I. Introduction
AnimaYume is a text-to-image model fine-tuned from Anima, a high-quality anime-style image generation model developed by CircleStone Labs. It builds upon Cosmos 2, a model developed by NVIDIA’s research team.
II. Information
For version 0.1:
This model is a preview version fine-tuned from the Anima base model using a custom dataset. Training was conducted across multiple resolutions ranging from 768 to 1280 pixels, with a primary focus around 1024. The goal of this release is to improve stability and minimize unwanted artifacts when producing high-resolution images.
Notes: All the example images at this version were generated at the resolution 1024x1536 or 1536x1024
For version 0.2:
This model is a continuation of AnimeYume v0.1. In this version, I improved the quality of my dataset and used several techniques to prevent oversaturation and low-quality outputs. Based on my testing phase, I observed that the prompt coherence is better than v0.1, and the model remains very stable when generating images at a resolution of 1536.
Note: I am still waiting for the final version of Anima and testing some methods to make my training process faster. I know the license might make the model less popular, but I only care about whether the model is good or not. I’m aware that many others use better licenses, but I’m too lazy to spend a bunch of money training a model from scratch.
For version 0.25:
This version was trained on Anima Preview 2. Due to several issues with the base model, such as overfitting, black/white borders, quality inconsistencies, and problems with artist tags, I decided to focus primarily on improving the model’s knowledge, reducing these issues, and making it as stable as possible.
Note: In this version, I did not attempt to improve the model’s style. I tried doing so, but it caused the model to forget some of its existing knowledge. The training process is similar to v0.2, but the dataset has been adjusted to better address the issues present in Anima Preview 2.
For version 0.3:
This version was trained using Anima Preview 2. It is an experiment with a new training method for the model. You can consider it as another branch of AnimeYume 0.25, developed in parallel. However, this version uses new techniques and a larger dataset compared to v0.25.
Note: In this version, I experimented with a new training approach, so the model is slightly different from v0.25. Additionally, all example images were generated using prompts shared with users on CivitAI to evaluate whether this new method.
For version 0.4:
This version was trained on Anima Preview 3 using a custom dataset. In this release, I improved prompt understanding and artist style. Based on my testing, some artist styles match my expectations, although I haven’t tested everything in detail since I’m currently quite busy :<. Additionally, I fixed several issues from Anima Preview 3 that also appeared in Preview 2.
Note: I’ve only tested with simple test cases, not comprehensively, so if you encounter any issues, feel free to let me know. I also used a larger AI computing cluster to speed up the training process :D.
All example images were generated using prompts shared by users on CivitAI, as I wanted to evaluate the model’s performance.
For version 0.5:
This version was trained on Anima Base v1.0 using my custom dataset (a mix of a small e621 dataset and Danbooru). In this release, I added many new characters and improved the existing ones. I also enhanced support for various artist styles, allowing the model to generate results that are much closer to the original styles. In addition, the model now understands some concepts and knowledge from e621, although the support is still limited.
Notes: I’ve only tested the model with a few simple test cases so far, so if you encounter any issues, feel free to let me know. This release can be considered a demo version showcasing my new training method, which focuses on preserving existing knowledge while adding new knowledge at the same time. The release also came sooner because I was finally able to use all the resources I had available :D
All example images were generated using prompts shared by users on CivitAI, as I wanted to evaluate the model’s performance using real user prompts.
III. File Information
This file contains only the diffusion model and does not include a VAE or text encoder. To use it properly, you will need to download those components from the link here
IV. Notes & Feedback
This is an experimental fine-tuned release, and I am waiting for the final version release to tune it :D
Your feedback, suggestions, and creative prompt ideas are always welcome, every contribution helps make this model even better!
V. Acknowledgments
Big thanks to narugo1992 for the dataset contributions.
Credit to Circlestone Labs and Nvidia for the fantastic base model architecture.
If you'd like to support my work, you can do so through Ko-fi!
Description
FAQ
Comments (36)
Well, after testing Anima Preview 2 for a long time, I’ve noticed several issues with this model. It seems quite biased, and in some cases the images are generated with black or white borders. Another problem I often encounter is that the quality scoring doesn’t work as expected. The artist styles also don’t seem to perform well.
Overall, Preview 2 doesn’t feel like a good checkpoint (at least from my perspective). While it performs better in some natural language cases, the trade-offs compared to Preview 1 are too significant.
I will release a version of Animayume for Anima Preview 2 soon, but it won’t be a major update. Due to the many issues, I’ve decided to stop tuning styles and just improve knowledge and prompt coherence.
Of course, this is just my personal opinion, I might have made mistakes during tuning or done something wrong. if you have any idea, feel free to share it to me.
Here is the comparation between AnimaYume v0.25 and Anima Preview 2: https://civitai.com/posts/27324346
i tried anima offcial preview 2 its often making a border or multiple panels for some reason
Maybe it's me, but based on all the comparisons you posted, Animayume looks sloppier than Anime preview 2, and the artist styles seem weaker.
