# Community Guide For Prompting

Version 1.2 - Last Updated June 02nd, 2024 - Last Backed Up June 02nd, 2024

Updated and Upkept by Lolime Schizo. (thanks Lolime Schizo, love from Doc Anon)

# Intro

Useful Information
How install, use & quick guide
How To Merge Models (Model Archive)
Stable Diffusion Goldmine
Models to experiment with
NSFW Prompting FAQ
Better Anime Hands

Move the .safetensors and .vae.pt files to the stable-diffusion-webui\models\Stable-diffusion folder.
You shouldn’t be using .ckpt files anymore, look for models with .safetensors, or make the .safetensor yourself.
This will automatically use the appropriate model with the appropriate file, but it can be manually selected if you put the VAE file on stable-diffusion-webui/models/VAE instead.
You’ll generally want to use a GPU to generate images. While Stable Diffusion can use CPU for image generation, it is very slow and taxing on your system, you’ll not be able to make images larger than the standard 512x512, maybe not even that.
All modern models you can find will more or less make the same stuff when given the same prompts, so there’s no longer a need for model specific prompts, not to mention the fact that AI proompters have gotten isolated, so there’s no longer a defining group-prompting aesthetic anymore.
To save time, you can make a “Style” with all the currently set prompts, both positive and negative. In order to use a style, just select it and press the paste icon. Try doing this with the negative prompts or baseline prompts for stuff you plan on prompting often, like a chuuba design.

# FAQ

# What model should I use?

Whatever your heart desires, there is no “should” here. But if you have to ask, the OrangeMix family of models, AutismMix and anima-pencil look really good. You could also use the leaked NAI model to check the old school proompting, but remember that shit is extremely outdated.

# My prompts come out faded!

You are not using a VAE file to colour correct the prompts. See Here for comparisons between different VAE files. Once you’ve selected your VAE file, download and move to stable-diffusion-webui/models/VAE (NOT models/stable-diffusion).
Once in place, open Stable Diffusion, go to Settings, look for SD VAE, reload if necessary, then change from Auto to your chosen VAE. VAE files MUST NOT be in the *.safetensors format, as the webui doesn’t recognize them yet.
NovelAI/Anything V3 VAE - Stable Diffusion 1.5 MSE VAE - Stable Diffusion 1.5 EMA VAE
Trinart Characters VAE - Waifu Diffusion kl-f8 anime VAE - Waifu Diffusion kl-f8 anime2 VAE
Don’t bother downloading the Anything V3 VAE file, it’s literally the Novel AI VAE file but renamed.

# **What’s the difference between .ckpt and .safetensors?**

They’re essentially the same, but safetensors are safer, crash less often and load faster.
By this point (June 2024), *.ckpt files have been effectively abandoned, as safetensors are just better in every way. Should you find a *.ckpt file, you can convert it into a safetensor by “merging” it with another model and setting the weights to 0 for the other model (thus only re-encoding the weights into a safetensor.
You must have an up to date webui to use them properly.

# How do I fix the fingers/hands?

Without using specialized networks, your best bet is positive prompting “(Accurate five-fingered hands:1.4)” and negative prompting “Bad Hands”
Try using a model that fixes broken hands, or making one yourself by merging a good hands model with the model you want to use. Using a model specialized for that kind of work (AbyssOrange, AutismMix etc) helps too.
Try using the Better Hands Lora to improve your hands, if you need a specific model with better hands.

# What do these sliders/buttons do?

I am no expert on this. While the info should be accurate, there might be a couple mistakes I am unaware of.
Take the technicalities with a grain of salt (the functionality should otherwise be correct).

# Models

The Stable Diffusion neural networks need biases and weights in order to ‘carve’ the image you want from random noise; different models have been trained on different samples to make images and thus have ‘learned’ how to make images differently.

A model trained exclusively on anime females would find it impossible to make a male character, since that dataset was never included and as such the model would not ‘know’ what a male character was. Specialized models are way more precise than generalist models, but are also way more limited in output. I.E: You’d be hard pressed to make a decent human with a furry model, but said furry model would be exceptionally good at making furry characters.

