💬｜general-chat

[ornate pilot]

People will write in negative prompts as ideals features just tp feel human again

[vestal dew]

There's a poem about Thumblongus. You likely haven't heard it. I posted some. Here is the whole thing.

In fourteen hundred ninety two, Thumblongus sailed the ocean blue. In fourteen hundred ninety three, he dreamed of taking the thumbs from you and me. In 1494, he wrote of this and so much more. In fifteen hundred twenty seven, dark magic held Thumblongus back from heaven. He sleeps and dreams of dreadful powers. His thumbs are mangled. He's coming for ours.

[wide juniper]

What’s an inpainting model? Why are they separate from their non-inpainting counterpart?

[vestal dew]

I posted a crude drawing of Thumblongus in general-images

[twilit olive]

How do I exaggerate facial emotions? Like prompt for a person so surprised, his jaw dropped to the floor?

I apologize I asked this on #📝｜prompting-help but I was wondering if someone here had any idea

[vestal dew]

I think you'd have to find a model or train for that

It's even hard to get tongue sticking out, much less jaw to the floor

I tried to get something like the scene of the person with tongue stuck to a pole...couldn't get the tongue out

[twilit olive]

you can try "sucking" as positive prompt and "penis" as a negative prompt for tongue out

[vestal dew]

I wanted someone with tongue stuck to cold pole

[twilit olive]

img-2-img or just with just text?

[fervent thunder]

the ai isn't that good yet to be able to create highly "abstract" things. It's more of a realistic/artist based one in my opinion, and even that has many limits, training those things yourself is most often the best way to get what you want :)

[vestal dew]

tongue sticking out isn't abstract

[vestal dew]

I tried both

maybe jaw to floow is

floor

[twilit olive]

Is theres a discord for stable diffusion where you show case your NSFW art?

[vestal dew]

I tried making a jaw to floor person in daz, then img2img in SD...no luck

[fervent thunder]

sticking out a tongue might not be abstract, but having the jaw dropping to the floor, in my mind to the AI, would be :P

[twilit olive]

A penis on a female body is not abstract too

[vestal dew]

Unstable Diffusion has NSFW sections

[twilit olive]

I am going to check it out thanks

[urban dome]

@wise stratus Is distilled diffusion still happening or was it not viable?

[fervent thunder]

Nooby question! Since all the models I'm using are based on 1.5, should i move on from using 1.4? Would i get better output?

[vestal dew]

you aren't using 1.4 if you use 1.5 based models

the model is the model

1.4, 1.5, 2.0, 2.1 are in the models, not the software

If you want to use 2.0 and 2.1 models, you need associated yaml files as well

seen?

[fervent thunder]

Oooh I thought i had like a1111 based on 1.4, and then the models from people who trained them on 1.5.

So i can go on my merry dum way as long as i'm not using models based on 2.x?

[vestal dew]

you can use those too as long as you get the yaml file

Mine wont run the 768x768 2,1 however

it will run the 512x512 2.1

[fervent thunder]

Interesting, thanks! Hm lemme see if i can remember some more questions.
The VAE thing! These are my settings:
https://i.imgur.com/6BFzpJA.png
Some models talk about needing VAE, some say it's baked in, some don't mention it. What should my default settings be?
And do i just google sth like "vae-ft-mse-840000-ema-pruned" and make it the default and call it a day?

[vestal dew]

that's what I do. However, it seems like some vae's work better with some models

I haven't experimented with it

[warm junco]

Set it to Automatic. When a model provides a vae you should download boath and rename them to match the name (Example.ckpt, Example.vae.pt). Put boath in the models folder.
Then the webui will automaticly use the right one for each

[vestal dew]

there's a setting to ignore the specified vae for models with vae next to them. Isn't that the same thing?

[warm junco]

yes its the same thing like automatic, but it means you can choose a vae and it will apply it for every model except these who had already one. So to choose Automatic is better

[fervent thunder]

Nice. So the default is set to automatic. Ah i see vae is a .ckpt file? Hm, ok gonna put it in the vae directory

[warm junco]

yea a vae that is named .ckpt goes into the vae folder

@fervent thunder then you have to go into the webui settings, under User Interface, there at the bottom are the Quicksettings, there you need to ad sd_vae,

[fervent thunder]

oh wait they have vae safetensors too! those are safer, i remember

[warm junco]

then apply and restart the webui

[mighty karma]

is there a way to know a LoRA's trigger words if i have just the file?

[fervent thunder]

https://i.imgur.com/6VO0D1R.png
Like, manually type ", sd_vae" in after sd_model_checkpoint?

[warm junco]

yes like: sd_model_checkpoint, sd_vae,

[fervent thunder]

oh cool it added the vae selection into the ui! 😄

so i still keep it at automatic by default? when do i manually change it?, when i know that the model wants it?

[warm junco]

sometimes its recommended by a model maker to use a specific vae. then you can change it there

let it on automatic if you have vae and model from the same author and boath have the matching name

if you merge a model with a model that needs a vae file. then the new merged model will also need the vae

[fervent thunder]

Don't think i'll be merging for a while
Thank you so much!
"if you have vae and model from the same author" What exactly does this mean? I think i only saw the creators of models say that VAE is either baked in or to download some specific VAE, but i don't think they provided any specific VAEs, right?

[warm junco]

here for example:

oh cant post images

[fervent thunder]

ye lol

[upbeat tangle]

anyone tried out the shutterstock ai ?

[hollow torrent]

Eyy guys; is the #dream command free of use or should i pay smth

[west vault]

hello guys, i need an info, i am using colab and i want to merge models, but in example code, model file directory is like this: https://huggingface.co/Linaqruf/anything-v3.0/resolve/main/Anything-V3.0-pruned-fp32.ckpt but when i choose a model from huggingface, the link is like this: https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0/blob/main/dreamlike-photoreal-2.0.ckpt only difference is resolve vs blob, how can i find directories like in the first example

[tired swan]

is there a website to browse prompts? i tried lexica/openart ai/arthub and didnt like em

[fervent thunder]

After recent update started to get this error while using pix2pix: torch.cuda.OutOfMemoryError
Worked great 2 days ago
Does anyone else have this issue?

[cloud granite]

Dreambot is going away, RIP

[silk fossil]

I am curious as to why SD has the "gaps" in its knowledge it has. It knows what Fauvism is, but it isn't great at obscure animals like skuas. What causes this?

[forest apex]

skua shows up in https://rom1504.github.io/clip-retrieval, so it does know it, idk

[silk fossil]

It does know, it just doesn't...well, I suppose it's unfair to say "it has missed the point" because it's a machine, of course it did. I tried "a brown skua in the style of an orthodox icon", but it doesn't seem to know how to make an icon with a bird instead of a man. I assume this is a limitation in the training data.

[forest apex]

Try a bird that would have more pictures of it online, maybe, specifying colors similar to a skua

[silk fossil]

I'll try that. This machine is pretty smart. It's quite the art student.

It's unfair for me to criticize it for not knowing animals, but it tends to draw birds almost as extremes - the medium sized birds are very small. The large birds are huge.

[forest apex]

hehe

you could also make an embedding of just skuas, and refer to that

[silk fossil]

Where can I read about embeddings?

[forest apex]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Textual-Inversion

[winter pike]

i personally use clip interrogator 2 if i want ideas of prompts: https://huggingface.co/spaces/fffiloni/CLIP-Interrogator-2

[forest apex]

that is one method of fine-tuning the model, the other being dream booth, but dream booth models need to be loaded instead of the current model, which is a bit cumbersome

[silk fossil]

I will read about it. Thanks!

[tepid dome]

hi i am looking for a job as a discord moderator if anyone has an opening let me know please

[silk fossil]

I heard they doubled the pay for that recently.

[wraith sky]

What is the pay for a mod (in general I guess) if you don't mind me asking?

[silk fossil]

zero dollars

[glad nova]

I'm a complete noob with stable diffusion, where do I start ?

[warm junco]

its based on the model you use

[silk fossil]

How can it learn an embedding with only 53 images? That's astounding.

[winter pike]

to use the bot use "/dream prompt" in dreambot channels, its gonna be deactivated over the weekend btw
locally you can just install InvokeAI, you will need a good gpu (my gpu is a GTX 1050 but im still able to generate despite being old)

[silk fossil]

Shit, I can barely recognize a warbler after seeing 53 images of them. This thing is way smarter than me.

[glad nova]

1060 here

[winter pike]

give it a try, its still worth it

its like 35 seconds per image at 512x512

20 steps

[silk fossil]

I'm taking about a 30s-1min per, 25 steps. Does number of steps matter?

[somber vapor]

discord bot and web site creating same quality material? so this means free vs credits?

[glad nova]

I heard something about giving the ai images to base itself on

How does that work ?

[winter pike]

thats img2img, it add noise to the base image and sd try to recreate it using the prompt you give to it

[silk fossil]

Where can I read about img2img?

[winter pike]

https://github.com/CompVis/stable-diffusion

at "Image Modification with Stable Diffusion"

[west hornet]

Whenever I try to generate an image it just gets stuck on 0%
How long should making an image typically take

[silk fossil]

Thanks!

[west hornet]

Nevermind it's just taking a while

[winter pike]

locally? depends of your gpu, the first generation takes a bit longer to load, at least for me

[forest apex]

Anyone seen this? https://www.reddit.com/r/StableDiffusion/comments/10lzgze/i_figured_out_a_way_to_apply_different_prompts_to/ - looks really cool

using a node based system to generate one image using multiple prompts for different parts of the image, then coherently merging the results

[warm junco]

yea its pretty amazing, but not works in the auto webui currently

[forest apex]

@warm junco have you tried out comfyanonymous's UI?

