbdjskd you can use this one, its updataed with base 1.0, the only caveat is that its a little more complicated and requires you to have a gui on your local computer.
@emanr is it not possible to use this on google colab 😭
@LastDelivery4801226 please explain how
@LastDelivery4801226 yeah you can definitely replace the model to the base, you'd just have to replace all the text with anima-preview3-base and anima-preview to the base version name e.g "anima-base-v1.0, or anima-base-v1.0.safetensors" in the related places. and it should work. most importantly is the hg face links corresponding to the model
"Hey team, first of all, thanks for this amazing model! I’m currently building a multi-stage workflow in ComfyUI using Anima as the base generator for composition and prompt adherence, then piping the latent to Z-Image/Flux for rendering photorealism.
Since Anima utilizes the advanced Qwen-VL text encoder, I've noticed that using a standard negative prompt with anime tags (like Danbooru) tends to degrade the composition. Instead, I started using a ConditioningZeroOut node for the negative slot, combined with natural language prompting for the positive text.
My question is: From a model-training perspective, does Anima prefer this unconditioned Zero-Out approach over a traditional negative prompt? And how well does the Qwen text encoder inside Anima handle pure natural language storytelling vs. traditional Danbooru tag weighting for high-quality details? Thanks again!"
using forge neo. running on rtx5090. image output is black. using correct encoder and vae. any idea what the cause is? happens even on fresh install :c
Remove all the checkpoint model included vae and encoder on the top and remount it, If the issue still same try add --disable-sage at the webui.bat
There's a few other comments on this too, but using Simple or Normal as the scheduler is another step to look out for. It fixed it for me.
Did you make sure to pick the right scheduler?
model works great but the result often comes out a it blurry, I want to get sharp lines and sharp detailing how do i achieve this?
Im currently genning at 1536(2048x1152) and then omnisr2x upscale when im happy with the result.
All default parameters: er_sde, 30-50 steps, cfg 4
i think you can use another lora like for highres or context detailer or something? and i heard sometimes if the resolution is quite big like you say 2048x1152 unlike the recomended setting, it'll get blurry because the model try to scretch things i guess.
im also new to these so maybe try adjust the resolution as the recommend setting, and the use like lora for highres and for the line i think there is like boldline lora out there that can reduce and increase the size of artline, and after that you can upscale it like usual
@rastandaa well in anima's own words I can generate at 1536^2 resolution so it's still within anima's parameters, im looking around to see what i can do
you don't, the model sucks. Increasing resolution will just get you bigger blurry blobs.
I'm convinced that the blurriness comes from an artistic style, since they tend to go away with lora, artist, and artstyle tags. I generate at 1280*1536 and the details have been crisp and coherent enough to not need FaceDetailer and Ultimate Upscaling.
@Milanor sadly the artists i like dont have a lora yet. what are some artstyle and detailing tags are you using ?
How does turbo-v1.0 compare to base-v1.0 with turbo lora?
Turbo lora is just a basic distillation lora trained directly on Base. The full Turbo checkpoint has its own aesthetic tuning and distillation. It's more complex than just base model + merged lora. In terms of style and quality, IMO the turbo lora v0.2 is quite noisy and "slopped" in a way that the full Turbo model isn't.
amazing
Do you have any other plans for the future? For example, removing this layer of adapter or replacing the model kernel with a newer version of Cosmos with larger parameters. The adapter in the model has really caused a lot of trouble for the community's development work. The development work for ControlNet (non-LLLite version) and IPA is difficult to carry out because of this adapter
Nvidia hasnt released Cosmos 3 Edge 4b yet
@Lynx2025 To be honest, I would prefer if a model version without an adapter were released in the future. It has really caused a lot of trouble, although I can understand why Circlestone Lab did it to minimize risks
I am confident that the LLM adapter isn't a problem, and that its existence has actually made the model much better than it otherwise would have been. If you don't train the adapter, then text encoder + adapter combo just acts like a frozen text encoder, same as almost every other DiT model.
