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
Partially retrained to adjust style, making it smoother and reducing artifacts while still keeping good details. Based on user feedback, most people will probably prefer this version.
FAQ
Comments (92)
Does Aesthetic version get knowledge update compare to base?
Doesn't seem so.
I still have issues with sizing (genitalia), but overall it's understanding of concepts remains loosely the same.
Yet another slop checkpoint that's worse than the basic checkpoint and even the first aesthetic1.0. Stop messing around with this and just make an "edit" version of the base model instead.
Yes, this version has rendered all my LoRa functions inoperative. The effect of every Lora has become extremely poor, making the picture look just like that of SD1.5.
@1164912321 Creating slope-based """"AESTETHIC"""" checkpoints is a dead-end approach to training neural networks - a crutch that ultimately breaks Lora’s reproduction, and no one is going to train a point-specific Lora on a slope-based checkpoint every single time. Either we need a new base model with more detailed and high-quality training, or the circus continues. I’m not even getting into the fact that these Slope checkpoints are essentially no different from one another when you compare the images generated from them.
1.1很好，比1.0好，甚至比我非常认可的b还要好。时至今日，我们终于摆脱了那要命的油亮AI感以及缺乏细节造成的ai感，并在细节和精细度上得到了一个良好的平衡。
短短一周进行了这么多次更新，辛苦了。我看到了团队的热情以及对品质的追求，让我不得不称赞。
个人的使用建议：
使用1.1或者b。
使用dpmpp_2m_sde_gpu+simple
beta57会带来极强的艺术性涂抹，但同时会降低画质（一般是可以接受的）。
euler会抹平一些非常细腻的细节，我建议不要用。
建议使用Lora。
质量提示词也很重要，建议按照例图里的质量提示词填写，去掉那些乱七八糟的质量词。
Noob question: What's the difference between the 1.1 aesthetic version and the base model? Will aesthetic deliver more standout performance in terms of prompt adherence and art style?
Quotes from description:
"Anima-Aesthetic
Fine-tuned for better consistency and a higher quality default art style."
"Partially retrained to adjust style, making it smoother and reducing artifacts while still keeping good details. Based on user feedback, most people will probably prefer this version."
As the name already implies, it's purely aesthetic changes giving a more consistent default art style.
You may think about it as a properly finetuned model as creators use much bigger dataset for it compared to normal CivitAI finetunes. As a result it have less bias and bigger diversity compared to other finetunes, but still less than in Base model. This is an optimal tradeoff for this average quality, I think.
question:
1. if I want to train lora, better use base model, aesthetic or I can use the Turbo model? which one is better?
2. aesthetic are turbo model with 1 cfg, or it's on base?
LoRA will be more compatible with version you are trained it on. You can choose the most basic one (base 1.0) as it have more checkpoints based on it. For future anima progress it's better to use newest models. Actually, you can try to train a lora on multiple versions to make a universal lora. Not sure there are an easy way to do so, but an AI can write a code for a specific trainer to do so.
We should follow the creator's instructions.
Anima-Base
The pretrained, unrefined base model. Maximum flexibility, diversity, and style adherence.
LoRAs should be trained using this version.
If you'd like, I hope the next BASE version can improve the understanding of the direction the tongue sticks out. Currently, if I use (tongue out to the side:2), the direction the character sticks out their tongue still doesn't change.
Try ：p it’s a danbooru concept
@necrophagism777 Thanks for the reminder, but what I wanted was for the character's tongue to stick out to one side when the mouth is open, :P It can only be used when the mouth is closed.
Could this mod be further optimized? This version has rendered all my LoRa functions inoperative. The effect of every Lora has become extremely poor, making the picture look just like that of SD1.5.
You should learn how LoRA works. They would work better for model they were trained on. More training of the model = less compatibility with older LoRAs. Problem is not in the model. Outdated LoRAs is the problem. If it's the only problem you have - it should be ignored by developers. There are no reason to stay on the same place. They should move forward.
After testing the Aesthetic v1.1, I personally prefer the Base v1 for my workflow. With my current image generation settings, the improvement in aesthetics doesn't quite compensate for the slight reduction in LoRA fidelity.
1. Character appearance: The Aesthetic version tends to depict characters with a slightly more mature appearance.
2. Aesthetic quality: It appears to render visual effects and light particles somewhat more prominently. However, compared to the Base version, the overall improvement does not seem particularly significant.
3. LoRA fidelity: Compared to the Base version, both character appearance and style seem to be reproduced with slightly lower fidelity when using LoRAs.