[warm junco]

nope

[forest apex]

going to give it a go in a little bit, seems like it would take some of the things I've been trying to generate much easier to get right

[oblique gazelle]

Where can I go as a personal user to ask various question about training?

[forest apex]

#🔧｜finetune might be able to help

[oblique gazelle]

okay thanks!

[runic magnet]

Hey, is there a differnce between safetensors and ckpt files for model? Stable can see and use both at the same time if put int omodel folder fro mwhat I saw

[west hornet]

what does negative prompt mean in stable diffusion?

[runic magnet]

thinks you dont want on picture

[west hornet]

oh alr. so if i provided a normal prompt AND a negavite prompt would i get a more accurate result?

[warm junco]

input words you dont want to see in the picture, like blurry, lowres, deformed, it will help a lot

[winter pike]

ckpt can contain malicious code, while safetensors are safe

and some people say that safetensors load faster

[winter pike]

yes, it does wonders

[peak orchid]

Anyone played around with LAION's 'Rosey' bot yet? https://huggingface.co/Rallio67/rosey_12B_instruct_alpha

[vestal dew]

What is a hypernetwork?

https://civitai.com/models/5198

It's called a model there, but the extension is .pt.

Is it used like a model, place in the models dir?

[warm junco]

it goes in the hypernetwork folder, and is selectable with the 3 button under the generate button

[vestal dew]

I see

Let me give it a try

[peak orchid]

it sits ontop of a model

[vestal dew]

I see

It goes in the modules/hypernetworks folder?

[warm junco]

models/hypernetwork

[vestal dew]

something wrong with this

when I copy it, then try to switch directories, the file folder shuts

had to open two separate windows...odd

[fervent thunder]

if i want to combine an actor its [ben affleck|chris evans] right? i feel like im only getting the likness of the first so i must be fuggin something up

[forest apex]

@fervent thunder (ben affleck:chris evans:0.50)

I was using pipes (|) but apparently it's supposed to be a colon separating the two items, both seem to work though

[old marten]

its brackets [] not parentheses?

maybe it works on both? I just always use brackets

[vestal dew]

I have the file in the hypernetwork folder

there is no button to select hypernetworks

there;'s a red button for extra networks, but that shows embeddings

[forest apex]

yea both seem to work

[vestal dew]

I wish these files on civitai had some kind of instruction. It says put the hypernetwork in models\hypernetworks...I've done that'

then what?

[fervent thunder]

never used hypernetwork but maybe it needs to be prompted with a keyword

like the name of the file

[vestal dew]

I think so

it shows tags. I will try that

I guess I have to make a file with all these tags

hmmm...it did nothing but change a plant and the expression on someone's face

it made a good image in the style in txt2img

didn't alter the overall look / style of an image in img2img

it worked on a different image in img2img

I see...less text is better...just reference the hypernetwork

[sick tangle]

There seems to be an issue on Dream Studio with Stable Diffusion v2.0-768. I tried using that model to generate, chose the correct parameters as usual with never having a problem until now. If i click dream it does nothing but charge me. Got charged 13.6 credits 4 times.

[forest apex]

try #1025467151206854736

@sick tangle also in the Stability AI FAQ they mention Use the Support Form inside of DreamStudio to report the issue and our support team will respond as quickly as possible to assist you.

[errant pumice]

have been trying to train a model with a face with no luck. Does anyone have any tip?

i'm using local install dreambooth in stable diffusion.

[wide juniper]

Is it true that some samplers gives better results for certain styles, or is it just a matter of efficiency?

[orchid wadi]

It is more like for certain image properties.

DDIM is still the king of things that are painted/drawn

Euler A is like best at photobashing by a lot, it makes things more like conceptual. So like "man wearing a diaper as a hat" get fullfilled by image having just those elements present in any form. Fairly good at cartoons also since it is rather aggressive with styling. Euler is like DDIM but faster and not as good. Then rest... Well... They do basically the same in my experience par for minute thing that takes effort to spot. Against same seed they even make same picture lot of the time.

However all but Euler A produce just better results given more steps. Euler A stops at like 40 and wont really do anything.

And sometimes the better and more refined results at higher steps simply get lost in the resolution of the image formation and gradients

You can have absurd amount of details being made in the latent space but fact is that only a portion of it is used.

[west hornet]

which sampling method should i chose? and another question, are there any recomended settings for the best result

[warm junco]

@west hornet EulerA, DPM++ 2M Karras, DPM++ SDE

[orchid wadi]

I use all of them. Like I basically run every prompt against all of them.

I get good results from like.. all of them par for PLMS... which I have yet to bothered to fucking learn

the SDE ones work really well for low steps like down to 5, excel at 10-15. The other samplers "get good" at 20, like Euler A starts to make sense after it. Basically all of them are fully defined at 50 step mark. Then it is a question of when the increase in details falls under the level of image formation being able to show the,

I have tested them to like 1000 steps. Why? Because why not?

Unless you are at VERY high scale, basically none of them do anything of significance past 150.

But for speed. Euler A, Euler, DDIM are fastest.

SDE ones take least steps

[west hornet]

alright thanks

[long sedge]

i get this error when trying to start stable diffusion: RuntimeError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 4.00 GiB total capacity; 3.40 GiB already allocated; 0 bytes free; 3.45 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Does anyone know where the pytorch files are located? (im trying to launch stable diffusion with the dreambooth addon)

[orchid wadi]

So bigger resolutions I personally would choose Euler or DDIM. For 768x768, hardly fucking matters.

[west hornet]

is there any way that i can refine my generation more? some things just look un natural and out of place in it

[orchid wadi]

You got 4gb of VRAM.

DB takes at least 12

[long sedge]

oh

[warm junco]

You cant use Dreambooth with 4gb vram

[west hornet]

i had this issue, edit webui.user.bat and add --lowvram after set commandline_args

worked for me

[orchid wadi]

I think at 512x512 at lowest it is 10g of vram

[forest apex]

8GB can work with some tweaking

[west hornet]

^

[orchid wadi]

Yeah I saw that... But the speed was so slow that even with near free elecricity it was not worth it.

[orchid wadi]

Prompt more and vary the prompt. Play with scale.

[west hornet]

scale?

[orchid wadi]

Also keep in mind that things are not what they seem to be in the image.

CFG scale

[west hornet]

what does cfg scale effect?

[orchid wadi]

Like you can put in to negative prompt every token that the model has for people. But still make a person prompting a shirt.

[orchid wadi]

To put it "Explain like I am 5" way. Image formation isn't a "static" process. The AI starts a journey. Based on your prompt and what it interrogates from the space. The CFG scale basically instructs the AI to find the tokens you prompted in more accuaracy and eliminte things that don't match.

The goal is prompt tokens - latent space tokens ≈ 0 the scale basically multiplies the prompt tokens

[west hornet]

alright. so if ive understood, if i increase the cfg scale the ai will be more intricate in generating the image?

[orchid wadi]

Because keep in mind that "face" is just not one thing. Face is broken down to components. Face -> eye, eye, nose, mouth, eye brows. With too high of scale the AI will just fetch the concpet of the "face".

[long sedge]

ayyy it worked ty :)

[orchid wadi]

As in sort of skin coloured icecream cone. (I have tested this a lot)

But at low scale the AI just basically does whatever it takes to negate your tokens.

Important to keep in mind is that the system is adversiary. The goal is to reach a scenario where the sum is 0 or near 0 with in a desired tolerance

[blissful pollen]

Hi

[west hornet]

yo

[blissful pollen]

I'm getting a new PC soon

Wondering what graphics card is good for SD

Like does SD work better with higher cuda or with more vram

[fervent thunder]

both are going to help, but the vram is the main thing i guess

[raven thorn]

is the 1.5 version actually more stable?

[fervent thunder]

no just used more widely

[orchid wadi]

8gb minimum, most training can be done with 12gb. If you want to just make pictures get whatever is the fastst 8gb card. If you want to do things other than than get whatever is cheapest for the amount of memory

[fervent thunder]

If you can afford it, Id say shoot for higher than 8 gig. future proof yourself a little. and it seems the Deep Floyd model may require a little more oomph but idk for sure

[old marten]

nah G

go for 24gb if u want real futureproofing

[orchid wadi]

What does that even mean?

Which is a better car? Volkswagen transporter or Koenigsegg Agera RS?

[old marten]

wrong. 1.5 is the real answer

2.x is filled with bugs

[fervent thunder]

I got 24 myself, bought a new system after xmas and it came with a 4090

UNLIMITED POWAH!

[old marten]

I couldnt find a 4090 for less than $1600

[orchid wadi]

Just make sure you keep in mind that.... power demands can be rather expesive. I did the maths and Colab is cheaper for me to run if I need more VRAM.

[old marten]

If they want entry level they can get a 3090 for $700 used

[fervent thunder]

idk this 4090 seems to run fairly cool

[orchid wadi]

Do give explain since I have no issues even at very high resolutions and I don't need any sort of highres fixes.

[old marten]

depends on the kind of prompts ur generating

2.x sucks at faces, celebs, artists, etc

[orchid wadi]

Which a crypto miner ran on 80C for years at 100% utilisation and now is trying to cash out.

[old marten]

and is good with landscapes

[fervent thunder]

2.x was a dud imho

[old marten]

nah G, a lot of ppl are upgrading to 40 series, many dont need that much vram

real vram guzzlers are those who use AI

12gb is good for gaming

[orchid wadi]

https://i.imgur.com/54Ino2C.png What if I want to make stuff like this? What model would you recommend?

[old marten]

protogen V2.2 anime

[orchid wadi]

Thats odd. Because I made it with just 2.1 768.

[old marten]

its not a very good picture tbh m8

[orchid wadi]

Ah iok...

