[Stable Diffusion] Prompt Sharing and Learning Thread

[Mr-Fox]

"Luxurious modern mansion" gives you this..


A tiny shed..

Spoiler: More

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

 

[me3]

I've tried all the fixes on Git & Reddit, so it does seem like a 'Nuke It From Orbit' approach is the only one left to me.

Fortunately I've enough disk space to store all the Checkpoint / LoRA etc files so I can drop them back in once done.

there's a 1.3.1 and it has some "fixes" relating to loras, no idea if they will have any impact for you, but if it's already broken you can't really make that much worse right

 

[me3]

"Luxurious modern mansion" gives you this..

View attachment 2667770
A tiny shed..

Spoiler: More

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

ssshh, don't tell anyone, there's a secret underground lair, that's just the totally believable entrence which noone find suspitious when large amounts of ppl enter and doesn't get full

 

[Mr-Fox]

Yes, i know, also sd1.5 model default is trained in 512x512 image, and sd 2.* on 768x768, but this is image for train, i start getting bad result on sd1.5 going upper to 1980x1080, but it seem related to aspect ratio, and for bad result i mean if you work on subject getting double. but if i work on landscape or interior, the result is good. Every without hi.res fix. And a lot of merging of my model with block weighted.
I think no one use the standard sd1.5 for generation actually.

To avoid multiple girls, try (1girl:1.5) and (more than one girl:1.5) (multiple girls:1.5) in negative. And to avoid getting only lying down poses in landscape format image use (standing:1.5) etc.

 

[Jimwalrus]

there's a 1.3.1 and it has some "fixes" relating to loras, no idea if they will have any impact for you, but if it's already broken you can't really make that much worse right

Did a git pull today and the latest version is already at 1.3.2
The Lycoris tab is separated out from the LoRA tab, and LoRAs work now.
Looks like all I need to do is work out which of the 400+ models in the LoRA folder are actually Lycoris and move them.
Still waaayy preferable to a clean reinstall!

 

[Jimwalrus]

Did a git pull today and the latest version is already at 1.3.2
The Lycoris tab is separated out from the LoRA tab, and LoRAs work now.
Looks like all I need to do is work out which of the 400+ models in the LoRA folder are actually Lycoris and move them.
Still waaayy preferable to a clean reinstall!

Thank fuck that's done, Clone Airways is ready to take to the ((((friendly)))) skies once more...

 

[Mr-Fox]

 

[Jimwalrus]

View attachment 2669635

The ripples on the water look superb.

 

[Sepheyer]

The ripples on the water look superb.

That and the tile on the pool's actual wall.

 

[Mr-Fox]

Thanks guys. Though.. A blonde bombshell with huge boobas .. And everyone is like.. Hey look at the water.. And the lovely tile work, must have been expensive.. Those guys building it, how talented..

 

[me3]

Been doing some testing to how training on different models affect the end result. Thanks to an update that added 3 hours to my already long training time it took a lot longer than hoped...
(each image is marked with the model used to generate the image)

Train on deliberate v2:


Trained on dreamshaper v6:


Trained on sd 1.5:


All training is done with the same settings, same dataset, captions etc

The one trained on and generated with sd1.5 is definitely the one closest to the dataset and overall it seems to give the "best" result.
Doesn't come as a completely surprise but i'm rather surprised at the amount of aging/wrinkling (and size/weightloss) in the others. I guess it can be used if people are trying to create that effect considering how bad the AI is at generating "middle aged" at times, it's either 20 or 80.
Not sure how helpful (or informative) this is to anyone else, but as i'd done it i thought i'd share just in case.

 

[Jimwalrus]

Been doing some testing to how training on different models affect the end result. Thanks to an update that added 3 hours to my already long training time it took a lot longer than hoped...
(each image is marked with the model used to generate the image)

Train on deliberate v2:
View attachment 2674951

Trained on dreamshaper v6:
View attachment 2674952

Trained on sd 1.5:
View attachment 2674953

All training is done with the same settings, same dataset, captions etc

The one trained on and generated with sd1.5 is definitely the one closest to the dataset and overall it seems to give the "best" result.
Doesn't come as a completely surprise but i'm rather surprised at the amount of aging/wrinkling (and size/weightloss) in the others. I guess it can be used if people are trying to create that effect considering how bad the AI is at generating "middle aged" at times, it's either 20 or 80.
Not sure how helpful (or informative) this is to anyone else, but as i'd done it i thought i'd share just in case.

Faacinating, and thank you for taking the time to do this. I always train on SD1.5, but only because people say it makes for a more flexible and accurate model - nice to see someone test it.