LoRA point have almost no sense. That's obvious LoRAs trained on older versions would have less compatibility with newer model. That doesn't mean model became worse - just model changed inside.
@Shio_N My comments regarding LoRA fidelity are subjective and based on my personal experience. User preferences will likely vary depending on their prompts, workflow, artistic style, and the specific LoRAs they use. Some users may find the Aesthetic version more appealing, while others may prefer the Base version.
In my workflow, however, the Base version produced more satisfactory results. With the Aesthetic version, I observed a few unexpected behaviors when applying my LoRA.
I am having the same experience. I'll be recalibrating my workflow and seeing if 1.1 is an upgrade or a sidegrade. My 1.0 setup is perfectly fine at the moment, but I can't refuse an opportunity to squeeze in more optimizations.
@Milanor It seems that the Aesthetic model was fine-tuned not only to enhance aesthetic quality but also to introduce subtle changes to the default art style. Personally, I was very satisfied with the creator's previous Highres/Aesthetic Boost LoRA because it had little to no impact on the model's original art style.
Judging from the community's reactions, opinions seem to be mixed. Some users prefer the new model, while others feel that the original model produces better results.
As for me, I'll probably either retrain my LoRAs using Aesthetic v1.1 as the base model or wait for a future fine-tuned model that better suits my workflow.
When I tested aesthetic v1.1 using a LoRA tailored to my style, I got results that were clearly better than v1.0 for me.
Specifically:
- Compatibility with LoRA is stable. I didn’t notice any discrepancies with v1.0 or any drop in image quality.
- While maintaining the improved image quality, the line art is cleaner than in aesthetic v1.0.
- The style alignment with my LoRA is even better, and concepts that were ambiguous in v1.0 are clearly expressed in v1.1.
Results may vary from person to person, but I think it’s worth trying v1.1 for users who are training LoRA on Base.
It's anima_based-v1.0 vs aeslop-v1.1
how does the diversity of the turbo version compare to the turbo lora?
Idk why, but Anima image generation on ComfyUI looks worse on Ubuntu/Kubuntu than Fedora Workstation and Windows 11.
Maybe because I had to use Pipx and couldn't do that python version 3.11 virtual environment command. Maybe the python version really affects it
Is Turbo using Base or Aesthetic?
I pretty like this Aesthetic v1.1 model, better than Aesthetic v1.0 in my opinion.
aesthetic-v1.1 looks more "ai slop" art style in my opinion, aesthetic-v1.0 seems more balanced
i love this model
Love this model but agree with others that aesthetic 1.1 is a bit of a disappointment. It seems to kill individual styles for a more generic slopified look.
peak
SOVL
I see great potential, fresh database is always good comparing to illustious and ponyxl. Quality of Anima 1.0 base models seems a bit better then illustiousXL. I feel mostly positive and hopefull.
But scene composition vise... Well its hard to get truly dynamic poses or angles, not impossible but harder then it should be.
Angle related tags do nothing most of the time or barely anything. Just overall a lot of stiffness for charecter poses. In PonyXL or Illustrtious motion was as easy just adding "dynamic pose/ dynamic angle", braindead and effective.
In anima for some reason i always have to type in tens of different pose related tags to each leg, hand, head + angles to get some sort of variety and motion.
NSFW related tags are also gutted for some reason, therefore half of the lora slots goes to character sliders. Tags like "Puffy nipples/ areolae size ect." do nothing no matter the weight. I cant believe that nobody surpassed PonyXL in this regard.
I wish these minor things to be impoved.
After some extensive testing with 1.1, I wasn't able to steer the model towards my target aesthetics (hehe). Unlike 1.1, which foundationally starts "cute" but is easily malleable, there's a mature look in 1.1 that's ingrained harder and needs some strong-arming to work with. This affects faces, poses, shadows, and body shapes. Aesthetics 1.1 also seemingly provided no (subjective) upgrades compared to 1.0. Rather than an improvement, it's more a sidegrade/slightly different style.
Between the two, I would lean toward Aesthetics 1.0 for flexibility, and Aesthetics 1.1 for a good starting style that better caters to many preferences.
They are both great models, and 1.0 has been absolutely phenomenal for my uses. Thank you for your hard work on these.
Alright, I’ve re-downloaded 1.0 after having deleted it... I’m going to generate some images for a comparison now. The 1.1 are great, but the success rate for pulling cards has dropped.
Are there any benefits to finetuning my style on aesthetic 1.1?