[old marten]

once you see what the good models can do u'll raise ur standards

[orchid wadi]

Ok how about for this 1.5 or 2.1 https://i.imgur.com/PNm8qYS.png

[old marten]

I dont recommend using base 1.5 or 2.1 in the first place. use custom models based on those

[fervent thunder]

yeah custom models all the way

[old marten]

protogen v22 anime is a good choice as well

[orchid wadi]

Well I don't want anime.

[old marten]

if u want realism use my own model blend

[orchid wadi]

I don't want waifus and loli

[old marten]

its not really anime, but good for hand drawn stuff which is what ur going for

[fervent thunder]

kinda dick move to imply we do bro

[orchid wadi]

Didn't say you do.

[fervent thunder]

its implied

[old marten]

if u want paintings over realism then thats the one I would use

[orchid wadi]

I said that the model dataset is based on people who do.

Well... Thus far I been able to get everything I want with just base model and TI.

[old marten]

then why ask

[orchid wadi]

Because you said that 2.1 is shit

[old marten]

no offense but those pictures are shit lol

they look like what midjourney did when it first came out

[orchid wadi]

Ok show me a good pic then

[old marten]

of what

[orchid wadi]

Well like those

[old marten]

I dont do guys except for myself

and I dont do cartoons

[orchid wadi]

Since that is what I spent a lot of time trying to get an embedding for and that is what I wanted.

Ok

Lemme look through my fucking massive set of images for something then.

So you do only realism?

[old marten]

pretty much. or cosplays

maybe scifi

[orchid wadi]

So realistic softcore prawn

I think I got the folder

Ok. This then https://i.imgur.com/LtVpehf.png

[old marten]

wtf is that

[orchid wadi]

Well... What model?

Would you recommend

It is a paper teddy pear obviously.

I think...

Look I generate lots of stuff, I can't be expected to keep track of it all...

[old marten]

I posted a screenshot of my promptbook, go see the examples there

thats about as much I can show without getting banned

[orchid wadi]

Where?

[old marten]

#🏞｜general-with-images

and like I said, I only mostly do modeling, no cartoons no landscapes no animals

[orchid wadi]

That is all incredibly boring and hollow looking stuff.

[old marten]

Well thats ur opinion. U like teddy bears I like hot girls

agree to disagree

[orchid wadi]

The kind of stuff deviantart overflows with as every cosplayer with onlyfans and access to few filters puts out.

I dislike realism. Realism is boring. If I want realism I go see a gallery with hyper realist paintings.

[west hornet]

how do i To create a public link, set share=True in launch().

like in what file

[old marten]

look for an address that looks like 127.0.0.1 or something

just copy and paste that into ur browser

[fervent thunder]

i think in the webui-user.bat

but i thought it was like dangerous to do that. or maybe its since been corrected

[forest apex]

less dangerous, afaik

[pulsar canyon]

I'm pissed.

Current Dreambooth fucking sucks and gives nothing but anime girls no matter the training references or prompting. And I am stuck using older dreambooth versions in order to get decent and accurate models.

Shit sucks hard

[blissful pollen]

Rx7900 20gb will work?

[rare shadow]

whenever i upscale images in auto's1111 webui the metadata gets wiped now, i remember it was preserved before, anyone have any idea how to fix this?

[blissful pollen]

And I have an 580 8gb card that I can add as a boost

But I prefer not

[pulsar canyon]

I posted an example of it in failed-diffusions channel feels bad man.
I think I got the new Dream to work but it is such a stupid work around. I sourced a new model using the old Dreambooth model I made from the older verison. At this point why use the new one at all

[orchid wadi]

I know very little about AMD. But Nvidia is the gold standard of all ML/AI stuff because they had a headstart on that front

[dense tendon]

What's the difference between generating 16 batches of 1 image each, 4 batches of 4 images each, or 1 batch of 16 images?

[warm junco]

@dense tendon
Batch Count = How much images your gpu generate in a row.
Batch Size = How much images your GPU generate at the same time.
If you dont have 12-24gb vram you cant go high with batch size xD

[dense tendon]

What's a good indication of whether or not I should use a larger batch size or not?

Torch used 3412 MiB of the reserved 4312 MiB VRAM, I'm guessing that means I should do larger batch sizes?

[warm junco]

if your vram can handle it. look at the taskmanager while generating. Also if you increase the image size you have to lower or disable the batch size

[worthy ledge]

Hello Guys,

how many images should i feed the AI to train a conecept LoRA?

[warm junco]

how much vram do you have ?

[dense tendon]

I have a 1660 TI, so 6GB

[old marten]

use the smallest batch size possible

[dense tendon]

Oof, okay

[warm junco]

normaly you dont want batch size but images with higher resolution or highres fix

i testet x512 with 8gb vram on 30 steps and i can generate 8 images simultan (its using 5,4gb vram with xformers)

[slim relic]

If I have a prompt like: "high quality, blue, girl"
does the order of the 3 words matter?

[warm junco]

yes

[old marten]

yup

tends to prioritize first words

[warm junco]

the ai reads from left to right

if you type 1girl, blue eyes, red eyes, it will be more images with blue eyes

[slim relic]

I see

good knowledge

what about 1girl, blue eyes, (red eyes:1.2)

xd

[warm junco]

it would be more red eyes

[grand fox]

Questioon how can I make a ckpt file?

[orchid wadi]

Right what is important to understand that diffusion works on layers. It first tries to find the thing you first prompt. So depending on your priorities you should prompt accordingly. Like generic prompts should go along the lines of: theme, composition, subject, verb, object, image properties, style, "filter" important here is to understand difference between style and "filter" which is a type of style. Example you can prompt "in the style of Mucha" and get the art style, but then "black and white image" to transform that in to black and white. Not necesassarily functional prompt.

But if you want like transform the subject of example Leyendecker's style to like a furry, then you should promp: illustration by leyndecker, furry... because what happens is that first the AI will fetch a generic representation of leyendecker's illustrations, on to which it attempts to place a furry.

Good tokens can be absolutely fucking ruined by putting them in incorrectly, especially if you use 2.1 model which doesn't use CLIP layers.

in 1.5 it is less meaningful since it navigates CLIP structure by finding the term that first describe the top most layer.

So you should try testing you promp by varying the order. Good way to do this on Auto is to just make a prompt file and loading the text file.

Because we don't really know or understand how the model truly works because the sheer size and amount of dimensions it has in it. So you never know what alternating prompts can bring up

[dry trellis]

Either by training one or merging two existing ckpt files.

[orchid wadi]

From what? For diffusers to Ckpt there is a script. CKPT is just archive format where the diffuser model is laid out in a specific way, you can open it with 7zip.

[silk siren]

Hello

Does automatic have an active discord server still? The link on their page is expired

[forest apex]

@silk siren afaik, no, but he pops up in here from time to time, and most of the support questions are for the webui

[fervent thunder]

can i get stable fussion on macbook?

[silk siren]

@forest apex I am trying to build a custom script but I am not very familiar with python, I can run the script in command line but I want to integrate it in automatic

[forest apex]

@silk siren you can try asking in #1011228667659178055 or browsing the github project's discussions tab

[silk siren]

ok

[muted temple]

I am using webui InPaint, and I select "Inpaint Area->Only Masked", but when I click "Generate", it regenerates other parts of the image, too. Not only what I have painted black using the "Inpaint" tab. What am I missing?

[toxic minnow]

Does anyone know about this issue? I'm generating images based on a sketch from paint.. they generate with amazing colors, but once they are done they are black and white only. Is there something I'm missing? I've put black and white in the negative prompt

the sketch is just a black pencil drawing on a white background

[warm junco]

thats normal. it just changes the stuff that is black masked

[warm junco]

img2img will try to use the colors you give it, so just black and white will give you only black and white

[toxic minnow]

I see.. So why is it generating images with color? they look really amazing, but is the only choice i have to screenshot while it generates and then use that to further imgtoimg generate

[warm junco]

while generating an image the image gets "denoised" starting from all colors the ai try to fit the img you give him so it will go into the black and white at the end

[prisma pendant]

hello

[vivid portal]

@karmic brook When is the dreamers contest decided?

[kind river]

Will there be a new version of SD that uses the LAION's new ViT-G/14 CLIP model?

[orchid wadi]

If there is a newer and more built up version of a stability's component. It will probably be in the next version of the model, as in 3.x because they'd need to start training from 0.

[fervent thunder]

Where can you see how many generations you have left in discord?

[silk fossil]

What is actually IN a ckpt file?

[wraith sky]

Midjourney has a feature where you can add multiple images in a prompt with or without text. Is there a comparable function for stable diffusion?

[orchid wadi]

https://huggingface.co/stabilityai/stable-diffusion-2-1/tree/main If you look in to those folders they have a binary files in them. These binary files are made of sets of data. Converting the diffuser to CKPT breaks down the binarys to smaller components that can be called in individually.

You can convert ckpt-diffusers or diffusers-ckpt just with a simple script that knows what pieces go where.

But ckpt is just a checkpoint

Basically it is a saved state of a model. No different in practice than good photocopy of a note. Photoshop backup file.

[old marten]

fuggin luuuv loadin up that vram

crunch those mf numbers

[vast pollen]

Remind me which type of models an average user should use. Pruned or not pruned?

Idk what the difference is besides file size

oh wait not pruned is for training nvm

[rocky marsh]

fml, trying to find the best model that could be near midjourney, so far not that good

[frank rune]

I'm having some trouble finding a viable dreambooth commit. What commit are you people using?

[winter pike]

did you tried the dreamlike diffusion model?

[frank rune]

that's a good one

[peak orchid]

A French university is banning all AI tools from being used, with the threat of explusion for violating the ban: https://www.reuters.com/technology/top-french-university-bans-use-chatgpt-prevent-plagiarism-2023-01-27/

The article title implies its just ChatGPT, but the article explains that its all AI tools:

The university said on Friday the school had emailed all students and faculty announcing a ban on ChatGPT and all other AI-based tools at Sciences Po.
Sciences Po, whose main campus is in Paris, added that punishment for using the software may go as far as exclusion from the institution, or even from French higher education as a whole.

