Hypernetwork for making cute and funny dollfie like anime face with realistic/semirealistic body and photographic images.

I create this hypernetwork on late january for personal fun and has been playing with it to create dollfie images and i think i will share this with you all.

Feel free to post your "cutest" image on the discussion section, i want to see it 😉.

trained on AbyssKotosMix (https://civitai.com/models/5245/kotosmix), i suggest you to use the same models. i think it also can be used with any anime based models (like anythingv3), also compatible with some character LORA.

Focus trained on dataset images of DDH-01, DDH-09, DHH-10, DHH-22, DHH-27, DDH-28 and my own dolls (mainly with cute face), 2000steps.

I will improve this hypernetwork and maybe create LORA version of it in the future.

I suggest to use cfg scale : 7 or lower.

This hypernetwork also good at img2img

Recomended prompt : 1girl, 3d, realistic, bangs, depth_of_field, eyelashes, motion_blur, solo, depth_of_field,bokeh, chromatic_aberration, (highres:1.05), (intricate detail:1.1025)

for negative just use the standard negative prompt or use EasyNegative https://civitai.com/models/7808/easynegative

example images used kotosmix at clip skip 2 as well as the Textual Inversions, bad-artist, bad-prompt-2, and bad-hands-5

I recommend setting the hypernetwork strength to 0.9

2nd image is a custom character i trained

Watercolour hypernetwork that is pretty useful for generating portraits with the soft painterly look on it. Can be used with both anime and realistic(maybe) models.

Would recommend testing to see what weights work for you. it might be a little too strong on anime models? might need to use lower weights for those while 0.8-1 on realistic models should be fine

Trained on mostly portraits so it might not be able to do backgrounds that well or default to background that is a splash of colour. For some reason likes to sprout flowers even unprompted.

Trained on the same data as the NEON ISOMETRIC model - but comes up with some vibrant different results as per the images

Coffee is nice: https://ko-fi.com/DUSKFALLcrew

--no joke, coffee is amazing, so if YOU DO like what we do, PLEASE send us some money for our bad addictions.

Hypernetwork for other model

hypernetwork of actress Eva Green. triggerword (not that important with hypernetworks tough) is 3v4gr33n, I suggest adding 'green' to your negative prompt

probably a bit overcooked, I suggest not using it at full strength

Hypernetwork for the model of the same name.

Can help generate better images.

Hypernetwork that creates gigantomastia breasts. Works ok with Uberrealistic porn merge v1.2 and Real BB 3.0. I've used clip skip 2 (for some reason) and ~0.8 multiplier for ok'ish results. Kind of needs some Lora's in conjunction. Use "gigantic breasts close up gigantomastia" as part of prompt.

It's the first thing I've trained so sorry for bad quality.

a hyper network trained by よし男's artwork.

(reference: https://www.pixiv.net/users/3584828)

only for study and self use

please do not publish or use for business.

请勿发表或商用

Author: Tongfan Wei (weitf@bu.edu)

In the style of Sazanami Mio, he mostly draws youmu so sometimes it erroneously gives characters white hair/ blue eyes ect.

It can work at high strengths (0.9ish) but the (worst quality, low quality) prompt being set high might make it grainy.

For whatever reason it struggles with hands more than the default AbyssOrange, which it was trained on.

Train after chilloutmix

After listening to a chilloutmix and some lora, I was struck by the distinctive makeup style of Thai women, which sets them apart from their Korean and Chinese counterparts. As with any cultural influence, it is important to appreciate and use this style in a manner that is both appropriate and respectful. It is imperative that this content is not exploited for the creation of pornographic or NSFW material, as such conduct would be highly inappropriate and disrespectful. Let us strive to use this content to explore and appreciate the unique beauty of Thai makeup, while upholding the dignity and respect of its cultural origins.

Trained on 82 unique NSFW images (164 after mirroring) that were edited to remove text, then tagged by deepbooru then manually tagged. 13 000 steps, 79 epochs, LR 5e-6, network structure: 1, 2, 2, 1, linear AF, no dropout, layer normalization.

Problems with faces and hands, but can produce good images if you find the right seed. All sample images produced with AbyssOrangeMix2_NSFW

Recommended tags: 1girl, large breasts, arms at sides, long hair, from below, cropped legs / full body, wide hips, thick thighs

Recommended negative tags: anklet, earrings, bracelet, jewelry.

This hypernetwork will help you to make your Senko-san be look like she was drawn by Rimukoro. This model was trained using AbyssOrangeMix2_sfw model so it should work fine with that specific model or with other relative models. Personally I would recommend you to use this hypernetwork with weight 0.7-0.8

All the content I make is and will be completely free on CivitAI.
But if you want to support me with a coin, then you can do it here.

Hypernetwork trained on screencaps of JoJo's Bizarre Adventure Stardust Crusaders on AbyssOrangeMix2_sfw with clip skip 2.
Optimal strength is 0.6, but you may go higher if you want more flat style.

Also you may adjust it with

Prompt: athletic, muscular, jojo no kimyou na bouken, araki hirohiko \(style\)

Negative prompt: 3d, realistic, kid, child, loli, young

Trained on 186 images of semi-realistic characters with highly detailed painted backgrounds. No specific characters or people in data-set. I recommend 0.7 strength.

All example images use AyoniMix V4 and V3 with hypernet strength at 0.7.

A hypernetwork for Modeus.

Trained from art by Mitsuba Choco, a visual novel kind of style

Trained on AbyssOrangeMix2_NSFW

For whatever reason it can generate grainy looking results, using "<hypernet:Mitsuba_Choco:0.9>, (masterpiece, best quality:1.2), extremely detailed, jewel like eyes" As your positive prompts and "(bad_prompt_version2:0.6), (worst quality, low quality:1.0)" as your negative makes the images a lot clearer. Alternatively, upscaling the images in extras helps a lot too.

Hypernetwork of everyones favorite milf. Due to the nature of source materials, this network is quite horny

Works always, 50% of the time, sometimes

A hypernetwork trained on a batch of images generated with a successful prompt using my phaninc model. Best to use futuristic and femborg in the prompt.

This is a hypernetwork of Emilia from Re:Zero.

You can select the hypernetwork from the Hypernetwork input field in settings.

This hypernetwork was trained on Corneo's 7th Heaven Mix.

For best results use values around 0.6-0.7 ,and using tag 'long hair'

This hypernetwork was trained on Warcraft 3 Reforged Units and Heroes portraits on AbyssOrangeMix2_sfw with clip skip 2.

