🌟 Introducing Karcher Mean Method Merge Model v0

Karmix-XL v0

📋 Overview

This article was created specifically for @StableDiffusion_CN

Latest update of main information in the article: 28.04.2025 18:55 (UTC+8):

Added the examples of Ultra-wide Aspect Ratio

Release Karmix-XL v0 - 28.04.2025 00:50 (UTC+8)

Past changes:

Edit from 27.04.2025 15:31 (UTC+8): More images have been added (the «landscape» aspect ratio with different variants of prompts, will be updated in future)
Edit from 27.04.2025 14:36 (UTC+8): Added little more example images. (examples for Natural Language, Natural Language + Artists Tags prompt)

Added ❗important technical information about model performance. Please check this out. Last edit: 27.04.2025 14:36 (UTC+8)

I'm excited to present a new merged model created using the Karcher Mean Method. At the moment, the current version of the model is extremely experimental and requires further tests and experiments with merging with different parameters. But overall, I believe it can have its own future as just a good merge (at least, I hope, haha)

🍀Main information about the model:

I would like to repeat that this is a highly experimental model, so it is version v0.

This is my first experience with serious model merging, so I am open to criticism regarding quality or other aspects and will always be happy to receive recommendations for improvement.

Currently, the model is provided 'as is', since it still requires improvements. However, it is already largely usable and can make good results, although certain nuances may occur.

Model is Epsilon-Pred

Advantages:

  • High-resolution latent support, up to 2048.
    (Please check the nuances regarding aspect ratios at high latent resolutions in the Model Performance section, update from (27.04.2025 14:36 edit) (UTC+8).)
  • In general, generations were performed at 1280×1856, 1860×1280, and 1536×2048, with acceptable to very good results.
  • Of course, it’s not perfect — in some cases, extremely high latent resolutions can worsen the output.
    Example: using the "middle finger" tag at such resolutions may result in extremely long, stick-like fingers.
  • Stylistic/conceptual improvement when mixing artist tags or using a single artist tag.
    Examples and comparisons with Illustrious v2.0 and other models are provided in the Example of Images section.
  • Different variations of prompt writing. It is possible to combine Danbooru Tags, natural language, and artist tags. Approximate orders and combinations will be described below.

Uncertain Points:

  • The model has not yet been tested for knowledge of newly popular characters.
    It is unclear which characters it recognizes and which it does not.

Problematic Points:

  • Possible anatomical issues (e.g., an extra or missing finger), though in my experience, this occurs infrequently.
  • The most critical and problematic aspect complicating model usage is detailed in the Model Performance section.
    (Update: (27.04.2025 14:36 edit) (UTC+8).)
    Please read it carefully to avoid problems when generating high-resolution images in a landscape aspect ratio.

Base Models Used (in the order of merging, 23.04.2025 model version (latest v0):

  • hesperidesIllustrious_v10
  • illustriousXLPersonalMerge_v30Noob10based
  • illustriousXL20_v20 (CivitAI version, not "STABLE" from OnomaAI HuggingFace)
  • OD39_E60_fp16

🛠️ Merge Parameters

Special thanks to our friend su momo for providing the script that enables custom alpha parameter tuning for TE/UNET and overall extensive information, and also thanks to 乙酰胆碱 for the OD39_E60_fp16 model.

The merge was performed using the following parameters:

1
2
3
4
5
6
7
--karcher-iter 30 
--karcher-tol 1e-7
--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

That is, the --alphas ratio for the models is:

  • hesperidesIllustrious_v10 - 0.23 ratio
  • illustriousXLPersonalMerge_v30Noob10based - 0.23 ratio
  • illustriousXL20_v20 - 0.23 ratio
  • OD39_E60_fp16 - 0.31 ratio

TE/UNET alphas in the appropriate models order.

⚠️ Performance Notes

Please note that this model is still in development and not perfect, but it demonstrates good results, especially at higher resolutions such as 1280×1856 and 1536×2048.

At 1024×1536 resolution, results are generally acceptable but may occasionally show issues with anatomy (particularly fingers).

The model supports a variety of options for understanding prompts:

  • Danbooru Tags
  • Artists Tags + Danbooru Tags
  • Artists Tags + Danbooru Tags + Natural Language or Artists Tags + Natural Language + Danbooru Tags
  • Natural Language
  • Natural Language + Artists Tags
  • Natural Language + Danbooru Tags
  • Natural Language + Danbooru Tags + Artists Tags

Pay attention!

According to my personal experience, when using artist tags along with natural language, it is best to write artist tags at the end, labeling them as "artist:". Example: Scene with a beautiful anime girl with white hair and red eyes. artist:konya_karasue
Example (so far, one) of the differences between using "artist:" and not will be given below.