[iron pendant]

in other news an English culinary school bans kitchen utensils, bowls, and ovens

[trail wing]

they're going back to raising their own chickens for eggs because mass production is just a gateway drug to electricity

[spring idol]

i want to know how stable diffusion works without math, but also without skipping steps, is there a video around the 10-15 minute mark that can do this explainer

i guess theres going to have to be a few sections explained separately, before it can be joined up in summary

[peak orchid]

Start with the basics, and check out visualizations of what's called 'gradient descent'

[spring idol]

gradient descent is the basics, or do i need to look into more

[peak orchid]

Its what currently underpins linear regression which is what AI algorithms are right now

If you visualize the data landscape as a 3 dimensional landscape, the gradient descent algorithm tries to find the deepest possible hole (or the the top of the highest mountain in the case of DeepDream)

Think of the data landscape as being a vast area with mountain ranges, canyons, and everything in between.

You don't need math skills to recognize how important the learning rate (also called step size) is, as its basically how large a step you take when navigating the landscape. Bigger steps can be quicker, but they risk making it impossible for you to reach the target location. Too small of a step can also have issues with getting stuck in small shallow holes before you find the deepest hole

Of course models like Stable Diffusion (and things like Textual Inversion) are trained with many more than 3 dimensions, but its easier to visual 3D landscapes than 512 dimensional landscapes

[spring idol]

this is too metaphor-y, i cant dissassemble it, jargon was introduced without explanation, thats the issue with readin around this space, i need a fundamentals class that isnt a lecture involving math, theres got to be some youtubers who condensed this

i was considering going down the ai art history rabbit hole but i didnt have several hours of unbroken focus

try explaining a fundamental concept of (this is technobabble:) "security architectures in campus routing and switching topologies" without having to unpack almost every term you use as shorthand you are expected to have learned in the fundamentals courses that came before, thats what i felt like when you said how important the learning rate (also step size is)

[wide juniper]

Is Anything v3 a leak of NAI and thus illegal to use? I want to credit it in my game but I don’t want to risk acknowledging something illegal

[silk fossil]

This will make everything just like it used to be, of course. AI defeated.

[patent violet]

Are there any tricks for ensuring there is no text in the output image?

[wide juniper]

Have you tried negative prompts?

[patent violet]

Yes, just writing "text, writing", doesn't help

[silk fossil]

Hot take: I like SD’s AI writing gibberish

[silk fossil]

I still don’t understand what they actually represent.

The knowledge of a neural network?

[limber trail]

how can I extend real existing images

[exotic terrace]

how can i make waifu versions of real girls

it always turns out sh1t.

[wide juniper]

How will the next stable diffusion version be better if everyone’s training in 1.5 and 2.1? Specifically, how does the tech improve as opposed to custom-trained models?

[astral goblet]

https://twitter.com/laion_ai/status/1618317487283802113 this is new

[fervent thunder]

something we can run in auto?

[slim quarry]

Hi frens 🙂

gm

If I have questions regarding wanting to create something from an example, where do I ask this?

Does SD work similarly as Midjourney, as in you can run the bot in a private channel in Discord? I'm completely new to this one 🙂

Thanks for pointing me in the right direction

[forest apex]

@slim quarry there is the dreambot on this discord, but I believe they are discontinuing it. mainly SD is run locally with a UI such as automatic1111 or InvokeAI. you can also generate for free using https://stablehorde.net/, which has a discord bot.

also there are sites like playgroundai.com and a few others which are free/low cost

[turbid thorn]

hey how do you set weights to your prompts on SD?

[brisk bloom]

select the word, press ctrl + up/down arrow. up to increase weight

thats for A1111

or just use () around the word

[stark hornet]

any of you know how to use embeddings in 1.5 ?
i m not able to train one, it completely ignores the training data pictures

[brazen fox]

Language models, efficiency, better and more training with improved captioning, among other things- not everyone has the funds to train a model from scratch after all, the custom models have to be built on something

[elder cradle]

Yes, that's why the original "owner" (publisher to huggingface actually), decided to private the repo

[fluid depot]

whats the best webui in yall opinions?

[rocky marsh]

yeah, it's not as good with details

[wide juniper]

I installed the ABG extension to remove backgrounds. Where do I find it in my webui?

[rocky marsh]

Did you reload your web-ui?

There is a button, Apply and restart UI

[wide juniper]

Yup

[rocky marsh]

Try clicking, Check for updates and then fully restart the cmd

[topaz drift]

Hello. Can anyone help me with a problem? torch.cuda.OutOfMemoryError: CUDA out of memory.

[wide juniper]

Where should it appear in my web ui?

[wide juniper]

Reduce batch size

[rocky marsh]

This error basically means, that your GPU doesn't have enough VRam, meaning your settings are too expensive or your GPU is just low performance

[topaz drift]

How I can do it?

[topaz drift]

Everything was fine until yesterday

[rocky marsh]

Not sure, probably in extras.

It might be a script

[wide juniper]

In your txt2img tab, set batch size back to 1

[topaz drift]

I never changed it. It was always 1

[forest apex]

Did you update your webui with git pull?

[topaz drift]

no...

I need write it to cmd in sd directory?

[wide juniper]

Ok then reduce image size

[topaz drift]

I had 512x512

[wide juniper]

Close other applications

[wide juniper]

Smaller

[topaz drift]

Nothing loads the Vram

[wide juniper]

What’s your gpu?

[topaz drift]

gtx 1650 ti

[wide juniper]

Yeah it needs an upgrade, even though it’s a good gaming gpu. More vram needed

Either upgrade or wait for the tech to improve in a few months

[topaz drift]

The problem is that until yesterday everything was working fine. The SD starts generating, generates one picture and gives an error with insufficient memory. But the memory is not released after that. And a couple of days ago I could generate 10 images.

[wide juniper]

I upgraded from 1660 Ti to 3090 recently and it still runs out of memory when my image size is too large

[warm junco]

use --xformers and --medvram in your webui-user.bat after the COmmandline_ARGs=
This will decrease vram usage alot

[topaz drift]

It's already done))

Any ideas besides replacing the gpu?

[warm junco]

update the webui

[topaz drift]

How I can do it?

[warm junco]

in the webui-user.bat you should insert git pull at the bottom

[topaz drift]

It was here

git pull origin master

Or I need write only git pull?

[warm junco]

only git pull

but it should be updated then

if it was there

[topaz drift]

I'm going to try

There is no tracking information for the current branch.
Please specify which branch you want to merge with.
See git-pull(1) for details.

@warm junco Do you need a full error message?

[warm junco]

okay then you have to put back your origin master

but it should work without it

[topaz drift]

I've already returned it. It doesn't work without it

@warm junco Why does the memory not free up after an error?

[warm junco]

bug, you have to restart the webui then

[topaz drift]

Restarting does not help. I have already tried reinstalling several times

[warm junco]

which tutorial did you followed?

[topaz drift]

https://github.com/serpotapov/stable-diffusion-portable

[warm junco]

why portable, so you have it on a stick?

[topaz drift]

no. It is easy to install

it is portable only for install

After install it looks like original SD

I was trying to install original Sd, but it doesn't help

@warm junco

[warm junco]

hmm then idk, maybe try it in #🤝｜tech-support

try something like: big statue standing in city,

[brisk bloom]

prompt what the statue is made of, or the subject. eg horse

[brave matrix]

Just posted the second episode in my Stable Diffusion based history series: The Vikings! https://youtu.be/XXANMhv4Bag

Would love feedback

[vestal dew]

ditch the background music or make it much fainter

sometimes it overshadows the voice

With your images, do more slight movement...like one image fading to another with a slight change, vs panning around the image

Around 2:11 is excellent because the panning has a purpose, then the image shifts to denote looking at the stars over time.

[wide juniper]

Since I have the GPU power for it, would generating at large dimensions be preferrable than 512x512? If so, does it cause it to add more things in it? Like more people in a crowd

[autumn compass]

You can raise the resolution a bit but if you do it too much you will have problems with the images like duplications and inconsistencies.
if you want a high resolution image as the final result I think the best thing to do is to try something like 768 x 512 or 960 x 640 and when you find a good image raise the resolution with one of the methods available, like hiresfix or sdupscale.

[proud nova]

Where can I find config for SD 1.5?

[forest apex]

it is in the webui\configs directory but you will rarely need to use it

[proud nova]

I need it for one model

Thx

[wheat hazel]

I think (maybe) I kinda found a way to "simulate" temporal coherence

[vestal dew]

Is it because my card is weak, or in general do long shots, full body shots, not show detailed faces? I just posted an image in general with images. Look at the face. Why can't I get something like this? https://images.unsplash.com/photo-1517466787929-bc90951d0974?ixlib=rb-4.0.3&ixid=MnwxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8&auto=format&fit=crop&w=686&q=80 It's a long angle, wide angle, showing the full body of a guy playing soccer. His face is clear

[forest apex]

@vestal dew did you use the Restore faces option?

