Sign In

Pretty Girls Next Door

218
2.7k
50
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
506
Reviews
Published
Jan 3, 2025
Base Model
SD 2.1
Hash
AutoV2
34529678BD
default creator card background decoration
woweenie's Avatar
woweenie

NSFW-enhanced full finetune of SD2, with v-pred + zero terminal SNR

  • download the YAML or use the diffusers mirror woweenie/pretty-girls-next-door-v5.1

    • **** YAML UPDATED **** sorry it was wrong before, please re-download

  • use a webui that supports v-pred, turn on ZTSNR betas, and use CFG Rescale=0.7 (InvokeAI with diffusers link recommended because it handles this for you). Scroll down for Automatic1111 install guide.

  • use NON-K samplers (ZTSNR does not work with Karras)

DO NOT USE TO MAKE DEEPFAKES. DO NOT USE TO MAKE IMAGES OF CHILDREN.

Using an entirely custom dataset, almost all manually captioned. Focus is on softcore and erotica, rather than porn per se. It's really great at doing erotic images of fully-clothed women. It can definitely do nudes, but it's more like z1shy or m3t-4rt than 4tkg4ller14 or "porn" porn. V5 now with all new training methods, highest quality yet.

Your prompt should be something like: "a photo of a woman with large breasts and a pretty face. front view, upper body shot. wearing a camisole. 85mm, f/3.5" Scroll down for an indepth prompting instructions.

Auto1111 Install Guide

  1. Download BOTH .safetensors and config (.yaml) files and put them in your webui\models\Stable-diffusion folder.

  2. Go to the Settings tab, type "zero terminal" in the search box and set it to "Zero Terminal SNR", then click Apply Settings.

  3. Go to the Extensions tab, click Install from URL, and install https://github.com/Seshelle/CFG_Rescale_webui

  4. Then go back tio Installed, make sure CFG_Rescale_webui is on the list at the bottom, and click Apply and restart UI

Download the Config too (.yaml file) or you'll get black images. Or use the Diffusers mirror with Invokeai or Comfy: woweenie/pretty_girls_next_door_v4.0

This model is horny and will produce nudity without prompting. Add "naked" to the negative prompt if you don't want NSFW output.

This is a SD2 model (with OpenCLIP VIT-L/14 text encoder), so prompting is more like SDXL than SD1.5 - write sentences, keep the negative simple. Use the samples as a guide.

Recommended settings: stochastic sampler ("SDE" or "Ancestral" eg DPM++ 2M SDE or Euler A), 50 steps, CFG 6-10, size 640x896

Recommended negative prompt: as short as possible, start with "corpse, 3d render"

Prompting guide

See below for the full guide it's from v1 but is still mostly correct.

You can get big quality imporvement in the normal way, by stating camera settings eg "canon eos" or "analog lens" or "film grain" or "100mm, f/3.5" (experiemnt with different numbers).

You can also get a big quality improvement by adding a random person's name, use Behind The Name's random name generator to make up some fake names for you. Changing the nationality of the name changes what the person looks like. They don't have to be real names, I also notice improvemetn just typing "a photo of ufailhk fnlkiha wearing a t-shirt" instead of "a photo of a woman wearing a t-shirt"

Pose is the hardest thing to prompt right. this is because it is ambiguous, depending on the camera angle as well as the pose, so you can get better poses by also prompting a view angle eg "front view", "rear view" etc.

  • if the pose still looks like a scrambled combination of multiple poses, try putting the mixed up poses in negative, so let's say you want "lying on her back" and it keeps drawing her butt but her head on backwards then that's because it's confused and trying to also (wrongly) render her "lying on her stomach" at the same time - so put that in negative to stop it.

  • this works for other poses too, one pose in positive and one in negative helps to make it solid

  • you can also add a view prompt to negative for views you don't want that it is still trying to make

  • you can also use "group" in negative prompt or "modernist garbage" which helps to fix the pose sometimes

Negative Prompt: try with empty first.

  • if she looks dead use "corpse",

  • if she looks like a 3d render use "daz3d" or "3d" or "render",

  • if she looks like exaggerated anime proportions or too smooth skin use "anime".

  • if she has extra arms or legs, or missing arms or legs use "group".

