[Stable Diffusion] Prompt Sharing and Learning Thread

[devilkkw]

Thank you. If your GPU can handle it, Hires Fix is superior to everything else. use either 512p or 768p and then use Hires mutliplier to do the rest. Start with 2 and if it works increase to 2.5 etc until you get cuda memory error. Then backtrack 0.1 until you don't get the error. Now you know your Hires limit. I use either

You must be registered to see the links

or

You must be registered to see the links

for upscaler.
For Ultrasharp choose to download zip and then unpack in Stable-Diffusion\stable-diffusion-webui\models\ESRGAN.
NMKD Superscale needs to be placed in this folder also. You need to restart webui you can do so in extensions/reload ui.
Read these posts:
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10333003 (big tutorial)
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10306143 (cuda error fix)

Really interesting, i read it two 2 ago, but for hires nothing work, my actually max is 1.2x and don't do so much on result.
For these reason i start use sd without any extra(hires,fix face) and make my merged model for having good(i hope) result without using it.
Thank you so much for suggestion, you look very expertise on this.
A fox for you.

Prompt: concept cgi photo, anthropomorphic, anthro, woman as fox
Negative prompt: (bad quality, worst quality, low quality:1.2) render
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 11,Size: 768x1024

I tested it on my custom checkpoint, dont know if work great in other.
It also good for making cosplay with changing "fox" with character you want, also do good result by changing "woman" in your favorite actor/actress.

Have a fun

 

[Mr-Fox]

Really interesting, i read it two 2 ago, but for hires nothing work, my actually max is 1.2x and don't do so much on result.
For these reason i start use sd without any extra(hires,fix face) and make my merged model for having good(i hope) result without using it.
Thank you so much for suggestion, you look very expertise on this.
A fox for you.
View attachment 2480843
Prompt: concept cgi photo, anthropomorphic, anthro, woman as fox
Negative prompt: (bad quality, worst quality, low quality:1.2) render
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 11,Size: 768x1024

I tested it on my custom checkpoint, dont know if work great in other.
It also good for making cosplay with changing "fox" with character you want, also do good result by changing "woman" in your favorite actor/actress.

Have a fun

Thank you very much for the cute fox lady, she looks amazing. Good composition, good details and fidelity. This is way better than what most can create even if they have much better hardware. It all comes down to knowledge, effort and a critical mind.
Wich GPU do you have and amount of Vram? The Hires multiplier is mostly for higher resolution. What makes the big difference is the Hires steps. So even if you can only do 1.2, this is not a reason to not use it.
If you use 15-20 Hires steps with 0.4-0.6 denoise with the 1.2 multiplier I bet it's an improvement over not using Hires fix.
One thing also that makes a big difference is to not use Restore Faces with Hi res, test out only postprocessing.
Either GFPGAN or Codeformer. I use GFPGAN mostly. Get the new VAE from this post.
In settings/Stable Diffusion " Enable quantization in K samplers for sharper and cleaner results ". Try out clipskip in Settings/Stable Diffusion it can do wonders.

 

[modine2021]

You must be registered to see the links

Sweaty muscular cheerleader in sports arena. Prompt by AIArtjak at civitai experience comment section, slightly edited by me and adding some lora's etc.

View attachment 2407094

OOOWEEE

 

[devilkkw]

Thank you very much for the cute fox lady, she looks amazing. Good composition, good details and fidelity. This is way better than what most can create even if they have much better hardware. It all comes down to knowledge, effort and a critical mind.
Wich GPU do you have and amount of Vram? The Hires multiplier is mostly for higher resolution. What makes the big difference is the Hires steps. So even if you can only do 1.2, this is not a reason to not use it.
If you use 15-20 Hires steps with 0.4-0.6 denoise with the 1.2 multiplier I bet it's an improvement over not using Hires fix.
One thing also that makes a big difference is to not use Restore Faces with Hi res, test out only postprocessing.
Either GFPGAN or Codeformer. I use GFPGAN mostly. Get the new VAE from this post.
In settings/Stable Diffusion " Enable quantization in K samplers for sharper and cleaner results ". Try out clipskip in Settings/Stable Diffusion it can do wonders.