# Sampling Steps

Simply how many iterations of the eldritch chaos magic the system goes through. More steps generally produce better images, though there’s a decreasing gain per each extra step.
20 steps is the usual ‘minimum’ to get a decent-ish picture, this is best used to find seeds that output good images, then bump the steps to a higher number on those seeds.
The “relevance” of each step is directly affected by the ETA Noise parameter (see below).
30-50 steps is a good compromise between quality and time. You can go higher with 100, even 150 steps, but you’re not getting much improvement per extra step at that point.

# Sampling Method

Stable Diffusion starts with random noise and ‘carves’ the actual image from the noise, this is controlled by a neural network that decides how much noise is removed each step and how is it removed. The difficulty and efficiency of these equations vary between sampling method, trading speed for image quality/accuracy or vice versa, though some of these methods are strictly better than others.
Euler and Euler a are the simplest method and as such, some of the fastest. Heun is an ‘improvement’ on Euler designed for accuracy, taking almost exactly 2x as long as Euler, but generally producing images more accurate to the prompt.
The a in Euler a stands for Ancestral. Ancestral-type samplers are more functionally related to each other than their non-Ancestral counterpart, adding noise in specific ways, diverging more heavily per step.
Euler a, Euler, LMS, DPM2M, DPM fast, LMS Karras, DPM 2M Karras, DDIM and PLMS take more or less the same amount of time, give or take a couple seconds.
Heun, DPM Adaptive, DPM2, DPM2 A, DPM++ 2S a, DPM++ SDE, DPM2 Karras, DPM2 a Karras, DPM++ 2S a Karras and DPM++ SDE Karras take around 2x long.

Some suggest farming for images with Euler or Euler a then remaking the seeds you liked with Heun instead, as this saves on the increased wait time with Heun proompting, others suggest proompting 128x128 images when seedfarming for increased speeds, but it all depends on the user and the intended proompt.
If you don’t know what to do just use Euler or Euler a.

# Seed

The first value the Sampling Method selected (see above) uses to generate an image. Identical prompts with identical configurations and identical seeds will make the same image. Different configurations but identical seeds will make similar images. Seeds are a hard boundary to your prompt, some seeds will block some of your prompt instructions, like making characters distinctly fair skinned, despite the presence of a ‘brown skin’ tag. Even when reinforced with (()) tags, the image would remain fair skinned.
It is generally advised that you generate tons of computationally cheap images (I.E: Low step count, faster sampling methods and smaller sizes) to find good seeds, then use these seeds to remake the image with better quality, size, etc.
If you don’t wanna bother with that, just set it to random (-1).

# CFG Scale

The CFG scale adjusts how much the image looks closer to the prompt and/or input image.
A low CFG will be more “creative” and drift away from the prompt, but with higher overall quality, while a high CFG will match the prompt more strictly, to the detriment of the image, increasing noise. In general, all high CFG images look trippy though this is not always the case
See here for examples. Notice how the low CFG added a lot of extra detail, while the high CFG hyper focused.
You want somewhere between 7-12, though don’t feel afraid to experiment.

# Clip Skip

Clips are the encoding steps the system uses to turn text into data to eventually make an image. For example, you give the system a prompt, which gets turned into sample 0, gets processed and turned into 1, processed again and turned into 2, then once more and turned into 3, which gets spit out as the processed output.
Clip skip refers to just how much of this process (and thus waiting time) you skip. ‘Clip Skip = 2’ means that, in this example, you’d use output 1, skipping outputs 2 and 3.
The earlier clips are usually less accurate, though models trained on Clip Skip can see some improvements on clip skipping.
NAI, Any V3, HD and AbyssOrange use Clip Skip, so just keep it at 1 or 2. Experiment if you want.

# ETA Noise

Controls just how much random noise is added per sampling step. Lower ETAs are faster and more accurate, but also more deterministic; these are better for low step counts (20-60). High ETAs are slower and more random, but allow for more “creative” images, these are better for high step counts (100+).
If you don’t know what you’re doing just keep it at the 31337 value.