There is a recent paper that shows that a stack of transformer layers to pre-process text embeddings before the DiT (basically what Anima's LLM adapter is) actually improves results a lot. And this is in fact exactly what Krea2 does in their model.
Very early on, after preview1's release, some misinformation started spreading like "almost all the knowledge is in the LLM adapter, the DiT is barely trained" which is provably not true. Seems like ever since then some people are convinced the adapter is breaking the whole architecture somehow.
Anyway I do plan to start looking into future improvements, and official controlnets are one thing I will be investigating.
@circlestone_labs I am not very familiar with this technology myself, but I came to this conclusion after seeing discussions among community developers in the group chat I joined. Since you can personally reply to me and explain your understanding, I am not sure what to say. I hope you can succeed!
Another point I would like to give feedback on is that Anima's current performance in drawing buildings and distant views seems a bit sad. Is this due to the small number of model parameters?
@circlestone_labs I never understood why people said you are a conceited hypocrite, but now I do. I'm actually losing respect for you.
Is it "Probably not true"? Is it true or not?
It didn't even matter at this point, but you had to take a stance. Did it burn you that people figured out the issue so quickly? Luckily, they told you to stop training the adapter.
Yes, even if it was necessary, the adapter sucks overall. I don't think that getting rid of T5 is a bad decision by itself, but not acknowledging that the adapter created other issues is just intellectually dishonest.
Just admit the problem. People can accept that you did the best you could with the tech available and the results are good enough.
@Daru_22 provably not probably. Provably means "able to prove".
@areinu Even worse.
@Daru_22 What is the issue with the adapter? I am genuinely asking because I actually don't believe it is causing any problems whatsoever.
When training the base model, there was an initial adapter alignment phase, where its output embeddings reached over 95% cosine similarity with the original T5. This is so close it can serve as a drop-in replacement for T5 in the base Cosmos 2b model. At this point, I could have left the adapter completely frozen for the main training phase, and it would have been nearly identical to training with T5, but with a lighter weight text encoder.
And I actually tested exactly that. Two initial training runs, identical parameters, one with adapter frozen, one with it trainable. Overall visual knowledge acquisition was very similar, validation loss curves were similar, just with a slight edge towards the trainable adapter version. Upon visual inspection, artist styles saw the biggest benefit from a trainable adapter, since the adapter is acting like a mini trainable text encoder, allowing artist tags to gain modified embeddings with semantically meaningful information in them, letting the DiT separate out and learn artist styles more easily.
Much deeper into the training run, I did another experiment with temporarily freezing the adapter, and validation loss continued to drop at the same rate. Showing that nearly all new visual knowledge is going into the DiT weights, as you would expect. This is why I say the misinformation of "almost all knowledge is in the LLM adapter" is disprovable. Admittedly I haven't shared these intermediate checkpoints directly showing this, but I have run the experiments. I'm not just blindly asserting something.
The adapter is acting very similarly to trainable CLIP in SDXL models. It is easy to "fry" it, exactly like it is with SDXL. If you do train it, you should use a lower LR than the DiT, again exactly like CLIP. This doesn't mean it's a problem, it only means a trainable text encoder is inherently more sensitive than DiT weights, since the text embeddings have an outsized influence on the whole generation process. If you want maximum stability, just freeze the adapter, then it's acting like the model has a frozen text encoder, same as most other DiT models. Which is why that is the recommendation, it's the safer and more conservative option.
Are the Turbo and Aesthetic versions merges of Base with the existing official Turbo and Aesthetic LoRAs or actually brand-new distillations/finetunes?
Valid question
They are completely new full finetunes, based on their own datasets and with new techniques.
@circlestone_labs Will you be making a "furry" based finetune in the future?
@circlestone_labs will you be updating the seperate turbo lora for the final version?
@circlestone_labs Why are this model's parameter names so different from those of other fully fine-tuned models? Also, when comparing the aesthetic version with the base version, the weight differences are concentrated entirely in the linear layers that are typically targeted by LoRA. Does this suggest that the aesthetic version was fine-tuned using LoRA and then merged back into the base model?
