[Stable Diffusion] Prompt Sharing and Learning Thread

[Jimwalrus]

I love happy accidents! I was messing around with 'Styles' which allows you to save a set of prompts/nrompts as a quick-select option. I'd created one based on my "StableAir Sexy Stewardesses".
I accidentally left it on while trying to generate a redheaded, green-eyed Taylor Swift:

Useful tip if you find a style you want to keep using - save it as a Style! Who knew!

Prompts in PngInfo as usual

 

[Sepheyer]

Getting Rid of Unwanted Hip/Thighs Lines

Seen those lines in your pics?


In collab with Mr-Fox, here is the negative prompt to address these:

((tanlines, tan lines, vcut, dickroot)).

It drastically reduces the number of these defects per set.

Spoiler: Result

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

 

[Sepheyer]

Different Results Yet the Same Prompt?

That's because xformers are enabled:

"When xformers is enabled image generation can be faster, but can also lead to image generation no longer being deterministic. That means if you generate an image with the exact same settings as before the new image may be slightly different."

You must be registered to see the links

 

[Sepheyer]

Tip for Inpainting

A quick explanation by Aitrepreneur channel:

You must be registered to see the links

 

[Sepheyer]

Edit: I am starting to think these pink spots are model-related artifacts. I barely see these using other models, but they are very common on PFG.
-
Would anyone have a theory what those pink spots (on the left hip) are? Lens flare? Sunspot? Reflection?

I'd love to yank these out during the image generation, curing these in photoshop/krita just sucks.

 

[Sepheyer]

CharTurner - Character Turnaround helper

The creator advises this is still raw'ish and WIP:

You must be registered to see the links

 

[Mr-Fox]

Would anyone have a theory what those pink spots (on the left hip) are? Lens flare? Sunspot? Reflection?

I'd love to yank these out during the image generation, curing these in photoshop/krita just sucks.

View attachment 2367549

It looks like a bit of mismatch.
When things like this happen I play around with sample steps and cfg scale. Increase the sample steps a couple of clicks and try again, then repeat until it's either gets better or different result. Also going back to the same sample steps and then testing different cgf scale can sometimes fix it. If nothing helps then go back to the same settings again and try small changes in the prompt. Rinse and repeat..
You can also try to just batch out lets say 10 with the same settings and see if at least one comes without the mismatched color.

 

[Sepheyer]

Removing Background

So, inside the Exntensions tab there is this "ABG_extension" script that removes the backgrounds once the image has been generated. Once you instal ABG then go to txt2img -> Scripts -> ABG Remover.

Naturally, the results depend on the granularity of the image. It will absolutely striggle with "anything hair". Otherwise it does a really good job.

The png prompt remains unaltered and you can always recreate the entire image, including its background, from the stripped background file.

Spoiler: More

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

 

[Mr-Fox]

Also you can send the image to "img2img". This is where you are using that image as a base or starting point and working with a new prompt specific to this image. You can try to tell the engine to remove the mismatch.
This video has some useful tips:

You must be registered to see the links

 

[Jimwalrus]

Would anyone have a theory what those pink spots (on the left hip) are? Lens flare? Sunspot? Reflection?

I'd love to yank these out during the image generation, curing these in photoshop/krita just sucks.

View attachment 2367549

A couple more steps during initial generation usually gets rid of them - they're mostly an artefact of the very small number of iters (~20) you normally use.
Also, I use nrompts of "bright spot" or "lens flare"

 

[Mr-Fox]

I have used seph's excelent schoolgirl prompt as an experiment subject.


I'm using URPM (UltraRealisticPornMerge) and wanted to see the result from this model/checkpoint. I also wanted to test it in the samplers that I have been using so far, Euler a and Heun.

Spoiler: Euler a and Heun

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

Next I wanted to test out using different Stars names with different hair color with the same prompt and "seed", getting the same style but different character. This gave very inconsistent with those samplers
wich led me to try out other ones.
Moving down through the list I then landed on "DPM++SDE". It gave some very interesting results...

Spoiler: DPM++SDE

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

I'm still trying out different tags in the prompt and exploring the functions. Very interesting stuff indeed.

 

[Sepheyer]

How to Compose Images

You must be registered to see the links

 

[Mr-Fox]

"Oopsee.." "one fell out.."

 

[HardcoreCuddler]

Spoiler: Guys I got it to work with this launch file....someone please stop me...

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

 

[Mr-Fox]

I have experimented with img2img, turning a render into a photo.
The subject is Yvette from darkhounds amazing renpy creation Holiday Island.

Original:


Generated:

Spoiler: Upscaled and some shop magic

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

 

[fr34ky]

Hey guys, after many weeks of trying and failing I could finally train a model with LORA with my 2060 8GB VRAM, still making some adjustments and learning how to get the best pictures for the training, but the great thing is that it takes only 15 minutes or less to train a model, so many new possibilities open with training! There are many girls that are not included in any model, like street fighter girls, or some other video games... Using this small LORAs we can add those girls to any model we want.

At the moment I've only trained LORAs with private people, so I won't be posting pictures here of them... still working to get the best results tho.

I could achieve the LORA model using this video from AItrepreneour.

You must be registered to see the links

Another cool thing is that he's not using automatic1111 which pissed everyone off with the constant buggy updates, so for training the best thing to do is use this new program he shows in the video.

The installation has many steps but it works 100%. Any doubts about how to install or any step just ask me

 

[Sepheyer]

Chokers are hot.

(Edited version, prompt below)

Spoiler: More

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

Spoiler: Prompt

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

 

[Sepheyer]

Shades are hot.

Spoiler: More

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

 

[fr34ky]

Chokers are hot.

(Edited version, prompt below)
View attachment 2369473

Spoiler: More

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

Spoiler: Prompt

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

I'm glad I'm not the only one with that soft stomach kink

 

[Mr-Fox]

Fiona Genesis 9 Daz Studio Character, img2img.

Spoiler: Prompt

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