# Hypernetworks

Subnetworks that add extra weights at specific points on the active model, affecting the image style. These are independent of the model and will bias the image towards whatever the hypernetwork was trained on. A hypernetwork trained on Sakimichan’s pictures will, when used alongside a model, make Sakimichan-style images. This lets you rework the model’s output without needing to retrain it completely.
Considered extremely obsolete after LoRas became a thing, though don’t assume that obsolete means useless, it just means that a similarly made lora would be more efficient.
If you have a hypernetwork, add sd_model_checkpoint, sd_hypernetwork, sd_hypernetwork_strength to your quick settings list under User Interface in Webui Settings to have the Hypernetwork and Hypernetwork Strength next to your model dropdown.
Hypernetworks go in: stable-diffusion-webui\models\hypernetworks
Install the Multiple Hypernetworks plugin if you want/need to use more than one, consider disabling the sd_hypernetwork, sd_hypernetwork_strength quick settings if you do.

# Embeds

A small collection of curated images turned into ‘custom tags’. Embeds can be trained on concepts or designs, allowing extreme control on how the model designs the image. Embeds overwrite how the image is designed towards its specific concept.
Hard to prompt concepts, (like lactation through clothes) can be turned into easily understood (by SD) embeds, which let you very simply, and without any extra specifics, prompt the embedded idea.
Prompting a specific character sometimes does not give you the character, like prompting Usada Pekora giving you a random anime girl half the time. Using a Pekora Embed (Trained on pekora images) will guarantee that any design used with that embed will output a pekora or pekora adjacent character.
You usually need around 150 images to make a proper embed.
Embeddings go in: stable-diffusion-webui/embeddings
To use an embed, simply prompt its exact file name (sans extension) as you would any other tag. You can find some embed examples Here.

# LoRas

Essentially Embeds that manipulate the model like a Hypernetwork would. In practice they are the same, but LoRas are WAY faster, more efficient, less prone to crashing, train FAST, are light (anywhere from 8 to 200mb), and can train all different aspects of a dataset simultaneously (such as both art style and character)
You can train a LoRa on just 6GB of VRAM, or use google colab. See here and here for more details, respectively.
You can find /h/’s lora collection here, here and here. You can find /vt/’s lora collection here, here and here. (Pepe Silvia LoRa is here). FlaVR Loras (Thanks MehloAnon) are here.
LoRas go in: stable-diffusion-webui/Lora

# Xformers

An alternate pathway to generate an image, generally being faster than not using xformers on older. However, this alternate pathway is only available for NVidia GPUs, and will require installing the NVIdia CUDA Toolkit, which facilitates parallel computing on GPUs via GPU acceleration. On weaker GPUs, Xformers is a guaranteed increase in performance, averaging out around 20% increase, but high end GPU users (series 4000 and up) have reported no noticeable increase in performance, or even a slight decrease. In short, xformers is essentially a pre-calculated set of instructions that let the model skip those instructions when making an image, which can help lower end proompters.
Previously, Xformers was non-deterministic, meaning any 2 prompts with the exact same settings would be subject to random system noise and would NOT be equal. Current Xformers IS deterministic.

# ControlNet

See here for more details.

# Experimental stuff

Stuff that’s not for beginners, requires more detailed knowledge or isn’t perfectly stable/usable goes here.

# Better Hands Lora

A lora that does the same as the defunct Realistic Hands hypernet, but better, and doesn’t use the deprecated Hypernet system, it fixes a lot of the finger spaghetti that’s usual for AI generation..
Add “good hands, best hands” to your prompt.

# Bad Image Negative Embed

Trained on thousands of bad proompts made specifically for this, it saves on the dozen negative tags replacing them with a single negative embed, giving you extra tag space for specifics. Trained for ElysiumV2, but works well with EerieOrangeMix, of the AbyssOrange family of models, as that one is derived from Elysium V2.