[vestal dew]

Yes

Soimeone told me about a photoshop pluging. I am going to try it

[wide juniper]

Right. But specifically I was referring to whether lower resolution results in details and objects that would not have been imagined by the AI

[slim quarry]

Does anyone know how to run SD on a mac so that I can use it for image 2 image processing for AI animated videos and deforum? I've checked out diffusion bee but it seems watered down? I have hundreds of frames i need to process and doing this manually is an impossible task. Many thanks! Cant find many videos on SD for mac. All help is appreciated! ❤️

[wheat hazel]

this happens because you render at small resolution

basically, your image is 512x512 and the head is about 1/12 of that in width, so 1/144 in area

so ~43x43 pixels

1.5 models have been trained to render 512x512 pictures at minimum

you can still render lower resolution pictures, but you'll get bad artifacts, ugly things,etc... because the AI doesn't really know how to draw really good details at lower resolutions

so, basically, you can fix your issue by either increasing the resolution of the picture 2kx2k for example (with the high res fix)

or you can mask the head on your 512x512 picture, and ask the AI to redraw it with "masked only" option enabled at 512x512

depending on the model you use, restore face option can deteriorate the details on the face, aswell

[raven thorn]

i feel like chatgpt does better prompts than me 💀

[warm junco]

Hey, look here for the installation of Automatic1111 webui on Apple:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Installation-on-Apple-Silicon
Would still recommend Diffusionbee

[trail wing]

I see a couple of photoshop plugins

isekaidev and Abdullah

does one have more tools in it

or are they basically the same thing

[slim quarry]

thank you! afaik diffusionbee doesnt support image to image for making animations right?

[warm junco]

idk but dont think so

[fervent thunder]

Is there some benchmark that compares the speed of Stable Diffusion on Apple M1 chip against a consumer NVidia GPU? I was suprised that GTX 1070ti is a lot slower than M1 ipad air.

[orchid wadi]

They are the files that make up the AI's world. Everything the AI knows is in those files. The stable diffusion system uses quite few different components and they all need their own "world" for the AI to understand things through.

[silk fossil]

But what is the structure of its knowledge?

[echo tulip]

I just notice that it's quite difficult to let characters to look at other directions,

I tried to generate an image about a person reading book

but instead the person is holding a book while looking at the camera.

I tried prompts like read book, look at book, face at book in positive and in negative i tried face at camera, and look at viewer, sadly it didn't work

[ruby scaffold]

Hey guys, where can i find ideas for promts?

[winter pike]

i personally upload images to clip interrogator 2 to have ideas for prompts: https://huggingface.co/spaces/fffiloni/CLIP-Interrogator-2

[ruby scaffold]

wow, thank you

[old marten]

its all magic dude

[silk fossil]

Based magic technology

[fervent thunder]

do most of you use dpm++sde for photo reals these days?

or has anyone found any particular sampler is better for photo real? like euler-a still ultimately the best or ddim idk

[orchid wadi]

I run all prompts against all samplers when generating stuff

They all produces good stuff par for Euler A which tends to make more... wild stuff.

[sleek igloo]

what are the best extensions for converting real pictures into anime syle pictures?

[stark hornet]

any of you around and willing to help me some with textual inversion / embeddings on 1.5 ?

[whole scaffold]

Yo guys, I've watched some trending youtube vid about major lawsuit against SD, it is real deal or some bs?

[slim relic]

I'm also learning about it. Perhaps we can share some experience

[slim relic]

Also, a really good reddit about it
https://www.reddit.com/r/StableDiffusion/comments/zxkukk/detailed_guide_on_training_embeddings_on_a/

[cold siren]

hey people does the ESRGAN-4x upscaler in Extras just make everything super ugly for you too?

[stark hornet]

i got the problem that everything works fine in 2.1 but not in 1.5 @slim relic

maybe we can try to run the very same settings and see if we both get similar results ?

that'd help me quite a bit already

[slim relic]

sry but my SD is occupied with an on-going training

which problem is it?

[stark hornet]

the pictures trained have nothing to do with the training dataset pictures

its just random gibberish

[slim relic]

did you preprocess the images?

mirror+captions

[stark hornet]

preprocess yes, no mirror, no captions, just like 2.1

[slim relic]

these helps a lot with accuracy

[old marten]

its real

but irrelevant

worst case scenario it halts progress for AI companies, whoopdedoo, I'm more than happy with 1.5 models and could get by with this for years

I have a feeling the courts are all gonna side with AI tho.

[old marten]

use an anime model like protogen v2.2

lmfao?

midjourney is way worse than some of the newer models

I feel bad for folks who pay for that when you can have SD for free and on top of that has way better generation + customization

[twilit bramble]

Hi guys, why is this site not working? I click search but nothing happens. https://stablediffusionweb.com/prompts

[whole scaffold]

Thanks for reply!! I feel the same, 1.5 is more than enough

[charred lintel]

when installing bert with git, do you need to change the directory to models/stable-diffusion ?

[cold siren]

modjourney has this like, preinstalled instagram sepia filter

[fervent thunder]

anyone here use dalle 2 lately?

[cold siren]

lol no why would anyone use a heavily censored cloud thing when we have SD

[fervent thunder]

because people into ai are into other ais

[winter pike]

i use it for outpainting

[fervent thunder]

i dont look at ai as a porn gen myself

just curious if theyve done any updates to the model comehu that you know of

[winter pike]

i use the free credits

[fervent thunder]

i havent used it in months myself, was just wondering if theyve done anything worth checking out beyond that inpainting lately

[winter pike]

i havent heard about updates, theyre probably focusing in chatgpt

[cold siren]

lol 🙄 thinking that censorship is only about "porn"

[fervent thunder]

whatever dude i was asking a general question you dont have to go all stable fanboy

chill the fuck out

[cold siren]

hey im no fanboy, im just cloud service hater

[fervent thunder]

so on the topic of stable... why do we have so many samplers when it seems nobody think anyone is better than the other?

like anytime i ask about whats the best photoreal sampler, people seem to say... all of them... which is fine, im not saying a sampler is going to be tied to photoreal, but it just seems silly theres so many when they are all just sort of ok.

[cold siren]

because nobody actually knows what they do 🤷

therefore cannot decide to remove some

[fervent thunder]

sounds about right

the dpm++ all at least work at lower steps right

I'm just trying to think of anything you could actually call out as a positive for different samplers

[trail wing]

that comment about censorship being about porn when confronted

[fervent thunder]

I'm still here franky if you want to make your case

[split narwhal]

@cold sirencensorship is very difficult when you have the SD locally ?

[trail wing]

you're not a judge

[fervent thunder]

neither are you

[trail wing]

but I don't want others to make their case

get it

so that's another clueless comeback from you

[fervent thunder]

me either, I give zero shits for your take on this

but I'll still listen

[trail wing]

anyways off to block

[fervent thunder]

buh-bye

[trail wing]

I understand some people just want to listen to their own bs, but don't make it a public room wank, cya

[fervent thunder]

people sure get defensive about it not being about porn

those who complain the loudest...

[split narwhal]

so what do you guys think of the lawsuit against SD and MJ ?

[cold siren]

that's the point 🤷

[split narwhal]

ok

having censorship is not the way to go

[cold siren]

anyway, has anyone ever used BREAK in the prompt?

[split narwhal]

no

[fervent thunder]

to what effect?

[cold siren]

that's the question 😕

[fervent thunder]

is it supposed to do something or youre just tossing a word out wondering what it might do

why not experiment with it and find out i guess

[visual oasis]

Hello, Does anyone recommend me an uncensored gpt chat collab?

[vale spindle]

hey quick question for anyone around (would be suuuuuuuper appreciated)

i havent messed around with img2img too much, but im playing with it today. Is there a specific setting I can tweak to make the img2img result be as close to the original image as possible? not sure if it works like that but please lmk

[brisk bloom]

lower the denoise strength to the point its nearly the same image

[vale spindle]

Thank you pure!!!

[narrow jackal]

How to use negative prompts?

it's this way?:--no(the stuff that you want remove)

[ruby scaffold]

In the negative field enter what you dont want to see

[narrow jackal]

@ruby scaffold I'm talking about the discord stable diffusion

[ruby scaffold]

ah, ok

[narrow jackal]

Só?

@ruby scaffold so, do you know?

[lucid siren]

😛 hello, i managed to install stable diffusion on my gpu-less pc and i'm getting 7s/it

[old marten]

I hate when people cite their speed on it/s

those numbers are meaningless since each task can use different iteration rates

[sand holly]

what is a DiffusionWrapper?

because i just started Automatic7777's webui and it immediately started downloading it but i never saw it doing that

[rain basin]

anyone just updated automatic1111 today? i'm getting much slower generations after, not sure what changed

like half usual speed

[sand holly]

since i dont know on what machine you are doing it its hard to tell what the issue may be

[gritty fox]

art = depression

BYE

[rain basin]

might be a false alarm, was basing it off the first generation after update

let me average it out over 100 or so and see

[sand holly]

ah ok

[frank rune]

Would someone mind sharing what dreambooth and webui commit you are using? I accidentally git pulled dreambooth :/

[sand holly]

Automatic1111's WebUI is pretty good

atleast you can easily install .ckpt files

and train SD even

[frank rune]

I'm using automatic's webui. Do you know what commit you are using? dreambooth is giving me poor results and I used git pull.

[sand holly]

commit?

never heard of that

whats that

[frank rune]

It's basically the github version of the repository. After starting the webui, it should output "Commit hash" followed by a bunch of alpha-numeric characters, in the command prompt/terminal. It's right before you see the message "Installing requirements for WebUI".

[sand holly]

never heard of it or seen it in person

i just downloaded the repo and ran the .bat file

worked for me wonders

[trail wing]

same speed as always

commit: 91c8d0dc

[rain basin]

thanks, figured it out, forgot add back --xformers to launch.py, need more coffee

[trail wing]

someone else put out SDA - Stable Diffusion Accelerated API

4x generation speed

Based on NVIDIA's TensorRT demo

[orchid wadi]

It is just another Tensor implementation.

You need a TPU capable GPU and about 12gb of VRAM.

