[Stable Diffusion] Prompt Sharing and Learning Thread

[Sepheyer]

Morning jog in the park. Shake shake shake..

View attachment 3249412 View attachment 3249413

I think I should explain. For a long time now I have been experimenting with mixing styles by using a lora that is anime or cartoon oriented with a photo realism checkpoint. Getting the extreme shapes and stylized look but with more realism. Lately, for a while I have been experimenting with using the refiner function in the same way. Meaning that I use a cartoon ckpt as the main ckpt and a photo realistic as refiner but I set the switch to happen earlier 0.3-0.5. This way I can achieve the same. However I seem to have found the limit of this concept. Doing a full body shot doesn't get me great result. Then I upscale in img2img with the photo realistic ckpt because hires fix doesn't work well with this process. I guess the alternative is to do something in img2img with controlnet to convert the comic/cartoon image into realism.

Why complicate things this much?

why the hell not?..
It's fun to experiment and try to push the envelope. This is what gives me the most enjoyment.

Example:

Without realistic "refiner"	With realistic "refiner" (0.4), and a faceswap (Jenna Jameson).
View attachment 3249464	View attachment 3249459

Indeed, i2i is the way for all kinds of shenanigans. I am still on SD kick, can't really wrap my head around SDXL - hence I use HS2 outrageous proportions and then i2i them into realistic using tile controlnet at around 80-90% strength:

Dunno if you've tried it yet, but if you need the motivation, then just do it. That fucker took my breath away. I compared it against depth and canny controlnets, and tile is by far the best.

 

[Mr-Fox]

Indeed, i2i is the way for all kinds of shenanigans. I am still on SD kick, can't really wrap my head around SDXL - hence I use HS2 outrageous proportions and then i2i them into realistic using tile controlnet at around 80-90% strength:

Dunno if you've tried it yet, but if you need the motivation, then just do it. That fucker took my breath away. I compared it against depth and canny controlnets, and tile is by far the best.

I have only experimented a little with using img2img for converting cartoon to realism but I'm on it now. I was really not happy with the result I got with Jasmine. I have had much better success with other images before so it's a valid method, just not for everything. As is usually the case with everything regarding SD.

Here's a couple different examples of using the refiner method:

( pirsusComicsStyle_hardStyle and refiner: photon_v1 switch 0.4)

You have been a naughty naughty girl... I'll go get the paddel.. giggidy..

 

[Dagg0th]

I have only experimented a little with using img2img for converting cartoon to realism but I'm on it now. I was really not happy with the result I got with Jasmine. I have had much better success with other images before so it's a valid method, just not for everything. As is usually the case with everything regarding SD.

Here's a couple different examples of using the refiner method:
View attachment 3249526 View attachment 3249539 View attachment 3249540
( pirsusComicsStyle_hardStyle and refiner: photon_v1 switch 0.4)

You have been a naughty naughty girl... I'll go get the paddel.. giggidy..

There is a LoRa that convert cartoonish images in t2i to 2.5D, after that it should be more easy to convert to realistic in i2i, that's the trick I used to make some hentai porn into realistic.
https://f95zone.to/threads/ai-art-show-us-your-ai-skill-no-teens.138575/post-11541639

I'm trying to find the Lora, but I think I already deleted it...

EDIT:
found it

You must be registered to see the links

Another trick is to create the cartoonish image, then HiRes it with a realistic model




after that, you can i2m for more realism I guess, more and more until it leaves the monitor for being hyperrealistic and comes back to life

BTW
Refiner never worked for me, don't even bother to use it anymore.

 

[Mr-Fox]

There is a LoRa that convert cartoonish images in t2i to 2.5D, after that it should be more easy to convert to realistic in i2i, that's the trick I used to make some hentai porn into realistic.
https://f95zone.to/threads/ai-art-show-us-your-ai-skill-no-teens.138575/post-11541639

I'm trying to find the Lora, but I think I already deleted it...

EDIT:
found it

You must be registered to see the links

Another trick is to create the cartoonish image, then HiRes it with a realistic model
View attachment 3249746

View attachment 3249747

after that, you can i2m for more realism I guess, more and more until it leaves the monitor for being hyperrealistic and comes back to life

BTW
Refiner never worked for me, don't even bother to use it anymore.

What do you mean by "hires it" ? Are you generating a cartoon image and then use it in controlnet lineart or similar with realistic ckpt? Or do you mean that you upscale with a realistic ckpt in img2img?

When I'm talking about refiner, I mean I use SD1.5 models and use a cartoon style for the stylized look and the extreme shapes and with refiner get more details and realism. The "trick" is to switch early.
I never use refiner for SDXL models though as it was first intended.

 

[Sepheyer]

What do you mean by "hires it" ? Are you generating a cartoon image and then use it in controlnet lineart or similar with realistic ckpt? Or do you mean that you upscale with a realistic ckpt in img2img?

