Home
Artists
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register
Partychan SimpleX Chat ThePornDude
Home Artists Posts Import Register
Join the new SimpleX Chat Group!

AI ART: Controlling layers of clothes (Patreon)

Published:
2023-01-14 19:22:58
Imported:
2023-01

Content

Using this technique you can generate quickly high quality images with characters wearing very specific clothes. The trick is to do something similar to what 2d artists do for nsfw artwork: start with the nude character and add layers of clothes on top.

This technique can be applied to more general things as well, like building a landscape shot: you can start with an empty beach for example and incrementally add things to it, making the AI sort of Photoshop for you. 


1. POSITION

First step is to run small quick images and find an image composition that you like, the body position in this case

Prompt:nude photo of korean woman, naked, ((standing on beach)), long legs, big hips, (blue and green hair) messy ponytail, posing

Sampler: euler_a, Inference Steps: 25, Guidance Scale: 4.4, Model: HassanBlend1.5, VAE: vae-ft-mse-840000-ema-pruned, Negative Prompt: lowres, drawing, Fix Faces: GFPGANv1.3

Size: 320x512


2. BODY

I picked the 3rd image by hitting "Use as Input" over it and increased the image size to generate a larger image next. I focused just on the body and ignored the faces for now.

Prompt: nude photo of korean woman, naked, ((standing on beach)), long legs, big hips, (blue and green hair) messy ponytail, posing

Seed: 1730434, Sampler: ddim, Inference Steps: 25, Guidance Scale: 4.4, Model: HassanBlend1.5, VAE: vae-ft-mse-840000-ema-pruned, Negative Prompt: lowres, drawing, Prompt Strength: 0.6, Fix Faces: GFPGANv1.3

Size: 512x832


3. FACE

I picked a body the same as before, by hitting "Use as Input". I also added a few more words to the prompt to try to get better faces. Then I clicked Inpaint and selected just the face area. The AI will keep the body that I selected and just modify the face

Prompt: nude erotic photo of korean woman, naked, ((standing on beach)), long legs, big hips, (blue and green hair) messy ponytail, posing, (portrait), (pouty lips)

Seed: 9613581, Sampler: ddim, Inference Steps: 25, Guidance Scale: 4.4, Model: HassanBlend1.5, VAE: vae-ft-mse-840000-ema-pruned, Negative Prompt: lowres, drawing, Prompt Strength: 0.6, Fix Faces: GFPGANv1.3

Size: 512x832


4.CLOTHES

I picked the image that I liked most and hit Use as Input.

To add clothes you have to do 3 things:
- edit the prompt to describe the new clothing item
- hit Draw and add a bit of color to help the AI
- hit Inpaint and select the area the AI should modify

After you do that the AI will generate variations of that clothing. When you like one, you can click Use as Input and you can move to the next item.


Prompt: nude erotic photo of korean woman, naked, ((standing on beach)), long legs, big hips, (blue and green hair) messy ponytail, posing, (red panties)

Prompt: nude erotic photo of korean woman, naked, ((standing on beach)), long legs, big hips, (blue and green hair) messy ponytail, posing, (purple top)

Prompt: nude erotic photo of korean woman, naked, ((standing on beach)), long legs, big hips, (blue and green hair) messy ponytail, posing, (yellow shirt)

nude erotic photo of korean woman, naked, ((standing on beach)), long legs, big hips, (blue and green hair) messy ponytail, posing, (black skirt)


5. FINAL TOUCHES (Optional)

To add a bit more variety and fix imperfections in the final results you can go through them again: hit Use as Input and turn down the Guidance Scale & Prompt Strength. No draw & no inpaint needed. The AI will now work on the whole input image and will add small variations to it.


Prompt: nude erotic photo of korean woman, naked, ((standing on beach)), long legs, big hips, (blue and green hair) messy ponytail, posing

Sampler: ddim, Inference Steps: 25, Guidance Scale: 2.3, Model: HassanBlend1.5, VAE: vae-ft-mse-840000-ema-pruned, Negative Prompt: lowres, drawing, Prompt Strength: 0.2, Fix Faces: GFPGANv1.3


I kept Fix Faces on during all images. That causes some variations to the faces from image to image. From what I can tell the AI alters the images a bit every time no matter what you do, faces will not carry 100% from image to image. Leaving Fix Faces off during the clothes steps and applying a Fix Face only on the last Final Touches step might work a bit better but I haven't tried it

Files

Comments

jimmy peterson fife

i know you're demonstrating the AI thing, but that model(s) is amazing.

Uatu

Could this be used to make texture files for VAM?

