Guide Overall basic guide to using Stable Diffusion WebUI

  • Welcome to the Fakes / AI / Deepfakes category
    When posting AI generated content please include prompts used where possible.

dkdoc

Superfan
Mar 11, 2022
63
513
639
0fya082315al84db03fa9bf467e3.png
This isn't a tutorial on "How to install" Stable Diffusion or the WebUI as there are plenty of tutorials already out there. This is a guide and overall "how to actually USE the software."

Most people here want to use the AI to inpaint for NSFW image creations, so I won't go too far into how to use txt2img. But here is an example of what you can do with it:

Please, Log in or Register to see links and images


Prompts:
Please, Log in or Register to view quotes

Negative Prompts:

Please, Log in or Register to view quotes

Again, do not expect to use one or two prompts and expect good results, be extremely detailed.

Now for the part everyone wants to do, inpainting.

First you are going to need an image to inpaint, I'm going to use an AI generated image of my fantasy character from the txt2img creations.

I'm going to paint around the crop-top she is wearing in order to expose the boob.


Please, Log in or Register to see links and images


I have left all of the settings as default except for the Width & Height sliders which I changed to match the aspect ratio of the original image. I included a basic text prompt of "nude woman" to showcase why you need to be more detailed with prompts.

The result was this, after 15 images. This was the "best" one.


Please, Log in or Register to see links and images


Now, I'm going to leave everything the same but be more detailed with my prompts. What I'm looking for is basically "medium sized boobs". So I'm going to be as descriptive as possible. I don't need to be concerned with parts of the image outside the mask, so I don't need to include prompts like "hands", "face", or "feet".

My new prompts were:

Please, Log in or Register to view quotes

My negative prompts were:

Please, Log in or Register to view quotes

The result, after 15 image generations, was this:


Please, Log in or Register to see links and images


As you can see, it's not perfect, but it's much better than the first result.

From here, you can continue to mask other parts of the image by clicking on "Send to inpaint" and masking out a new area of the image. You can continue doing this until the image looks the way that you want it to.

If you want all of the clothing removed, then mask the clothing in parts. Don't attempt to do everything in one go, it very rarely results in a good image. Instead start off with the shirt or pants, then when you have a good result move on to other parts of the body. Once everything looks nice then you can paint over the entire body (except the face) and continue generating images until you get one you like.

For my prompts I generally follow a guideline of:

  1. describe the main focus of the image (a woman)
    1. if using txt2img, what do you want the face to look like?
      1. Shape
      2. Shape of her nose, mouth, eyes?
        1. Small, large, wide nose?
        2. Full, small lips?
        3. eye shape & color
      3. Hair
        1. Color, hairstyle, what's it doing?
        2. Flowing down her back, covering her boobs, flowing in the wind?
    2. What body type should she have?
      1. Skinny, fat, chubby, curvy, athletic, toned?
      2. Skin color?
        1. Skin complexion?
          1. Does she have freckles?
          2. How much and where are they?
      3. Does she have tattoos?
        1. What's it look like
        2. Where is it
        3. What color is it?
    3. Hands
      1. What are they doing?
      2. holding anything?
      3. gesturing?
    4. Feet (if in view)
    5. Orientation of the body.
      1. Is it facing the camera or not.
      2. profile (side) view?
      3. shot from the back, above, below?
      4. standing, sitting, posing?
  2. Shot Composition
    1. Camera used? Footage type?
    2. Clear, hazy, atmospheric, foggy, grainy?
    3. Realistic, illustrated, CGI, digital art?
    4. Everything in focus, out of focus, depth of field?
    5. Moody, dim, bright, even, studio, cinematic lighting?
  3. Background (if replacing it or generating a new one)
    1. WTF is going on back there?
      1. Should there be furniture?
        1. Material? Color?
      2. Floor? Buildings? Sky?
Everyone wants to know what prompts are used to generate great images, but not a lot of people write them down and keep track of them. So what I suggest doing from now on is to create a .txt file and record the prompts you are using to generate images. Whenever you have a batch of images that turn out really well, KEEP THOSE PROMPTS! If you have a batch of images that don't look like what you want, don't put them into the .txt file.