Keep hypernetwork strength about 0.6 for decent results. Best results you may get with square images, but another aspect ratio also supported.

See my prompt examples in showcase generation info.

Please share your images in review section down below.

This is a hypernetwork of Miku Nakano from The Quintessential Quintuplets.

You can select the hypernetwork from the Hypernetwork input field in settings

This hypernetwork was trained on Corneo's 7th Heaven Mix.

For best results use values around 0.6-0.8

This is a hypernetwork of Itsuki Nakano from The Quintessential Quintuplets.

You can select the hypernetwork from the Hypernetwork input field in settings

This hypernetwork was trained on Corneo's 7th Heaven Mix.

For best results use values around 0.6-0.8

This is a hypernetwork of Hayasaka Ai from Kaguya-sama: Love Is War.

You can select the hypernetwork from the Hypernetwork input field in settings

This hypernetwork was trained on Corneo's 7th Heaven Mix.

For best results use values around 0.5-0.7

This hypernetwork can generate images in a cartoon, anime, or realistic styles. Realistic style with robot look sometimes hard to achieve! Also hyper net can produce NSFW arts.

The hypernetwork is trained on anime style model, so other models in cartoon or anime styles should be fine.

Activation tags - jenny_wakeman

For cartoon look - 2d, cartoon, cartoonlike

For more realistic/anime look - detailed, masterpiece, anime

Personally I use "Euler a" sampler and CFG scale - 7. Clip skip - 2.

Also don't forget to play with hypernet strength I prefer somewhere in the 0.6-1

On some models tag "realistic" might break stuff or generate ugly girls, so you might wanna use lower values - (realistic:0.2) or (realistic:0.4)

Taliyah hypernetwork for AbyssOrangeMix2

This hypernetwork was trained on AnythingV3, so it should work pretty well with any or almost any other anime-based model. Also shows pretty solid results with Protogen, so feel free to experiment.
Showcase images was generated with AnythingV3.

My prompts for experemintation was pretty simple, so i recommend to begin with this parameters:

• Prompt:

  masterpiece, your_object, dark_background, flat illustration, vector, very detailed

  Also try
  masterpiece, your_object, dark_background, very detailed, icon, centered, realistic, cinematic lighting

• Negative:

  low quality, lowres, text, error, (cropped:1.2), worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, black and white, monochrome, censored, (out of frame:1.3)

Please share your images in review section down below.

Hypercum-4 is a hypernetwork that allows user to create an NSFW image of target character getting cumshot or "bukkake" in Japanese porn terms. By default, AI seems trying to avoid from making an image of cumshot with such large amount of semen ejaculated around the character target, only covers on most areas eg. breasts and thighs but less focus on covering other areas.

This hypernetwork were trained with couple of edited anime/hentai artwork with large, thick creamy cumshot as datasets in order to make AI trying to mimicking the moment. Trained at around 5000 steps with a gradient level of 10 and trained with Berry_HD model.

Currently, this model only works with Berry_HD and Berrything_HD, requires hypernetwork strength to be set between 0.055 to 0.255. Default setting value will just make image result too horrible to see.

Warning: This model is focused solely on NSFW image generation. You must be at least 18 years old or above before using this model. I an not responsible if anything happens during the usage of this model below the legal age.

Trained from art by qqqrinkappp from the various booru sites.

Model trained on anything v4.5

From testing I've found that a hypernetwork weight of around 0.8 seems to work best, depending on your settings you may want to play with this value.

In order to get outputs closer to the original style I've added "starry_sky_print, starry_sky, (traditional_media:1.2), (light_particles:1.2), depth_of_field, milky way, blurry" as prompts

A hypernetwork trained on tattoo designs, the keyword "phantatt" is very useful, I also tend to use "american traditional" or "neotraditional"

This hypernetwork can generate images in a cartoon, anime, or realistic styles. It's not great, but hey, it works! The original cartoon style can sometimes be tricky.

The hypernetwork is trained on model Any v4.5, but other models, especially in the anime style, are fine too.

Activation tags - FionnaTheHuman, fionna_the_human_girl

For bunny ears don't forget to add - aimal_ears, bunny_ears, animal_hood, bunny_hood, white hood

Also don't forget to play with hypernet strength I prefer somewhere in the 0.6-1

There is room for improvement, and I might do a version 2 if this hypernetwork becomes popular.

Trained on cyborgs, mecha and techwear for you to build your own drone army with.

thanks to aah! for curating the collection I trained the hypernetwork on.

Trained on 122 fantasy art images, specialised in horns and facial ornamentation

A model that allows you to portray someone whom the guards are often afraid of. A personality that is firm in will and character. COMMISSAR!

use here : https://civitai.com/models/5160/project-photo-beta-20-lite

Just a hypernetwork trained on monochromatic gothic RPG art inspired by ClassicRPGArt's checkpoints on this site. All examples are on top of the protogen 2.2 model although it was trained from the base stablediffusion

A model that allows you to create space marine. Unfortunately, it doesn't always work well. But some of her creations look very pretentious. Experiment, combine models, use different checkpoints, all this will allow you to get interesting results.

In most cases it works relatively well, it was not tested with other models! (Have Fun)

Project Unreal Engine 5

This version not finished!

Dont forget use tagger

Donate : https://www.buymeacoffee.com/hotfantasyai

My first model. Trained to create only cybernetic faces or heads in the style of Adeptus Mechanicus. This creation is far from perfect. The images you receive may lack the usual colors of the color. But I think I was able to achieve the necessary style. I am also open to criticism and would welcome any comments.

Sam Does Arts style (Hypernetwork) for Anything v3 compatible models and different versions.

This is just a reupload of my previous Sam Does Arts hypernetwork.

If you like this model, please add a small review or a sample of your work to share with the others all the possibility of this model.

Aroma Sensei style (Hypernetwork) for Anything v3 compatible models and different versions.

This LoRA model is a request of some Civitai users.

If you like this model, please add a small review or a sample of your work to share with the others all the possibility of this model.

The hypernetwork is trained to generate Anna Miller from Metro Exodus on the Deliberate v1.1 model https://civitai.com/models/4823/deliberate

Token word : "annamedus"

It doesn't work very well, but it was enough for my tasks.

