#💬｜general-chat

1 messages · Page 47 of 1

lapis rain
#

hello all, is there a tool or service that can connect my images? I have 100+ images with faces, some are similair and some dont. wonder if there exist and service that can match them for me

bleak matrix
lapis rain
#

a service that can tell if a face looks similair to another face

#

want to create folders for each similair face

#

so i can categorize them

empty iris
#

Does anyone know if there's a nice open source text2video generator out there? similar to runway gen-1?

shut hedge
queen tendon
#

Hey guys, I've only took a few days away and now I see all kind of posts with lora tag, what is lora? is it like dreambooth?

astral goblet
#

fast and easy to train

queen tendon
astral goblet
#

theres also LoCon, LoHa too

queen tendon
#

any idea if its possible to do lora using 8gb vram or colab?

astral goblet
#

i think so yes. it's designed for low vram training

queen tendon
#

If we make comparisons with dreambooth, do we get almost the same quality or dreambooth still wins?

astral goblet
#

and you can merge loras too

queen tendon
#

I see.

#

That's really good, my hdd almost reached its limit with dreambooth 😅

astral goblet
#

i'm still researching this all too so , take what i say with a grain of salt. i might be misunderstanding some things but its hard to tell. so much to take in

queen tendon
#

exactly, you miss one week, and you get back and its like a new world here.

astral goblet
#

2.1 has controlnet models now too

queen tendon
#

is it any good? I still use 1.5

#

I tried 2.0 and it gave me really bad result.

#

and I heard the 2.1 is not that much of a big difference.

#

if only we had 1.7 : (

waxen pasture
queen tendon
#

yeah, the workflow is changed.

waxen pasture
#

training is a lot more important, but not enough people are training for it.

#

So peopple just assume it's bad.

queen tendon
#

people like to get quick results so it makes sense.

waxen pasture
#

I did make an embedding that can improve results, but it's for digital art specifically

queen tendon
#

not to mention 2.0 trained with wrong values.

waxen pasture
#

Yeah, I have never used 2.0, I only use 2.1 occasionally.

queen tendon
#

are there any rumors about a new model?

astral goblet
#

openclip instead of the old closed off clip models that can't be further refined due to licensing restrictions

#

so it's a brand new pretrained model on 2 and forward

#

not rumors. Actual work being done. 3 is getting the ground work laid. datasets are being built

queen tendon
#

Doubt they will release it to the public.

waxen pasture
waxen pasture
queen tendon
waxen pasture
#

That doesn't mean they will never release it to the public.

queen tendon
#

Hopefully I'm wrong and you are right.

sleek osprey
#

I see the laoin 5b images are of different sizes. How did SD train with different size images?

waxen pasture
#

I hope so too.

peak apex
#

where i can make images in this discord channel ?

sleek osprey
#

pretty sure

surreal perchBOT
#

@peak apex

FAQ: How do I generate images?

Currently there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or running Stable Diffusion locally using your own system resources! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using the website or #🤝｜tech-support for any problems you encounter while installing it locally!

peak apex
#

is there a dailys credit ?

#

how to enable NSFW ?

drowsy vigil
#

Has anyone tried to sell images they created with stable diffusion?

astral goblet
# peak apex how to enable NSFW ?

Many of the commercial services don't sell the ability for NSFW generations because then it opens them up to all the regulations and laws around hosting and producing pornographic images. Look into stable hoard. i don't use it so i dont know my way around it but i've heard they've made it simple to use.

#

community cloud of stable diffusion hosts that donate their compute power to the hoard for free

untold moss
#

So - is there a plan to solve the "hands" problem?

hearty dune
#

Guys, could anyone help with automatic1111 on collab with controlnet?