[lucid siren]

changing the output size changes the image a whole lot

i though it would give me the same image or similar, but it's very very different

[mortal pagoda]

how was this generated

without being blurred?

https://twitter.com/heartereum/status/1618973299739951105

[brave matrix]

hey everyone! Just posted my second AI HISTORY video, this one on the Vikings! The script is entirely generated by ChatGPT, and all the visuals are Stable Diffusion. Would love some feedback! https://youtu.be/XXANMhv4Bag

[brisk bloom]

I liked it, besides the jittery side scrolling, which was off putting.

[solar heron]

hello, can you send me the server to md?? please

[brave matrix]

yeah i agree! I need to figure out how to improve that

[brisk bloom]

amazing work to achieve, non the less. well done

maybe more frames in between to reduce it

[brave matrix]

you, or maybe will just do scrolling vertically not horizontally? Here is my first video that is just vertical: https://www.youtube.com/watch?v=Vvukxgxi5LA&t=6s&ab_channel=AIHistory

[cedar kayak]

any mod here ?

is art that for example shows nipples, but when a nipple is replaced with a black box/bar , or pixelated or blurred or even some emoji on top of it for here, still considered NSFW ? or would that be allowed? As i not want to do anything wrong

cheers

[brave matrix]

Wondering if I could get some recommendations on the best model to use for my History videos above? I am currently using Dreamlike Diffusion

[twilit peak]

hey all — im looking around but haven't found anything. can you use image urls are prompt parameters in SD, similar to how midjourney works?

[somber sand]

are there any resources for those of us who are looking to create projects / find others to work on them alongside?

[weary jacinth]

So I used SD with the novelai leak originally, haven't been keeping up for the last month or two

What are the best models right now and where would I find them?

[lilac reef]

Anyone in here use kohya?

[wide juniper]

Does Photoshop consume lots of VRAM?

[lilac reef]

Yes, and even its own 2-8 gig pagefile like dealio

[wide juniper]

What's a pagefile?

[vital crest]

Why are there only 9 dreambots? Didn't it used to be 25?

[brisk bloom]

see #📣｜announcements

[vital crest]

I'm gone for a few weeks and the whole sever changes xd

[delicate olive]

Hi from Dominican Republic Guys.👋🏽🇩🇴

Gm.☀️🏝

[vestal dew]

Fusion prompting?

[glad skiff]

Hi all, is there any open-source tool/notebook to generate a NeRF of a landscape scene from a single image?

[vestal dew]

NeRF?

[still glacier]

Did anyone manage to run xformers 0.0.16 using Auto's main branch ?

[glad skiff]

Neural radiance field, a 3D object

[vestal dew]

https://www.youtube.com/watch?v=wS0e5uwy5M0&ab_channel=Graen

[unkempt hatch]

Sometimes I'll crop a section, e.g. 512x1024 for the inpainting, so the mask can see e.g. the entire torso to help it decide where the abs should go, but it seems when I do this, I get like, 20 layers of abs.

Is this because SD just can't handle the ratios, even with inpainting?

[vestal dew]

Click on your possible virus? No.

[wide juniper]

is there any way to use different models in stable diffusion in parallel at the same time?

[vestal dew]

You can merge models

use hypernetworks or LORA's

textual infusion

[bleak matrix]

Good morning, everyone. How are we all today?

[vestal dew]

all to add elements to whatever model you're using

[wide juniper]

I want to work in different styles for different projects at the same time

[vestal dew]

I don't know of a way to use two different models. Someone else my.

may

Hypernetworks, LORA, infusions - all let you add a tag to call up the style

I can make an image of a man. Then I can add the keyword from the hypernetwork to get the same image in the style of the hypernetwork

that's for custom styles

for well known styles, you can add it to the prompt, like "In the style of Dr. Seuss" or some other famous artist

[wide juniper]

True. I want to use anything v3 for anime though, and protogen for realistic stuff

[vestal dew]

I think you'd have to merge them or add an anime style to protogen with a hypernet, lora, or infusion

you can then add weights. Let's say your anime style is called ANIME, you could add (ANIME:.5) and (REALISTIC: .5)

[wide juniper]

Ah, nevermind lol I don't want them to influence each other, however subtly

[vestal dew]

So you want to do something totally anime, then something else realistic?

[wide juniper]

Yeah. As I said, different projects

[vestal dew]

then change models

You can select a different model each time you generate

You can do that in A111, in StableArt inside Photoshop, in Simple Diffusion, lots of UI's

while it's making one, work on and then queue the other

[wide juniper]

I do that, but switching costs time. Maybe it's just me but it takes a minute

[vestal dew]

no, it does take time

Hypernet doesn't. It adds a style on top of a model, but that style wont be just like another model

A111 the model loading is on the front end

Easy Diffusion it's on the back end, when the generating starts.

[warm junco]

switching models takes seconds

[wide juniper]

NansException: A tensor with all NaNs was produced in Unet. Use --disable-nan-check commandline argument to disable this check.
Just add --disable-nan-check to webui-user.bat right? I don't even know what it means lol

[warm junco]

make sure there are on an ssd

[vestal dew]

Mine take minutes

on a HDD

My SSD isn't so large...I use it just for system

[warm junco]

ah okay, yea i bought a 1tb ssd (not m2) a while ago for virtual machines. but now most models take the space there xD

switching takes like 5-10 seconds

[vestal dew]

If I wanted multiple styles with one model, I'd do Hypernet or LORA or text infusion

With Stable Art in Photoshop,. I got a tongue sticking out, but either with a vampire mouth or the tongue as a forked snake tongue

tongue stuck to pole.,..like the scene from ChristmasStory...I'll see if I can get that in photoshop

[winter pike]

just 4 dreambot channels left

gotta make one last generation

[slim relic]

If I want to adapt the style of an anime artist, what is the best approach - Embedding, Hypernetwork, or Dreambooth?

[warm junco]

Embedding or Lora i would say

[slim relic]

aight

I'm getting horrible results with embedding T-T
Idk if it's because of the model or parameters

[ruby lake]

https://google-research.github.io/seanet/musiclm/examples/
Googles new MusicLM Model. Samples at link to listen to.
I'm quite impressed, a lot better than the best out there

[rigid bough]

couple hours left to dream

[winter pike]

embeddings arent the best option tbh, were slowly transition from textual inversions to LoRAs

[winter pike]

Check this, i hope this anonymous group let us play with it. Apparently is a implementation of MusicML, not sure at all https://noise2music.github.io/

[ruby lake]

Interesting, I think the one from Google is much better, but this still looks interesting

I've been telling all my family for a year now that 2023 is when good generative music will arrive (comparable to the quality of images from AI)

Hopefully we can see one more iteration this year so I'm not a liar haha

Its getting close

[winter pike]

i wonder if OpenAI have an ai music paper coming too

i love that ai music is getting recognition lately

[ruby lake]

Well it was trash up until recently hahaha

But with how fast things are moving it won't be long before its near perfect

[rigid bough]

https://discord.gg/e9KnrzJt?event=1068609173987790889

[hidden shale]

does stability has any plan on training a open source language model like chatgpt?

[sullen abyss]

hi , can i ask a question here?

[warm junco]

if its not a #🤝｜tech-support or a #📝｜prompting-help then yes

[sullen abyss]

It's a question about CreativeML Open RAIL M License , Can i publish the images on twitter ?

Or they are under copyright

[warm junco]

do you mean images created with Stable Diffusion?

you can publish them

[sullen abyss]

Yes , the models use CreativeML Open RAIL M License, so i was wondering if i could use them for whatever i want as long as they dont violate any of the terms like "Defame , miss information etc...)

[warm junco]

yea that should be okay. i dont know whats covered in the license but i saw a lot of images posted on reddit, twitter, pixiv, and other websites. I think you should also check if the site allows ai generated images before posting. (the big sites do this)

i think you can even sell ai generated images xD

[sullen abyss]

when i post i always post with #AI GENERATED

To respect the law that was set a few months ago

[sullen abyss]

Yes you can as long that you specify that they are AI gen

that's what i've read

[carmine horizon]

Hello everyone!
I have a quick question which echoes a lot of existing discussions, but on a precise point.

I've been trying out character embeddings for the last days with automatic1111. My aim is to manage, if that is even possible, to be able to reuse a whole character (not just a face and, if possible, a full char under multiple angles) in prompts. Like a protagonist of a story/pen and paper rpg.

I'm having mixed successes with TI embeddings (but I suppose I just need to learn/study more).

Today, I read about LORAs here. I didn't even know they existed. So my question would be..

For that specific purpose (a full char, different angles/poses/action scenes, for storytelling purpose) would switching to LORA training (using A111;Dreambooth) be more useful/give better results than embeddings?
It looks so, if I understand what I'm reading here correctly, but I thought I might as well ask directly.

Thank you! (and sorry in advance, EN isn't my native language)

[warm junco]

An embedding fully trained on one char with a lot of pictures should get you good results, in the Anime Community single chars get trained in embeddings. But now Loras a rising up. I think if your embeds are not that good ask someone who made a good TI to learn. Or try Loras they are great too, maybe even better.

[winter pike]

lora would be a better option, but its "new" and theres too little tutorials compared to textual inversions. The other option would be dreambooth which has been tested a lot of times and gives great results

[carmine horizon]

Hmm.. Thank you! I've had a lot on wonderful advices re:TI a few days ago so I'll try to apply them, see how it goes and maybe ask again should I fail! Thanks for your answer!

[sullen abyss]

Thx for the help btw

[carmine horizon]

Yeah, but I heard a 3060 with 8gb VRAM wouldn't be enough at all to use DB. Am I right? Might give LORAs a try in between 2 attempts at TI 😊

[warm junco]

np 🙂

[warm junco]

you cant do Dreambooth with 8gb i think

[carmine horizon]

Really? Well, I'll definitely give it a try then! Thanks!

[warm junco]

you can try it but heard 12gb is the minimum. also your training time will be a lot higher, (train at night) 😄

[carmine horizon]

Do you mean for LORAs? Or Dreambooth?