Organize your prompts how you want: Faces, bodies, boobas, vaginas, arms, legs, hands feet, compositions, camera settings, backgrounds, etc.

- Have fun
 

dkdoc

Superfan
Mar 11, 2022
63
513
639
0fya082315al84db03fa9bf467e3.png
Below are various positive prompts that I've been using to create SD images. I'll post the Negative prompts later.

Feel free to add ( ) to them to suite your needs.

Face Prompts:

Shape - heart, diamond, oval, round square, rectangle

Eye prompts:

Shape - almond, close set, protruding, downturned, wide set, hooded, upturned, round, monolid

Color - blue, hazel, brown, green, gray, amber, violet

Mouth Prompts:

Lips Shape - full heavy upper, wide, round, heavy lower lip, thin, bow-shaped, heart-shaped, downward-turned

Lips Color - natural, <any lipstick color>

Nose Prompts:

Shape - roman, nubian, straight, crooked, fleshy, button, hawk, snub, big

Nostrils - small, big, wide

Hair Prompts:

Hairstyle - long, straight, spikey, bob, shoulder-length, perm, curly, cornrows, chignon, bun, french twist, ponytail, braid, french braid, braids, pigtails, long wavy, wavy, afro

Hair Color - rose, lilac, platinum, light blondes, golden blonde, ash blonde, medium blonde, caramel, ash brunette, chocolate brunette, dark brunette, strawberry, copper, red brown, violet, black, red, brunette, blonde

Overall face descriptors:

Details - shaved, flowing, detailed, beautiful, small, wide, cute, symmetrical, extremely detailed, HD, clear, gorgeous

Body Prompts:

Body Shape - rectangle, triangle, hourglass, inverted triangle, round

Body weight - skinny, supermodel, curvy, athlete, athletic, muscular, abs

Skin - white, black, tan, pale, tan-lines, tan lines, tanned, sunburn, dark, skin pores, real skin texture, wet skin

Complexion - perfect, freckles, light freckles, heavy freckles, scars

Details - detailed, beautiful, perfect, small, wide, cute, symmetrical, extremely detailed, realistic, HD, clear, topless, gorgeous

Actions/Pose - sitting, standing, standing up straight, posing, superhero, pose, sexy, sultry, explicit, porn, onlyfans, kneeling, on knees, on all fours, teasing, tease, lying down, laying down, leaning, leaning over, lying back, laying back, arched back, hunched back, hunched over, pinup, pin-up, facing camera, profile, side, facing away, backside, back, facing front, facing side, side profile, legs spread, legs apart, legs in the air, cross legged, legs crossed

Breast Prompts:

Breast Shape - asymmetrical, symmetrical, bell-shaped, athletic, round, east-west, slender, relaxed, side-set, tear drop, conical, close-set, retrousse, perky, saggy, sagging

Breast Size - a-cup, b-cup, c-cup, 34c cup, d-cup, dd-cup, small, medium, large, full, enormous

Nipples - hard, perfect, beautiful, soft, pale, ghost, dark, detailed

Breast Details - perfect, displaying, exposed, hands in front, hands cupping, natural, fake, bimbo, hyper realistic, realistic

Hand Prompts:

Fingernail type/shape - short, long, manicure, manicured, french manicure, natural, rounded, wide, squoval, cut out, trapeze, almond, edge, lipstick, ballerina, square, pointed, stiletto, mountain peak, rounded lipstick

Fingernail color - natural, red, black, green, blue, white, purple, orange, violet, french tipped,

Hand Details - perfect, realistic, extremely detailed, individual fingers, 8K hands

Navel/Bellybutton Prompts:

Type - outer, inner

Piercing - navel piercing, navel jewelry

Vagina Prompts:

Vagina Details - tight, exposed, perfect, pussy, bottomless, trimmed, small, small lips, perfect lips, detailed, bush, landing strip

Butt Prompts:

Butt Shape - square, round, heart, pear, inverted

Butt Size - large, small, toned, wide, big, enormous

Butt Details - perfect, cellulite

Hips/Waist Prompts:

Details - wide, small, perfect

Feet Prompts:

Details - perfect feet, individual toes

Clothing Prompts:

Tops - straight across, sweetheart, semi-sweetheart, v-neck, scoop, asymetric, bateau-boat neck, sexy scoop, spaghetti strap, jewel, halter strap, off-shoulder, high neck, queen anne, deep cut, halter, square, t-shirt, crop, tube, bra, bustier, sports bra, sportsbra

Bottoms/Panties - boyshort, briefs, thong, cheeky, highrise, brief, slip shorts, hipster, bikini, g-string, tanga brief, high cut, french cut, control top brief, panties, garter

Legs - stockings, garters, garter-belt, garter belt, fishnet, jeans, shorts, pants, skinny jeans, yoga pants

Feet - sneakers, shoes, stilettos, high-heels, high heels, pumps slipper

Outfits/Bikini - sport suit, wetsuit, halter, tankini, plunge, off-shoulder, tie-front bikini, sport, high-neck bikini, micro-bikini, bikini, bandeau, string bikini, one-shoulder, longline bikini, rash guards, microkini, skirtini, one-piece, monokini, swim dress, burkini

Color - <literally any color>

Design - floral, intricate, extremely detailed, trim, sheer, lace, silk, velvet, cotton, polka dot, mesh

Tattoo Prompts:

Location - arm, sleeve, hand, thigh, ankle, feet, shoulder, back, hip

Shape - large, small, heart, fish, dolphin, floral, flower

Color - black, red, blue, green

Shot Composition:

Camera - low quality, high quality, DSLR, RAW, portrait, medium shot, full body shot, wide angle, 24mm, 35mm, 50mm, 90mm, smartphone

Resolution - 8K, ultra HD, extreme detail, digital photo, hyper realistic, stunning, clear image, sharp focus, depth of field

Lighting - cinematic, low, dim, dark, bright, overexposed, over-exposed, even, clear, majestic, regal, studio, 3-point, harsh, professional, very dark, epic, dreary

Filters - post processing, filters, instagram, onlyfans, RGB noise, film grain, sepia, black and white, greyscale, B&W

Photo Type - illustration, drawing, cartoon, anime CG, 3D, CGI, realistic, photorealism, photorealistic
 

dkdoc

Superfan
Mar 11, 2022
63
513
639
0fya082315al84db03fa9bf467e3.png
Negative Prompts:

Face - cross eyes, cross-eyes, closed eyes, lazy eye, lazy eyes, beard, moustache, mustache, missing teeth, missing tooth, toothless, no teeth, strabismus, heterochromia, duplicate faces, bad eyes, multiple faces, blurry eyes

Body - extra limbs, extra arm, bad body, man, barbie doll, hairy chest, amputee, amputated, hands in front of body, hands covering body

Breasts - lopsided, extra breasts, extra nipples, deformed, unequal breast size, hands over breasts

Hands - extra fingers, extra hands, fused fingers, fused hands, mutated fingers, mutated hands, distorted hands, distorted fingers, twisted fingers

Navel - multiple bellybuttons, multiple belly buttons

Vagina - penis, dick, balls, testicles, oversized vagina, distorted vagina, mutated vagina

Feet - extra feet, extra toes, fused feet, fused toes, mutated feet, mutated toes, distorted feet, distorted toes

Overall Negative Descriptors:

deformed, bad anatomy, disfigured, mutant, mutation, mutated, broken, cutoff, cut off, cropped, out of frame, stretched, stretch, wide, extra wide, missing, disformed bad proportions, plastic skin, plastic, weird, odd, alien ,abnormal, fused, blending, airbrush, airbrushed, airbrushed skin, photoshop skin, face out of frame, conjoined twins, multiple heads, group, awkward poses, unusual poses, cropped, duplicate limbs

Composition:

Camera - low quality camera

Resolution - blurry, out of focus, pixelated, pixeled, RGB noise, film grain, hazy, haze, blur

Lighting - uneven lighting, harsh lighting, harsh shadows

Filters - watermark, greyscale, black and white, sepia, B&W, highlights, filters

Photo Type - CG, render, cartoon, illustration, painting, digital art, poorly drawn, bad art, art, anime
 

dkdoc

Superfan
Mar 11, 2022
63
513
639
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes
If you are trying to generate a completely original image of the celebrity in question you just need to identify the celebrity then describe literally everything else about the photo.