Vamvader

I love your plugins but I've found your AI tutorials very helpful at getting to grips with the concept, so thanks! I couldn't get the one you use to run though but found that NMKD Stable Diffusion works great for me, although the interface does not seem to have quite the same options as the one you're using there are (slightly clunky) ways to get similar results. Great stuff though!

SPQRAeternum
>>99630153

thank you! never heard about that one. there's Automatic1111 too if it helps, that's I think the most popular & powerful one but it's more difficult to install. I had like different 5 problems with it and eventually gave up on trying but it might be just my system. SDUI for me has been double-click, install, done

Vamvader
>>99630153

NMKD was also a one click and you're on your way install. For some reason I kept getting CUDA memory errors with SDUI so I removed it as NMKD was working well enough anyway. Not sure if there was some sort of python conflict with the multiple installs; didn't really look too much into it and just went back to NMKD. Had similar errors when I tried to install a newer version of NMKD (I'm on 1.8.0 and I love it) so who knows? Anyway, thanks again for the great tutorials, really helpful stuff!

Mark Hunter

I am not sure what I am doing wrong. I cannot get the inpaint and draw tool to work right. I was wondering what I was doing wrong, so I went step by step with your instructions and all I get is a blob of color across her chest and not a top.

SPQRAeternum
>>99630153

oh, I get stuff like that too, that happens when the GPU runs out of memory and it's because you're doing an image that's too large. And once it happens you have to restart it. In that session it keeps happening even if you decrease the image size so it might seem like not working at all when it happens, there are some instructions in the message that helps: turning off turbo mode (takes longer to make images but it frees up 1gb of the gpu usage), setting a smaller size to the image (768x768) and using smaller size input images

SPQRAeternum
>>99645120

did you also set the other settings? I didn't write them down for the clothes but they were the same as above: Sampler: ddim, Inference Steps: 25, Guidance Scale: 4.4, Model: HassanBlend1.5, VAE: vae-ft-mse-840000-ema-pruned, Negative Prompt: lowres, drawing, Prompt Strength: 0.6, Fix Faces: GFPGANv1.3 The blob thing happens when the Guidance Scale & Prompt Strength are not set right. It's not an exact things though and maybe I just got lucky with those settings. For different prompts you might have to adjust those up or down a bit for the AI to focus. You might also have to adjust the prompt to be simpler. Adding too much stuff makes the AI think of too many different things. For the clothes part even in my example i think something like "photo of woman wearing purple top" would have been enough and probably better

Vamvader
>>99630153

I was getting the error even at the smallest settings though with SDU for some reason. NMKD works a little differently; sometimes I'll get the error if I try to run "generate" too quickly after my previous prompt has finished but I don't need to restart anything, I'll just wait a few seconds and try again and no error (if the config I've picked is reasonable anyway). I guess it's still flushing the VRAM or something. I don't think I have a turbo mode in there either but it's fast enough anyway. 1024 x 1024 images upscaled 100% with face correction on and over 0.5 weighting to the initial image tend to come out great with no errors, and NMKD warms you if your prompt is too long. What you don't get in NMKD is the ability to pick from your generated images within the GUI, you have to pick the image manually but it's not a deal breaker for me.

SPQRAeternum
>>99630153

ypu, that's what I was saying: if you get the error once, I think it will keep showing for that session no matter how low you make your settings. So it might seem like no settings work. Practically you just have to start with small images and keep going higher and figuring out what your pc can handle that way. At least that's for GPU mode. I never tried CPU mode, that's slower but might work with any size of image myself I'll stick with the web browser based GUIs, I think these will get updated more often as it's very easy to do features & fixes for them and there's also teamwork. The C# software-ish ones I don't think will keep up, the skill required to do cool things like that is much higher and it's more difficult for people to contribute to those projects

Mark Hunter
>>99645120

Thanks! That got it to work. I haven't got it perfect yet but that seems to have gotten me on the right track and now I just need to play around with it for a while.

Michael Jordan

Unrelated to the AI Tutorials (which are great), but in Alive 30 the touch trigger options have been removed... Any reason why?

SPQRAeternum
>>99691527

yes, it's in the post for Alive 30 :) I made ExtraTriggers for stuff like that, it has more updates & features and it's also free in case you want to build something and share it. In Alive it became obsolete and it was hindering progress on more important things

SPQRAeternum
>>99645120

nice! yeah, it's a lot of trial and error and it varies a bit based on the image & the prompt. I'm yet to make sense of those values myself, I go random with them until I find something that works and just fine tune it from there also very important is to uncheck Random while testing changes. This will make the AI work on the same variation each time you run it and you can see exactly how the changes you make influence the results. with Random on, it's hard to tell if your changes help or not