Details
Files
qwen_image_vae.safetensors
Mirrors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwenImageEdit_qwenImageEditVAE.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
Qwen_Image-VAE.safetensors
Qwen_Image-VAE.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwenImage_qwenImageVAE.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
Qwen_Image-VAE.safetensors
qwen_image_vae.safetensors
qwenImageVAE_qwenImageVAE.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae (1).safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
Qwen Image VAE [BF16].safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
Qwen_Image-VAE.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
vae.safetensors
qwen_3_4b.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
vae_qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwenImage_qwenImageVAE.safetensors
qwen_image_vae.safetensors
vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
Krea_2_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwenImageVAE_qwenImageVAE.safetensors
qwenImageEdit_qwenImageEditVAE.safetensors
qwen_image_vae.safetensors
qwenImageEditFP8GGUF_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwenImageGGUF_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
eventHorizonAnime_ehanimacleanlineV10.safetensors
qwen_image_vae.safetensors
qwenImageVae_v10.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
jdxanima_illustration.safetensors
qwen_image_vae.safetensors
seeKrea2_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
flashpointHexical_v30Anima_txt.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
chosenMixAnima_v10.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
Qwen_Image-VAE.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
akanezora_v055BFP8INT8.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwenImageVAE_qwenImageVAE.safetensors
qwenImageVae_v10.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
Qwen_Image-VAE.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwenImageGGUF3QKM10G_vae.safetensors
rinFlanimeAnima_v14.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
eventHorizon_krea2V10.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
animaMayhem_v10.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
krea2Turbo18For_v1.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
toonimaSeries_toonimaThirst.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
toonimaSeries_toonimaZest.safetensors
qwen_image_vae.safetensors
miaomiaoAnimeReality_ani11.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
qwen_image_vae.safetensors
anima_baseV10_txt.safetensors
Mirrors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
text_encoder-bf16.safetensors
qwen306BBaseAnima_1.safetensors
qwen_3_06b_base.safetensors
model.safetensors
model.safetensors
model.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
model.safetensors
model.safetensors
model.safetensors
qwen_3_06b_base.safetensors
model.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
model.safetensors
qwen_3_06b_base.safetensors
qwen306BBaseAnima_1.safetensors
qwen_3_06b_base.safetensors
qwen_3_600m.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_600m.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
model.safetensors
miaomiaoHarem_anima11.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
reedAnimaXXX_v11_txt.safetensors
qwen_3_06b_base.safetensors
anima_baseV10_txt.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
anima_baseV10_txt.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
miaomiaoHarem_anima13_txt.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