看起来是比较糊的画风...
state of the art when it comes to anime, holy moly the amount of prompt adherence is crazy, you just have to mess with it
After testing the aesthetic-v1.1, v1.0, and v1.0b models for a while, I find their aesthetic quality truly stunning; with the right prompts, they can achieve a highly sophisticated look that is perfect for illustrations. However, their narrative capability remains limited, and there are still shortcomings regarding prompt adherence and realization. They are excellent if your goal is simply to create beautiful images without strict requirements for specific details, but achieving a compelling narrative requires a significant investment of time and effort, often involving compromises on specific details. Regarding prompt adherence, the best "Anima-based" model I have used so far is waiANIMA_v10Base10, followed by animaika_v44. Additionally, the aesthetic series exhibits some subtle compatibility issues with LoRAs; using LoRAs designed to enhance aesthetics, lighting, or dynamic poses can slightly disrupt the composition, leading to a partial loss of control over the image. Overall, the aesthetic series serves as a fantastic base model—kudos to the creator!
A small suggestion: regarding composition, while the general placement of characters or props can currently be controlled by combining directional terms (left/right/top/bottom) with depth layers (foreground/midground/background), controlling the scale ratio between characters and the background remains difficult. The more detailed the prompt's description of a character, the larger that character appears; conversely, detailed descriptions of the scene cause the character to shrink and the scene to expand. Unlike models like Illustrious, which have relatively fixed perspectives, Anima’s compositional structure is highly variable. Consequently, to fine-tune the composition, users are sometimes forced to cut descriptive prompts or deliberately add "filler" text just to adjust the weightings. Could a "scale" concept be incorporated during development? For instance, if the total canvas space is defined as 10, users could specify values in the prompt—such as "Character A scale: 7, Character B scale: 4, Character C scale: 15 (off-screen)"—to achieve greater control over the composition.
i am curious, will the model get new data once it had big update? like if there are 2.0 version, there will be more data set trained or what?
I wonder why the model likes to change the word "brought" into "brught". I know the model is bad at text, but it is just weird that it is almost consistent into changing that specific word.
based anima is not at all with text, in fact one of the best local models as I heard
Really like Aesthetic 1.1, the default style is much better, and works well with the lora/artist styles I have tested.
I cant use at all . I download with matrix and nothing happen
Yoy need to use Comfy workflow instead.
You can also use Anima with InvokeAI or Forge Neo
CIRCLE STONE LABS AND COMFY ORG! DO AN OFFICIAL NSFW ANIMA MODEL AND MY LIFE IS YOURS!
It's probably one of the most nsfw model already by default.
@necrophagism777 i'd still say that illustrious and noobai are better with nsfw as of right now with the huge selection of loras they have
LMFAO.. it's already NSFW :DDD. How much more NSFW do u need ?? :D
@evilstormhot109 okay yeah youre right, i am just salty that anima only uses danbooru/gelbooru tags and not other booru-themed sites. the nsfw on it is actually amazing
@F4K3xN1GHTM4R3
Text encoder understands natural language better than tags, just saying modern military armor gives you some fantasy slop, instead use something like plate carrier, ammo bags, camouflage clothing, knee pads, and combat boots and you get something close to actual modern military oufits.
Step 1: follow these links.
Easy NSFW Anime = https://civitai.red/models/2668799/cyberrealistic-anima?modelVersionId=2996716 - use his linked WF for Comfy
Easy NSFW Semi-Real Anime = https://civitai.red/models/2668799/cyberrealistic-anima?modelVersionId=3068504 - use the same WF for Comfy
Step 2: Both go in the Diffusion models folder, take ~10 min to get setup and get the hang of.
Step 3: Don't look back.
how do you deal with artist names and artist logos on the image while using the turbo version? negatives dont do anything with cfg 1
There are Comfyui nodes for CFG-less negatives, NegPip or NAG
@Dioxide thanks a lot, that helped :D
you can also use "Big lama Remover" custom node in combination with watermark detection to remove watermark/logos/text.
can i use it in forge NEO?
Yes, I'm using it now - just grab the 3 items... anima turbo model, text encoder, and VAE.
@civitaithrowaway28398 thanks so i have to download turbo not aesthetic, and can you tell me which text encoder and vae do i need to download
am I stupid or does this not work with WebUi? ive inputted everything correctly but whenever I try to load the VAE I get an error
only works with forge neo, but i never tried it myself, i use comfy
you mean stable-diffusion-webui? That thing is ancient, hasn't been updated in years. Switch to something else, there's tons of new projects. I personally use SwarmUI.
working great in forge neo
@formanyreasonzzz619 oh shit I never knew about forge neo, im gonna try it out ty
@k1logr4m yeah lol, i never sought out anything else because this has been working out for me, ill check out other ones ty
hello where can I find vae for anima?