I give a chance to it, need some try to understand it better, also thank you for upscaler you linked, work really great.
For hires you use latent? or some different?
The setting is as you suggest, also VAE is included in my model.
I take some time to see how hires work on my model.
My gpu is rtx3060 laptop 6gb vram, so i don't have to pretend so much.

 

[Mr-Fox]

I give a chance to it, need some try to understand it better, also thank you for upscaler you linked, work really great.
For hires you use latent? or some different?
The setting is as you suggest, also VAE is included in my model.
I take some time to see how hires work on my model.
My gpu is rtx3060 laptop 6gb vram, so i don't have to pretend so much.

For Hires Fix I use the upscalers I linked to. After restarting the webui they will show up in the same list as latent.
I use the "8x NMKD Superscale 150000" mostly. You have the same amount of Vram as Sepheyer. He posts really great images and is also very helpful with answering questions and giving tips. He's the one who created this thread also btw. I need to check out your merge checkpoints more, dude you have 5. Amazing what one can do with a little laptop..

 

[azrazr]

Respectfully, two questions.

- is it possible to download a modified and improved stable diffusion, simply copying the files and folder to the Pc?? taking into account the speed of the improvements, perhaps there can be a considerable gain in performance and quality (without wanting to sound like a leech)






- is this considered illegal or inappropriate ?? what would be the desirable limits ???

 

[Sepheyer]

Respectfully, two questions.

- is it possible to download a modified and improved stable diffusion, simply copying the files and folder to the Pc?? taking into account the speed of the improvements, perhaps there can be a considerable gain in performance and quality (without wanting to sound like a leech)

View attachment 2482181




- is this considered illegal or inappropriate ?? what would be the desirable limits ???

View attachment 2482177

"Underage" portrayals are governed by F95's rule 1.7:

• 1.7. Content (a game, image, video, story, animation, etc.) depicting nudity or sexualization or sexual/erotic activity involving prepubescent character(s) is prohibited. A "prepubescent character" is one that visually shows no sign of puberty, or if non-visual is otherwise clearly intended to represent a prepubescent character. 2D and 'unrealistic' depictions are generally looked upon more favorably. As this involves some discretion, if you are unsure ask a Staff Member.

To me the image is clearly not a "prepubescent character" because the tits are a clear sign of puberty. And since my opinion doesn't really matter the rule 1.7 states that in case of a dispute the Staff Member is the one with the final opinion.

Additionally, your intentions matter as stated here: "if non-visual is otherwise clearly intended to represent a prepubescent character".

Note, the requirement is to meet pre-puberty threshold, not the age of consent. The character doesn't need to be 21 years old, per the site's rules. So, if you insist that the image is of a say 18-year old, you are fine. If you insist she is a 4 year old, then that's a rule break.

 

[bobsen]

When I'm using img2img to try to add details, or apply a different model to an existing picture, it almost always freaks out and draws nipples everywhere. Anyone fixed this?

 

[Mr-Fox]

When I'm using img2img to try to add details, or apply a different model to an existing picture, it almost always freaks out and draws nipples everywhere. Anyone fixed this?

Do you mean fix a generated image with inpaint? You need to elaborate and be more specific. For example what do you mean by "apply a different model" ? If you are talking about changing the appearance of a "character " into something else, this might be the job for pix2pix. I have not messed with that one yet myself so I would need to direct you to youtube tutorials.

 

[Mr-Fox]

Instruct pix2pix. (Turn a character into anything you want with text).

The checkpoint:
Fair warning large filesize 7+Gb

You must be registered to see the links

You must be registered to see the links

(faster)

You must be registered to see the links

You must be registered to see the links

(includes ot tips and more in depth about pix2pix).

 

[devilkkw]

For Hires Fix I use the upscalers I linked to. After restarting the webui they will show up in the same list as latent.
I use the "8x NMKD Superscale 150000" mostly. You have the same amount of Vram as Sepheyer. He posts really great images and is also very helpful with answering questions and giving tips. He's the one who created this thread also btw. I need to check out your merge checkpoints more, dude you have 5. Amazing what one can do with a little laptop..