⚠️‼️New important information about performance (27.04.2025 14:36 edit) (UTC+8)

A strange behavior of the model was noticed when generating NSFW content in landscape aspect ratio when using high latent resolution (1856x1280 and few lower)

When trying to generate an image in this aspect ratio and resolution using the Euler (scheduler Normal) sampler, anatomy most often "breaks".
Solution - reduce the resolution or change the sampler to a series of DPM++ (2M, 3M), scheduler to sgm_uniform and CFG 2.5.
It is recommended to change the sampler, as changing the resolution may not help much.

At the same time, under exactly the same conditions (1856x1280, Euler Normal), no such problems are observed during SFW image generation.

Hypotheses: referring to the documentation of Illustrious v2.0 from CivitAI, it says that the recommended sampler is DPM++ 2S Ancestral CFG++, Scheduler is Normal, and CFG Scale is 2.5.

It is likely that the model has received some kind of separation into "modes" (?) . Large latent portrait and landscape resolution for SFW work better stylistically with the Euler Normal (most likely taken from other models), but landscape for NSFW are on DPM++ series with CFG 2.5, since ILLv2.0 requires it.
I used DPM++ 3M SDE with Scheduler SGM Unifrom and CFG Scale 2.5 and the effect of normal anatomy was obtained.

Examples of this with SFW/NSFW will be given in this article, as I do not want NSFW to be present in the main article:
⚠️ Problematic SFW/NSFW Examples with different Samplers

The parameters that were used to generate images:

  • Software:ComfyUI
  • Steps:28
  • Guidance Scale:4.3
  • Sampler: Euler
  • Scheduler: Normal
  • Without any hires.fix/After Detailer

🖼️ Examples of Images

  • Artists Tags + Danbooru Tags


[1536x2048] | [1280x1856] | [1024x1536]

colored eyelashes, tooasagi1, toosaka1,nekojira, artist:dino_\(dinoartforame\), quasarcake, rolua, rella, void 0, chen bin, fymrie, villainchin, sushispin, from above, Very dark environment, dim lighting, shadow, dark background, night, black theme, ,1girl,solo ,animal_ears, grey_hair, sleeves_past_wrists, apron, maid_headdress, cat_ears, long_hair, white_background, cat_girl, hair_ornament, sleeves_past_fingers, blue_eyes, cetacean_tail , smile, bangs, white_apron, fang, long_sleeves, :d, frilled_apron, striped, heart, maid, simple_background, open_mouth, solo, shirt, frills, virtual_youtuber, very_long_hair, bow, looking_at_viewer, blue_dress, blue_bow


[1280x1856]

yuuka \(blue archive\),mari \(idol\) \(blue archive\) \(cosplay\),kaede_\(sayappa\) ,alphonse_\(white_datura\),z3zz4,flippy \(cripine111\), gomzi, fajyobore,re0n, rei_\(sanbonzakura\),ogipote,ask_\(askzy\),yoneyama mai,[betabeet],Overhead lighting, soft shadows, diffused light, low contrast, highlights on hair, skin highlights, fabric folds shadowing, subtle reflections, light from above and to the left, gentle illumination, no harsh shadows.1girl, solo, cosplay,one eye closed, garter straps, black thighhighs, detached sleeves, looking at viewer, alternate costume, black headwear, spoken heart, blush, black panties, dress, thighs, top hat, best quality, very aesthetic, absurdres



[1024x1536]

[1280x1856]

[1536x2048]

z3zz4,wanke,mashiro_kta,ohisashiburi,karyln,fymrie,ibuki \(blue archive\), 1girl, yellow halo, black tail, solo, blonde hair, yellow eyes, double v, black dress, long hair, demon tail, white shirt, white background, open mouth, demon wings, black horns, demon horns, smile, simple background, blush, pinafore dress, pleated dress, side ponytail, black wings, skin fang, collared shirt, long sleeves, looking at viewer, red ribbon, feet out of frame, nail polish, no text, best quality, very aesthetic, absurdres


[1280x1856]

[1536x2048]

amamiya kokoro,[haku89],[[modare,[piromizu]]],[[nixeu,reoen]],[zygocactus,miv4t,rella,konya karasue],[[kakao rantan,kakao_\(chocolate_land\)]],Perfect_lighting, close-up, 1girl, solo, {{> o}}, {{:p}}, tongue out, virtual youtuber, blue hair, yellow eyes, ahoge, hair ornament, ribbon, long hair, fang, skin fang, hair ribbon, looking at viewer, hair bun, black ribbon, apron, two side up, double bun, portrait, hair between eyes, x hair ornament, frills, maid, dress, black dress, year 2024,masterpiece,best quality,absurdres,very aesthetic,very awa


  • Artist Tags + Danbooru Tags + Landscape Aspect Ratio

[1856x1280]

kaname madoka, artist:konya_karasue, 1girl, solo, arrow \(projectile\), black background, bow, bow \(weapon\), buttons, center frills, chest jewel, choker, clenched teeth, clock, cross-laced shoes, determined, drawing bow, dress, energy beam, firing, floating, floating clothes, floating hair, footwear bow, frilled skirt, frilled sleeves, frills, full body, gears, gleam, gloves, hair bow, holding, holding arrow, holding bow \(weapon\), holding weapon, kneehighs, light particles, light rays, looking ahead, looking up, magical girl, mary janes, miniskirt, petticoat, pink background, pink dress, pink eyes, pink gemstone, pink hair, pointy nose, puffy short sleeves, puffy sleeves, red bow, red choker, red shoes, shoes, short dress, short hair, short sleeves, short twintails, skirt, skirt under dress, socks, square neckline, teeth, twintails, two-tone background, waist bow, weapon, white gloves, white skirt, white sleeves, white socks


