[Stable Diffusion] Prompt Sharing and Learning Thread

[me3]

Not really sure what you mean by "shit quality". Also, just flooding the prompts with "quality prompts" might not have the affect you think, specially not if they are conflicting. Clean up your prompt and see what is actually need, very often you don't need that many things to increase the details/quality, same goes for negative prompts. Everyone has a habit of just using a set of negative prompts they always use but the truth is that in many cases you won't actually need man/any.

As for the highres, it pretty much just increase the canvas and needs to fill it with something usually that means reusing the "primary" subject in your prompt. Enforcing the singular nature of that might work or you can try drawing out the background first and then the main subject when much of the space is already filled. Adding other details to fill background might work too.
If you purely want to make the original image bigger you can use img2img to scale it. Use highres in a much smaller scale first, then use img2img to outpaint or purely upscale it there

 

[Jimwalrus]

I need some help. I want high quality picture. 512x512 every picture comes alright. But the picure is small and quality is shit as shown below.

View attachment 2664469 View attachment 2664505

Then i send it to extras to upscale it to 2048x2048. It comes bigger but the quality is still shit. When i put size 1920 x 1080 this happens. View attachment 2664525
They are all beautiful but i just want one girl in the picture. Hires.fix does not help or i dont know how to use it.

For Hires.fix, just set it to x2 (or as high as your GPU will allow) - don't upscale in Extras, do it as part of the initial generation whenever you can*. Set the Denoising strength at no more than 0.4 (0.25 - 0.33 seems to work best for non-photorealistic, 0.1 - 0.25 for photorealistic)
Double the number of gen steps for the hires steps.
That should be a pretty good starting point, you can play with settings from there.

*It will try to upscale using the prompts, thus giving you some 'fine-tuning' steps on top of the gen steps.

 

[Jimwalrus]

For Hires.fix, just set it to x2 (or as high as your GPU will allow) - don't upscale in Extras, do it as part of the initial generation whenever you can*. Set the Denoising strength at no more than 0.4 (0.25 - 0.33 seems to work best for non-photorealistic, 0.1 - 0.25 for photorealistic)
Double the number of gen steps for the hires steps.
That should be a pretty good starting point, you can play with settings from there.

*It will try to upscale using the prompts, thus giving you some 'fine-tuning' steps on top of the gen steps.

And this is something I seem to keep pointing out - people aren't 1:1 aspect ratio. If you're wanting a standing person, set your initial image size as 512 pixels width and go for a height that's some multiple of 64 pixels above that.
Try 512x960, upscaled by x2 (to 1024x1920), 40 hires steps, denoising of 0.2.
Please post your results, she's very nice!

 

[Sepheyer]

And this is something I seem to keep pointing out - people aren't 1:1 aspect ratio. If you're wanting a standing person, set your initial image size as 512 pixels width and go for a height that's some multiple of 64 pixels above that.
Try 512x960, upscaled by x2 (to 1024x1920), 40 hires steps, denoising of 0.2.
Please post your results, she's very nice!

You've been pounding your fist on this since like post #7 here:

https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-9986405

 

[Sharinel]

I need some help. I want high quality picture. 512x512 every picture comes alright. But the picure is small and quality is shit as shown below.

View attachment 2664469 View attachment 2664505

Then i send it to extras to upscale it to 2048x2048. It comes bigger but the quality is still shit. When i put size 1920 x 1080 this happens. View attachment 2664525
They are all beautiful but i just want one girl in the picture. Hires.fix does not help or i dont know how to use it.

A lot of it depends on the model you use. Here's how the exact same prompt and seed looks on a variety of my models


As you can see models seem to interpret it very different, I have no idea why Stylejourney want's to give her the boobs of a small giantess.

Changing it to 512x768 has the following outcome



And swapping them to 768x512 changes focus to close in



Again, impressive from Stylejourney showing us the Triple Breasted Whore of Eroticon 7 (showing my age here)

And finally if I go to my normal for Landscape, which is 960x540 I start getting the issues with multiple people. You would then need to start going in to the prompt and changing things - for instance try solo:1.2 or 1girl or both

 

[Sepheyer]

Landscape solo girls can be done reliably with Comfy UI using its compose workflow. One image in three/five should give you want you want.

https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10706510

 

[Mr-Fox]

Not really sure what you mean by "shit quality". Also, just flooding the prompts with "quality prompts" might not have the affect you think, specially not if they are conflicting. Clean up your prompt and see what is actually need, very often you don't need that many things to increase the details/quality, same goes for negative prompts. Everyone has a habit of just using a set of negative prompts they always use but the truth is that in many cases you won't actually need man/any.

