■This is an experimental ground for Anima LoRAs.
The LoRAs I create serve more of a role in stylistic tuning rather than reinforcing single concepts.There is no inherent superiority among the LoRAs; they simply serve different purposes. You can find the description for each LoRA in its respective tab.
■If you notice pixel patterns in your generated images, try the VAE below. It removes the pixelation typical of the qwen_vae and makes the images much sharper.
https://civarchive.com/models/2718533/hdr-vae-anima-qwen-image
■If the VAE looks a bit too saturated for your taste, try using this VAE merge node to blend it with the original qwen_vae to dial it in. A 0.5 merge should be just right.
https://civarchive.com/models/2487550/vae-merger-for-comfyui-anima
■I would like to share the possibilities of anima with everyone.
My wish is for many people to discover basemodels with potential and to see their possibilities unfold even further. I would be happy if I can help make that happen.
■The anima architecture has been heavily modified and retrained from scratch—almost like a radical overhaul—but the base architecture is Cosmos-Predict2-2B. To put it simply, you could compare it to a lightweight version of Flux 1.
I think it’s a great architecture capable of generating high-quality images despite its small size.
I have a very good impression of Anima when it comes to color vibrancy, lighting, and prompt adherence.
Its knowledge regarding NSFW concepts, characters, and styles is at least on par with SDXL. The overall feel of using it is also quite similar to SDXL.
Furthermore, it is lightweight and easy to train. I believe large-scale training is also a viable option.
Since we are provided with a base model that already contains a vast amount of knowledge, we rarely need to teach the model completely unknown concepts from scratch.
Basically, our work mostly comes down to adjusting styles or reinforcing minor concepts and characters. You can achieve great results without having to do any heavy lifting.
The developer has also ensured transparency by sharing the training settings, which means you are less likely to stumble over black-box issues.
Because of this, the burden of training is low, making it an option well worth nurturing together as a community.
■The LoRA has a stronger influence when used with shorter prompts.
Since longer prompts often carry their own inherent style, the LoRA will primarily take on a supporting role to refine and unify the overall image.
■https://github.com/gazingstars123/Anima-Standalone-Trainer
This is a great tool that allows for easy training, even on Windows. It has everything you need for LoRA training.
Even if you run into any issues, you should be able to get your LoRA training up and running by troubleshooting with ChatGPT or Gemini.
If you find the tool useful, you might want to consider making a donation to him. It allows the developer to focus more on development and deliver even better tools, which ultimately benefits you in the long run.
■If you have any questions, please feel free to ask!
日本語での質問も大丈夫ですのでご気軽にお声がけください～
If you enjoy my work and feel like making a donation, I would be incredibly grateful. Even the smallest token of support goes a long way in helping me lighten the load.
Whether I receive donations or not, I will keep doing what I love and pursuing my interests just as I always have. However, simply knowing that this kind of support exists brings me immense joy and serves as a huge driving force for me.
Description
This is a style LoRA for Sunrise's masterpiece OVA, "Mobile Suit Gundam Thunderbolt."
It has been trained on 1,750 carefully selected and curated images.
■While it's difficult to put this LoRA's unique style into words, it can be described as:
High-quality animation and intricate detailing.
Top-tier professional compositing that elevates the artwork through post-effects and digital coloring, while maintaining a muted, mature color palette and a heavy, gritty atmosphere.
With this LoRA, you can truly experience the culmination of Sunrise's technical mastery.
Recommended Settings
LoRA Weight: A weight of 1.0 works perfectly fine, but dropping it to 0.8 or lower will likely yield more stable and consistent results.
■About the Files & Backups
To stabilize the style, this main LoRA is a merge of several different training steps.
I am also sharing the unmerged backup LoRAs that were used as ingredients for this merge. Included in these files is a version trained all the way up to 4250 steps, so I highly recommend experimenting with them to find the exact strength and look you prefer.
■Thoughts on the Anime
This anime is a luxurious masterpiece that showcases the absolute peak of Sunrise’s craftsmanship. It carries the sophisticated, mature atmosphere of a well-aged wine.
The sense of style is outstanding, packed with deep, niche, and geeky elements. For those who resonate with this aesthetic, it is an absolute masterpiece. If you love a hard-boiled atmosphere, the heavy, gritty anime style of the 1990s, and incredible music, you will likely fall in love with this series.
While it features some shocking scenes and might not be for everyone, anyone drawn to its style and mood will find it incredibly worthwhile. Even if the overall vibe doesn't completely click with you, the master-class, veteran professional animation alone makes it worth the watch.
FAQ
Comments (14)
looks great!!
Thank you! It’s great to know someone else appreciates this style!
@hjhf yeah this one and vampire hunter D are awesome!
Here's some styles on top of my head maybe for inspiration?
Kaleid Liner Prisma Illya
Bible Black
Yusuke Murata Manga(There's some great lora for it on site but only for Ill)
@besporyadkin966
Thanks for the info!
Playing eroge and visual novels is actually a big hobby of mine, so those sound like highly interesting styles to look into. The classic "エロゲ塗り,eroge-nuri" (eroge coloring style) from the 2000s is my absolute favorite art style of all time.