miaomiaoHarem_anima13_txt.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
silvermoonmixAnima_v10.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
anima_baseV10_txt.safetensors
miaomiaoHarem_anima14_txt.safetensors
qwen_3_06b_base.safetensors
anima_baseV10_txt.safetensors
qwen_3_06b_base.safetensors
model.safetensors
model.safetensors
pearlySimplyAnimaStart_v10_txt.safetensors
qwen_3_06b_base.safetensors
qwen_3_06b_base.safetensors
omnitoonANIMA_v11_txt.safetensors
comicBookAnima_animaCbV15_txt.safetensors
qwen_3_06b_base.safetensors
estrogenima_v01_txt.safetensors
smoothmixUltimateAnima_anima_txt.safetensors
animapulseAnima_v09_txt.safetensors
animapulseAnima_v11_txt.safetensors
qwen306BBaseAnima_1.safetensors
jdxanima_semi_txt.safetensors
jdxanima_illustration.safetensors
anima_baseV10_txt.safetensors
unholyDesireLunar_v20_txt.safetensors
lijghtline_l1Harmony_txt.safetensors
zodamix_anima10_txt.safetensors
v01dANIANIME_baseV1_txt.safetensors
akanezoraMultiPrecision_v05A_txt.safetensors
vergardANI_v10_txt.safetensors
animaReverieXL_animaReverieXLV10_txt.safetensors
raehoshiAnima_v10_txt.safetensors
hsAnima_v10_txt.safetensors
PVCStyleModelMovable_anima10_txt.safetensors
cutiefuranima_V10_txt.safetensors
cutiefuranima_V11_txt.safetensors
animadronesV19_v1_txt.safetensors
funkyMixANIMA_v00_txt.safetensors
anima_baseV10_txt.safetensors
anima_baseV10_txt.safetensors
toonimaSeries_toonimaZest.safetensors
hakoANIMA_betaV20Base10_txt.safetensors
eventHorizonAnime_ehanimacleanlineV10.safetensors
riMixIllustriousAnima_riMixAnima_txt.safetensors
JANIMA_v10_txt.safetensors
blendermixAnima_v1_txt.safetensors
miaomiaoAnimeReality_ani11.safetensors
chosenMixAnima_v10.safetensors
auralisAnima_v10_txt.safetensors
nijce_1_txt.safetensors
miaomiaoHarem_anima10.safetensors
miaomiaoHarem_animaBase_txt.safetensors
oneObsessionAnima_v30_txt.safetensors
miaomiaoHarem_anima12_txt.safetensors
miaomiaoHarem_anima13_txt.safetensors
oneObsessionAnima_v20_txt.safetensors
miaomiaoHarem_anima14_txt.safetensors
miaomiaoHarem_anima8Step10_txt.safetensors
miaomiaoHarem_anima15_txt.safetensors
oneObsessionBranch_matureAnimaV1_txt.safetensors
miaomiaoHarem_aniAnimeColoring10_txt.safetensors
terraRisingUnity_v30Unity_txt.safetensors
plaijfull_in1_txt.safetensors
pearlySimplyAnimaStart_v10_txt.safetensors
theAWondermix_v2_txt.safetensors
loxsMixPixanimania_v10_txt.safetensors
waiANIMA_v10Base10_txt.safetensors
nyaIrisAnima_base1V20_txt.safetensors
lilithsDesireAnima_v10_txt.safetensors
animafranken_v10_txt.safetensors
cyberrealisticAnima_v10_txt.safetensors
cyberrealisticAnima_v20_txt.safetensors
cyberrealisticAnima_v30_txt.safetensors
cyberrealisticAnima_v40AestheticV10_txt.safetensors
cyberrealisticAnima_v42TurboV10_txt.safetensors
cyberrealisticAnima_v50AestheticV11_txt.safetensors
cyberrealisticAnima_v60AestheticV11.safetensors
aixiaoni_animav10_txt.safetensors
reedAnimaXXX_v10Base_txt.safetensors
reedAnimaXXX_v11_txt.safetensors
reedAnimaXXX_v12_txt.safetensors
hakoANIMA_betaV10Base10_txt.safetensors
blendermixAnima_v1_txt.safetensors
syntonij_i_txt.safetensors
miaomiaoBlocks_anima10_txt.safetensors
rinAnim8drawAnimaAnime_v10Exp_txt.safetensors
gemCollectionAnima_moonstone_txt.safetensors
reedness_v10_txt.safetensors
oneObsessionAnima_v10_txt.safetensors
silvermoonmixAnima_v22_txt.safetensors
akanezoraMultiPrecision_v065B_txt.safetensors
akanezora_v055BFP8INT8.safetensors
masterANIMAAnimaster_masterANIMAV2Prev3_txt.safetensors
oneObsession3D_v10Anima3d_txt.safetensors
vrchat2176_v10_txt.safetensors
darkbubble_animaV10_txt.safetensors