As for the highres, it pretty much just increase the canvas and needs to fill it with something usually that means reusing the "primary" subject in your prompt. Enforcing the singular nature of that might work or you can try drawing out the background first and then the main subject when much of the space is already filled. Adding other details to fill background might work too.
If you purely want to make the original image bigger you can use img2img to scale it. Use highres in a much smaller scale first, then use img2img to outpaint or purely upscale it there

I agree with most of what me3 said and he's on point about the prompt. However I have to strongly disagree that hiresfix is only increasing the canvas size.
This is simply not true. The point of hiresfix and what makes it superior to normal upscaling with extra tab or img2img is that it is part of the generative process and adds new pixels to the image while normal upscaling can't invent pixels that are not already there, so normal upscaling only increase the canvas size but hiresfix actually increase the image quality.
There's an add-on that's on my to do list to try out that is called

You must be registered to see the links

that is supposedly increasing the actual image quality just like hiresfix does and it has sliders and settings for color and maybe light etc. I have not tried it yet so I can't comment how it stacks up against hiresfix or if it could be a compliment. It is for img2img tab so you need to generate an image first so you might as well use hiresfix for this and then use Loopback Scaler to increase the visuals even further. I will have to test and see if it can be used successfully this way.

 

[Mr-Fox]

I didn't know that I have anything on Bleeter / LI - they are prolly leaching my twitter. I dont really know what these are.

I post collages on twitter under Efgypt

You must be registered to see the links

. Scratches my itch for landscapes and IRL commentary.

View attachment 2664327
(Comfy UI prompt included.)

Bleeter=Twitter Life Invader=Facebook (GTA V).

 

[Sharinel]

Bleeter=Twitter Life Invader=Facebook (GTA V).

Each individual word was in English, but I have no idea what you are saying here!

 

[Mr-Fox]

Each individual word was in English, but I have no idea what you are saying here!

I was only joking a bit. Instead of saying Twitter I said Bleeter that is the name in GTA V and Life Invader is Facebook..
It wasn't my intention to confuse people or make poor seph worry that someone was copying his content.

 

[me3]

I agree with most of what me3 said and he's on point about the prompt. However I have to strongly disagree that hiresfix is only increasing the canvas size.
This is simply not true. The point of hiresfix and what makes it superior to normal upscaling with extra tab or img2img is that it is part of the generative process and adds new pixels to the image while normal upscaling can't invent pixels that are not already there, so normal upscaling only increase the canvas size but hiresfix actually increase the image quality.
There's an add-on that's on my to do list to try out that is called

You must be registered to see the links

that is supposedly increasing the actual image quality just like hiresfix does and it has sliders and settings for color and maybe light etc. I have not tried it yet so I can't comment how it stacks up against hiresfix or if it could be a compliment. It is for img2img tab so you need to generate an image first so you might as well use hiresfix for this and then use Loopback Scaler to increase the visuals even further. I will have to test and see if it can be used successfully this way.

you're misunderstanding what i meant and poorly explained in terms of trying to keep it simple.
I meant canvas in the sense of area you have available to draw/put the image in, and that it wouldn't just blindly "redraw" the same image using more pixles.
Why i said it had to be filled with "something". If you stick within the bounds of its drawing "attention", highres will mostly increase canvas and image in a matching way, but if you go beyond that, canvas will keep increasing but the image won't instead addition new "images" will be placed along side within the new canvas (sticking with poor explanations)

(Not tested or looked into this in any way, but wouldn't surprise me if the highres works on a similar logic to how outpainting does as both are using the original prompt (outpaint can use and edited one as well though) to fill in the expanded area. Anyway irrelevant side note in this...)

 

[Mr-Fox]

you're misunderstanding what i meant and poorly explained in terms of trying to keep it simple.
I meant canvas in the sense of area you have available to draw/put the image in, and that it wouldn't just blindly "redraw" the same image using more pixles.
Why i said it had to be filled with "something". If you stick within the bounds of its drawing "attention", highres will mostly increase canvas and image in a matching way, but if you go beyond that, canvas will keep increasing but the image won't instead addition new "images" will be placed along side within the new canvas (sticking with poor explanations)

(Not tested or looked into this in any way, but wouldn't surprise me if the highres works on a similar logic to how outpainting does as both are using the original prompt (outpaint can use and edited one as well though) to fill in the expanded area. Anyway irrelevant side note in this...)

My point was only that hiresfix is generating new pixels that ads to the image and increase the image quality and not only increase the image size, while normal upscaling in extra tab or img2img can't invent or generate new pixels that is not already in the source image so it doesn't increase the image quality, only the size. This is why hiresfix is superior to normal upscaling. Normal upscaling doesn't make the image sharper or ads more detail but hiresfix does by generating new pixels. I hope I explained it clearly.

 

[devilkkw]

Guy's i had strange issue maybe someone can help me.
Updated Nvidia driver to 535.98 and strange problem come:
showing image take about 40 sec, generation is faster but pushing in view is really slow.
Before update driver all process was done in about 11 sec. but now generation take 10 sec to reach 100% and 40 sec to show.

I don't have change anything, only driver.

Also checked now i'm able to generate image at high res without going on cuda error, and as strange seem pytorch memory is variable:

before driver update pytorch max was about 6Gb, and never change.( max resolution before getting out of cuda was 1152x896, result in 11 sec)
after driver update i reach 12Gb, but it change seem based on resolution of image i set.(i reach 1920x 1260, not cuda error, result in 3 min, where 35 sec is generation, and rest of time is for pushing in viewer)

Can someone have done some test on last driver? have you same issue?