#

does it have to be in the img2img tab? and do you need an input img in the img2img tag, besides the controlnet input img that is

rose granite
proud nova
#

That's the only thing you can do about that. We'll see what the future brings

rose granite
#

can somebody give tip how to make good prompt?

agile pebble
proud nova
rose granite
#

when I use very detailed classic prompt it draw what I want but lovres

proud nova
#

Then use high res. fix

rose granite
#

does it work with local automatic 1111?

proud nova
#

It will make your image twice as large and will add some details

proud nova
rose granite
#

this is option or extension?

proud nova
#

Can't send images here...

rose granite
#

ok, I`ll google it

proud nova
#

Check it out

covert wyvern
#

Does anyone here use Deforum to create animations? I'm curious if you have any advice about keeping the prompt subjects in frame. Sometimes the prompt objects kind of disappear and all you see is a solid color or a few areas of color.

proud nova
#

I tried it, but can't say I have experience with it

hearty dune
#

how do you upload a depth image to controlnet in automatic1111 colab? i mean, not uploading a picture and getting the depth map, but uploading a depth map directly

vast ingot
hearty dune
vast ingot
#

but in both case it should be that, yeah

#

you download them, put them in your drive in the right folder

#

and it will see them on next start

hearty dune
#

Again, Thanks!

hearty dune
vast ingot
#

or maybe I'm missing something

hearty dune
#

I mean in controlnet... will it still wofk in text2img?

vast ingot
#

yeah. I mainly use controlnet on txt2img personally

#

it's different

#

it doesn't "seed the colors" in

#

meaning you can have a lot different content in the picture

#

but it still follows

#

(can't link pics here)

hearty dune
#

i see, in img2img, which input "comes first"? if i upload a specific picture of a tree, upload a picture of a man to controlnet and prompt to generate a robot, will it retain the man's shape still?

hearty dune
vast ingot
#

the init image in img2img will be the initial noise. it will be the initial image that SD will modify, keeping it very close (or less depending on the controlnet settings you use)
same in txt2img, it will let the controlnet image just be a control, a way to push outputs in a given direction, but it won't force the first picture entirely as initial noise

vast ingot
#

in that case, the picture of the tree would be the initial noise, it would be the picture you would get if you set it to 0 steps, or denoising to 0

#

the picture of the man, in controlnet, would depend on the controlnet model you use. but it would basicaly force the output to be in the shape of a man

untold moss
#

has anyone figured out why hands are such a problem?

#

is it really that complex a problem?

vast ingot
#

and the prompt robot would force the image to create a robot as much as it can.

#

so in the end, it would morph the pixels of the tree, and try to make a robot in some kind of green color inside the shape of a man

vast ingot
# untold moss is it really that complex a problem?

it's quite complex yeah. lots of intersections, positions, angles to see them from, ... and detail level changes a lot in the sources.
controlnet makes a great tool for that though, it makes great hands using depth map

untold moss
#

Thanks Guizmus.

vast ingot
untold moss
#

So I tried using controlnet on hugface and ... it was bad. it was like SD1.5 quality. How finicky is controlnet?

vast ingot
# untold moss So I tried using controlnet on hugface and ... it was bad. it was like SD1.5 qua...

there are quite a lot of details. you can go very far in it. the main tips I could give after some hours on it :
1/ go into the settings tab, and push "controlnet quantity" higher. you can have depth + canny + normal map at once for example, and it brings more fidelity
2/ don't hesitate to use use an inpaint mask with it too, so that some part of the picture don't get modified at all.
3/ try to play with the "threshold values" on canny mode, and the "preview annotator result". This will give you an idea of the details that were picked up by controlnet, and that will be forced on your output. you can change those values for more or less fidelity to get better quality

untold moss
#

ok

fervent thunder
vast ingot
# fervent thunder do you ever make 1.5 embeds?

not that much, no. I did some text inversion long time ago, but mostly felt in love with fine tuning full models. I work on a bigger dataset on my side, but LoRA don't seem to lend the quality I want, compared to classic dreambooth

fervent thunder
#

i was curious if you knew if training 1.5 embeds you need to use the standard model

#

i was trying to make an embed using a realisticvision model since thats what im using and it comes out all funky

waxen pasture
fervent thunder
#

where do you even make loras in auto?

waxen pasture
fervent thunder
#

oh right

vast ingot
#

I did a little guide on that with lots of links

#

all training methods are presented

waxen pasture
fervent thunder
#

ok so I guess you cant train on anything but the base model then

#

I thought maybe I was doing something wrong with a vae or god knows what else haha

waxen pasture
#

It's best to train on the base model and it will still work well on other models.

