- Dec 21, 2020
- 1,523
- 3,589
- Aug 4, 2020
- 2,383
- 3,049
wtf I had no idea there were installable versions of stable diffusion, I thought it was all paywalled and shit
thanks
thanks
- Oct 29, 2017
- 812
- 2,167
Probably as you say it has to be with the training data, I got into this same problem some days ago trying to get some interesting positions, but training also is one of the hardest things, lots and lots of bugs, impossible to train with dreambooth or lora in a 8GB VRAM card even when they claim it is possible, even after watching hours and hours of tutorials. The only way I could do training was with embeddings and it takes me 2 hours for a 10 picture training!!
- Aug 4, 2020
- 2,383
- 3,049
- Dec 21, 2020
- 1,523
- 3,589
Using Script to Explore Image Variations
Say you want to see how an image looks when you tweak one of its parameters, for example steps, like this:
To do that you use script (not a real script, it doesn't require you to program anything):
Say you want to see how an image looks when you tweak one of its parameters, for example steps, like this:
To do that you use script (not a real script, it doesn't require you to program anything):
You don't have permission to view the spoiler content.
Log in or register now.
- Aug 4, 2020
- 2,383
- 3,049
oh you were just referring to that thing, thanks
also
can you clarify how you use this? I assume you run it using something like webui-user.bat --xdeformers --medvram then enter in your prompts, negatives then generate? I do that but it doesn't do anything, it doesn't make my laptop scream either which I expected.
It's using up boatloads of VRAM but no % utilization on CPU or GPU at all, not even on the integrated
EDIT: got it to work, that last --medvram argument did the trick. It waasn't throwing an allocation error like the first time I ran it but it was still not doing anything, weird
EDIT AGAIN: nevermind it crashed lol, allocation error, I need the --lowvram argument
sad fucken days man XD
Last edited:
- Aug 4, 2020
- 2,383
- 3,049
well, looks like this ain't for me folks... --xdeformers and --lowvram was not enough...
F
F
- Sep 15, 2021
- 853
- 3,177
I'm surprised it's not worked as you appear to have 4GB of VRAM which should be enough with --lowvram.
Unfortunately you won't be able to generate anything larger than 512x512 or do any upscaling as they're what really eat VRAM and are incompatible with --lowvram.
Three options are:
- if you have a USB-C port on your laptop, you could buy an external GPU enclosure and an RTX 3060 12GB. Ok, that's ~$400, but it should do it. Or,
- go to unstability.party, by far the best online NSFW SD instance. You can't do img2img, training, or images of celebs etc though, so it is still rather limited. Or,
- install NMKD instead. The low VRAM settings in there will still limit you to 512x512 etc, but all the config is sorted for you. I got it to work on a 2GB GTX 960(!).
- Dec 21, 2020
- 1,523
- 3,589
Better Fingers
This is a theory right now, but it feels specifying nails, (i.e. red nails, opi nails, etc.) reduces the frequency of fused fingers.
Yet to properly test, rather early observations at this point.
This is a theory right now, but it feels specifying nails, (i.e. red nails, opi nails, etc.) reduces the frequency of fused fingers.
Yet to properly test, rather early observations at this point.
- Aug 4, 2020
- 2,383
- 3,049
thanks for the advice but this seems like a lot of investment for very little return.
i have very little time to run these generations anyway so it's not much worth it for me. I'd rather stick to animating in HS2 which I didn't plan to give up anyway.
though I am intrigued by the whole img2img thing because I'd like to get some old normal photos upscaled...but I'm not willing to put them thru the internet for that
thanks again
that makes sense in a way, since the AI would then pay extra attention to meet the specification for the fingers while in its learning, fingers with painted nails would...USUALLY not be fused or distored lol
- Dec 21, 2020
- 1,523
- 3,589
Don't Use "Without", "No", etc.
This is a theory right now, but if you don't want underwear, then put "underwear" into the negative prompt rather than use "without underwear" in the positive prompt. For yet better effect, focus AI's attention using brackets, so the negative prompt looks like this:
This is a theory right now, but if you don't want underwear, then put "underwear" into the negative prompt rather than use "without underwear" in the positive prompt. For yet better effect, focus AI's attention using brackets, so the negative prompt looks like this:
(((underwear)))
- Dec 21, 2020
- 1,523
- 3,589
Avoid Geek Keywords in Prompts
So, I actually tested the following. The PFG set doesn't really understand if a photo is taken with 20mm, 200mm or 600mm lenses, neither it seems to understand apertures. So 2.8 is the same as 16.