 

[Mr-Fox]

I need some help. I want high quality picture. 512x512 every picture comes alright. But the picure is small and quality is shit as shown below.

View attachment 2664469 View attachment 2664505

Then i send it to extras to upscale it to 2048x2048. It comes bigger but the quality is still shit. When i put size 1920 x 1080 this happens. View attachment 2664525
They are all beautiful but i just want one girl in the picture. Hires.fix does not help or i dont know how to use it.

I think she is cute and you did a very good job with your images, even if it's not the desired result. It only means that it's a WIP (work in progress). I agree with me3 though, that it's best to simplify the prompt and only add tokens as you need it while generating images and after each change and add things towards the end goal. With that being said, SD has some clear tendencies and it's very time consuming to create a prompt from scratch so it's very understandable that people tend to reuse parts of their prompt such as the typical negatives that is so common.

I did some testing myself and have reworked the prompt a little but not too much. I was trying to force a whole body shot by adding weight to "high heels" but SD had other ideas.. Well played is all I can say..


Also, apparently while streaking in a museum it's best to take off those noisy heels..

 

[me3]

Guy's i had strange issue maybe someone can help me.
Updated Nvidia driver to 535.98 and strange problem come:
showing image take about 40 sec, generation is faster but pushing in view is really slow.
Before update driver all process was done in about 11 sec. but now generation take 10 sec to reach 100% and 40 sec to show.

I don't have change anything, only driver.

Also checked now i'm able to generate image at high res without going on cuda error, and as strange seem pytorch memory is variable:

before driver update pytorch max was about 6Gb, and never change.( max resolution before getting out of cuda was 1152x896, result in 11 sec)
after driver update i reach 12Gb, but it change seem based on resolution of image i set.(i reach 1920x 1260, not cuda error, result in 3 min, where 35 sec is generation, and rest of time is for pushing in viewer)

Can someone have done some test on last driver? have you same issue?

If you got 12gb vram you should have been able to generate much larger images than that before. I could create images adding up to 2k (width+height) on 6gb, can get to almost 1,5k on just 2gb.
Something must have kept you from using all the vram before

 

[Mr-Fox]

Guy's i had strange issue maybe someone can help me.
Updated Nvidia driver to 535.98 and strange problem come:
showing image take about 40 sec, generation is faster but pushing in view is really slow.
Before update driver all process was done in about 11 sec. but now generation take 10 sec to reach 100% and 40 sec to show.

I don't have change anything, only driver.

Also checked now i'm able to generate image at high res without going on cuda error, and as strange seem pytorch memory is variable:

before driver update pytorch max was about 6Gb, and never change.( max resolution before getting out of cuda was 1152x896, result in 11 sec)
after driver update i reach 12Gb, but it change seem based on resolution of image i set.(i reach 1920x 1260, not cuda error, result in 3 min, where 35 sec is generation, and rest of time is for pushing in viewer)

Can someone have done some test on last driver? have you same issue?

It doesn't sound as an issue but rather a positive if you can reach higher resolution.. You are talking about seconds while some people sitting here and having to wait 30 minutes... What you say is "pushing into view" is probably only the hires steps finishing. If you can reach a higher resolution it will of course take longer and it will require more vram.
If everything is working and you get nice images I would just let it be. It wasn't clear to me what you perceive as an issue.
Just my opinion.

 

[Sharinel]

Also, apparently while streaking in a museum it's best to take off those noisy heels..

You haven't been to a museum recently, that's probably an exhibit

 

[Mr-Fox]

You haven't been to a museum recently, that's probably an exhibit

Modern Contemporary "art"..

 

[me3]

I think she is cute and you did a very good job with your images, even if it's not the desired result. It only means that it's a WIP (work in progress). I agree with me3 though, that it's best to simplify the prompt and only add tokens as you need it while generating images and after each change and add things towards the end goal. With that being said, SD has some clear tendencies and it's very time consuming to create a prompt from scratch so it's very understandable that people tend to reuse parts of their prompt such as the typical negatives that is so common.

I did some testing myself and have reworked the prompt a little but not too much. I was trying to force a whole body shot by adding weight to "high heels" but SD had other ideas.. Well played is all I can say..

View attachment 2665572
Also, apparently while streaking in a museum it's best to take off those noisy heels..

Could try to delay when she's added to the image[what:when], with mention of shoes you have a "head to toe" prompt so might work. Has a habit of cutting off at the knees though and it might push her too much into the picture so we get that very nice facial mutilation that's all to common

 

[Mr-Fox]

Could try to delay when she's added to the image[what:when], with mention of shoes you have a "head to toe" prompt so might work. Has a habit of cutting off at the knees though and it might push her too much into the picture so we get that very nice facial mutilation that's all to common

Very good idea. I'll leave that to ronimikael to figure out though as it's his project. I hope that we all have given him the guidance he was looking for. I gotta say I love how awesome and responsive this thread has become. Everyone is being very helpful and has good tip's and ideas.