# Bad Hands Negative Embed

Trained on proompts with bad hands, gives a slight help against the hand spaghetti menace.
Trained for AnimeIllustDiffusion, but can work with other models. Works better at high CFGs (<=11). Just drop on the embeddings folder and use “badhandv4” as the negative embed.

# Novel AI

Leaked Novel AI model from around October 2022, don’t expect the same level as the paid NAI service, since that one is still getting trained.
This is kept here for referential purposes, do not use NAI for prompting, as it is extremely outclassed in every way. It’s Good for recreating the OG prompts but not much more nowadays.
Fully deprecated as there are better NAI-based and non NAI-based models that keep getting updated.
If you don’t have a good prompt for a model, these tags can work as a baseline to get an idea.

Source

Start your positive prompt with “masterpiece, best quality,” use the NAI negative prompt to reduce the chances for body horror (unless that’s your thing).

Negative Prompt	lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, bad eyes OR lowres, polar lowres, bad anatomy, bad face, bad hands, bad body, bad feet, bad proportions, (bad leg), (more legs), worst quality, low quality, normal quality, gross proportions, blurry, poorly drawn, text, error, missing fingers, missing arms, missing legs, short legs, extra digit, 2girls,
Dr. Citrea Lolime	Positive: 1girl, solo, twintails, medical eyepatch on left eye, petite, black streaked hair, blonde hair, light green hair, multicolored hair, Negative: facemask
Donna Vina D’Merlot	Positive: 1girl, solo, purple hair, long hair, grape hair ornament, french braid, blunt bangs, purple eyes, half closed eyes, mole under eye, medium breasts, mole on breast, (((formal))), business suit, white shirt, black jacket, pencil skirt, cleavage
Minna Vanilla	1girl, solo, white hair, medium hair, blue hair bow, heart ahoge, blue ribbon, hair ribbon, hair ornament, grey eyes, medium breasts
Rika Paprika	1girl, solo, delinquent, tomboy, dark red hair, short hair with long locks, ahoge, sunglasses, orange eyes, gradient eyes, arm tattoo, shoulder tattoo, (muscular), medium breasts
Sena Bonbon	1girl, solo, tall female, light brown hair, brown skin, short hair, two hair buns, freckles, orange eyes, large breasts, hair ornament
Juwun Sugarbloom	1boy, solo, tall male, otoko no ko, blonde hair, gyaru, long hair, heterochromia, green eye, purple eye, tiger boy, tiger ears, no breasts, fit, crotch bulge, headphones, bulge, oversized clothes,
Nyana Sextyn9 (Original Nyana)	(Positive: masterpiece, best quality, 1girl, solo, dark skinned female, blonde hair, colored inner hair, long hair, cat ears, animal ear fluff, yellow eyes, makeup, small breasts, long nails, headphones, tank top, midriff, black shorts, facing viewer, looking at viewer, standingNegative: flat chest, headphones on neck, more headphones Note: I’m having better results using dark-skinned female, (light skin)
Alunette Mignon	1girl, solo, villain, tall, black hair, blue hair, colored inner hair, low-tied long hair, messy hair, red eyes, eyeshadow, small breasts, white dress, gold trim
Abigill Anemoia	1girl, solo, short female, magenta hair, long hair, long bangs, one eye covered, head fins, blue eyes, small breasts, glasses, anchor hair ornament,
Coconut Blanc	1girl, solo, pale skin, long hair, white hair, white eyes, medium breasts, ((detective)), brown coat, buttons, light brown turtleneck dress
Jubilee Junagold	1girl, solo, bright red hair, very long hair, ((braided ponytail)), green eyes, pointy ears, large breasts, yellow dress, long dress, front slit, cleavage, navel, long sleeves
Nayvell Milkdrop	1girl, solo, orange hair, long hair, white eyes, flat chest, ski goggles, pinafore dress, orange dress, yellow neck ribbon, white shirt Note: Add loli for the original Nayvell concept, or don’t for Nayvell’s current form.
Marie Mehlo	1girl, short woman, thick eyebrows, yellow eyes, blunt bangs, fluffy hair, long hair, curly hair, white hair, small breasts, wide hips, ornate, yellow shirt, corset, gothic lolita, high-waist leather shorts, stockings
Araminta Aminthe	Positive: 1girl, solo, mint, dark green hair, medium hair, ponytail, dragon horns, white eyes, medium breasts, green tail, (dragon tail), underboob cutout, dark green shirt, gold trim, bare shoulders, elbow gloves, midriff, armored skirt, dark green skirt, arms at sides, head-on, standing Negative: jacket, sideboob, pointy ears, wings, side slit, elf
Raimi Ricotta	1girl, solo, mouse ears, mouse tail, pixie cut, parted bangs, dark brown hair, big eyes, rectangular mouth, blue eyes, flat chest, yellow hood, yellow cape, white blouse, puffy short sleeves, brown belts
Cara Malone	1girl, solo, tall female, medium breasts, mature female, pale skin, orange hair, brown hair, dark brown hair, straight hair, long hair, (((hair slicked back))), half-closed eyes, mascara, gray eye, green eye, scar on cheek, muscular, cross necklace, jacket on shoulders, tank top, leather belt, gray pants, emotionless, angry