For most celebs a simple, "A photo of <insert celeb name here>" will do. Then fill out the rest of the prompt space with more detail.
 

RB_1875

THE Latex Simp
Oct 13, 2022
186
3,214
622
0fya082315al84db03fa9bf467e3.png
are there any particular models that need to be installed to get this working properly or should the base version of stable diffusion be able to do what you want?

also any advice on how to stop it from stopping because it's run short of memory? I'm running a 2070 super but it just overflows the VRAM immediately and stops
 
Last edited:
  • Like
Reactions: dkdoc

dkdoc

Superfan
Mar 11, 2022
63
513
639
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes

I use
Please, Log in or Register to see links and images
, but any NSFW model from
Please, Log in or Register to see links and images
will do. Follow the install instructions on the model page if there are any included. As far as SD versions, I'm still using 1.5 and the version of WebUI that was released in November. I haven't updated it because I'm too lazy + it already works and the mentality of, "If it ain't broke, don't fix it."

Advice for running out of memory would be:

1. lower your settings
2. lower your batch sizes
3. lower the output image resolutions

I have a GTX 1080 and I can generate 5 batches of 3 images per batch up to 512x776 resolution. If you are having Out of Memory errors, it's probably one of those 3 things I listed above.

My settings are almost always left on their default values
I never raise the batch size above 3 because I can't handle more than that due to VRAM
I generally keep the output image resolutions below 512x900
 

seasonedkek

Diamond Tier
Mar 17, 2022
33
948
1,039
0fya082315al84db03fa9bf467e3.png
For txt2img there's this thing called Hires fix which helps if you generate images significantly bigger than 512x512. Is there anything like that for inpainting?
 

dkdoc

Superfan
Mar 11, 2022
63
513
639
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes

For inpainting there is "Inpaint at full resolution"

When checked, that tells the AI to follow these steps:

1 - Analyze the masked area
2 - upscale the masked area to the designated image output resolution (default is 512x512 unless changed)
3 - create content for the masked area at the full image output resolution
4- downscale the masked content to the size of the mask

Highres.fix does help in generating larger images, but that's not really what it's intended for. Essentially what the option actually does is to prevent artifacts and distortions when creating images larger than the default 512x512 size.

So if you have an image placed into the WebUI but the subject is standing in an awkward angle or they're sitting down and slightly hunched over, it can lead to an oddly shaped area that you need to mask. Like if the subject is standing with their arms down by the sides and their body is turned 45 degrees away from the camera. That leads to a mask that isn't what the AI would consider "normal". A "normal" mask would be "body facing the front, no obstructions".

The AI, without proper prompting, might generate a very distorted side of the mask like a large and warped shoulder without a "normal" mask area. One side of the body could come out small while the other is overly large. There might be hip/genital area where the belly button should be if the subject is sitting down because the AI doesn't "know" the exact pose of the body based off of the mask size and shape.

You can fix that by being extremely detailed with your prompts, but also knowing which images are going to produce good results and which aren't. The more difficult the image/mask area, the longer it's going to get you a decent result. Sometimes it will be impossible, sometimes it'll take over an hour and dozens of image generations to get 1 "good" recent result.

There are times when I'm generating images with good lighting, good camera quality, and a normal pose where I only have to generate one image and I'm done. The entire process from loading the image in to the AI to saving the produced image takes less than 2 minutes. There are also times where I've tried to work on a problematic image and it's taken me 2 hours to get a "half decent" result on a single image.

There isn't really a magical "always type in these prompts" or "always use these options" settings to get the image you're looking for. You have to play around with things. The more you use the software, the better you'll get at it.
 

