How to make beautiful women in Stable Diffusion

Introduction

I'm no expert in Stable Diffusion as it's only been out for a few months as of this writing and we're all still learning about it. My main area of interest when it comes to SD is making non-nude sexy glamour photos of women, usually standing in their underwear, like the image below.

Although I have limited experience working with nudes, many of the techniques here will probably apply, though you may want to use a nude-friendly model. Ditto for Anime girls which I also have limited experience with as well. Also, some of these images I've generated (like the one above) were using a custom Dreambooth model that I've trained. I'm by no means an expert in Dreambooth and to be honest, a lot of the positive results I've gotten from it were from days of trial and error, and I can't really guarantee if my methods are reproducible, though I may add something on it in the future.

Please bookmark this guide as I'll add new tips in the future.

Love this guide? Here's how to show your support

Please consider joining my Patreon if you love this guide. You can also follow me on Twitter.

Use img2img

When I first started with SD, I struggled with crafting the right prompt. I'm still hardly an expert but after stumbling onto img2img, I used it for the vast majority of the artwork I made.

When you're telling the AI to make something using txt2img, you're basically telling the AI "Please draw from scratch according to my prompt". With img2img, you give it a picture as well as your prompt and tell the AI "Please use the image I've provided as guidance to draw a picture that resembles what I'm describing in my prompt." The denoising strength indicates how much or how little of the original image to "destroy" before trying to draw what's in your prompt (in lay man's terms.) The default setting of 0.75 basically says to destroy about 75% of the original image (evenly throughout) before trying to draw what you've described in your prompt. The lower you set this, the more the final image will resemble the original (with all else being equal, especially since your prompt can also provide guidance which will consequently be amplified by a higher denoising setting.)

IMHO, img2img is much easier for a beginner to work with rather than trying to come up with prompts and hoping the AI draws what you want. Here, you give it a picture of something (like a hot girl) that you want to make a similar picture of. What to put in the prompt really depends on what you find most important in the picture, like, say, her smile, the clothes she's wearing, the background, the lighting, etc. Play around with it as there's no single way to go about this; use the "Interrogate CLIP" and "Interrogate Danbooru" tools to get suggestions for keywords to put in your prompt and adjust their weights up (by enclosing them in brackets) like this if necessary.

Keep in mind sometimes complex backgrounds can interfere with the image generation so you may want to remove the background using a service like Removebg first.

Another great thing about img2img is that if the resulting artwork isn't what you like but it's a step in the right direction, you can resend it to img2img again to hopefully generate a picture closer to what you want. You may need to adjust your prompt (including the negative keywords) to steer it in the right direction, one step at a time.

Choose the right model

Hopefully, you're not using SD 2.0 for generating beautiful women. Instead, use some of the following as a start:

  • SD 1.5 is great as a general purpose model and can generate some good-looking women but it's more of a Swiss-army knife.
  • F222 is 1.5 fine-tuned on some softcore images but it works great even for non-nudes as well, though you may need to put words like (nude), (tits), (naked), etc. in the negative prompt.
  • Hassanblend I think this one includes F222, some Anime models, and some other NSFW models but it seems to be the best choice for nudes as well as skimpy clothing (like underwear.)

You can also browse the countless others out there on Civiai and find one to your liking.

Mixing models can also be helpful if you want to combine different styles though it can sometimes be hit or miss. I've had the best luck combining models that have specific trigger words and using multiple trigger words from different parent models when generating with the combined model. Say you have a model for cute faces with a trigger word of (cute face) and one for sexy legs with a trigger word of (sexy legs). If you mixed both models hoping to generate pics of girls with cute faces AND sexy legs, make sure you include "(cute face), (sexy legs)" (without the quotes) in your prompt; if it's not as strong as you'd like (common after mixing models), you may want to amplify them, like "((cute face)), ((sexy legs))".

Give her some makeup

This one I happened on by accident. It seems adding makeup keywords gives your girl a new look that's much greater in magnitude that what you might expect by giving her a different brand of makeup (or lack thereof.) I suspect it's because some of the images the SD models were trained on pertaining to makeup were advertisements for those brands of makeup and they all had their unique styles. Here's a grid of all the different brands of makeup I've tested in my prompt and the resulting images:

Here are the brands of makeup and related products I've tried including in my prompt:

Sephora
Lancôme
L'Oréal
Chanel
Dior
Maybelline
Estée Lauder
Clarins
Givenchy
Revlon
Covergirl

Now I'm not an expert on makeup brands so there's probably many more out there that I haven't even tested. But this is a good starting point, and you can even copy and paste it into a wildcard file like makeup.txt and use the __makeup__ variable in your prompt to select a random makeup if you're using the Dynamic Prompts addon.

Leverage Anime models

I have limited experience working with Anime models but I've had good results with the following:

From my experience, Anime models are better at creating detailed clothing like shirts, skirts, underwear, etc. as well as generating limbs and especially fingers/toes correctly more often than the other models. But if you want to have your cake and eat it too, you can do so with img2img.

First, generate pictures of females to your liking using the Anime model. Then you turn it into a (fake) photo by running it through img2img. Just use the "Send to img2img" button which will also autopopulate the prompt, negative prompt, CFG, steps, dimensions, etc. Keep everything the same but change the following:

  • Denoising to 0.5
  • Model: pick the model you'd like, whether it's F222, Hassan, SD 1.5, or whatever.

Then generate a couple of images and see if they're what you like. If not, feel free to adjust the settings (like your prompt, denosing, etc.)

Here's an example of an Anime picture I ended up turning real by using a denoising of 0.5:

Edit

Pub: 16 Dec 2022 19:24 UTC

Edit: 17 Dec 2022 20:07 UTC

Views: 2093