I'm testing it, but don't know why, now hires is 1.1 max. no problem for testing. seem work good on photorealistic, but for fantasy or anime, i get better result without it. need experimenting more.
I'm glad you want to test my merge, i suggest use my last (

You must be registered to see the links

) because is a merge of all other at different ratio and is
really simple to use(i focused my merge on this way, you can see how simple is with last 3 sample images on civitai model page).
Also if you have a time, giving feedback is appreciated, and let me know where come some problem, i'll try to fix.

I also work on .py script, actually for change merging formula, but work with size of tensor, shape and layer is little hard.

I saw some post of "Sepheyer" and i'm amazed with some of it.I hope see more work.
Is there any 3d for model comparison? i usually check

You must be registered to see the links

for fast comparison, but many model missed.

Thank you again for patience and for giving good suggestion to improve my skill.

 

[atheran]

Okay, for the past couple of days I've been playing around with CUI and if not for a couple of things, I'd probably switch over to it completely from A1111. But I'm not known to be smart when trying to speak google, so.. anyone around has found a working face restore workaround for CUI? That and highresfix are the most serious issues I have with it.

 

[devilkkw]

This is a little test without and with hires.fix

Spoiler: Prompt

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

Spoiler: No Hireres

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

Spoiler: With Hires

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

Not much difference,but i need to exeperimenting more.

 

[Sepheyer]

Okay, for the past couple of days I've been playing around with CUI and if not for a couple of things, I'd probably switch over to it completely from A1111. But I'm not known to be smart when trying to speak google, so.. anyone around has found a working face restore workaround for CUI? That and highresfix are the most serious issues I have with it.

What's CUI? Never heard of it.

 

[atheran]

What's CUI? Never heard of it.

ComfyUI, apologies.

 

[Sepheyer]

Ravenclaw wanna-b.

So, there is this cool feature: { blonde | brunet } really drops variety on you and lets you cover a wide range of things you want to see. Do take a moment to see what the prompts start looking like with this feature being there:

Spoiler: Prompts

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

Spoiler: More

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

Spoiler: CUI Prompts + Workflow

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

 

[devilkkw]

Ravenclaw wanna-b.

So, there is this cool feature: { blonde | brunet } really drops variety on you and lets you cover a wide range of things you want to see. Do take a moment to see what the prompts start looking like with this feature being there:

Spoiler: Prompts

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

View attachment 2482890

Spoiler: More

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

Spoiler: CUI Prompts + Workflow

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

great results. i use wildcards inteased of {value|value}, i'm on automatic1111. Don't know what is comfyUI, i also use nmkd, but my favorite is a1111.

 

[sharlotte]

In relation to the two images below, which I posted in https://f95zone.to/threads/ai-art-show-us-your-ai-skill.138575/post-10343577 , how can I 'inpaint' or any other technique to correct the leg/feet and the missing arm. I can use inpaint without any issues to modify eyes, faces, hands and background but can't, for whatever reason, do the same for these. Any hints appreciated. I'm sure i'm missing something obvious.

 

[Sepheyer]

In relation to the two images below, which I posted in https://f95zone.to/threads/ai-art-show-us-your-ai-skill.138575/post-10343577 , how can I 'inpaint' or any other technique to correct the leg/feet and the missing arm. I can use inpaint without any issues to modify eyes, faces, hands and background but can't, for whatever reason, do the same for these. Any hints appreciated. I'm sure i'm missing something obvious.
View attachment 2483049 View attachment 2483050

What I am about to say is not a fix. Rather a philosophical approach where the fix is not to change the corrupt image, no matter how good it is, but to generate new ones.

So, as an experiment, can you generate another 50 images to see if you will end up actually discarding these two images because you found better ones?

I am saying this because in my experience every available fix is too high touch and, the worst part, doesn't guarantee successful outcome.

 

[bobsen]

Do you mean fix a generated image with inpaint? You need to elaborate and be more specific. For example what do you mean by "apply a different model" ? If you are talking about changing the appearance of a "character " into something else, this might be the job for pix2pix. I have not messed with that one yet myself so I would need to direct you to youtube tutorials.

I prompted a picture using the standard stable diffusion checkpoint, then I changed to RealisticVision to get it to look more lifelike. At that point it freaks out, and changes it into an eldritch horror instead.