■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 LoRA trained via knowledge distillation on Anima, using Krea 2 Turbo as the teacher model.
I am also sharing the actual training scripts and codebase used for this distillation.
With this LoRA, Anima can generate images with the aesthetic and quality of Krea 2 Turbo.
While fine details remain challenging, it offers noticeable improvements in overall atmosphere and a significant reduction in structural artifacts.
Because photo-realistic datasets were included during training, it also handles a semi-realistic, 2.5D aesthetic quite well.
Relying only on Danbooru-style tags might bias the output toward an anime look, so try adding natural language phrases that guide it toward a realistic style.
The primary artifact is a distilled LoRA tuned for CFG 1, 12 steps, using PCM. Please use it directly on the Anima base model.
If you want to apply it to other Turbo models, blending it at around a 0.5 model merge ratio might help average things out and make it easier to work with.
Alternatively, you can use the 'krea2 distillation-only' LoRA included in the ZIP archive. I can't guarantee optimal results across every arbitrary model, but since it is a standard LoRA, it should produce noticeable improvements.
Anima sometimes feels a bit weak when it comes to background rendering; using this LoRA improves that to some degree. That said, I personally don't think Krea 2 itself is exceptionally strong at background micro-details either, so this inherits that ceiling of Krea 2 Turbo. However, anatomical accuracy and overall structural coherence definitely see a solid upgrade.
---
### Motivation & Technical Concept
The idea started with wanting to achieve Krea 2's visual quality inside Anima. Since they share the same VAE architecture, I realized that instead of distilling raw output images, we could teach Anima how Krea 2 Turbo actually denoises and generates. After bouncing ideas off an AI assistant, I set up the training pipeline.
When people hear 'knowledge distillation,' they might assume Anima inherited every bit of Krea 2's semantic knowledge base—that isn't quite the case. It is better understood as learning Krea 2's generation methodology.
Think of it like Anima running an intermediate latent generation step, and Krea 2 stepping in to perform an internal hires-fix/refinement. By supervising that continuous corrective step, Anima effectively learns how Krea 2 polishes a piece.
It's analogous to an art class where a student draws a rough draft and the instructor makes direct corrections on top of it: 'Ah, that's how the teacher shapes the strokes—that makes sense.'
The scale wasn't massive, and the process turned out to be surprisingly straightforward. It was trained on roughly 1,200 captioned samples split evenly between photorealistic and anime datasets (50/50). Even with that compact dataset, it yields remarkably clean results with no major issues.
The CFG 1, 12-step PCM distillation is my own custom pipeline—it is distinct from generic off-the-shelf Turbo methods.
---
### Archive Contents (Backup ZIP)
* Krea 2 Distillation Only LoRA: Highest fidelity and stability. Requires standard generation settings (~30 steps, CFG 4.5), so it is slower.
* Krea 2 Distillation + CFG 1 LoRA: Optimized for ~30 steps at CFG 1. Roughly 2x faster than standard generation, offering great practical usability.
* Training Code: The complete distillation scripts and training routines.