💜
i wanna see this model get updated with new characters once in a while
What I want is an Anima XL or something, because while for it's size it punches WELL above it's weight class, it does show it's limitations, not so much in the characters, but backgrounds and environments, I would like to see a 6B version of Anima with a 4B LLM. It would address like 98% of the market of people with enough compute to run DiT model as with running the model in K_M_Q5 which retains like 95% of the quality of the model you can run that on a 6GB card... I honesty am baffled at why they made it so small in the first place, and using a 0.9B LLM... WHY THOUGH!!! 4B would have been a massive improvement and 4B has been demonstrated to be enough to be getting along with with Z image and Krea2, yeh there are still some limitations, but you can mostly push around them, but an 0.9B LLM... Why, you can load the LLM get the conditionings and then unload the LLM as long as your prompt doesn't change you don't need to load the LLM again, and loading a 4B model doesn't take THAT long.
And again you run a FP8 mixed 4B LLM you can run that on a 6GB card without needing to offload to system memory, which means the 1060 is still usable, so why make it so small that it compromises what the model can do, both the DiT and the LLM.
The only down side is it will take significantly longer to train lora's, like that is a genuine down side, but it's the only downside really, yeh generations with a turbo model would take like like 20 - 40% longer, although likely the turbo lora with a 6B model could actually get great results in 8-12 steps, rather than Anima as it stands which seems to really take 12 - 16 steps with the turbo lora to get mostly equivalent results to a base render, so... really the time difference wouldn't even be significant.
Like don't get me wrong Anima is still is most areas a massive improvement over SDXL based models, but I'm just disappointed knowing it could have been so much better and just blown anything SDXL out of the the water in every respect.
@Etheoma I do agree with a lot of what you said, but it is worth noting that there was a reason given by tdrussell as to why Qwen 0.6B was used (near the end of the discussion): https://huggingface.co/circlestone-labs/Anima/discussions/67
Basically he tried using Qwen3.5 2B instead and there was little improvement, and with longer training times, he decided not to use it. Whether it's because of the LLM adapter I don't know (some other discussions seemed to imply this but there was never a direct answer), but that is at least the reason given for the small text encoder.
Greetings, since the release of base 1.0, I have trained many style-based LoRa classes, but based on my experience and what I've learned from the community, it seems that style reproduction is generally only around 80%. What causes this? Could you explain?
Hello, I am a translator whose English is not easy to use. If you are comfyui, you can try this node called Conditioning Zero Out, which has the function of resetting negative prompt words to zero. I trained Lora myself and used this node. I found that it has a great improvement on my training style. It may be related to my training method
@Bocchisaki Thank you for your help. Could you please provide a sample workflow?
And could you briefly describe the improvements you've observed?
We have many anime checkpoints. What makes this worth yet another few GB of space on my hard drive?
Anima is better than Illustrious and Noob AI. Its a replacement and not "just another". The next step so to speak. Even now its outperforming those two quite easily with how consistent it is and how easy it is to train and get solid picture with.
It doesn't suffer from pony/illustrious VAE compression, so it can do full body and multiple characters.
Anima has a better overall experience, supporting natural language and multiplayer images, and the ecological development of this model is very fast
It's not better on knowledge than NAI/CHEN (some styles and franchises doesn't work) but is better on consistency and final quality, you can do multiple characters on NAI/CHEN too but is a little harder to prompt, can have "bleeding" and the quality can be bad without hiresfix, Anima doesn't suffer this and natural language works, another good thing is that full-body shots looks a lot better.
Why do my generated images keep turning into split scenes? For example, the same character appears on both the left and right sides performing different actions, or the image is generated as a multi-panel comic instead of a single scene.
aesthetic v1.1 is much better than other new versions.
I recommend Aesthetic v1.1 or Base v1
Quick question please when tagging for dataset training should i put backslash to the brackets like when generating or leave without, for example: shrug (clothing) or shrug \(clothing\)
No, the backslashes are only for inference as parentheses get interpreted as emphasis instead. You don't need to escape parentheses in training
You've done a great job! I'm excited for the next version. Thank you!
Details
Files
anima_aestheticV11.safetensors
Mirrors
anima-aesthetic-v1.1.safetensors
anima_aestheticV11.safetensors
anima-aesthetic-v1.1.safetensors
anima-aesthetic-v1.1.safetensors
anima-aesthetic-v1.1.safetensors
anima-aesthetic-v1.1.safetensors
anima-aesthetic-v1.1.safetensors
anima_aestheticV11.safetensors
anima-aesthetic-v1.1.safetensors
anima_aestheticV11.safetensors