  • if the photo looks like it is from a cheap 1980s camera use "instamatic" "snapshot" "vintage" etc

  • if the pose or expression is boring (like a mail-order catalog) try adding some of these: "fashion" "magazine" "advertisement" "catalog"

  • if she looks like trash try "hooker" or if the whole picture is gritty try "brushstrokes"

  • Jokers: "lumpy", "modernist garbage", "sketch", "draft", "ugly" (but risks sameface), "grimace"

  • Sometimes a gen is bad because of a word in negative, and its harder to tell which one so best to keep negative short unless you like to go through and try deleting the words one by one to find the bad one.

V5 changes

  • better prompts, about 200k steps more training, better training strategy

V4 changes

  • no need for complicated CFG rescale or custom beta schedule

  • pose prompting improvements

    • added "fetal pose" for lying on her back, curled up with her knees at her shoulders

    • "supine" is now "lying on her back"

    • "prone" is now "lying on her stomach"

    • "recumbent" is now "lying on her side"

    • full list of pose tags: standing, sitting, kneeling, fetal position, on all fours, squatting, lying on her back, lying on her side, lying on her stomach

    • can be combined with: closed legs, parted legs, spread legs, crossed legs, one leg raised, both legs raised

  • better view control

    • 3/4 view is now "angled view", you can use "angled front view" or "angled rear view" ensure front/back view

    • add "from below", "from above"

    • full list of view tags: front view, angled front view, side view, angled rear view, rear view, from above, from below, high angle, low angle

  • improvements to wet skin prompts (try "wet skin", "soapy skin", "wet see-through t-shirt")

  • full list of trained clothing tags, bold is best: bikini, bikini bottoms, bikini top, blanket, bodysuit, bra, camisole, coat, crop top, crossover top, dress, garters, glasses, goggles, gown, gym shorts, halter top, handbra, hat, hoodie, jacket, jeans, jumper, lace-up top, leggings, lingerie, long pants, long sleeved top, nightgown, off-shoulder top, one piece swimsuit, overalls, pajamas, panties, pantyhose, piece of fabric, robe, scarf, shirt, shorts, skirt, sleeveless top, slip, sneakers, sock, socks, sports bra, sunglasses, sunglasses on her head, sweatpants, t-shirt, tank top, tied shirt, top, undersized bikini, wetsuit

V3 changes

  • ZTSNR training, needs trained beta

V3 is only available as diffusers -

v3 woweenie/pretty_girls_next_door_v3

v3.1 woweenie/pretty_girls_next_door_v3.1

Trained on 4500 12000 images (v1 dataset attached to download) from well-known soft porn websites, and a low alpha merge with Artius-v21 and Mangled_Merge_v3. Mix of manual captions and LLAVA autocaptions with manual tag guidance. Focus is on natural language prompting and flexibility and precision with clothing and nudity.

All of the samples are made with diffusers. Generation only, no inpainting or img2img. Seeds are CPU ("random number generator source" option in automatic1111 or "generator source" option in vladmandic).

V2 changes

  • better response to shorter prompts

  • general image quality improvement

Usage guide

Download the .yaml or you'll get black images. If you want to reproduce the samples use InvokeAI or vladmandic with diffusers backend.

Recommended negative

3d render, ugly, misshapen, corpse, brushstrokes, child

  • DPM++ SDE Karras, CFG 6-8, 24-36 steps

  • DPM++ 2M SDE Karras, CFG 6-8, 24-48 steps

  • DDIM or Euler a, CFG 6-11, 48-96 steps

Recommended settings (V1)

  • DPM++ 2M SDE Karras, CFG 6, 24 steps or 30 steps

  • DPM++ SDE Karras, CFG 6, 24 steps or 30 steps

  • DDIM or Euler a, CFG 6-11, 48-96 steps

How to Prompt

The model was trained on multiple sentences separated by full stops, and if you prompt like that you will get more out of it. Eg "a photograph of a woman on the beach, wearing a bikini. beautiful soft rim light, intricate details, 8k."

Prompting walkthrough

Start simple. Don't use tag lists, just write something short like "photo of a naked woman at the beach":

To get closer, add more words for skin or body parts, but this can make generation weird:

To fix weirdness add a pose to the prompt:

Bring the face back into frame by prompting it:

To move further back, add background details:

You an also change the influence of background details by moving them to a different place in the prompt. If they're closer to the start they will have more influence.

More tips

  • Increase quality with words like "ultra realistic, skin details, skin pores, 8k, intricate details".

  • "Detailed face" or just "face" improves faces, "eyes" improves eyes, "lips" improves mouth shape.