Trained based off of the work of Kyouda Suzuka, Images taken from safebooru. Trained on anything v4.5

Example images were made with hypernetwork strength on 0.9 on anything v4.5 with clip skip on 2.

Note that the further the subject is from the viewer the thinner the outline is

InCase style (Hypernetwork) for Anything v3 compatible models and different versions.

This is just a reupload of my previous InCase hypernetwork.

If you like this model, please add a small review or a sample of your work to share with the others all the possibility of this model.

Training on Randy StarFru1T art maker.

Arty took from Art Station. NSFW welcome, made on AbyssOrangeMix2_Hard model

Good promts: A completely random word can give you a good result

Negative promts:(worst quality, low quality:1.4), bad finger, bad hands, bad anatomy,

Trained on the most expressive illustrations of off8mile. I didn't expect the results to be as... visceral as they are, often more exaggerated than even the source material, but the hypernetwork strength acts as a great emotion modulator, whether you're looking for seething hatred, contemptuous disgust, maniacal glee, cataclysmic despair or twisted obsession.

On behalf of the artist of the training images, I should reiterate that this hypernetwork and any images generated with it are strictly for non-commercial purposes only.

This was my first try of a Vtuber Template

Sample images are 4k 16:9 wallpaper res, so need to click to expand

This is a hypernet trained primarily on some of my favorite artists works which brings together a really cool and unique style.

Cosmicangel is how to trigger it, I also advise using "by PeterMohrBacher" with it, like so either to the end or for more abstract to the front:

(cosmicangel, art by PeterMohrBacher)

And add other artists you like or replace Peter Mohrbacher with others to try out interesting mixes.

I HIGHLY recommend using DPM Adaptive, it's the absolute best sampler for it.

Example prompt output:

(photorealistic,masterpiece),(ornate knight in a castle),(glowing, neon ),moody lighting, colorful,ornate,(cosmicangel , art by PeterMohrBacher),

Steps: 38, Sampler: DPM adaptive, CFG scale: 7, Seed: 3129750051, Size: 909x512, Model hash: 58b1e17931, Model: ImagiMostgradv25Proto53gradv3, Hypernet: cosmicangel, Hypernet hash: 26d331600d, Batch size: 8, Batch pos: 0, Denoising strength: 0.54, Eta: 0.3, Clip skip: 2, ENSD: 31337, Hires resize: 1280x720, Hires steps: 64, Hires upscaler: 4x-UltraSharp

Use lowres in negative prompt and clip skip = 2.

To get male characters use the 1boy in the prompt and set the weight of the hyper network lower than 1. Using other words such as man, without additional descriptors, may result in female characters.

This hypernetwork was trained on Anything V3 so it should work with almost any anime-based model.

Every sample image was made with AbyssOrange_hard.

I recommend 0.6 strenght for most generations.

There is a few versions of hypernetwork, and each of it have different shading and affect on whole image with different strength.

Base dataset was 4K screenshots from Batman v Superman cropped to 512, so with more strenght of hypernetwork images will colorized more like color palette of this movie, and have more film grain with bokeh.

Sometimes it may conflict with LORA's styles so you may need to decrease strength of hypernetwork or LORA, also order of use matters in new Automatic 1111 interface, so maybe you will ned to experiment with stuff a lil bit.

Also if you have any info how to train hypernetworks, LORA's and TI properly, please share your knowledge.

Slim Models v.1 - No more "Big breasts"

This hypernetwork works with female poses. You must to use the tags skinny, small breast etc.. on your main model It will be create more accurate images and poses of small breasted girls.
It doesn't works full face portraits.
It works with all the models, but works best with photo realistic ones.

Trained with more than 200 erotic photos of small breasts models, posing to the camera.

Negative prompt: busty
Hypernetwork Strength: 0.5 - 1

Optional: vae-ft-mse-840000-ema-pruned.ckpt (https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt)

This hypernetwork was trained on AnythingV3, so it should work pretty well with any or almost any other anime-based model.
Showcase images was generated with AnythingV3 and AbyssOrangeMix2_hard.

My prompts for experemintation was pretty simple, so i recommend to begin with this parameters:

• Prompt:

masterpiece, character_name standing menacingly on a dark background with lightnings behind him, thick black outline

• Negative:

(bad_prompt_version2:0.7), low quality, close up, old, lowres, text, error, (cropped:1.2), worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, black and white, monochrome, censored, (out of frame:1.3)

Euler A, 512x512, CFG 7-8

As soon as you understand for yourself how to get generations with consistent quality you will able to experiment wIth it properly.

...or you just can stop reading this lame stuff and go for punching this "generate" button.
Anyway, feel free and enjoy!

Abstract description.

need portrait keyword.

trained on deliberate v1.

>0.5 for balanced and best result<

should kill multicolored hair at 1 strength.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Many thanks for such a great Widowmaker and Mercy embeddings goes to Master_Corneo

Also i really appreciate luisa_pinguin for great guide about hypernetworks training that helped me beat my fear of training tab.

So, about hypernetwork. It was trained on some Midjourney generations so it makes anime generations slightly more semi-realistic and adds some bloom and blurred background with some soft light.

• Trained on Anything.v3 with clip_skip 2 so it should work with almost any other anime-based model.

• Showcase images was generated with AbyssOrangeMix2_hard and clip_skip 2.

In general this hypernetwork is a child of luck and many tries, so if you know exactly how to train it properly and what all the parameters means, please share your knowledge.

magazine cover,fake cover keyword

don't generates coherency texts, but can helps you with cover design.

384x576 default cover size, 576x384 will generate something like thumbnail.

This hypernetwork just help to gen realistically and filled-text magazines.

This Hypernetwork is only to be used for personal curiosity and to obtain beautiful anime character designs. Please do not sell images made with this.

This was created with Axsens artworks, I do not own any of the rights to the original artworks. This was made 100% for fun and not for monetary gain.

Well now that this is out of the way, this model is really good at giving colorful images with space and magic.

________________________________________________________________________

NEW LORA HERE : https://civitai.com/models/8365/axsens-lora
Way more reliable than this hypernetwork at larger resolution.

________________________________________________________________________

Best used with lower weight, I recommended 0.5 weight max, 0.2-0.3 being the sweet spot I found.

by Axsens add to your prompt.
light rays, light particles, magic really help to get the style out.
Use uncensored, nipples for NSFW.

Also added my version 4 which introduce scales artifacts sometimes but carries more of the style for people who might want to give it a try.