Prisma Illya has such a cute art style. I’ve never watched or played any of the Fate series, but I’ve always been curious, so I’ll definitely check it out. As for Bible Black, I wasn't familiar with it, but the quality looks incredible and it has a really cool vibe. I'll make sure to look into that one too.
Yusuke Murata's artistic skill is on a whole different level. His style would definitely be fascinating to look into.
This gave me some amazing inspiration. Thank you!
@hjhf ooh really?? I adore Prisma Illya's style, I will hold my breath if you'll decide to train that one!
@hjhf also highly highly recommend original Fate/stay Night novel! It's now on steam too finally officially translated but 18+content is cut(It wasnt very good although)
@besporyadkin966
I’ve actually been wanting to play the original Fate/stay night for a long time too!
I did play Tsukihime, but that was ages ago. I've always known that the Fate spin-offs are super popular, and I've been really interested. But honestly, since I haven't played the original Fate/stay night yet, I kept telling myself I should play the original first before exploring the rest—and now years have passed, lol.
Your suggestion might be the perfect push I needed.
Prisma Illya looks really cute and seems like a great candidate for a LoRA, so I'll definitely check if I can make one! Right now, I'm fully booked with building datasets for various 90s titles, so the priority might be a bit lower for a while. But thanks to our conversation, it's officially on my to-do list, so please hang tight and wait for it!
@hjhf Im very happy! And will wait patiently!
I'm kinda confused. So are these enhance loras or a style loras? I use the 2D_aesthetic_base and 2D_aesthetic_base_EX but it gave me different artstyles every time i generate
Thanks for the message! Also, I’m sorry if my replies are a bit long!
■To answer your question directly, they act more like enhance LoRAs.
Here is a detailed breakdown of how they work and the differences between them.
■Overall Tendency and Training Data
They are trained exclusively on 2D style images, but the dataset includes a wide variety of artists. Because of this diversity, the generated art style will naturally change depending on the seed and the tags you use.
Basically, I excluded any 2.5D or semi-realistic 3D styles. Even if some were included, I only picked ones that were either created by Japanese artists or clearly respected that specific aesthetic.
I’m aiming for a clean look overall, so I’ve kept visual novel styles out of this one.
Since my goal was not to copy specific artists, I did not use artist tags. The captions only consist of character names and general tags.
I built this LoRA because I wanted to achieve high-quality, aesthetic outputs without needing to micromanage the style. I wanted the tags themselves to naturally guide the generation toward an ideal look.
■How Tags Affect the Style
Keeping the above in mind, if you don't use specific style prompts, the general tags you enter will dictate the resulting style.
* For example, if you use gothic-related tags, the model will likely pull from the styles of artists in the dataset who excel at gothic themes.
* The style might also shift based on body features (like breast size) or whether the prompt is NSFW/nude, as it adapts to the vibe of those specific tags.
* Plus, because of the dropout I used during training, you might notice some randomness in the art style if your prompt is short and doesn't specify a style.
* If you don't like this fluctuation, I highly recommend specifying your desired style directly in your prompt to keep it consistent. I'm not sure if it will work, but...
■The Difference Between "2D_aesthetic" and "EX"
The main difference lies in how strictly the images were curated.
*2D_aesthetic_base_EX (More Diverse):
I built the dataset for EX first. It consists of about 5,000 images carefully selected from many 2D artists based on both my own visual preference and AI aesthetic scores. If you want more flexibility and a wider variety of styles, EX is the better choice.
You might be able to get a sense of the vibe from the sample images as well. The EX version, in particular, contains images that resemble a wider range of artistic styles.
*2D_aesthetic_base (Highly Curated):
I created this by strictly filtering the EX dataset down to about 1,500 images—only keeping the ones I personally felt were truly exceptional. Because of this, it has a stronger bias toward my personal aesthetics. It aggressively tries to generate highly aesthetic images and leans heavily toward sharp, crisp styles.
Just a heads-up: my explanation of how the LoRA works is based on my own feel and observations from training. I haven't fully mapped out every interaction between the tags and the model myself, so there's definitely room for trial and error.
In my opinion, that diversity is the defining strength of this LoRA.
■I hope this information is helpful! If you have any other questions, please don't hesitate to ask.
i do hope you will branch out to Krea 2.
Thanks for the message! I’m currently focused on dataset construction and training for Anima, so Krea2 hasn’t really been on my radar. To be honest, it wouldn’t even have been an option for me if you hadn’t brought it up.
That said, if there are any of my previous LoRAs that you’d like to see a Krea2 version of, let me know—I’d be happy to consider testing it. While I wasn't originally planning to move to Krea2, I’m open to giving it a shot if there’s genuine interest.
As a side note, I’m still working on the Mega ManX dataset.I'm gradually gathering enough materials that I'm satisfied with.
Since I’m juggling a few other projects and training sessions, it might take a while, but it’s a unique concept that’s really worth the effort. It’s definitely one of the projects I’m committed to finishing. I might even set up the dataset in a way that allows us to test it out on Krea2 down the line.
vampire and retro would be knockouts, i think anima and krea 2 will be the modern kings of anime ai.
@heatwoodzachary763
Thanks for letting me know! I might test out those Krea2 style LoRAs at some point. I still have some Anima style LoRA training that I haven't started yet, but I might be able to run the Krea2 tests in parallel while I'm building datasets for those, so I'll give it a try.