 

[JakoDel]

Hi, just getting my feet wet with stable diffusion after trying it out once months ago, does anyone know what are the most photorealistic checkpoints now? gonna bother Mr-Fox as well

PS nice blondie, but not as nice as the water

PPS send help for my poor 8gb 1070 Ti, she's screaming in pain

 

[Jimwalrus]

Hi, just getting my feet wet with stable diffusion after trying it out once months ago, does anyone know what are the most photorealistic checkpoints now? gonna bother Mr-Fox as well

PS nice blondie, but not as nice as the water

PPS send help for my poor 8gb 1070 Ti, she's screaming in pain

I've found the most versatile photorealistic checkpoint to be CyberRealisticv3.0
There are many others, but this seems the best all-rounder.
It's also particularly good at portraying celebrities realistically.

 

[Sepheyer]

Hi, just getting my feet wet with stable diffusion after trying it out once months ago, does anyone know what are the most photorealistic checkpoints now? gonna bother Mr-Fox as well

PS nice blondie, but not as nice as the water

PPS send help for my poor 8gb 1070 Ti, she's screaming in pain

Also, my $0.02 is that the sampler matters a lot. I.e. Heun does result in great images, pretty much regardless of a model you use:

Spoiler: More

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

 

[me3]

Adding to the two posters above, going with the more commonly downloaded "realistic" models will keep you pretty well covered, minor differences, one might suit your needs better in one case, then another the next time. If you use lora/embeddings you might need to use a specific model for "best results". Depending on what you consider "best results" at that time.

I strongly advice staying away from low tier "merges" that's infesting things though. Generally those are poorly done and are leaching off of other peoples work, often without even bothering to give credit to those they are stealing from. And you end up having to fight issues with your renderings that you shouldn't have to deal with simply because others have taken shortcuts etc.

Depending on situation/imagery, i've had pretty good results with these, just to start off with:

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

If i'd have to pick just one i'd probably agree with Jimwalrus and go with cyber. It's being updated/worked on pretty often and has inpaint model too. So if bandwidth or hdd space is concern, i'd start there, and as Sepheyer mentioned, sampler matters too, so try different ones and see which you like.

 

[Jimwalrus]

Adding to the two posters above, going with the more commonly downloaded "realistic" models will keep you pretty well covered, minor differences, one might suit your needs better in one case, then another the next time. If you use lora/embeddings you might need to use a specific model for "best results". Depending on what you consider "best results" at that time.

I strongly advice staying away from low tier "merges" that's infesting things though. Generally those are poorly done and are leaching off of other peoples work, often without even bothering to give credit to those they are stealing from. And you end up having to fight issues with your renderings that you shouldn't have to deal with simply because others have taken shortcuts etc.

Depending on situation/imagery, i've had pretty good results with these, just to start off with:

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

If i'd have to pick just one i'd probably agree with Jimwalrus and go with cyber. It's being updated/worked on pretty often and has inpaint model too. So if bandwidth or hdd space is concern, i'd start there, and as Sepheyer mentioned, sampler matters too, so try different ones and see which you like.

Absolutely.
Also, never underestimate the usefulness of an X/Y/Z plot of different samplers, models and steps.
Yes, it will take some time (esp on a 1070) but you will gain some really strong insights.

One word of caution though - don't add too many variables. For instance, including 6 samplers, 10 models and eight different numbers of steps will give you... ...hang on a sec... ...carry the one... ...something something... ...480 images to create.
If each one takes 4 minutes that's <gives up, opens calculator app> 32 hours of torture for your GPU!

 

[Mr-Fox]

I'm also stuck with a GTX 1070 atm. It does take ages to generate good images but at least it can be done and I have seen people make do with even less vram, creating awesome images. I like Clarity over most other checkpoints as it is consistent and have a somewhat predictable and good respons on the prompt, however there are many others that is definitely worth a try.
For sampler I would recommend either DPM++ 2m karras for semi realism and DPM++ SDE karras for photo realism.
The choice of upscaler is as important as sampler maybe even more.
I'm fond of NMKD Superscale, there's also other dedicated versions of NMKD for specific uses such as face focused for portraits etc. If you are not after highly detailed photo realism but rather hentai, anime or semi realism, 4x-UltraSharp is good and very popular. It gives a smooth image with sharp contours. In the end, nothing can replace a really good and well made prompt.
It's the most effective and powerful part of the generative process.

 

[Sepheyer]

I've been on this

You must be registered to see the links

model kick, it even replaced Clarity for a quick minute:

 

[Mr-Fox]

Revisiting an oldie, with curtsy of the eminent Sepheyer. You can find his original post and prompt here.

Spoiler: Prompt

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