# Assorted Prompts list

Non-specific tags that were made for one model or another, but with current models being more universal, the specificity is no longer necessary.
All of these were designed for AnythingV3, but will work with the more advanced models.

Dr. Citrea Lolime	Positive: 1girl, solo, petite, minigirl, blonde hair, light green hair, multicolored hair, lemon, lime, black hair, streaked hair, twintails, long hair, yellow eyes, medical eyepatch on left eye, flat chest, labcoat, light blue dress Negative: facemask
Donna Vina D’Merlot	Positive: -1girl, solo, purple hair, long hair, grape hair ornament, french braid, blunt bangs, purple eyes, half closed eyes, mole under eye, medium breasts, mole on breast, (((formal))), business suit, white shirt, black jacket, pencil skirt, cleavage
Minna Vanilla	-1girl, solo, white hair, medium hair, blue hair bow, heart ahoge, blue ribbon, hair ribbon, hair ornament, grey eyes, medium breasts
Rika Paprika	-1girl, solo, delinquent, tomboy, dark red hair, short hair with long locks, ahoge, sunglasses, orange eyes, gradient eyes, arm tattoo, shoulder tattoo, (muscular), medium breasts
Sena Bonbon	1girl, solo, tall female, light brown hair, dark skinned female, short hair, two hair buns, freckles, orange eyes, large breasts, hair ornament, OR 1girl, solo, tall female, dark skin, light brown hair, hair buns, short hair, hair clips, orange eyes, freckles, large breasts, wide hips, skinny
Juwun Sugarbloom	1boy, solo, tan, ((white tiger)), petite, otoko no ko, blonde hair, long hair, messy hair, tiger ears, eyeshadow, glitter, green eyes, purple eyes, heterochromia, headset, rainbow scarf, oversized hoodie, thigh highs, bulge, tiger tail
Nyana Sextyn9 (Original Nyana)	(old nyana) 1girl, solo, cyberpunk, dark skinned female, oppai loli, shortstack, blonde hair, colored inner hair, long hair, cat ears, animal ear fluff, yellow eyes, makeup, large breasts, long nails, headphones (redesign) 1girl, solo, dark skinned female, blonde hair, colored inner hair, long hair, cat ears, animal ear fluff, yellow eyes, makeup, small breasts, long nails, headphones Note that these two are for the original ethot and eventual teen-sized nyana, not the current bratty loli nyana.
Alunette Mignon	1girl, solo, villain, tall, black hair, blue hair, colored inner hair, low-tied long hair, messy hair, red eyes, eyeshadow, small breasts, white dress, gold trim
Abigill Anemoia	1girl, solo, petite, magenta hair, long hair, long bangs, hair over eyes, ((anchor hair ornament)), head fins, blue eyes, glasses, flat chest, blue hoodie, open jacket, frilled swimsuit, white swimsuit, hip vent, fishnet pantyhose
