SD likes lower res images, models are trained on 512x512 and 768x768 images, if the image is really low res you can upscaled to 720p.
Dark images are really bad, brighten them up a bit in PS, Gimp or any other software.
Clothes colors like teal, hot pink or other that really stand out can be bad, you can try to change to similar to skin color with PS or GIMP.
All is done on the inpaint tab that is used to mask a area that will be replace by the ai using the various settings.
The big choice is Inpaint Area or Whole picture that will determinate how most of the others behave will get to this later.
Resize Mode:
This is linked to Inpaint Area if Whole picture must be Crop and Resize or will distort the image, for Only Masked Can be either Just resize or crop and resize
Mask Mode:
Self explanatory, will work on selected or not selected area
Masked Content:
The goal is to keep the image as close to the original as possible but if we have too much of the original you get residues like clothing or colors, but letting SD go crazy will lose shape and pose.
Fill: let the AI choose the colors based on the image. I use second in case Original not giving good results.
Original: starts with was there can lead to having clothing colors or outlines but keeps the original shape better. I start with this.
Latent Noise: fill the area with random color, pretty useless for this.
Latent Nothing: Uses a Smudge to paint the area, use as last resort.
Denoising strength:
Usually 1 is fine if getting weird pose, angle, proportion lower to 0.92.
When using Fill may need to be as low as 0.85.
Inpaint Area:
Whole Picture: Uses the whole picture as base, on resolution should be careful higher resolutions are very slow to process and can break the model if too wide i would stick to Width max 768 and the Height keeping the aspect ratio.
Only Masked: Enables to use high res picture. The only area worked is going to be the mask plus some around and ignore the rest. It generates a piece image and try to fit the result. Can be nice to keep a lot of detail and kind of get better skin blend with the original.
When using this, resolution should be 512x768 more on this later.
The "Only masked padding, pixels" should be high like 100-200 it uses that many pixel to get the context of the image and get better results. So if you mask the torso but low pixels is not going to know what shape and how to use in the image. But also vary on the image resolution a image with 1024x1024 maybe fine with 200 but if the image is 3000x3000 even the max 256 value might be too little.
Sampling Method and Sampling Steps:
Method is the algorithm that is gonna do work i like Euler A for fast results to see if is working, now i default to DPM++ SDE Karras, or DPM++ SDE, but there is no correct answer, some people love Heun, DDIM.
Steps is how many times is going to go through the algo, a low value get bad results but is fast, a high value is slow but better, a balance of around 40-60 for photorealism seems fine, you can use 20 just to get images faster to see if is working.
Width and Height:
If you are using Inpaint Area Only Masked this will shape the crop image to insert a smaller width and bigger height gives more context because most pics are portrait, 512x768 is the ideal because is in the range of the trained models.
If you have Inpaint Whole Picture this crops your image. You can adjust sliders so that the red area doesn't crop your image wrong. most models start to break if you set width than 768, bigger height usually fine. If generation is taking forever, or get weird things something is wrong.
CFG Scale:
Control the AI "creativity" i keep 4-7, depends if is a more standard "basic" image i keep 7, but going lower allow the AI to try describe the image, i found going lower than 4 is more like gambling sometimes it hit, sometimes it gets nothing.
Drawing the mask:
Draw around the clothes to be removed don't need to be pixel perfect the models are pretty good knowing what to keep and the settings help with that, precision can be helpful if trying to preserve something like a tattoo or accessories.
Don't keep a lot of cleavage, is usually not realistic and the AI going to try to fit something different and will look fake.
If using Only Masked somethings masking more can be better to get more context.
Don’t draw two separated masks, must be one continuous.
One part and piece at the time
Don’t try the full body at once, in full clothes or big clothes remove parts first, do under the breasts, legs, crotch, turn the pants into panties then naked.
Prompts:
Positive prompts I keep between 2-5 keeping to a minimal, something like
a naked woman, medium breasts, you can try to add stuff like high detailed skin, 4k,dslr, sometimes adding the skin tone can help get better like pale, tanned, but ignore describing the image like is txt2img.
Also can add ‘average figure’ or ‘natural figure’ so waists don’t get too small.
Start with minimal then add more later to refine.
Negative prompts I use some basic stuff:
semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, ugly mouth, ugly eyes, missing teeth, crooked teeth, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, anorexic, emaciated
The standard for Realistic Vision also works:
(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts,ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck
Negative prompts can be used to block results too, add big boobs if breasts size generated are too large.
Ending
If results are good you can generate 5-20 pictures at the time depending on GPU power, you can also use the generated image to continue removing details getting the denoising down to 0.2-0.4 and masked content to original to smooth skin with a good model like realistic vision, deliberate or other.
Settings to start
You must be registered for see links
, 
or
if you like the work!
109133544_640049343558873_7351988731756829007_n.jpg
109133544_640049343558873_7351988731756829007_n.jpg
Bad 2.png
Bad 1.png
nwdn_file_temp_1607962050122.jpg
nwdn_file_temp_1607962296544.jpg
File_000.jpeg
123511560_196898008478232_8109666748119250501_n.png
126043621_459694905011929_7264696757238030879_n.png
yfxhxdp6ch861_12462321.png
139737847_182549930273910_6378192271380528596_n.png
00004979.png