[warm junco]

All, even embeds can take long, depending on the settings used

[winter pike]

you can try thelastben dreambooth in google colab

[carmine horizon]

Thank you, I'll give it a try! 😊

[tall hare]

can you use loras trained on 1.5 for 2.x models?

[carmine horizon]

Yeah. I can feel 8gb is a bit limiting (for TI) and training time is high. But I'm in no hurry as long as, in the end, it has a good chance of working 🙂

[warm junco]

i dont think that works

[finite sundial]

Hmm.. GTX 1660 (6gb), running --precision full --no-half in the options since my card has a weird nvidia-ish problems and doesn't render images without those. Worked fine using earlier, now though getting CUDA out of memory issue. Any way to clear the memory and what are the consequences of that?

[bold sand]

Hey anyone there

[sullen abyss]

hey i was wondering i have an Rtx 4090 and im worried for stable ai to damage it , is there any risk ?

The question might seem really stupid

[ruby scaffold]

I think that the video card will be very stressed if you constantly train models on it

but during generation it is unlikely to suffer

[winter pike]

playing current gen games for hours would do more damage than sd tbh, but gpus are made to support certain grade of stress so dont worry

[sullen abyss]

What is "training models"?

[sullen abyss]

okay thanks , i was worried because i knew mining wrecked them and i was wondering if using the gpu for this kind of thing could have the same effect

[ruby scaffold]

well, creating your model from scratch for example

[sullen abyss]

oh ok

60C when generating is ok ?

[ruby scaffold]

under 80 its ok

[sullen abyss]

okay okay

Thanks a lot

Have a nice day and thanks for the help

[mighty gazelle]

is using inpaint to remove shutterstock watermark illegal?

[wide juniper]

No, but using it without license is

(but nobody's gonna give a shit)

1. Will I get into trouble if I use someone's AI art without permission?

2. Also, will I get into trouble if I use Emma Watson's likeness and made an anime out of it, without calling her Emma Watson?

3. Some people are claiming that some samplers do better at realistic skin textures. True/false?

[mighty gazelle]

I have a fear next time I git pull
everything could get erased and a message will flash on cmd

Thank you for using Stable Diffusion, due to recent legal setbacks users will now have to subscribe to generate images on a token basis and all nsfw will be filtered

[orchid wadi]

And what stops you from rolling back?

[fervent thunder]

how close are we from an animation anime model?

[nocturne mason]

Maybe best to make a backup install for safe* keeping.

[remote lake]

Hey guys 🙂 do you know of any tutorials for noobs in both coding and AI on how to start? if one wanted to dig deeper and create own applications or just use the API for fun stuff 🙂 thanks for any inputs

[mighty gazelle]

imma keep a backup on a flash drive and hide it in a retina scan locker

[nocturne mason]

https://tenor.com/view/jerry-seinfeld-i-like-the-way-you-think-gif-14890068

[dapper island]

Hey everyone! I just uploaded a new video on my YouTube channel showcasing some AI-generated art.

https://www.youtube.com/watch?v=edAAu7Fw1u8&ab_channel=BAGIGU

[warm junco]

you should use --xformers. cuda memory gets cleared by closing the webui cmd

[solar flame]

And --medvram should help too

[dire pawn]

Hello!

I have a query

Is dreambot free, and if so then does it have a limited number of generations per person?

[ruby lake]

What law? Where?

[nocturne mason]

Right?

Most supposed "law" is literally written in fictitious conveyance of language

[old marten]

Never git pull

I got 2 philosophies in life

1. Never pull out

2. Never git pull

[brisk bloom]

anyone know what the up/down arrows are meant to do on txt2img, to the left of batch count/size?

I think theyre new?

[waxen pasture]

Dreambot just got shut down for a bit I think, you missed it :( But from my understanding it was 100% free I think

[dawn yarrow]

free man unlimited

[warm junco]

They are to switch widht and hight with each other

Best new feature

[brisk bloom]

ah right. didnt notice them changing lol

yeah nice feature

I guess cause i was on 768 square

[warm junco]

Haha yea and if you start the webui its 512x512 so no one knows whats changing xD

Its perfect for my 512x768 Portrait and then switch to horizontal

[brisk bloom]

definitely useful

would like a button to match resolution to image size for img2img though. that would be so cool

[warm junco]

That would be nice!

[nocturne mason]

https://tenor.com/view/impress-based-american-psycho-very-based-impressive-very-based-gif-21665669

[rancid lance]

dreambots are gone forever?

[elfin peak]

Just curious if anyone knows. As I'm making comparisons using S/R and X/Y/Z, the output is a grid. But when the grid is gets bigger, the images on the grids gets smaller, making it tougher to see the difference.

Is there a grid save setting that makes the final grid image original sized, making a huge grid?

[nocturne mason]

This dream bot is for ever.

[cloud granite]

Yes

There used to be a total of 1.1 million images in this server, now it's down to about 266,084

[toxic minnow]

I would love a series of wallpapers with hundres of little pompt words listed all over, to spark my imagination whenever I'm zoned out and thinking about ideas

[fervent thunder]

so what we can use dreambot ever again or just not for rn?

[old marten]

I have a prompt book with pictures if you wanna take a look at it

[toxic minnow]

a prompt book? like a physical book? sounds cool

[old marten]

No, it's digital I would show you a screenshot but my computer's busy running dreambooth

[fervent thunder]

what’s the reasoning. they took dreambot off?

[cloud granite]

I was not really expecting DreamBot would be removed, they only started removing it a week after I joined here

[old marten]

It's because of the lawsuit

[cloud granite]

Oh no

[old marten]

Real power users run it locally anyways

Sabrina, do you specialize in older vehicles because you're old

?

[fervent thunder]

i don’t have a pc that sucks

who’s doing the lawsuit ?

[toxic minnow]

So the DreamBot model reservoir is still in circulation I guess?

[cloud granite]

I am not that old, but however I have found an interest since I was young.

[river maple]

did they actually say that?

[old marten]

No its just speculation trolol

[cloud chasm]

It would be impossible to chase down all the checkpoints, and their variations.

[sacred iris]

Anyone know good dreamboat alternatives ? I've tried out some web implementations but none are as robust as dreambot

[cunning temple]

me over at midjourney with 100+ alternate accounts:

[astral goblet]

keep in mind that the lawsuit is in america, and stable diffusion is in uk. so like, if it wins, what can they do really?

not shut anything down

they didn't sue in UK for good reason

fair use is fair use there. the courts would've said that right away. no middle men litigious bs required

i guess microsoft owns github so they could force microsoft to not host that particular git, as if that matters

[fervent thunder]

https://mingyuan-zhang.github.io/projects/MotionDiffuse.html

How long until stability AI trains this paper with Stable difussion dataset?

[astral goblet]

that just generates a motion vector for a human style skeleton. it can then be applied to a model. I think what you want is Stability to relase an ai that generates 3d animated models. Perhaps that could be done! This sytem wouldn't benefit from training on laion though. I think it would be used along side a model that textured a model. I think that can be done with Stable diffusion already

it'd be really neat to see if a system could create a nerf around those motion vectors, but again that'd be separate from the motion generation and trained far differently

i think the software is here already, but memory requirements might be beyond 24gb for many of these models

https://www.reddit.com/r/StableDiffusion/comments/1075z6u/stable_diffusion_can_texture_your_entire_scene/ tools like this already exist. wouldn't be far out for a procedurally generated human model to be textured by stable diffusion, and then animated with vectors

[wide juniper]

1. Will I get into trouble if I use someone's AI art without permission?

2. Also, will I get into trouble if I use Emma Watson's likeness and made an anime out of it, without calling her Emma Watson?

3. Some people are claiming that some samplers do better at realistic skin textures. True/false?

[fervent thunder]

which sampler?

I've never seen a good defense/reason to use any particular sampler. other than like those ++ sampler work better at lower numbers of samples like 10-15

[astral goblet]

if someone else created the art, it's their intellectual property and you require license. fair use is still fair use though,

2. if you're obviously infringing on emma's image without her consent, yes. It may also be against rail-m license of SD and all models refined from their base models

3. selection and confirmation biases exist quite a bit in the AI space. your best bet is to do your own research, since a 6month old technology has no masters

[peak orchid]

1. Yeah, someone can still do a DMCA takedown among other things against you if you take their work and claim it as your own. AI art is no different. 2) Making fan art of people can sometimes land you in hot water, regardless of whether its AI assisted or not. Its even more dubious if you plan to try and make money from the likeness of someone.

[dawn yarrow]

1 dont even try

[narrow jackal]

Intellectual property doesn't exist

[dawn yarrow]

then explain how i own your every thought

[narrow jackal]

Haha property is something scarce

[dawn yarrow]

intellectual property exists as an idea like santa

hes pretty popular

[narrow jackal]

@dawn yarrow haha

[dawn yarrow]

😉

[narrow jackal]

@dawn yarrow do you use stable diffusion through Pc?

[dawn yarrow]

omg i wish but gpu is lacking

stuck with collab at best

[narrow jackal]

@dawn yarrow completely free(unlimited)?

[dawn yarrow]

taggy aggy

[narrow jackal]

@dawn yarrow what?

[dawn yarrow]

tag me again so i can lay in peace i suppose

idk man

i want image to image with gimp masking really bad

6 genereations and its gold

brb

[narrow jackal]

@dawn yarrow ok, have a good night

[astral barn]

What do you guys do?

[peak orchid]

That's because I wrote your code so that it owns every one of his thoughts, as part of one of the experiments for this simulation. Just don't compare backstories as I got lazy and just used the same one for both of you

[dawn yarrow]

progamer has it under wraps

[astral barn]

The problem with Midjourney is that it's too expensive. In order for the server to become free, the authors will have to pay dearly for the servers they bought. This is necessary for maintenance.