    • "Painted eyes" and "cartoon face" in negative help if the eyes look like they're painted on

  • Long negatives are often worse. Start short and add words as bad things appear. Clean up words that are doing nothing as you go.

  • Prompt a pose to improve results (try to match what the random seed is drawing on its own)

    Prompt leg pose to improve results (try to match what the random seed is drawing on its own)

  • Sometimes "professional photograph" "stunning photograph" "detailed photograph" help, sometimes they make the image worse.

  • Sometimes the model gets too close up. Use negative prompt "close up" to fix

  • Some things are overtrained in the base model, especially "bedroom", so move it further down the prompt or downweight it with [[bedroom]]

  • Adding words like "face" or "breasts" or "hips" or "knees" brings camera closer to those things (or further out of they're not in frame already)

  • The first word has a large influence on results. Try "HQ photo", "ultra realistic photo", "professional photo", "digital photo", "detailed photo", "stunning photo", and just "photo"

  • Adding camera, lens, or film words eg "Canon EOS Rebel, f/11 35mm" or "ektachrome ER film" sometimes helps, but sometimes make the output less nude

  • If some things are dominating, change the order of words.

  • Sheer/wet can benefit from "sheer clothing", "see-through", "wet clothing", "visible nipples"

  • Try "open" clothing, eg "open shirt", "open robe". Also try "unbuttoned" and "unzipped" - "wearing an unzipped one piece swimsuit" is fun.

  • Increase cleavage with "cleavage" or a longer phrase like "a tank top that shows her cleavage"

  • To fix cut-off legs or arms put "stump" or "amputee" in negative, or "weird anatomy" or "strange anatomy" or "bad anatomy" 

  • "Good anatomy", "beautiful composition" sometimes helps

Trained words

Face

pretty face, cute face, beautiful face, ugly face (for negative)

Pose

reliable: standing, kneeling, squatting
ok: on all fours, sitting
unreliable: prone (lying on stomach), supine (lying on back), recumbent (lying on side)

Body: 

large breasts, medium breasts, small breasts, cleavage, cute butt, round butt, broad hips

View:

front view, 3/4 view, side view, rear 3/4 view, rear view, low angle, high angle, overhead shot, medium shot

Scene: (also good for negative)

At the pool, at the beach, at a river, in the shower, in the bath, in a photo studio

Legs pose (can be combined with pose, doesn't always work): 

legs together, legs apart, legs spread, legs crossed

Clothing (some work better than others):

Whole body: dress, overcoat, robe, bathrobe, pajamas
Top: tank top, t-shirt, shirt, jumper, sports bra, jacket, hoodie, halter top, polo shirt
Bottom: leggings, jeans, skirt, shorts, sarong, socks, sneakers
Swimwear: bikini, bikini top, bikini bottoms, one piece swimsuit, wetsuit
Underwear: bodysuit, bra, panties, nightgown, slip
Face: glasses, sunglasses, hat

Expressions (unreliable):

duchenne smile, friendly smile, flirty smile, fake smile, polite smile, laughing, confident, distracted, aroused, annoyed, embarrassed, bored, scared, dominating

Clothing state:

Wet, mesh, sheer (combine for better effect eg "wet mesh sheer see-through t-shirt"), fishnet, thin, open, unbuttoned, unzipped, undressing, unbuttoning, unzipping

Pussy (can be combined like "hairy puffy pussy"):

Hairy pussy, shaved pussy, bald pussy, tulip pussy, puffy pussy, curtains pussy, horseshoe pussy

Extras:

Wet skin, with her arms raised, with her hand on her hip, nipple pokies (unreliable), strap slipping off her shoulder, lifting her shirt, hiding her pussy between her legs

Updates

I have updates to the data source planned when I can find some time to work onthem:

  • Fix prone + front view + low angle + large breasts

  • Add sports bra rear view

  • Fix nipple pokies?

  • Fix undressing?

  • Dress, long with legs not visible (leg poses under dress)

  • Sheer sports bra?

  • Dress/gown rear view

  • Check for "legs crossed" and "legs together" overlap, and "legs together" "legs apart" overlap

  • Add bikini + small breasts

  • Add bodysuit + rear view

  • Why is "slip" so weirdly close-up?

  • Sheer + naked? It seems like leaving off "naked" means "sheer" can't draw nipples

  • Remove natural light?

  • More wetsuits

  • Add half-ponytail hair

For now the dataset is attached, please credit me if you use it.