When I'm talking about refiner, I mean I use SD1.5 models and use a cartoon style for the stylized look and the extreme shapes and with refiner get more details and realism. The "trick" is to switch early.
I never use refiner for SDXL models though as it was first intended.

Yea, ComfyUI makes it very easy to first render a cartoon with outsized proportions, then pull realistic on it. I used to do it, and then realized that HS2 lofi renders works just as good as a cartoon but gives more control.

 

[Mr-Fox]

Yea, ComfyUI makes it very easy to first render a cartoon with outsized proportions, then pull realistic on it. I used to do it, and then realized that HS2 lofi renders works just as good as a cartoon but gives more control.

I was looking at his image in png info to re-engineer what he was talking about and saw he is using a different extension that allows a pos and neg prompt. I simply use hiresfix. I tried kohya hiresfix but didn't like it as it completely changes the image composition and also added some unwanted strange visuals. I was not aware of this hires extension.

 

[Sepheyer]

A workflow:

3D render (HS2)	Clothes via gfx editor	Background via gfx editor	Img2Img

 

[me3]

Not really been getting the "cooperation" i'd like from the AI with this, loads of elements it either doesn't want to include or it turns them into an extra leg or arm for some reason...
Since it doesn't want to do both the pants/salvar and a see through "skirt" i guess we're sticking with the see through, less is more right...
Headband becomes a tiara, but probably not where eyes are going anyway. At least the hair is at a fairly correct length
Prompt is 99% from a wd14 caption, which i guess could be some of the problem. Probably shouldn't look too closely or count things

Spoiler: prompt

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

 

[Dagg0th]

What do you mean by "hires it" ? Are you generating a cartoon image and then use it in controlnet lineart or similar with realistic ckpt? Or do you mean that you upscale with a realistic ckpt in img2img?

When I'm talking about refiner, I mean I use SD1.5 models and use a cartoon style for the stylized look and the extreme shapes and with refiner get more details and realism. The "trick" is to switch early.
I never use refiner for SDXL models though as it was first intended.

What I mean is that in A1111 you can hi.Res using another chkpnt.
Probably it need to be enabled settings, you can add more prompt for Hi.Res and add more realistic tokens, previously unavailable in cartoonish chkpnts.

 

[Mr-Fox]

What I mean is that in A1111 you can hi.Res using another chkpnt.
Probably it need to be enabled settings, you can add more prompt for Hi.Res and add more realistic tokens, previously unavailable in cartoonish chkpnts.
View attachment 3250543

I found it in "UI alternatives". Thanks

 

[Mr-Fox]

Still using the refiner "method" I posted about in these images but I will try the hiresfix route.

Spoiler: More

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

 

[sharlotte]

With the original prompt using SDXL - prompt inside

 

[sharlotte]

same ckpt, using the revised prompt:
Not sure where the logo comes from as I did not use controlnet but she's way too skinny ;(

 

[sharlotte]

And now for combined prompts,still same CKPT

 

[Mr-Fox]

I have tested using hiresfix with a different ckpt that is realism oriented as the eminent Dagg0th posted about.
Then I did some fixes with inpaint and a small edit in photoshop and then upscaled it.
A little "cheating" but not that much.

 

[Delambo]

Man, I'm struggling to get anything I want in ComfyUI. Does anyone have a half way decent image of a 2 character sex pose that I can look over the workflow of?

 

[namhoang909]

I have tested using hiresfix with a different ckpt that is realism oriented as the eminent Dagg0th posted about.
Then I did some fixes with inpaint and a small edit in photoshop and then upscaled it.
A little "cheating" but not that much.

View attachment 3253671

look at those perfect fingers
Question:
I saw there are Ydetailer and DDetailer which replace Adetailer in ComfyUI, which one is better?
I have 3060 with 12gb vram, will that work for SDXL models? I want to generate and up-scale images, ..(post processing) not training.
Here is

You must be registered to see the links

, I haven't found the Ydetailer download link though.

 

[sharlotte]

Have not tried either in Comfy so cannot tell you.
As for SDXL, sure i've got the 3060 as well and there's no issue. In CUI it's running super fast Here some of the pics I generated yesterday just for a change from ladies

. In A1111, it's a bit slower, especially if you use controlnet, hires, adetailer at the same time but it's working fine.

 

[Synalon]

I've been trying to get some kind of "cabaret stage" type background on this for a day or two now, anybody else want to try?

 

[devilkkw]

I've been trying to get some kind of "cabaret stage" type background on this for a day or two now, anybody else want to try?

View attachment 3256695

I'll try with some other evenet's, but without using as background, it give me good result.
Prompt is simply: subject, type, perfoming on American got talent.
This is the base prompt, one exemple is: blonde braid haired woman, singer, performing on American got talent.
Also try some other most know show (the voice for example) and seem work well.