rinFlanimeAnima_v10_txt.safetensors
rinFlanimeAnima_v14.safetensors
illustrijMa_1_txt.safetensors
oilecule_animaV10_txt.safetensors
cutiefuranima_V20B_txt.safetensors
cutiefuranima_V20A_txt.safetensors
dessertModelsAnima_mochi_txt.safetensors
dasiwaAnima_luminousLabyrinthV1_txt.safetensors
lvAmerimanga_v10_txt.safetensors
copperlijghtDREAMS_1_txt.safetensors
animaMayhem_v00SnakeSkinBoots_txt.safetensors
animaMayhem_v10.safetensors
specustrious_anima_txt.safetensors
GoosemixANIMA_v11_txt.safetensors
anijpaint_v1_txt.safetensors
qwen306BBaseAnimo_0.safetensors
indigoFurryMixAnima_v10_txt.safetensors
miaomiao3DHarem_animaLH3D10_txt.safetensors
pieModelsAnima_cream_txt.safetensors
pieModelsAnima_cottage_txt.safetensors
filigreeAnima_v20_txt.safetensors
comicBookIllustrious_comicbookV15_txt.safetensors
comicBookIllustrious_comicbookV16_txt.safetensors
silvermoonmixAnima_v10.safetensors
silvermoonmixAnima_v21_txt.safetensors
silvermoonmixAnima_v23_txt.safetensors
anima_baseV10_txt.safetensors
akanezoraMultiPrecision_v055B_txt.safetensors
characterirlBy_v10_txt.safetensors
cyberfalafelFalafel_cyberfalafelAnimaV10.safetensors
miaomiaoRealskin_anima10_txt.safetensors
miaomiaoRealskin_anima11_txt.safetensors
toonimaSeries_toonimaThirst.safetensors
minijma_1_txt.safetensors
mijcrox_1st_txt.safetensors
unrealvisionAnimaNSFW_animaV10_txt.safetensors
toonimaSeries_toonimaSplash_txt.safetensors
nyaIrisAnima_base1V10_txt.safetensors
qwen_3_06b_base.safetensors
miaomiaoAnimeReality_ani10_txt.safetensors
miaomiaoHarem_anima11_txt.safetensors
hassakuAnima_v01_txt.safetensors
animij_s1_txt.safetensors
terraRisingUnity_20TerraRisingAnima_txt.safetensors
silvermoonmixAnima_v20_txt.safetensors
theAWondermix_v1_txt.safetensors
animij_ai_txt.safetensors
fantasiaAnima_v10_txt.safetensors
animafranken_v11_txt.safetensors
animafranken_v12_txt.safetensors
reedAnimaFurry_beta10_txt.safetensors
nectarineNymphsModel_animaV1_txt.safetensors
surreijl_1_txt.safetensors
blenderstyle_animaV10_txt.safetensors
loOmaij_thread005LeanIn_txt.safetensors
loOmaij_thread003Connect_txt.safetensors
loOmaij_thread002Closer_txt.safetensors
loOmaij_thread001Spark_txt.safetensors
masterANIMAAnimaster_animasterBase10V1_txt.safetensors
something_v10_txt.safetensors
jdxanima_anime_txt.safetensors
MiaomiaoHaremAni25D_v10_txt.safetensors
lvwesternmixwc_v10_txt.safetensors
animeBulldozer_anima_txt.safetensors
kirazuriAnima_v40AnimaBase1_txt.safetensors
kirazuriAnima_v30AnimaBase1_txt.safetensors
animapulseAnima_v10_txt.safetensors
lvwesternmixcartoon_v10_txt.safetensors
anima_baseV10.safetensors
Mirrors
anima-base-v1.0.safetensors
anima_baseV10.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima_baseV10.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima_baseV10.safetensors
anima-base-v1.0.safetensors
BaseV10.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima_baseV10.safetensors
anima_baseV10.safetensors
anima_baseV10.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima_base_bf16_v1.0_by_circlestone_labs.safetensors
anima_baseV10.safetensors
anima_baseV10.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
AN-anima_baseV10.safetensors
anima-base-v1.0.safetensors
anima.safetensors
anima-base-v1.0.safetensors
anima_baseV10.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima_baseV10.safetensors
anima_baseV10.safetensors
anima-base-v1.0.safetensors
anima-base-v1.0.safetensors
anima_baseV10.safetensors
anima-base-v1.0.safetensors
anima_baseV10.safetensors
Anima_Base_V10.safetensors
anima-base-v1.0.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.