But! Since the actual photos taken with these lenses normally have noticeably different styles, the styles do get reflected. Such as you see a lot of low angle or high angle with 20mm and eye level with 600mm.
Overall, it seems the PFG model strongly reacts to notion of style and not at all (hyperbole here) to geeky terms.
So, I actually tested the following. The PFG set doesn't really understand if a photo is taken with 20mm, 200mm or 600mm lenses, neither it seems to understand apertures. So 2.8 is the same as 16.
But! Since the actual photos taken with these lenses normally have noticeably different styles, the styles do get reflected. Such as you see a lot of low angle or high angle with 20mm and eye level with 600mm.
Overall, it seems the PFG model strongly reacts to notion of style and not at all (hyperbole here) to geeky terms.
- Dec 21, 2020
- 1,523
- 3,589
Playboy? Dead*.
*Unless the men were really buying it for the articles.
You don't have permission to view the spoiler content.
Log in or register now.
- Jan 4, 2022
- 233
- 865
PFG 1.11 with VAE-ft-mse-840000-ema-pruned
Without VAE
Difference with your original
Last edited:
- Dec 21, 2020
- 1,523
- 3,589
My PFG is a different flavor, it is the checkpoint version from HF, not the Safetensor from CivitAI. Here is the:
You must be registered to see the links
.Also, you have a different seed. This is me:
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3014654993, Size: 512x512, Model hash: 5a369d04a0, Model: pfg
This is you:
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2910245781, Face restoration: CodeFormer, Size: 512x512, Model: PFG_1_11
I have A1111 webui with auto-update. What are you running?
Edit: Imma grab CivitAI's version to see whats up.
- Jan 4, 2022
- 233
- 865
I use PHPinfo of your photo ( you are wrong here
)Same here
downloaded your PFG, result is 100% like 1.11 Safetensor
Last edited:
- Dec 21, 2020
- 1,523
- 3,589
From "oh wow" to "fuuk" in under three seconds:
You don't have permission to view the spoiler content.
Log in or register now.
- Oct 29, 2017
- 812
- 2,167
Hey guys, to get the underboob effect you just need to add 'underboob' on the prompts, and then you can add details to your preference.
The key to get the underboob is in the model you use, the model Anything 4.5 will give you that effect without any effort, there are other models that may give you that effect, like samdoessexy, if you add a lot of ((())) but, the real deal to get this effect are all the 'anime style' models. In case you want something more realistic, you'll have to make your own merges with any of the models that give you the effect and play around with the merging multiplier.
This is a game changer for me, now I know that to get any 'position' it's as easy as looking for a model that has more of that stuff and merging it with a model I like. I've been looking a lot for this, I was kinda frustrated with the generic standing position that always looks good.
You can get the details from this picture if you want to see the prompts, I made it simple so you can see that it's not prompt magic, but the other stuff that I mentioned before.
Even thought this picture sucks, it will be useful for the learning part, the good stuff is on the other thread in case you want to see.
The key to get the underboob is in the model you use, the model Anything 4.5 will give you that effect without any effort, there are other models that may give you that effect, like samdoessexy, if you add a lot of ((())) but, the real deal to get this effect are all the 'anime style' models. In case you want something more realistic, you'll have to make your own merges with any of the models that give you the effect and play around with the merging multiplier.
This is a game changer for me, now I know that to get any 'position' it's as easy as looking for a model that has more of that stuff and merging it with a model I like. I've been looking a lot for this, I was kinda frustrated with the generic standing position that always looks good.
You can get the details from this picture if you want to see the prompts, I made it simple so you can see that it's not prompt magic, but the other stuff that I mentioned before.
Even thought this picture sucks, it will be useful for the learning part, the good stuff is on the other thread in case you want to see.
- Dec 21, 2020
- 1,523
- 3,589
Some of these images are to die for.
Retouched version:
Also, IRL this image would cost at least $1,000 to make - between the model, the gear, the travel, etc. And SD churns them out at 37 seconds a piece (on my 5 year old laptop). FML man.
You don't have permission to view the spoiler content.
Log in or register now.
Also, IRL this image would cost at least $1,000 to make - between the model, the gear, the travel, etc. And SD churns them out at 37 seconds a piece (on my 5 year old laptop). FML man.
- Dec 21, 2020
- 1,523
- 3,589