Installation of the hypernet :

Copy the Axsens_v7.pt file into the \stable-diffusion-webui\models\hypernetworks folder.

optical illusion keyword for best result,ex:

optical illusion of a liminal space

1 strength for totally explosion, 0,6 for coherency.

can do cool effects in deforum/animations

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

strength: 0.8 + Don't force too many negatives.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

only for portraits or closeups, far shots will be messy.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Hypernetwork for generating images stylized as waven chibi characters. Actually produces good results even in different proportions of strength.
Trained with danbooru tags.

If you like my work and want to buy me a coffee :')
(updated link to ko-fi)

There two versions "1.0a" and "1.0b", they are parallel results of same dataset. Both will produce good looking images and it's final version of this hypernetwork.

A little tip, use 'turn pale' in negative and clip skip 2.

It's a hypernetwork for vaporwave screenshots. Prompt with "vaporwave screenshot." Add "computer window" for bonus -a e s t h e t i c-.

Trained on Ratnika.

trained on 20 minecraft-ish lush field images

about keywords, green,nature,lush field, mountain, everything related will boost the effect.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Trained on Ratnika.

trained on 50 kawaii/cyberpunk images.

about keywords, if u wanna cyberpunk, you need the key cyberpunk, otherwise will be only kawaii.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Yo dawg I heard you like models, so I trained a model to make models so now you can model while you model.

Trained on Anythingv3 using around 250 pictures of plastic models, the majority made by Kotobukiya.

Trigger word: plasticmodel

Also: mecha_musume

PROTIPS

• I’ve only tested it with Anything v3. Please let me know results from other models.

• You’ll probably need to play with prompt weights to get it to cooperate.

• Increasing the steps helps too.

•Helpful prompts for maximum effect: macro focus, bokeh, photography

• It's not good at non-plain backgrounds

Creates some really wild monster designs based on the iconic bestiary of the Shin Megami Tensei games.

Experimental = All results can be bad.

Recommendation, 0.6,0.7 strength.

tested in protogen photoreal, no need for keywords.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

A hypernetwork I trained on anime/painterly style post-apocalypse and cyberpunk images. I was going for a Lain/Dorohedoro/Texhnolyze kind of feel. This works great on semi-realistic models and anime models.

Recommended tags:

Location/setting: cyberpunk, cityscape, city lights, streets, alleyway, outdoors, slums, apocalypse

Colors: muted, desaturated, brown, gold, silver, gray (this is also great at doing very colorful images as well!)

Artists/style: Your favorite anime artists or anime shows/movies from the 80's/90's ;)

Best range of hypernet strength is 0.5 to 0.7 imo.

Can do NSFW, but no NSFW was included in the training data, so use at your own risk.

yes, this style is too specific to be shared, but maybe u can do something cool with that.

Recommendation, 0.7,0.8 strength.

tested in protogen photoreal, no need for keywords.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Recommendation, 0.7,0.8 strength.

tested in protogen photoreal, no need for keywords.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Recommendation [0.4,0.7 strength], this hypernetwork is actually strong.

tested in protogen photoreal.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

any strength should work [realistic 0 <strength> 1 stylized]

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Recommendation: 0.8 strength and portrait as keyword

tested on ratnika model.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Recommendation: 0.8 strength

sometimes the character will wear a police/suit costume, because of laetitia dataset

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

this version will be more closely to original composition(strenght 0),but also is weak compared to original version.

this version can gen images more closely to original artist:https://civitai.com/models/3672/luisap-please-shad-no

trained using another custom dataset.

-i recommend to use anything v3,berry or animeish related model.-

Hypernetwork strength: you can use 1 without be afraid, but u can try do change to 0,7 or 0,8 if u are getting realistic generations.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Like the title says. Sexy xenomorph.

If you clicked this page, that's on you pal.

Turns generations into inklings.

Tested using ratnika.

for stylized faces.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

needs tarot card, in prompt, like : "tarot card,the dark knight, batman"

default resolution: 384 x 576

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

gens sci-fi/hard surface

Needs the keyword "structure" ex: a cyberpunk structure, a building structure, a turret structure.

trained on protogen 53xphotoreal

this hyper is good for making textures too, see the last example.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

i'm learning to make better hypernetworks, also, please suggest a new hypernetwork one in comments!

needs the key cyberpunk, but also works without.

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

works fine in 0,8 strength, will make wings more strong in 1, and will try to hide them in 0,5.

tested in berry mix at 512x512

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

This is a hypernetwork trained mostly on red rope shibari bondage for anime girls. Has very good results for anime, with some weighting you can get photorealistic rope but I wouldn't recommend it.

To make the hypernetwork work you have to include "Bondage_Shibari_V3" as a prompt.

Prompts that work well:

• Woman in bondage

• leg up

• crotch rope

• shibari

• red rope

• arms behind back

• ahaego

• bound breasts

• box tie

• frogtie

Can do gags, ball gags, cloth gags, and blindfolds, but it's not the strongest.

Can also specify position, including:

• On back

• laying on back

• laying on stomach

• on stomach

• standing

• from front

• from behind

• from side

• front/back/side shot

All images were tagged NSFW, but only some were tagged "nude, naked" so if you want something like lingerie bondage or shibari over closthes, include "nude, naked" as a negative prompt and "shibari over clothes" as a positive prompt.

(Third example image took a lot of img2img to make)

note from fuggy, the original poster: I have left Civitai due to corrupt and authoritarian moderation and will not be uploading here anymore, if you want new models, update, etc. Please go here, I'm sorry for the inconvenience

note from Civitai: Original images posted for the model were reported multiple times for depicting children. Images were removed and the user was prompted to upload images that were age-appropriate, as the model is to showcase a piss fetish, changing the age of the individuals in the images shouldn't be an issue.

Hypernetwork for generating images of Negev from Girls Frontline, created 48 images from Gelbooru, 2000 steps, flipped images enabled, and made with Anything 3.0 in mind. I hope it's of use :3

Trained on vanilla SD1.5 but works better with art-focused checkpoints like Elldreth's and others from this website. Training dataset contained game portraits, key art, and some cherrypicked fanart.