@bionagato FYI, from these images i am not using any artist style, i just prompt normally like tag and natural language. Moreover, i just want to know how my model look like and which field it enhance :v
@duongve13112002, artist tags is one of the core features of the Anima on par with prompt understanding. It'll be good if you added a few examples how the model handles them, because judging by previews currently the model looks like some Illustious mix
@duongve13112002 This is my humble opinion (for something you're giving us for free), because you asked for some opinions:
Look how the floor disappears and becomes solid white in Animayume:
https://image-b2.civitai.com/file/civitai-media-cache/21bac3e4-2386-4c25-83a5-2ad52f55f35e/original
In this image, Animayume's hand is worse and the lighting in the background is strange compared to Anima Preview 2 (but the eyes are much better in Animayume):
https://image-b2.civitai.com/file/civitai-media-cache/790ce66b-8332-4e43-9150-878421c2d168/original
In this image, the composition of Anima Preview 2 seems better, and the girl's posture and position look more natural:
https://image-b2.civitai.com/file/civitai-media-cache/b572f586-29e0-4216-9818-249176fc2bb2/original
Here, the background and lighting are more natural in Preview 2 (but the eyes and details are better in Animayume):
https://image-b2.civitai.com/file/civitai-media-cache/9959d091-edd2-4070-b9a5-6cc71a4f331b/original
In my opinion, Anima Preview 2 is much better and more stable than Preview 1 (especially with multiple characters), and (based on your own images) it is a bit better than Animayume in composition and lighting. Animayume has better eyes in almost all images and stronger color but for my personal taste, I prefer more the colors in Anima Preview 2, as they seem less ai-like.
Something I dislike (and that basically screams that the image is AI generated) is the glowing inner hair in almost all girls even with low illumination in animayume, see these images:
https://image-b2.civitai.com/file/civitai-media-cache/59188095-0706-4b46-afbe-ee9aed865c3c/original
https://image-b2.civitai.com/file/civitai-media-cache/9959d091-edd2-4070-b9a5-6cc71a4f331b/original
https://image-b2.civitai.com/file/civitai-media-cache/790ce66b-8332-4e43-9150-878421c2d168/original
@bionagato oh thank for your comment i will check later about the color and the another currently this problem releated on my preference :v. Moreover, i have another checkpoint but the color is dumb so i am not release in here :v. May be a small tunning style can cause this problem. If you want a version without tunning style i will release it on huggingface :v
Feedback, noticeable biases in v0.25:
1. Brighter. And the lighting is weird. (thus feels sloppier, i guess)
2. Cute/loli faces
(1) can be solved by RDBT lora. (2) seems to be hard to overwrite by just prompts.
@ikekph5 Oh may be this problem releated to tunning style. So in the next version i will i will remove it about the loli faces i am not sure the core problem because my dataset contains various age range.
Um I love Anima Preview 2 more these looks like AI from 2025
Day 1 testing of v0.25 with my Preview2 trained character Lora. My first impressions:
The checkpoint is (once again) very style neutral which for me is a major plus. It means you can use it with Loras and it will respect their style.
I also noticed subtle quality bump with better represented details and fixes to anatomy like bad hands/fingers/feet. Character eyes also look better in wider shots.
When comparing side-by-side with Preview2 I would say 7 out of 10 times I prefer the output from v0.25.
Overall very solid first version for Preview2.
It wasn't trained on preview2.
I love AnimaYumeV0.25, it's a really great model!
It's disappointing that the .25 version leans more towards creating loli girl images than mature female. I've tried many anime models, many diffirent tag in both POS and NEG Prompt but none of them have been able to create images of mature women.
Are you tried using ((aged up))?
Can you maybe include a tad bit of e621 data for concepts? A lot of e621 is low quality in my personal opinion, but there is for sure a lot of it that is high quality and could make it learn more concepts maybe just do some level of filtering
Edit: People will get mad at me of course, you realize that a lot of this applied to human characters too?
Literally every preview image looks better on the original.
the preview image of v0.25 may not good compare to base model but it work great with lora than base model.
As always great checkpoint. Help's preview 2 s bit. I kind of wish that they went with a bigger texting encoder model as Qwen3 0.6b is decent But does struggle with more complex prompts, especially if not using stability or distilled LORAs or have high steps and CFG.
I was wondering if your checkpoint is compatible with low step LORAs? I have found they are bit more stable with base models like Flux Klein 9b and Z image turbo are both distilled versions of their base variant and the quality and prompt adherence is much better (However they use Qwen3 4b and Qwen 3 8b as the text encoder). So I definitely think distilled is great. Unfortunately you don't get as much variety though.
Usually low step means dmd2. I have not seen any dmd2 so far. Training on latent directly is not low step distillation.
I was going to do dmd2, but I think it costs too much. Time + poor LoRA compatibility + many artist styles will be nuked.