Coconut Blanc	1girl, solo, pale skin, long hair, white hair, white eyes, medium breasts, ((detective)), brown coat, buttons, light brown turtleneck dress
Jubilee Junagold	1girl, solo, bright red hair, very long hair, ((braided ponytail)), green eyes, pointy ears, large breasts, yellow dress, long dress, front slit, cleavage, navel, long sleeves
Nayvell Milkdrop	1girl, solo, loli, orange hair, long hair, white eyes, flat chest, ski goggles, pinafore dress, orange dress, yellow neck ribbon, white shirt
Marie Mehlo	1girl, teenager, short woman, (thick eyebrows), yellow eyes, blunt bangs, fluffy hair, long hair, curly hair, white hair, (small breasts), (wide hips), ornate, yellow shirt, corset, gothic lolita, high-waist leather shorts, stockings negative prompts: (loli), (big breasts), (flat chest), skirt lift OR 1girl, short woman, thick eyebrows, yellow eyes, blunt bangs, fluffy hair, long hair, curly hair, white hair, small breasts, wide hips, ornate, yellow shirt, corset, gothic lolita, high-waist leather shorts, stockings Negative: flat chest, medium breasts, from below, skirt lift
Araminta Aminthe	1girl, solo, mint, dark green hair, medium hair, ponytail, dragon horns, white eyes, medium breasts, green tail, (dragon tail), underboob cutout, dark green shirt, gold trim, bare shoulders, elbow gloves, midriff, armored skirt, dark green skirt
Mixie Mixin	(1girl, solo focus:1.4), (Accurate five-fingered hands:1.4), (dark-skinned female:1.7), white beret, brown hair, long bangs, spikey hair, (short hair:1.3), (green eyes:1.2), (white shirt:1.4), (brown cardigan:1.6), (coat over turtleneck:1.2), tomboy, (brown skirt:1.4), pleated skirt, shirt tucked into skirt
Raimi Ricotta	1girl, solo, mouse ears, mouse tail, pixie cut, parted bangs, dark brown hair, big eyes, rectangular mouth, blue eyes, flat chest, yellow hood, yellow cape, white blouse, puffy short sleeves, brown belts
Kusuri Kouyaku	1girl, solo, mature female, tall female, skinny, ultra long hair, messy hair, wavy hair, dark green hair, parted hair, parted bangs, long bangs, hair over one eye, dot pupils, narrow eyes, bags under eyes, mascara, yellow eyes, small breasts, grey sweatshirt, striped thigh highs
Neo Pawlino	1man, solo, simple background, white background, cowboy shot, looking at viewer, ((pixie cut)),(((brown hair))), (((multicolored hair))), (white hair), pink hair, ((dog collar)), tshirt, pants, ((dog ears)), chocker, black eyes, messy hair, shy,