[1856x1280]

kaname madoka, ultimate madoka, artist:konya_karasue, 1girl, bouquet, bow, city, cloud, dress, floating, flower, frilled dress, frilled sleeves, frills, full body, grass, hair bow, high heels, holding, holding bouquet, hugging object, lake, long dress, long hair, looking at viewer, open clothes, open dress, open mouth, outdoors, petals, pink bow, pink flower, pink hair, pink sky, pink thighhighs, power lines, shoes, short sleeves, sky, smile, solo, sun, sunrise, thighhighs, two side up, very long hair, waist bow, white bow, white dress, white shoes, winged boots, yellow eyes


  • Danbooru Tags + Ultra-wide Aspect Ratio [DPM++ 3M SDE SGM Uniform]

[2048x1024]

[3072x1024]

kaname madoka, ultimate madoka, artist:konya_karasue, 1girl, bouquet, bow, city, cloud, dress, floating, flower, frilled dress, frilled sleeves, frills, full body, grass, hair bow, high heels, holding, holding bouquet, hugging object, lake, long dress, long hair, looking at viewer, open clothes, open dress, open mouth, outdoors, petals, pink bow, pink flower, pink hair, pink sky, pink thighhighs, power lines, shoes, short sleeves, sky, smile, solo, sun, sunrise, thighhighs, two side up, very long hair, waist bow, white bow, white dress, white shoes, winged boots, yellow eyes


  • Artist Tag + Natural Language + Danbooru Tags + Landscape Aspect Ratio

https://s.iimg.su/s/27/wnSisLzEyoE6GwYcmS7YkMLFAngT2hAXb3uaQYoJ.png [1856x1280]

{akemi homur, akuma homura, kaname madoka, ultimate madoka}, artist:konya_karasue, Overall scene with a very touching and romantic atmosphere. 2girls, black dress, bow, dress, eye contact, hair bow, long hair, looking at another, multiple girls, open mouth, pinned, purple hair, short sleeves, simple background, smile, straddling, strapless, strapless dress, tears, very long hair, white background, white bow, white dress, yellow eyes, yuri


  • Natural Language

[1280x1856]

The anime depicts an adult woman with long black hair, wearing a white short-sleeved shirt, leaning on the balcony railing at night, gazing out at the street. The cinematic view from side. The background is in the anime style, with dark blue tones, a night scene, city lights, a distant view, and a panoramic perspective, creating a melancholic atmosphere. The woman is dressed in casual


  • Natural Language + Artists Tags

[1280x1856]

The anime depicts an solo adult woman with long black hair, wearing a white short-sleeved shirt, leaning on the balcony railing at night, gazing out at the street. The cinematic view from side. The background is very photorealistic, with dark blue tones, a night scene, city lights, a distant view, and a panoramic perspective, creating a melancholic atmosphere. The woman is dressed in casual. artist:toosaka_asagi,artist:nekojira, artist:dino_\(dinoartforame\), artist:quasarcake, artist:rolua, artist:rella, artist:void_0, artist:chen_bin, artist:fymrie, artist:villainchin, artist:sushispin,


  • Natural Language + Artist Tags + Landscape Aspect Ratio

[1856x1280]

A magical girl Kaname Madoka from mahou shoujo madoka magica anime with pink hair and a frilly dress floats in a cosmic space, aiming a glowing arrow at a unknown giant skeleton with a bow \(weapon\). The scene has the distant view from back. Her gaze is aimed at skeleton. Around her spiral large, transparent clock faces, hinting at a theme of time. The scene uses vibrant pinks, purples, and deep space tones, with a detailed, anime-style fantasy aesthetic. artist:konya_karasue


[1856x1280]

A magical girl Kaname Madoka from mahou shoujo madoka magica anime with pink hair and a frilly dress floats in a cosmic space, aiming a glowing arrow at a unknown giant skeleton with a bow \(weapon\). The scene has the distant view from back. Her gaze is aimed at skeleton. Around her spiral large, transparent clock faces, hinting at a theme of time. The scene uses vibrant pinks, purples, and deep space tones, with a detailed, anime-style fantasy aesthetic. artist:konya_karasue


🔄 More images in progress

📝 Feedback

If you have any questions or suggestions, you can contact me - @chemxigua (Telegram).

🔄 Future Work

Work is ongoing to refine the model further and address the current limitations. Stay tuned for updates.

Edit

Pub: 25 Apr 2025 12:04 UTC

Edit: 28 Apr 2025 11:04 UTC

Views: 795