In my case low and disteller loras, downgrade images with my lora charcters((
Compared to the official anima-preview2, version 0.25 offers more stable quality and better overall results. It captures the artist's expressive style perfectly—it’s a fantastic model. If Anima continues to receive updates, I was wondering if you could release more versions like 0.25 in the future: versions that don't tweak the art style, but focus on fixing model issues and improving stability?💕
Illustrious is still better in quality and performance, even Lumina contains more and newer information than Anima.
The whole point of Anima is that it uses a Qwen LLM (a small one) for prompt comprehension. It adheres to intricate prompts much better than the models you mentioned. It's sort of like a miniature anime version of a Black Forest Labs model.
takes big fat rip off this doink
bait used to be believable..
This is just a "preview", but the model is already close to IL quality. It also understands natural language and tags well at the same time. Additionally, IL requires using a ton of different Lora to implement a concept.
@_Jarvis_ Yes, I know this isn't the full version and it will be better in the future, but what if someone ditched their customized Illustris and switch to this version and get worse quality and performance? Just put up with it? When the proper version comes out, I'll say everything's super cool.
Preview 2 but better, very nice
All in all, I’m tired of Illustrious model because it’s so outdated. I want new checkpoints that work like the ones in NovelAI, they have a lot of updated features, even though you really need to understand how they work. I also dislike their user interface.
So my question is: does Anima, Lumina, or any other model come close to at least a NovelAI checkpoint?
I haven’t tried any other checkpoints yet, so I can’t tell what their differences are. I hope some kind people here can briefly explain them. I’m tired of downloading checkpoint after checkpoint just to see if they work the way I imagine.
you can personally DM me if u feel the need some privacy.
EDIT:
i have installed comfyui for the first time of my life, and use the checkpoint, so far it's been great.
now i want to know the prompting style, should it work with danbooru and "flux" style prompting? sorry for noob question, i just need baby steps.
I haven't used NovelAI so I can't draw comparisons there. For Anima prompts, you should start by checking the Anima huggingface page which has reasonably detailed guidelines. Anima allows you to mix danbooru tags with natural language. Aesthetically, it's easier to produce a "nice" image in a specific style using Illustrious/NoobAI finetunes with LoRAs thanks to long term community support, but it seems like you'd prefer working with Anima + natural language prompts.
A reasonable starting point is "masterpiece, best quality" for your positive prompt and "worst quality, low quality, score_1, score_2, score_3" for your negative prompt. When using Anima, I like to describe the parts with danbooru tags then connect them together with natural language afterwards. I usually follow the flow of subject -> composition -> details (same with the SDXL-based models here).
Finetunes are going to be highly subjective and people are still figuring things out. For Anima Preview 2, I'm currently leaning towards this one or AnimaIka. For Anima Preview 1, I've had the best results stacking rdbt v0.12 onto AnimaYume v0.20 or base Anima. You can't really get around experimenting since everyone is looking for different things out of models.
bro how is illustrious outdated, its still currently today the best anime model in the world. until a better model than it comes out then it will be outdated, how is it possible to be outdated when nothing better exists yet
AnimaIka is rdbt v0.12 + animayumi v2.
The author of rdbt was mad about this base model in their post. Because the creator refused to give any credit to the original model authors.
Please support real trainers who spent lots of GPU hours and $ fine-tuning the model, instead of mergers who don't even give original models credits.
Since you speak about illutrious you must know how to tag, Just use natural prompting and you can add quality tags and stuff like illutrious if u want, check anima lora and the prompt they use for their image and there is a website with artist tag and stuff https://thetacursed.github.io/Anima-Style-Explorer/ might help you if you are looking for a style, I agree that Anima is definitly good and it's the first model that made actually use illutrious a bit less. NovelAI is not better than illutrious in my opinion not yet not with raw prompting at least, but novelai strenght was never their model but their tools and stuff like that, good news is that with a bit of learning and research you can reproduce a lot of feature of novel ai like style reference etc, thought i think novel ai still better in that regard. And for reference if u wonder about people reaction Anima have very limiting right of course for free and hobby user it's nothing but for creator that invest time and money and effort Anima rules are a bit annoying and it's pushing lot of the community away from it x) just general information if u are curious about people reaction.Good luck x)
Anima is the closest to NovelAi. I encourage you to use it as a fellow nai v4.5 user.
@Frankainstein Nai 4.5 absolutely demolishes Illustrious in raw prompting...it uses a T5 text encoder hence why anima went the TAG and NLP route. The dataset they put together and training they used is absolute SOTA shit. Illustrious has no text capabilities but lemme guess, you're obsessed with visual aesthetics? NAI V4.5 is still the most advanced anime model by far. Noobai is the most advanced open-source anime we have, v-pred is deadly when you know what you're doing. Illustrious is probably the "prettiest" but also has worse prompt adherence than noobai with no E621 dataset, that's it.
As for NovelAIs UI, does it not just have a API?