Bloodline666

Superfan
Mar 15, 2022
28
547
737
0fya082315al84db03fa9bf467e3.png
Hi I was wondering what parameters to use if I want to feed stable diffusion several images of my own OC and have it generate new ones based off prompts and the images fed
 
  • Like
Reactions: lightweaver

dkdoc

Superfan
Mar 11, 2022
63
513
639
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes

You need to either be extremely detailed with your prompts or train your own model with textual inversion to do that.

If you decide to train your own model you need 5-20 images for it to learn from + a good graphics card (10GB+ vram) for it to work properly. If you don't have a GPU good enough, you can use Google Collab or Dreambooth to run the computing power for you, but I haven't tried them.

This is a good tutorial for training a custom face using textual inversion using your own GPU:

 

seasonedkek

Diamond Tier
Mar 17, 2022
33
948
1,039
0fya082315al84db03fa9bf467e3.png
"Medium" usually turns out to be way too big, I do something like (small:1.5), (medium:1.3) for more realistic results
 

xXNerdCrusherXx

Fan
Mar 15, 2022
5
73
328
0fya082315al84db03fa9bf467e3.png
im getting this strange stuff generated when i switch away from the default checkpoint. is there something im doing wrong? it worked first time i launched it but now it just comes out like this every time
Please, Log in or Register to see links and images
 

Bloodline666

Superfan
Mar 15, 2022
28
547
737
0fya082315al84db03fa9bf467e3.png
so iv been looking at other peoples prompts and seen alot words in () with prompts like this (ugly:1.3), (worst quality, low quality:1.3) whats the logic behind the () and numbers?
 
  • Like
Reactions: dkdoc

dkdoc

Superfan
Mar 11, 2022
63
513
639
0fya082315al84db03fa9bf467e3.png
Please, Log in or Register to view quotes

adding a ( ) to a prompt tells the AI to pay more attention to that prompt. There are specific values to ( ), (( )), and ((( ))) but I don't know them and I'm too lazy to remember them.

You can use them for both positive and negative prompts.

So, for example, if you tell the AI - "a wooden table with a red apple placed on top of it with a white background". But it gives you a photo of "a wooden table with a random piece of fruit placed randomly in the scene with a red background" you have a problem.

Normally, you would just dial up the CFG scale to make it pay more attention to the prompts, or increase the sampling steps, or do both of those things. But what happens when you already have a CFG scale of 15-20 and a sampling rate of 30-70 and it is still giving you bad images?

That is where the ( ) prompt comes in.

So a new prompt would be something like - "(a brown wooden table) with a (red apple) placed on top of it with a (white background)". You don't have to do it that way, you could just put the ( ) on (apple).

Using multiple ( ) just tells the AI to pay more attention to that specific prompt. So (((apple))) is essentially telling the AI "Give me a goddamn apple in the picture". Likewise, with negative prompts ( ) = pay extra attention to not include that. (( )) and ((( ))) = pay extra double fucking attention to not include that.

Be careful with the (( )) and ((( ))) prompts though because they can sometimes mess up your image. Especially if you are trying to put filters or post processing on the image. I've found that adding in "film grain" will mostly give me a ultra clean looking image. (film grain) is a bit better, but (((film grain))) sometimes makes the image look like an obnoxious amount of grain with multiple high-pass filters going on.

You can also set the AI up to randomly generate a set of prompts with the [ ] and / inputs.

For example, - "skin color [white/black/tan/pale]" or "hair color [blonde/black/red/brunette/white/purple]"

That will tell the AI to pick one of those skin tones or hair colors and use it for an image. The / input will tell the AI "chose one". However, if you type in the same prompts with , instead of / it will tell the AI to combine those.

So "skin color [white, black, tan, pale]" or "hair color [blonde, black, red, brunette, white, purple]" will give you a multicolored skin tone and hair with gradients.

Same with clothing if you type in like "lingerie with [red, gold, black]" will give you a multi-colored lingerie. Using the / will make the entire lingerie be either Red, Gold, or Black.