How to consistently get very stylized results:

• Use Elldreth's lucid mix

• Use "[{subject}:abstract background:0.5], expressionism, digital art, colorful background" prompt, replace {subject} with anything

• Parameters: 20 steps, CFG=2, DPM++ 2M Karras sampler

• Some usual negative prompts

See my comment below for examples of this method

note from fuggy, the original poster: I have left Civitai due to corrupt and authoritarian moderation and will not be uploading here anymore, if you want new models, update, etc. Please go here, I'm sorry for the inconvenience

note from Civitai: Original images posted for the model were reported multiple times for depicting children. Images were removed and the user was prompted to upload images that were age-appropriate, as the model is to showcase a piss fetish, changing the age of the individuals in the images shouldn't be an issue.

Fine tuned hyper network that can create Pipkin Pippa! This is my first time creating a hyper network, but I am happy with the results. I hope this model is useful or enjoyable :3

I created this model with 88~ images scraped from Gelbooru, and use Anything 3.0. I used this guide thanks to Aitrepreneur I would have been too stupid to figure it out otherwise.

things gonna be glowing and stylized

trained on ratnika, 512x512

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

Illumination attention hypernetwork

strenght: 0.6/0.8

if u wanna buy me a kofi :3 https://ko-fi.com/luisapinguin

This hypernetwork turns all girls into dick girls. It was trained on Anything-V3.0, but can also be used with other anime style models with varying degrees of success.

Feb 11 2023: This hypernetwork has been upgraded to Futa frontal 2 (LoRA), which creates higher quality 2.5D images. I recommend you download that LoRA instead of this hypernetwork.

Use instructions:

For best results use with models that focus on anime and use Danbooru tags. Good models that I've found to use it with are: Anything-V3.0 / 4.5 (4.5 for 2.5D, 3.0 for 2D), AbyssOrangeMix2, SimpMaker 3k3, and a mix of Anything + HassanBlend.

It performs poorly with models that have a photorealistic focus, such as: f222, SXD, HassanBlend, URPM, SD1.5.

Recommended hypernetwork strength: 0.8

The training data was 512x640 resolution so expect the best results when using that resolution, but it still works just fine at 512x512.

For prompting you need to include Danbooru tags since the hypernetwork was trained with WD14/Danbooru captions. Unlike Anything-V3.0, leaving the prompt text box empty will give terrible results. At minimum you need to prompt for "1girl, futanari, penis", and the more specific you get with your prompt the better the results will be.

Embeddings that were created with photorealistic people also work well as long as they aren't overtrained.

You will need to use a VAE in order to not get desaturated images.

See below for more example prompts.

Training info:

The hypernetwork was trained on 186 high quality anime style images captioned with WD14/Danbooru tags at 512x640 resolution. In order to reduce the horror that is multiple penises growing out in multiple directions, the types of images were restricted to mostly frontal images. This helps make the penis generation more consistent, but reduces the flexibility in poses you can ask it to do. Very large penises/boobs were filtered out to give the generated images more realistic sized anatomy.

Install:

For Automatic1111, install like any other hypernetwork by putting the .pt file in the stable-diffusion-webui/models/hypernetworks folder. Make sure to enable the hypernetwork in the Settings tab, and set the hypernetwork strength to 0.8.

Example prompts:

Many of the preview images used these prompts. For exact generation details, click the white ( i ) button on the bottom right of each preview image. All images were made at 512x640 resolution and upscaled by 2x with hires fix turned on.

Prompts:

• 1girl, cute thin futanari, (huge long penis:1.2), erection, testicles, waving at camera, smiling, head tilted, sitting on a wooden park bench, crop top, exposed navel, (legs spread:1.1), collarbone, ponytail, close-up penis, blurry background, relaxed, 8k, high detail

• 1girl, futanari, (penis:1.2), erection, futanari masturbation, on a bed in a bedroom, sunlight, soft lighting, blushing, grin, 8k, high detail, chiaroscuro

• 1girl, futanari, (penis:1.1), erection, lifting skirt, outdoors, sunlight, soft lighting, blushing, 8k, high detail, chiaroscuro

• 1girl, futanari, (penis:1.1), erection, nude, cum everywhere, projectile cum, futanari masturbation, trembling, breasts, mouth open, blushing, 8k, high detail, chiaroscuro

• 1girl, futanari, (penis:1.1), erection, nude, cum, ejaculation, futanari masturbation, breasts, nipples, mouth open, blushing, 8k, high detail, realistic

• 1girl, futanari, (penis:1.1), erection, nude, cum, ejaculation, futanari masturbation, breasts, nipples, mouth open, beach, ocean, sky, sunlight, 8k, high detail, realistic

• 1girl, futanari, (penis:1.1), erection, bikini, bikini aside, mouth open, beach, ocean, sky, sunlight, 8k, high detail, realistic

• 1girl, futanari, (penis:1.1), erection, on a city street, wearing a sweater, lifting up skirt, embarrassed expression, 8k, high detail

• 1girl, futanari, (penis:1.1), flaccid, on a crowded city sidewalk, wearing a sweater, lifting up skirt, embarrassed expression, 8k, high detail

• 1girl, futanari, (penis:1), (flaccid penis:1.2), foreskin, outdoors, lifting up skirt, view from below, embarrassed expression, 8k, high detail

• 1girl, futanari, (big veiny penis:1.2), erection, nude, cum dripping, big breasts, red ponytail, cyberpunk city

• 1girl, futanari, (penis:1), sitting, legs spread, jean shorts, crop top, nipples poking through shirt, 8k, high detail

• 1girl, futanari, (penis:1), Asuna, sitting, legs spread, nipples poking through shirt, 8k, high detail

Negative prompt: lowres, low res, text, error, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, text, username, blurry, robot eyes, out of focus, censored, out of frame, b&w, mask, selfie

Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Size: 512x640, Model hash: f0d37826, Model: Anything-V3.0-pruned-fp16, Hypernet: futa33, Hypernet hash: daff235c, Hypernet strength: 0.8, Denoising strength: 0.4, Hires upscale: 2, Hires upscaler: None

trained on ratnika, 512x512

trained on ratnika, recommended with 0.5/0.7 strength

trained on ratinika, 512x512

This model was trained on 21 images of actress Ana de Armas. It was trained in sd1.5, so works best in photorealistic models, and also works best in close shots. For longer distances, upscaling and inpainting seem to work okay. One outift in particular in the training data can overpower the images, so if you notice some metallic or rainbow vertical banding, lower the hypernetwork strength. I get results that I like around .7 in Protogen, though most of my samples were at full strength in an older model of mine

Edit: Added a few Astronaut Anas made with Protogen x5.3 to fill out the 20 image limit.

needs "movie sequence" or "((movie sequence))" in prompt, and maybe 0.5/0.7 strength

i'ts not stable, can fail in mostly times.

trained on ratnika, the default is 512x512

no trigger needed, automatically do pixel art.

trained on ratnika, but works in every stylized/cartoony model

[realistic model can be messy and bad as hell]

Include "Ahegao tongue drool" in your prompts. I find that adding "extra tongue" as a negative prompt also helps a bit on reducing the occurrence of monstrosities with double mouths and tongues.

