Introduction
Raejin Anima is a specialized fine-tune of the Anima base model, meticulously crafted to elevate its visual fidelity. This model addresses specific limitations found in the original, aiming to deliver a more polished, aesthetic, and balanced output while preserving the core strengths of the base model.
Training Details
v0.1 (Experimental Alpha)
Base Model: anima-preview2
Method: Full fine-tuning using the
kohya/sd-scriptstrainer.Stability: LLM Adapter was disabled during training to mitigate catastrophic forgetting, ensuring the model retains its foundational knowledge while adopting new stylistic improvements.
Dataset: Trained on a curated set of 2,2k high-quality images, utilizing a hybrid captioning strategy of both booru-style tags and natural language to enhance prompt adherence and stylistic depth.
Technical Requirements
[!IMPORTANT]
This model does not include a built-in Text Encoder or VAE. To achieve the intended results, please ensure you have downloaded these components separately.
Download Text Encoder & VAE: HuggingFage page
Recommended Settings
Sampler: Euler a or ER SDE
Scheduled type: Normal
Steps: 32
CFG: 4-5
Resolution: 832 x 1216, 1216 x 832, 1152 x 896, 896 x 1152, 1344 x 768, 768 x 1344, 1024 x 1024
Positive: masterpiece, best quality, absurdres,
Negative: low quality, worst quality, jpeg artifacts, bad anatomy, bad hands, bad proportions, mutation, deformed, extra digits, fewer digits, missing arms, missing legs, signature, watermark, username
Usage Notes
This model is currently in its experimental phase and has not reached its full version. If you encounter any issues or have suggestions for improvement, please feel free to share your feedback.
Support the Development
If you enjoy using this model and would like to support its development, consider:
Sending a tip of Yellow Buzz on this platform.
Buying me a coffee via Ko-fi
Acknowledgements
A special thanks to the community and the developers who made this project possible:
Circlestone Labs: For providing the incredible open-source Anima base model.
Kohya_ss & Contributors: For the comprehensive and powerful open-source training framework.
License
Description
Initial Release
FAQ
Comments (10)
its pretty good suprisiznly good on prompt adherence
It got that sdxl look, is that normal or is it bc of the dataset?
That’s normal, this is still an experiment with the training parameters, and it’s also using my old dataset
I noticed that there is a tendency to make monochrome/ black and white results
Thank you for the feedback! To help me investigate this, could you share the specific prompt and settings that gave you monochrome results? It would help me find a fix for the next update. In the meantime, you might want to try adding monochrome or greyscale to your negative prompt
@Raelina sure ,it was this: @rossy, 1girl, red hood retro days \(nikke\), \(goddess of victory: nikke\), adjusting her bra, standing next to a mirror, tall, wearing a loose open jacket, short jeans open, armpits, smirking, shadows, sunbeam, bedroom, game cg, Highres
Another problem that I noticed is huge like letterbox on half the image : @rossy \(yowayowachimpo\), 1girl, red hood retro days \(nikke\), \(goddess of victory: nikke\), adjusting her bra, standing next to a mirror, tall, wearing a loose open jacket, short jeans open, armpits peek, smirking, shadows, sunbeam, bedroom, game cg, reflections, shadows, side view, Zigzag
Shadows,Highly contrasted color palette
,cel-shading, Highres
I will upload both images so you can see the full results
@Neon_signs Thank you for providing the images that show the issue. After testing with the workflow from the image you shared, I found that the problem comes from the way the prompt is written. Here are the solutions I found after several tests:
Image 1 – Black border or black areas on the sides
To fix this issue, remove “simple background” and “white background” from your negative prompt. Also add quality tags at the beginning of the positive prompt, such as “masterpiece, best quality.” After making these changes, the issue no longer appeared in my tests.
Image 2 – Monochrome or black-and-white output
To fix this issue, simply add “masterpiece, best quality” at the beginning of the positive prompt.
For the best results, follow this tag order in your prompt:
[quality/meta/year/safety tags] [1girl/1boy/1other, etc.] [character] [series] [artist] [general tags]
Alternatively, you can start by using the same negative prompt that I used in the preview showcase.
@Raelina interesting, the reason why I avoid those tags are usual because it tends to alter the original artist style, but I will indeed try removing those in the negatives ,maybe trying masterpiece and highres only since they are more common on danbooru,thank you so much for the help lady🙌🏻, can't wait to new uptades
nice, do you amke this locally?
Thank you, I don't train them locally. all my models are trained by rent a cloud GPU such as RunPod. Full finetuning requires high-end GPU and large VRAM. I only use my local machine for generating the image showcases