Karmix-XL v0

Karmix-XL v0 is an experimental anime-style text-to-image model based on SDXL (Epsilon-Pred). It focuses on high-resolution latent support, stylistic flexibility, and the ability to interpret varying prompt structures ranging from Danbooru tags to natural language.

Sample images demonstrating the model's capabilities with various prompts.

Note: These results are raw text2image, no Hires.fix, AfterDetailer, Inpaint, etc. applied. The portrait resolution is 1280x1856, landscape is 1536x1024.

The software used is ComfyUI, uploaded images contain workflow in images metadata.

Model Description

This model was created using the Karcher Mean merge method, specifically utilizing a modified version of the script originally developed by win10ogod

The model is a merge of models with a 1536x1536 base (Illustrious-XL 2.0), 1536+1024 mixes (Illustrious + NoobAI) and a 1024x1024 base (trained OD39_E60, based on NoobAI 1.0)

Source Code:

The modified source code used for this merge is available here:

https://github.com/chemwolf118/Karcher-merge

Merge Recipe (v0)

The merge was performed using the following alpha ratios:

Model Component | Ratio (Alpha)

• hesperidesIllustrious_v10 | 0.23

• illustriousXLPersonalMerge_v30Noob10based | 0.23

• illustriousXL20_v20 | 0.23 | CivitAI version

• OD39_E60_fp16 | 0.31

Detailed Parameters:

This model was merged with karcher-iter 30 and karcher-tol 1e-7. The specific layer weights were:

--alphas 0.23 0.23 0.23 0.31

--alphas-te 0.28 0.18 0.18 0.36

--alphas-unet-in 0.15 0.24 0.15 0.46

--alphas-unet-mid 0.15 0.24 0.15 0.46

--alphas-unet-out 0.28 0.18 0.18 0.36

Usage & Recommendations

Important: This model is experimental ("v0"). You may encounter different behaviors depending on resolution, sampler settings and artist tags written.

Resolutions and Samplers

The model supports high-resolution latent generation (up to 1856px).

Tested Resolutions: 1280×1856, 1856×1280 and standart 1024x1536, 1536x1024.

Note: In some cases, it is possible to do 1536x2048 and 2048x1536, but in very rare cases and not with all prompts.

🚧 Model Limitations and Performance Note

Although the model has the ability to generate directly in 1280x1856, it still has its limitations and nuances regarding the quality generated in such a high resolution.

Since several models with an admixture of NoobAI are used, which is based on 1024 rather than 1536, using middle finger/full body in some cases can cause an elongated anatomy depending on tags/pose/angle.

Also, a known problem exists where anatomy may breaks in landscape orientations when using the Euler sampler.

Use Case | Recommended Sampler | Scheduler | Notes

• Portrait | Euler | Normal | Works generally well.

• Landscape | DPM++ 2M or 3M | sgm_uniform | Crucial: Using Euler in high-res landscape (e.g., 1856x1280) often causes anatomy breaks.

Recommended Settings

• Resolution / Aspect Ratio - 1024x1536-1280x1856 for portrait aspect ratio and 1536x1024 for landscape

• CFG - 3.5-5 (4.0-4.3 recommended)

• Steps - 28-30

• Sampler - Euler

• Scheduler - Normal

Prompting Styles

Karmix-XL v0 is flexible and supports multiple prompting structures. You can mix and match styles as shown below:

1. Danbooru Tags

Standard tag-based prompting.

> 1girl, solo, cat_ears, maid, white apron, smile, looking at viewer, indoors

2. Natural Language

Descriptive sentences similar to standard SDXL usage.

> The anime depicts an adult woman with long black hair, leaning on the balcony railing at night, city lights in background.

3. Mixed: Natural Language + Artist Tags

Use natural language for the scene and tags for the style.

> Scene with a beautiful anime girl with white hair and red eyes, intricate dress. artist:konya_karasue

4. Complex Mixes

You can combine all methods:

> A cinematic shot of a warrior girl, forest background. 1girl, armor, sword, holding weapon. yao liao wang,xilmo, (fkey:0.5), meinoss,dikko

Note to prompting: if you encounter a problem with the appearance of unnecessary "creatures" or objects in images when using artist tags with specific nicknames (like if tanti_men_(dragon) suddenly causes the appearance of dragons, chuck_(harfmoondark) causes the appearance of chuck_(pokemon), etc.), then it is recommended to use "artist:" for such tags, for example artist:chuck_(harfmoondark).

Version History

• v0 - Created 23.04.2025, Uploaded to HuggingFace 28.04.2025 | Link

  • Initial release.

• v0-CivitAI - Uploaded at 5.12.2025

Other Versions

Other experimental versions with different merging methods/parameters can be found here:

https://huggingface.co/NullAxis/karmix-merge-experiments

(Note: These versions are prototypes and do not reflect the final quality of the future main release).

NullAxis is an organization that was created as a branch to work on experiments with Karmix and various model fusion methods, and now it has a website nullaxis.dev. Information about activities, such as existing works and upcoming plans, will be published here as they appear.

Acknowledgments

Special thanks to the contributors and developers whose work, advice, and support made this project possible:

• su momo — The main contribution to the improvement of the Karcher-merge script, advice and help.

• win10ogod — The author of the Karcher-merge method script, with which Karmix was created. GitHub | HuggingFace

• 乙酰胆碱 — The author of the beautiful OD39_E60 model used in Karmix.

• DammK YaZuXaWa — The original source of information about the Karcher-merge method from win10ogod and various help in process. GitHub | HuggingFace | CivitAI

• Kurosu Chan — Help with various questions and evaluation of the initial article for Karmix. GitHub | Gist | HuggingFace

• Stable Diffusion CN — Special thanks to the entire SDCN community for everything.

• Laxhar Lab — for the NoobAI XL

• OnomaAI — for the Illustrious XL

• bloodsplash — for the Hesperides XL v1.0 | CivitAI

• dims2 — for the Illustrious XL personal merge v3.0 | CivitAI

• and others — Community support & testing.

The initial article has now been archived and information will be provided in the HuggingFace model card.

If you have any questions, suggestions or need feedback, please create a discussion on HuggingFace or contact us by email at [email protected].

License

This model is released under the Fair AI Public License 1.0-SD (Illustrious License).

It also incorporates components from NoobAI, which carries its own specific usage terms. Please ensure compliance with both:

Illustrious License terms: https://freedevproject.org/faipl-1.0-sd

NoobAI License terms: https://huggingface.co/Laxhar/noobai-XL-1.0/blob/main/README.md#model-license