Just started fiddling with hypernetworks after many Dreambooth models. I think it's quite difficult and more time consuming than DB training (for now) as there's a lack of coherence in the tutorials I've read/watched.

I tried using hypernetwork on faces and it sucked, so I'll keep using Dreambooth on those. But I was curious if hypernetworks can be used on facial expressions instead of subjects. Turns out it can, though v1 still feels half-baked. I still need to experiment on different layer structures and learning rates.

0.7 strength is the best one

This hypernetwork was trained on Anything-V3.0-pruned-fp32 [1a7df6b8] at a res of 512x768. The training dataset included primarily NSFW images and so the model leans heavily toward that type of content. All of the versions were trained at their respective listed training steps and a learning rate of 0.00005. I personally like the results from v1-16000 the best.

For best results please use with the bad-artist embedding.

Recommended keywords: meru, merunyaa, black hair, demon girl, demon tail, freckles, horns, mole, pointy ears, red skin, orange eyes

Recomended negative keywords: multiple moles, multiple horns, glasses, mole on breast, multiple tails

Please download the attached keyword list for additional training keywords.

I'd love to see what people who are better at prompting than me can produce with this! Please post your beautiful creations! Post a comment if you have a recommendation for a character I should train next.

force strength to 0.5,0.7 or similiar.

needs venom in prompt

0.5 Strenght! needs keyword sonic and maybe, with brackets in some actions ;

ex: a portrait of sonic,((oil painting)),((by leonardo da vinci))

you can force more the diagram with [diagram] and newspaper with [explanation] prompt.
good for game asset!

adds detail slightly,force the prompt attention,and smooth without destroying the original composition [0.5 strength]

** trained in photo realistic compositions

A hyper network trained on marble statues intended to give the statue look to humanoid characters. Works well with both realistic and anime models.

If you have a hard time getting it to work try using statue and made of marble as activation triggers somewhere in your prompt

this version is more "stronger" and "artistic.
Also, emotions can affect the effect, happy, sad, angry.

tips: emotional prompt can change the effect [ sad,angry, happy ]

Lada 2107 car hypernetwork, for your eastern europe memes

As the name implies, this hypernet is one made using M-da S-tarou's art. It was made using mostly some of their newer stuff, though I did add older ones in as well.

The Neo in the name is because this isn't the first M-da S-tarou embed I've made, but (I think) it's the one that works best. I typically run it at low values (.25 or .5), as the eyes tend to go wonky when it's too strong.

The example images are largely genned on my own model mixes, one of which (Anything and Everything Mix Ver 1.5) can be found here.

This hypernetwork was trained on Anything-V3.0-pruned-fp32 [1a7df6b8] at a res of 512x768 and it's dataset included NSFW images. All of the versions were trained at their respective listed training steps and a learning rate of 0.00005. I personally like the results from v1-5500 the best.

For best results please use with the bad-artist embedding.

Recommended keywords: princess ruto, ruto, blue skin, zora, earrings, fin, fish girl, fin ears, purple eyes

Recomended negative keywords: multiple earrings, 3d

Please notice the attached keyword list for additional training keywords.

488 x 576

needs keyword: PlayStation 2 cover box art

change your strength variety! .6, .7, .8

Trained on 448 x 576,recommendation >>448 x 640<<

>>1 strength = yugioh.<<
>>0,5 strength = tarot/mtg<<

This is a test hyperrnetwork to see what WD 1.4 (and models based on SD 2.1) is capable of.

This hypernetwork is heavily inspired of Black Souls and was trained on some Sushi Yuusha Toro's artworks (Black Souls's author/artist/dev) and somes Black Souls 2 CGs.
It was trained on the first epoch of Waifu Diffusion 1.4 (Anime) [8cbce067] at 512x512 with a learning rate of 5e-5:149,5e-6:10000,5e-7:20000.

In some cases, adding "sketch" or/and "jaggy lines" to the prompt can improve result.

I recommand to not use WD 1.4's quality modifiers (best quality, normal quality,etc...) with that hypernetwork especially as a negative prompt (worst quality,lowres,etc...)

This hypernetwork was trained on Anything-V3.0-pruned-fp32 [1a7df6b8] at a res of 512x768 and it's dataset included primarily NSFW images. Training data was manually tagged. All of the versions were trained at their respective listed training steps and a learning rate of 0.00005. I personally like the results from v1-6000 the best.

Recommended keywords: princess ruto, afrobull, fish girl, fin ears, purple eyes

Recomended negative keywords: multiple earrings, tail, open mouth, blush

Please notice the attached keyword list for additional training keywords.

Recommendation: 0.5 and 0.7 strength!
this negative prompt:
lowres, bad anatomy, bad hands, bar censor, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, out of focus, censorship, censored, ugly, old, deformed, amateur drawing, odd, fat, morphing, black and white, anime, ((simple background)), (((depth of field)))

This is new of my models from: https://rentry.co/PinguinStableCollection

it was trained with a new method [ tanh / xaviernormal ] so it's why the file is 5 mb for this hypernetwork.

works better on semi realistic models, like ratnikaMix, berry, etc.

This hypernetwork was trained on Anything-V3.0-pruned-fp32 [1a7df6b8] at a res of 512x768 and it's dataset included NSFW images. Additional checkpoints will be uploaded soon at various training steps. Edit: nvm apparently I won't be uploading additional checkpoints because the site won't let me (devs pls fix, thx). All of the versions were trained at their respective listed training steps and a learning rate of 0.00005. I personally like the results from v2-6000 the best.