# OrangeMix

A family of merged models, these are NOT trained, but rather a merge of existing data, and they look pretty damned good.
OrangeMix models are optimized for high quality, highly detailed, and painfully complex images with realistic details that are unable to be drawn naturally (like making a real life flower in a drawing), with around a 50% chance of making normal hands (instead of the almost guaranteed failure with other models). Keep your prompt as simple as you can. Due to their high potential, they can be somewhat finicky and can be hard to use for newcomers, tinker around until you get a feel for ‘em.

It is suggested by the creator to use DPM++ SDE Karras instead of Euler or Heun, but experiment for best results.

Use the Assorted prompts as your baseline, and tweak until you find a prompt you like.
_Base models are strictly SFW, _Half models are soft NSFW, _Night models are hard NSFW, the rest are untagged and can range from SFW to hardcore.
The NSFW models can make SFW images but not the other way around.

OrangeMixs

# Negative Prompt

The creator suggested negative prompts have some extra tags to avoid children, loli, petite, tans, muscular images and NSFW, remove these accordingly to fit whatever you intend to prompt, or use the NAI default (mind the issues).

nsfw, (worst quality, low quality:1.3), (depth of field, blurry:1.2), (greyscale, monochrome:1.1), 3D face, nose, cropped, lowres, text, jpeg artifacts, signature, watermark, username, blurry, artist name, trademark, watermark, title, (tan, muscular, loli, petite, child, infant, toddlers, chibi, sd character:1.1), multiple view, Reference sheet,

# MehloAnon’s AbyssOrange Prompts

Talent	Positive	Negative
Sena Bonbon	tall female, light brown hair, (tanned skin:1.1), short hair, (double buns:1.4), (hair buns:1.2), freckles, orange eyes, large breasts, (hag:1.2), orange eyes	(tan line:1.2) ,blond hair, single hair bun, ponytail, man hair bun
Nyana Sextyn9	dark-skinned female, blonde hair, colored inner hair, long hair, cat ears, animal ear fluff, yellow eyes, makeup, small breasts, long nails, headphones, tank top, midriff, black shorts, (headphones on neck:1.1)	flat chest, headphones on neck, more headphones
Marie Mehlo (is perfect)	short woman, thick eyebrows, (gold eyes), (yellow eyes), blunt bangs, (fluffy hair:1.1), very long hair, (curly hair:1.2), (white hair:1.2),(silverhair:1.05), (small breasts:.9), (medium breasts:1.1), wide hips, thick thighs, (ojousama:1.1), (noble women:1.1), (fluffy hair:1.2), ornate, yellow shirt, high-waist leather shorts, stockings, corset, white corset	(loli:1.2)
Araminta Aminthe (older)	mint, (dark green hair:1.1), medium hair, ponytail, 2 short soft dragon horns, (white eyes:1.2), medium breasts, green tail, (dragon tail), short horns, golden horns, brown horns, rounded tail end, pointy teeth, (human ears:1.2), underboob cutout, dark green shirt, gold trim, bare shoulders, elbow gloves, midriff, short dark-green skirt, thong straps, triangular breast covering, black thigh highs, black gloves, exposed shoulders	sideboob,side slit, jacket, elf ears, (pointy ears1.2), floppy ears, wings, more then 2 horns, twisted horns, oni horn, mini horns, branching horns, tail end bobble , multiple tails, doubled horns, merging hair horn, (Lamia:1.1)

# Assorted Models

Not necessarily modern, up to date, or even pretty, but these might give you a starting point.

# Pastel Mix

A stylized model designed to emulate pastel-like art. Very, very pretty.

# Any Pastel

A mix of Pastelmix and AnythingV4. Looks very nice.

# AnythingV4

A merged model made with Anything V3 and AbyssOrangeMix2, which, ironically was made using AnythingV3 (and NAI, NAI SFW, Gape60 and Basilmix). Shows promising results.

# CoolerWaifuDiffusion

A merged model of WaifuDiffusion and Cool Japan Diffusion with zero NAI or NAI-based mixes.

# Hitokomoru Diffusion V2

A trained model based on the art of Japanese Artist Hitokomoru.

# Dream Shaper

Intended as a model to make high quality portraits that look like actual paintings, can also make good backgrounds and anime-style images.

# Xynaptix

An Elysium derivative generalist model like Hentai Diffusion that is very sensitive to artstyle prompts, such as “(watercolor (medium), oil painting (medium):1.2)”. Very pretty and great at making hands by itself. As of V2 has a VAE included, so you’ll need to disable your personal VAE before using.

# Phantom Diffusion

A model trained on art by japanese artists that have openly stated their opposition to AI image generation. See the link for more aesthetic details.

# GuoFeng3

A realistic model intended to make Chinese aesthetic images. Can make anime characters but with a notable pseudo-realistic aesthetic.

# Sonic Diffusion V3

Now you too can prompt sonic-style chuubas. God help us all.

# ControlNet (WIP)

GET OUT OF HERE PROMPTER, NOTHING TO SEE HERE.
YOU SAW DIDN’T YOU? TO THE DOLPHIN PIT WITH YOU!
I can’t prompt to save my life nowadays, just check a youtube video on prompting with controlnet.

# AI CHARACTERS

Character.AI has been nuked to the ground for “muh safety”, don’t expect much outta these, if anything (not to mention the characters haven’t been used much at all, so no training):