[dawn yarrow]

what do i do?

seems like a broad question

im um disabled

for a living

mentally nutz yep thats it

[astral barn]

It's kind of boring. Bluewillow generates not exactly what you want. I've already tried adding various hints. prompt, the result is the same.

[dawn yarrow]

do image to image!

with transparent masking in photoshop

[astral barn]

I'm sitting like an asshole and waiting for an update.

[dawn yarrow]

well you might be sitting for weeks ?

stable diffusion on pc

cmon it works wonders

[astral barn]

It's possible to sit on your ass for weeks.

I have an android. On a PC, it takes a lot of effort to install. It's a pity that it's not in the exe file. Ready. You have to compile it first.

It's a pity that this is not a ready-made exe file.

[forest apex]

There is an installer for automatic on github that does most of the work - https://github.com/EmpireMediaScience/A1111-Web-UI-Installer/

[peak orchid]

I have acquired a large enough dataset to potentially give us an MJ quality model. While I work on getting the training up and going, I am considering trying to train with the best SD 1.5 content as well on the SD 2.1 model

[narrow jackal]

@peak orchid do you know how to use stable diffusion through android?

[peak orchid]

My earlier experiments show that training SD on MJ content can result in higher quality content than both MJ and SD, so now I need to do that on a bigger scale

[peak orchid]

Web browser via whatever UI you are using (gradio makes URLs you can use), or download a remote connection app. Be warned though that both options have can have security concerns

[narrow jackal]

@peak orchid ok, thank you

[forest apex]

you can also try using something like stable horde

[waxen pasture]

please tell me someone else has had this issue or knows what might be causing it, I am banging my head over this #🔧｜finetune message

[forest apex]

@waxen pasture did you update your webui after your first good results? something may be broken in the UI itself

[waxen pasture]

That's what I am thinking, but no one else has said anything

And I don't know which version I got the good results on

[forest apex]

I haven't been able to train dreambooth since around december, but I don't know that everyone is having the same problem.. you can try and look at the date of the embedding and revert back to the commits around then... or install to another directory and do that

[waxen pasture]

true

but it's getting too late to do that today

I am trying one last thing, about to see if it's working

I do not think it's working

I tried using gradient accumulation

[obsidian geode]

Hello guys, is there a way to expand the image when the image is only a face to showing a little more of the shirt?

[waxen pasture]

outpainting

[obsidian geode]

so I put my face in a bigger rectangle then I outpaint?

[waxen pasture]

if you are using a1111 there are many methods of outpainting, like using the outpaining scripts in the scripts section in img2img, or extensions like openOutpaint

[brisk bloom]

check the date on an image that you created which was good. Then do "git log" to see the commit versions and revert to that

[waxen pasture]

Thanks, I will do that tomorrow if I can find the time. Will that remove everything in my generations folder? Should I back all my models and images up first?

[brisk bloom]

it wont change models folder

[obsidian geode]

Should I do it progressively, like I add shirt at the end of my prompt and I add 30px vertically eac time then I do outpaint on the bottom?

[waxen pasture]

what about "outputs" folder?

[waxen pasture]

That's probably a good idea. See if you can find any outpainting tutorials on the interwebs.

[obsidian geode]

Alright!

[brisk bloom]

fairly sure it wont touch that either. its the sme process as updating, but just for an older version

no models or outputs are touched during updates

[waxen pasture]

I see

[brisk bloom]

https://youtu.be/0fkGd9wIhrA?t=84 this video shows the process

[waxen pasture]

Also I think I know the exact issue that's happening but don't know a fix other than trusting the process of the training, because what I think is happening is that I have it set to save every 50 steps, and it's creating the exact same embedding every 50 steps instead of continuing training on that embedding. Just a hunch though, might not be right

[waxen pasture]

thanks

actually, I could probably try it now

I've decided to do a clean reinstall, it's probably about time for that anyway

[fervent thunder]

speaking of fresh install, are any extensions in auto regarded as must have? i did a new install a bit ago and havent bothered to dl any

[brisk bloom]

systeminfo is really good to have

[forest apex]

Image Browser is very useful

[fervent thunder]

there's so many of them. is there anything that actually helps with fidelity/quality of image?

[lilac reef]

What does it mean when CLIP says something like this when interrogating a photograph? a sci - fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi

[fervent thunder]

maybe it's singing the Breakup Song by Greg Kihn Band

[lilac reef]

No, there is a reason I just forgot what it was.

Can fix it too

[fervent thunder]

fair enough. the song is actually uh uh uh uh uh, for some reason i thought it was fi fi fi fi fi fi. old song nobody even knows what the hell im talking about 😂

[brisk bloom]

maybe it doesnt understand sci-fi and wants science fiction?

[lilac reef]

I am not telling it anything that is it interrogating my image.

1 out of 15

[astral barn]

Dreambot is being shut down by the gentlemen. Clear.

It's a pity, of course, it was a good bot.

Press F. R.I.P.

I suggest we all get together. And via chatgpt. Create an alternative to Stable diffusion. It will be free for everyone and without restrictions.

[jagged cave]

@flint sparrow what was the prompt you used for the "metsmorph" image?
dont really care about model or anything. but the prompt may help me greatly creating a model with similar sorta transformation art stuff!

[warm junco]

Stable-diffusion is free for everyone. Since the code is OpenSource everyone can use it. But to generate images for free in a discord for thousands of people alot of GPU and Server Hardware is needed. And thats a Cost they may want to reduce.

[astral goblet]

the notion that stable diffusion is "restricted" is just insane propaganda stirred up by the unstable diffusion discord people, who wanted to create hype for their fundraiser

[forest apex]

There's already stable horde which has a discord bot. Donno if it could handle hundreds of users unless a lot more people contributed GPUs, though

[narrow sluice]

Uh-huh, how are they doing? Do they have a billion dollars worth of venture capital funds?

Sounds kinda weird, biting the hand that feeds you.

[elder cradle]

Are there still dudes that say that AI can make nsfw images of them without their consent, thus wanting it banned?

[ornate pilot]

Ayo where did all the dreambots go?

[forest apex]

check #1047610792226340935

[ornate pilot]

Tthnx

[wide juniper]

Is Dalle also a diffusion model?

[forest apex]

yup

[fervent thunder]

RIP Dreambot (again) Gone but never forgotten.

[wide juniper]

What's the point of forward diffusion process if it's just about adding noise? Why can't diffusion models just begin from plain noise?

[ornate pilot]

They just updating it - read the thingo at the very bottom folder on the left 👈🏻 it’s all written in there

[fervent thunder]

Under what header?

[ornate pilot]

Dreambot discussion then dreambot chat

Scroll up a bit

[simple solstice]

what's THE guide to install SD on linux?

[ornate pilot]

#1047759610737610784 message

[fervent thunder]

Brill. I hope it remains free but obviously its CPU intensive so understand if they charge people. Midjourney probably makes a lot of money charging their users.

[ornate pilot]

I actually don’t like midjourney’s results, stable diffusion can be rough in comparison but I like the process of working with the Ai’s strengths and limitations to get desired results

[fervent thunder]

I prefer Stable Diffusion too probably because I can run it locally lol

[ornate pilot]

If you can run it locally why are you looking for the bot servers in here lol

I can’t run it locally, i have a windows 8 potato laptop and an iPad 😆lol

[fervent thunder]

I have a 4gb laptop and the reason I like the bot in here is so I can make an image when I am out and about. Plus I learn from the prompts people use.

[empty barn]

Use A1111's webui.

[fervent thunder]

Try with the links at the bottom of #1014939219904450590

[ornate pilot]

I went through like 20 versions of this to get this expression on her face that I wanted 👉🏼 https://www.flickr.com/photos/rickmolon/52658278375/

[placid flax]

what should be best to create NSFW, you know like those on Pixiv ?

[ornate pilot]

I wouldn’t get a result like that on midjourney

[fervent thunder]

Thats a great image.

[ornate pilot]

Thnx

[fervent thunder]

I dont make them as I dont have a use but a quick google brought this,it may help - it may not. I dunno https://stealthoptional.com/tech/stable-diffusion-nsfw-how-to-make-nsfw-images/

[compact quiver]

Does anyone know of a good colab link that will allow you to training a model and merge it with other models. I'm using Fast-dreambooth, but when I click on checkpoint merger there's nothing in the dropdowns to merge with.

[placid flax]

thnx

[fervent thunder]

Sites like https://civitai.com have hundreds of models, maybe check there too. Not sure if there is anything that is specifically NSFW but could be worth a look.

[unreal eagle]

did the 98% booster come out yet?

also will HANDS be fixed with any models or AIS or diffusion models or GANS ever???

why are hands so messed up

[dark marsh]

hihgg

[finite sundial]

Sorry for taking so long to answer. That command will clean up the memory? Since memory isn't cleaned by default after webui is closed, at least for me.

[odd isle]

Has anyone trained a latent diffusion model here? I'm trying to generate butterflies from the "huggan/smithsonian_butterflies_subset" ds but simply cannot
In pixel space it is quite easy however when working in a latent space it totally fails, have you any idea?

[finite sundial]

And the thing says that there's 5Gb allocated and 5.14Gb reserved by pytorch with 0 being free.

[warm junco]

No the Command will decrease the vram usage while generating images. It prevents that you get that cuda out of memory errors that often

[finite sundial]

Didn't really help unfortunately. Memory's flooded full and is not being cleaned~

[warm junco]

Did you editet your Webui-user.bat ?

Make also sure that only The webui is running.
If your vram gets flooded once, only closing the Programm causing it to get cleared again

[finite sundial]

Yeap. It has three command arguments in it and that's all. All the git/python paths are in webui.

[warm junco]

Also youse --medvram that could help too