Recommended keywords: portrait of midna, midna, 1girl, blue skin, forehead jewel, front ponytail, grey skin, orange hair, yellow sclera, pelvic curtain, hood, cloak, hooded cloak, hooded cape, cape, hood up, anklet, tattoo, leg tattoo, body writing, orange eyes, yellow eyes, purple lips, two-tone skin, neon trim

Recomended negative keywords: unibrow

This Hypernetwork uses the same dataset as my Aesthetic Gradient "djzCyberSocietyV0" (SD1.4/1.5) with additional textual metadata applied.

djzCyberSocietyV21 - trained for 30,000 steps

Recommended for use with the Textual Inversion Embeddings and/or Custom/Fine Tuned Models SD2.0/2.1

All examples are using a fine tuned 2.0 model combined with my Textual Inversion "djzCyberSamuraiV2" as they are designed to work together, however will operate with other 2.1 models/embeddings.

Hypernet designed to create sexy biological monstrosities and living orgies.

Edward Hopper was an American realist painter and printmaker known for his iconic and evocative images of American life, landscapes, and cityscapes. He is one of the most influential and widely recognized artists of the 20th century, with a career that spanned more than five decades.

Hopper's work is characterized by his keen eye for detail and his ability to capture the mood and atmosphere of his subjects. He is best known for his paintings of everyday life in America, often depicting lonely, isolated figures in empty rooms or on deserted streets.

___________

About this Model:

This model was trained using SD 1.5, and good results not guaranteed on other models.

It seems like using DPM++ 2S a Karras produces the best results, and the images shown above were done with a hypernetwork strength of 0.75.

How to Install:

Place in ./stable-diffusion-webui/models/hypernetworks/ folder. Select the hypernetwork from the Settings tab under "Stable Diffusion."

A Hypernetwork made of Kafu from Cevio AI

Shared by Maru on Unstable Diffusion

Haru from Cevio AI

Shared by Maru on Unstable Diffusion

Rime from Cevio AI

Shared by Maru on Unstable Diffusion

IMPORTANT INSTRUCTIONS!

This model was trained on SD base 1.5 version BUT It does also work for 1.4 as they both share the same Clip encoder.

Install instructions.

Simply place the invisible.pt file inside the \stable-diffusion-webui\models\hypernetworks folder. Load the model inside the Automatic1111 interface under settings hypernetwork.

Use instructions.

Use 1.0 hypernetwork strength for best results.

Use DPM++ SDE Karras sampler with 15 steps and CFG of 4.0.

Make sure and always include the word invisible somewhere in the prompt. For people always preface the subject with invisible, example "invisible man walking", "invisible girl playing in the backyard", etc...

VERY IMPORTANT! Always describe the background in some detail or you WILL get a very generic boring background.. So for example DON'T just say "an old invisible man". DO say "an old invisible man inside a rustic hut".

For some fun, use the 1.5 inpaint model with this hypernetwork on your photos. Simply prompt invisible and use the inpaint tool to inpaint any exposed skin and watch the magic happen.

It's a hypernetwork of Wappah.

IMPORTANT INSTRUCTIONS!

This model was trained on SD base 1.5 version BUT It does also work for 1.4 as they both share the same Clip encoder.

Install instructions.

Simply place the chimp.pt file inside the \stable-diffusion-webui\models\hypernetworks folder. Load the model inside the Automatic1111 interface under settings hypernetwork.

Use instructions.

Use between 0.55-1.0 hypernetwork strength, more strength will give a more real chimpl look while .55 gives a more human form chimp look. I find .7 works well enough.

Use DPM++ SDE Karras sampler with 15 steps and CFG of 6.0.

Make sure and always include the word chimp somewhere in the prompt. For people always preface the subject with chimp, example "chimp man walking", "chimp girl playing in the backyard", etc...

VERY IMPORTANT! Always describe the background in some detail or you WILL get a very generic boring background.. So for example DON'T just say "an old chimp man". DO say "an old chimp man inside a rustic hut".

Some fun info. People have been sleeping on hypernetworks and I plan to change that. Hopefully the flexibility of this hypernetwok will show everyone their true potential. Because this model is a hypernetwork it can be used in conjunction with ANY model based on the 1.4 CLIP architecture. That means this model will work on any custom 1.4 or 1.5 model, like the modern disney model, or classic disney, etc… for example, let's say you want to load classic disney as base. Well simply load the classic disney model, make sure and preface every prompt with classic disney. As per instructions of the model. Then follow up with my “chimp” tag as instructed once you have loaded the hypernetwork. So the prompt should look something like this “classic disney. chimp girl playing in the backyard.” Make sure and adjust the hypernetwork strength to .5 for a more cartoon look or .7 for a realistic chimp look. Have fun folks!

https://huggingface.co/doctorderp/water_elemental/tree/main

IMPORTANT INSTRUCTIONS!!!

This model was trained on SD base 1.5 version BUT It does also work for 1.4 as they both share the same Clip encoder.

Install instructions.

Simply place the water elemental.pt file inside the \stable-diffusion-webui\models\hypernetworks folder. Load the model inside the Automatic1111 interface under settings hypernetwork.

Use instructions.

Use between 0.55-1.0 hypernetwork strength, more strength will give a more transparent elemental look but starts to overfit. I find .7 works well enough.

Use DPM++ SDE Karras sampler with 15 steps and CFG of 7.0.

Make sure and always include the word water elemental somewhere in the prompt. For people always preface the subject with water elemental, example "water elemental man walking", "water elemental girl playing in the backyard", etc...

VERY IMPORTANT! Always describe the background in some detail or you WILL get a very generic boring background.. So for example DON'T just say "an old water elemental man". DO say "an old water elemental man inside a rustic hut".

Some fun info. People have been sleeping on hypernetworks and I plan to change that. Hopefully the flexibility of this hypernetwok will show everyone their true potential. Because this model is a hypernetwork it can be used in conjunction with ANY model based on the 1.4 CLIP architecture. That means this model will work on any custom 1.4 or 1.5 model, like the modern disney model, or classic disney, etc… for example, let's say you want to load classic disney as base. Well simply load the classic disney model, make sure and preface every prompt with classic disney. As per instructions of the model. Then follow up with my “water elemental” tag as instructed once you have loaded the hypernetwork. So the prompt should look something like this “classic disney. water elemental girl playing in the backyard.” Have fun folks!

https://huggingface.co/doctorderp/muppet/tree/main

IMPORTANT INSTRUCTIONS!!!

This model was trained on SD base 1.5 version BUT It does also work for 1.4 as they both share the same Clip encoder.

Install instructions.

Simply place the muppet.pt file inside the \stable-diffusion-webui\models\hypernetworks folder. Load the model inside the Automatic1111 interface under settings hypernetwork.

Use instructions.

