[Stable Diffusion] Prompt Sharing and Learning Thread

fr34ky

Active Member
Oct 29, 2017
812
2,189
I've got an RTX 3060 with 12GB of VRAM, I'll give it a try this evening (once Mrs Jimwalrus is asleep!)

Edited to add: For whatever reason I cannot get it to InstructPix2Pix to show in my SD WebUI. Very odd.
For the time being I'll have to leave it at that, sorry.
Be sure that you have the last automatic1111 version installed.

I found out yesterday that that in order to make that 'git pull' to work, you need to clone repository in github when you make the installation of Automatic1111 for the first time, if you downloaded the zip and pasted the files in a the folder to make the installation, git pull if won't work if you added it on the .bat file so you won't have the updated version.

I'm saying this cause I've just updated my automatic1111 and that pix2pix suddenly appearead in the webui, so clearly that was the problem.
 
  • Like
Reactions: Jimwalrus

Sepheyer

Well-Known Member
Dec 21, 2020
1,570
3,766
The workout sets have a lot of unusable pieces (at least for me). But sometimes there are these gems that make the whole thing worthwhile.
Model: PFG ( )
You don't have permission to view the spoiler content. Log in or register now.
You don't have permission to view the spoiler content. Log in or register now.
00034-2724991221-sexy woman from Sweden. _face perfect, European. European features._(full hea...png
You don't have permission to view the spoiler content. Log in or register now.
 
Sep 10, 2022
227
267
Hey guys, sorry to bother but recently I tried to use this model to make some nsfw MHA art, but my art never look as good as the image from OP, any ideas on how his looks so good?


edit: here are some examples
edit2: apparently removing the negative prompts made it better, still bad though
 
Last edited:
Sep 10, 2022
227
267
You need to be sure you use the trigger words and you MUST use negative prompts.

Here are some pictures I created with that model as an example, I added some extra juice but the important part are the negative prompts.

View attachment 2349760


You don't have permission to view the spoiler content. Log in or register now.

You don't have permission to view the spoiler content. Log in or register now.
your look really great, but comparing to the link examples look nothing like the style of bha (the objective of the model), also I tried to use your prompt and I received this error: OSError: [Errno 36] File name too long
 
Sep 10, 2022
227
267
You need to be sure you use the trigger words and you MUST use negative prompts.

Here are some pictures I created with that model as an example, I added some extra juice but the important part are the negative prompts.

View attachment 2349760


You don't have permission to view the spoiler content. Log in or register now.

You don't have permission to view the spoiler content. Log in or register now.


Edit: sorry, I didn't pay much attention to your post, I was too horrified by one of the examples, after this post I checked your prompts and there are 2 problems:

This nosense was in the prompt, this probably made so ugly, just erase this from your prompt: <lora:horikoshiKouheiMyHero_v1:1>
Other problem is that you didn't choose the correct model, I saw that you were using a model called 'anythingfurry_1'

This are the complete details I got from your picture:

<lora:horikoshiKouheiMyHero_v1:1> ochako uraraka from boku no hero academia, horikoshi kouhei
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 294318830, Size: 512x512, Model hash: 3edd27788a, Model: anythingfurry_1, Clip skip: 2
isn't <lora:horikoshiKouheiMyHero_v1:1> how the SD knows that its supposed to use the this lora model instead of another? this is what happens when I click on it in it's tab, or just having on my folder is already enough? and yeah I used the wrong model but I used the Anything one (just like the creator of it) and it was still bad
 
Sep 10, 2022
227
267
I made a generic prompt in seconds with just 20 steps just to show you that it can look normal and not that ugly thing.

You can add more steps details and try different render method, and more prompts, they usually post pictures that have a lot of work to make the model look as best as possible for the downloaders.

I've just copied the prompt from a picture in the comments for this one:

View attachment 2349836
You don't have permission to view the spoiler content. Log in or register now.
can't use the prompt that you used, I get this error OSError: [Errno 36] File name too long
 

Sepheyer

Well-Known Member
Dec 21, 2020
1,570
3,766
Hey guys, sorry to bother but recently I tried to use this model to make some nsfw MHA art, but my art never look as good as the image from OP, any ideas on how his looks so good?


edit: here are some examples
edit2: apparently removing the negative prompts made it better, still bad though
Any chance you can benchmark if your installation/setup is correct by running one of the prompts that are published in this thread? If you get the same images as the respective post contains, we'll know that your setup is fine and we merely need to troubleshoot your model. Otherwise, we'll know that there is something up with the setup. I just don't have the MHA model downloaded and installed, sorry.
 

fr34ky

Active Member
Oct 29, 2017
812
2,189
can't use the prompt that you used, I get this error OSError: [Errno 36] File name too long
As promised, I tested it and could achieve the results.
03233-378664745-horikoshi kouhei, (1girl), uraraka ochako, absurdres, bangs, blush, bodysuit, ...png

To be honest I took all this time to do this because I wanted to understand how to use those Lora, since it's a topic I'm interested in, not to be nice or anything. You only have to follow the instruction the guy give in the model, he says this:

How to use LoRA's in auto1111:
  • Update webui (use git pull like here or redownload it)
  • Copy the file to stable-diffusion-webui/models/lora
  • Select your LoRA like in this video
  • Make sure to change the weight (by default it's :1 which is usually too high)
Go and read that stuff and watch the video he suggests, if you do that exactly you'll be able to use it.

My recommendations:
- download the model he used in combination with the lora, it's called anything v4.5
- Just copy the prompts and other details from this file -> go to png info and drag the picture in your webUI and you'll have everything there.
- update your webUI if necesary, this is very important
- follow what he does in the video
- what he mentions about 'the weights' it's very important too, you can see he edits part of the Lora prompts on the video, that is key.

If you don't touch the weights you will get something like this:
03237-3161306128-horikoshi kouhei, (1girl), uraraka ochako, absurdres, bangs, blush, bodysuit,...png
 
Last edited:

Sepheyer

Well-Known Member
Dec 21, 2020
1,570
3,766
Model Comparison

Same prompt, different models.
You don't have permission to view the spoiler content. Log in or register now.
PFG

01035-4177220714-ultrarealistic photography, 4k, 8k, 16k, gtx, 135mm lens.__gorgeous woman._((...png
You don't have permission to view the spoiler content. Log in or register now.
 
Last edited:

Sepheyer

Well-Known Member
Dec 21, 2020
1,570
3,766
Gents, what prompt would give me tits sticking out from a crop teeshirt? I am looking for extreme underboob where the nipples are fully showing. Any ideas?
 
  • Thinking Face
Reactions: Jimwalrus

fr34ky

Active Member
Oct 29, 2017
812
2,189
Random comment about something I'm testing: You can create a .txt with prompts and then iterate the entire txt file, that way you can try many concepts while you are doing other things, then come back and look if there is something interesting.

It's a nice way to test new models, if you have some standarized prompts you always use, you can see how the new models responds to each one of them.
 

Jimwalrus

Well-Known Member
Sep 15, 2021
1,042
3,984
Gents, what prompt would give me tits sticking out from a crop teeshirt? I am looking for extreme underboob where the nipples are fully showing. Any ideas?
I've had a good try at this, but am also having no luck.
I've tried forcing exposed nipples, calling it an "extreme cropped top" and many, many others. Even getting underboob has proven impossible.

For reference, I started with the prompts from post #51 and tried a couple of models (PFG and F222).
I think that kind of outfit may be insufficiently present in the original training data. Time for a bit of Textual Inversion perhaps?
 
  • Like
Reactions: Sepheyer