Another new day a new resumption

[Scarol]

I see a lot of differend models, but I installed version: 1.9.4 I think does it matter if I download and use a SDXL Lightning or do I have to download SDXL version?

[cerberus]

I see a lot of differend models, but I installed version: 1.9.4 I think does it matter if I download and use a SDXL Lightning or do I have to download SDXL version?

No, all checkpoint versions should be fine with V1.9.4, it's what I'm using and I'm not having problems. Lightning and Hyper versions are good, they usually need less steps to get good results. Turbo's are good as well but the licencing is different, you aren't allowed to use SDXL Turbo checkpoints for commercial purposes without getting a specific licence for that purpose, so I tend not to use those.

Does this mean your problem is fixed?

[Scarol]

I actually don't know what I did but I can run the generate again, I have download differend files and it works again, and i put --lowvram, i don't know what the generate would do I use --medram or delete that word.

[cerberus]

I actually don't know what I did but I can run the generate again, I have download differend files and it works again, and i put --lowvram, i don't know what the generate would do I use --medram or delete that word.

I've had that sometimes, you get a weird error, kill everything, restart it and it works fine. It's a complicated piece of code that runs a lot of stuff in parallel split between the normal processor/memory and the graphics processor/memory, sometimes I think it can get things our of sync and get itself confused.

--lowvram causes the program to be split into lots of modules, only one of which resides in memory at any one time. This allows the process to run on machines with limited VRAM/memory but has quite a big impact on performance.

--medvram causes the program to be split into three modules, only one of which resides in memory at any one time. This is faster than --lowvram but uses more memory/VRAM.

Not using either of these allows SD to use all the memory it needs or can get> It will use VRAM first (which is what you want as it is faster) and if it needs more it will grab some normal RAM. This should give the fastest results.

From the spec of your laptop that you supplied earlier, you should have at least 8GB of VRAM, this should be enough that you don't need to use either --lowvram or --medvram.

On the right of the A1111/SD bowser page below the picture and the details of the prompt and settings there is a line that starts "Time Taken", this indicates how long it took to generate the image. At the other end of this line is some memory stats labelled A, B and Sys. A is the memory used on the first pass (SD is a two pass process) B is the memory used on the second pass and Sys shows the amount of VRAM available and how much has been used. If you do some runs comparing --lowvram, --medvram and neither, you should be able to tell which performs the best.

The link below show all the available arguments and their meanings

https://github.com/AUTOMATIC1111/stable ... imizations

[Scarol]

I deleted the text and it also worked.

I made this one:

00023-5553338699.png

Its strange with SD, if you see the pics on my scarol fakes u see it's a bit more blurry while this is very sharper. I think I used a sharper VAE checkpoint, I think I still learned the half of SD, but it gets better.

[PA Jack]

Sharp or blurry, who cares? This last photo is simply great.

[cerberus]

I deleted the text and it also worked.

I made this one:

00023-5553338699.png

Its strange with SD, if you see the pics on my scarol fakes u see it's a bit more blurry while this is very sharper. I think I used a sharper VAE checkpoint, I think I still learned the half of SD, but it gets better.

That's a very nice image, I'd be happy with that. Well done, getting this far with SD is an quite an achievement, it's not an easy straightforward program to set up and use, but you've got a good basis to build from.

Regarding the other images that are a bit more blurry, it's probably a matter of fine tuning, a few more/less steps, a change of sampler or scheduler etc. Most of the checkpoint models I've tried have been capable of good quality results, but I've had to experiment to find the optimum settings. The ones I've stopped using tended to have restricted poses available, I want a bit more variation in position and viewpoint etc. Anyway, I look forward to seeing what you come up with next.

[cerberus]

I have to admit that a completely different scenario was about to play out in my mind, perhaps I shouldn't go into too much detail about that, but a ploughed field wouldn't be deep enough for my idea!

Lol, you thinking about some mudstomping her cheating boyfriend into the bog? Surely a way to go

Well not exactly, it was just the way she was hitching up her skirt that looked ominous!

[Scarol]

Thx Cerberus! I'm happy I found a better checkpoint. But indeed looking to others but this one I like very much!

[cerberus]

in the first group, the redhead looks like an elegant lady, the last could be a cosplay scene

in the second, the redhead is almost ready to pull out
but the star is the first shot... in my mind she's the captive of a cult that sacrifices damsels to the Old Gods, so extracting her will be an adventure! but she has a lot of sinking still to do

Elegant, which one? That's not a word that came to mind for any of those three, still it was late and I was having to type around a cat.
Yep, we are more it tune on the second set. Not sure about the cult, but waiting, trying to sink slowly, don't want to rush things, etc.