Make sure and use 0.7 hypernetwork strength for best results. Using strength of 1.0 will lead to background distortion.

Use DPM++ SDE Karras sampler with 15 steps and CFG of 7.0.

Make sure and always include the word muppet somewhere in the prompt. For people always preface the subject with muppet, example "muppet man walking", "muppet girl playing in the backyard", etc... For animals put the word muppet at the end of the prompt, example "a cow in a green field muppet". With longer prompts you will want to put the word muppet closer to the beginning of the prompt for better effect. Sometimes putting the word muppet smack dab in the middle of the prompt is best, play around with it for optimal effect for your prompt.

VERY IMPORTANT! Always describe the background in some detail or you WILL get artifacts when generating an image that has no info on the background. So for example DON'T just say "an old muppet man". DO say "an old muppet man inside a rustic hut".

Some fun info. People have been sleeping on hypernetworks and I plan to change that. Hopefully the flexibility of this hypernetwok will show everyone their true potential. Because this model is a hypernetwork it can be used in conjunction with ANY model based on the 1.4 CLIP architecture. That means this model will work on any custom 1.4 or 1.5 model, like the modern disney model, or classic disney, etc… for example, let's say you want to load classic disney as base. Well simply load the classic disney model, make sure and preface every prompt with classic disney. As per instructions of the model. Then follow up with my “muppet” tag as instructed once you have loaded the hypernetwork. So the prompt should look something like this “classic disney. muppet girl playing in the backyard.” Make sure and adjust the hypernetwork strength to .5 for a more cartoon look or .7 for a realistic muppet look. Also you can muppify your own face! Simply prompt muppet in image to image at .6-.7 denoise strength and watch out, HAHA! Have fun folks!

This hypernet is capable of 2 different image effects:

1. Dispersion effect (remember the cool BLIP effect from Avengers end game?)

2. Ink Smudge effect

Dispersion Effect

Get Thanos to work for you! Let people Blip!

Use "in the style of dispersion" in your prompt to trigger the effect

A sample prompt is:

photo of beautiful woman standing, 8k,4k,highres,masterpiece,in the style of dispersion

I have not tried it on non-human generations. It didn't work as well in a few generations involving a cat.

Ink Smudge Effect

During testing, I realized that it can also generate what is called "Ink Smudge Effect"

See the samples, png info retained. Just add below to the prompts:

highly detailed, bright colored graffiti, in the style of dispersion, soft focus

Installation

Simply copy the download the file into your hypernetworks folder, select it while generating images and you are done.

Georges Seurat was a French post-impressionist artist who was famous for pioneering the distinctive painting technique known as pointillism. His most famous work, A Sunday Afternoon on the Island of La Grande Jatte, featuring many people sitting and enjoying a day at the park, is found at the Art Institute of Chicago.

This model is particularly well-suited for Seurat's favorite subjects of boats, beaches, seasides, sailing, and so on, and it handles people pretty well too.

This is a 1-1.5-1.5-1 Hypernetwork trained on SD v1.5-pruned-emaonly.ckpt.

Hypernetwork based on a selection of high quality images of Henri Rousseau's work. Rousseau was a French post-impressionist painter known for his vivid jungle scenes.

While this doesn't mirror his style precisely, which would be hard for any neural network to do, it does create some really stylish looking foliage. 0.75 strength seems to yield the best results.

This is a SD v1.5 model trained on v1-5-pruned-emaonly.ckpt.

"Hi mister farmer, I need to be milked now. Would you help me? Then I promise I'll milk you after. Pleeeease?"

This is not a model, but a hypernetwork. This means that you can use it within Automatic1111 Web UI to influence any other model that you have loaded.

The prompts you enter which exist in this model will affect your generation.

The example picture was done using HassansBlend1.4 (4cf12f5d), at 120 steps (because more ears were added at later steps), cfg 12, seed 4274011886 with Euler a.

I used the prompt:

full body photo of sexy (young woman), red hair, freckles, light blue eyes, age 30, (cow ears:1.3), (on knees), ((kneeling)), ((milk bucket)), field,  spread legs, spread pussy, barefoot, naked, completely naked,smooth soft skin, (huge breasts), sagging breasts, large areolae, perky nipples, big dreamy eyes, photograph, photorealistic

And the negative prompt:

horns, cow horns. dildo, penis, 3d, cartoon, (disfigured), (bad art), (deformed), (poorly drawn), (extra limbs), strange colours, blurry, boring, sketch, lacklustre, repetitive, cropped, hands, greyscale, text, watermark

The negative prompts may not matter much, but I inserted the horns, dildo and penises because this hypernetwork is tested on many models and some would insert dildos etc, which I did not want in my example picture.

The hypernetwork should be able to do ahegao and lactation. And it knows the Pornhub/OnlyFans performer Octokuro because some of the training images were of her.

It is trained on mainly NSFW images, but should be able to do completely SFW cosplay as well.

The training set is available here if somebody want to help me collect photos for a full model:

https://gofile.io/d/h5LsQb

or

https://workupload.com/file/Aj2YkkEFVtr

A wordlist is here so you know what to enter in your prompt (the underscore is, as usual with the Web UI, not required):

https://workupload.com/file/PByeJaspAHu

Computing time: 13,5 hours on a Geforce 3060 12GB with OC @ 1927 MhZ.

Hello, this is one small experiment i did to prettify the Ugly results i was getting in the 2.1 version.

The images are with and without the Hypernetwork enabled.

It pops up the style the more tags you use.

I trained it with mostly booru tags so you guys know whats up. (Dont expect much)

Look the examples for ideas of the tags.

This is a test for a new hypernetwork of SamDoesArts based on 1.5

Its still in training i leave it here just for testing purposes

So.. here is my first "model"

PLEASE SHARE YOUR GENS! I WANT TO SEE WHAT THIS IS CAPABLE OF!

THANKS!!

Its a fully custom hypernetwork trained on a bunch of Images generated with the help of JH model

You can try less step count versions for subtle results

It should work in a bunch of different models.

I've tested it on base SD1.5, InkPunk, and Booru based models with mixed to good results.

Originally posted to HuggingFace by Hamaru and shared on Reddit.

Hypernetwork trained on some Midjourney V4 portraits.

Euler a and DPM++ samplers work best. CFG scale at 7 and low step count (<50) work well.

Prompts should include words like "portrait", "octane render" and "highly detailed" for best results. Avoid using face restoration like GFPGAN or CodeFormer if possible