#

My embeddings are all mostly pretty good.

fervent thunder
#

fair enough. i was just getting into this embeds a bit. inspired by those reddit posts of that guy putting himself in historical scenes. doing that with myself. but i thought since i used a realistic vision model an embed of myself would be best from that model too but i guess not lol

waxen pasture
#

Train it on the base model and use it on the realisticvision model

fervent thunder
warm wind
#

Hi all - I have a specific application and not sure where to start. In Architecture, we frequently create construction documents in packages that are not suited to rendering. However clients frequently ask for renderings, and much time is wasted converting black and white axio drawings into colorful renderings to please the client. Are there any good SD applications that would keep the general structure of a drawing, but be able to change the style to a rendered look? Or even something to feed a 3d model into and then have it supply some textures that are generally popular in the architectural renderings category of internet images.

waxen pasture
vast ingot
vast ingot
tulip tangle
#

controlnet not working...after giving command...showing this in command prompts..

#

controlnet not working...after giving command...showing this in command prompts.. Loading model: control_sd15_canny [fef5e48e]███████████████████████████████████████████| 16/16 [00:04<00:00, 5.74it/s]
Loaded state_dict from [D:\stable-diffusion-webui\extensions\sd-webui-controlnet\models\control_sd15_canny.pth]
ControlNet model control_sd15_canny [fef5e48e] loaded.
Loading preprocessor: canny
0%| | 0/16 [00:01<?, ?it/s]
Error completing request
Arguments: ('task(4eqh3zlvc0ora2o)', 0, 'robot doll in a futuristic city', '', [], <PIL.Image.Image image mode=RGBA size=512x512 at 0x28199DDF880>, None, None, None, None, None, None, 20, 0, 4, 0, 1, False, False, 1, 1, 7, 1.5, 0.75, -1.0, -1.0, 0, 0, 0, False, 512, 512, 0, 0, 32, 0, '', '', '', [], 0, True, True, <scripts.external_code.ControlNetUnit object at 0x0000028199E1FBE0>, False, False, False, False, '1:1,1:2,1:2', '0:0,0:0,0:1', '0.2,0.8,0.8', 150, 0.2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, '<ul>\n<li><code>CFG Scale</code>

#

couldn,t add the whole response

dreamy igloo
#

An update on the copyright situation, involving suspected systematic removal of images. Many of the reposted images with altered logos bearing resemblance to a particular TV show franchise survived after I removed the logos. However, there are still occasional, random, but systematic removal of some of my posts after that.

At first I thought it was because the name of the show, or a name that resembles the name of that show appeared in the write up/title of post. So I added dots(.) and hyphens (-) in between the letters to avoid my post being spotted and removed/reported. It worked for a day or two. Then in a day or two I post new parodies, and they get reported/removed despite the altered title/text. So today I try to remove any text or title related to the show. The only way for anyone to know the reference to the show is if they are familiar with the main character. Yet again, today, 1 of my post gets removed within about 5 minutes of posting on instagram. There was never an official complaint sent to me, and I have no idea on why some of my post get targeted.

The posts I made a few days ago bare an altered title in the write up, contained character images that look very close to the main character of the show, and they survived. But today, I made several attempts to downplay the resemblance, yet somehow the new post get struck in 5 minutes or so, its like they have a bot that remembers the images in the post, and strikes at them whenever I repost it, despite me trying to change all kinds of words to avoid detection. Yet the old post that bare greater resemblance survive all along.

#

I tested more today by posting a second post with a set of images that also bare noticeable resemblance to that show, but started off posting it with a write up that bare no mention to the show. Despite having a recognizable character, that post stayed on all the hashtag spaces and did not get reported/struck down.

I am worried that there is a bot looking for posts that textually mention anything close to the show, and reports them. And that instagram remembers the associated images, so that even if I re-word the post, they will just simply strike it down again.

But I am unsure as to how they find and gauge what to report/strike down. My older post bare more resemblance, with both text/images close to the show. Yet they survived for days now. I am just worried that they will keep tightening the noose, such that any new material I post, so long as there is even a slight hint of resemblance, will be reported/struck down.

#

If they keep doing that on instagram. I'd consider using twitter for this set of images/stories involving this franchise. I am suspicious of a systematic crack down of a.i. art related to this franchise since I've seen all kinds of super-hero a.i. art surviving, but I've seen no one post this franchise (maybe there were, but were similarly quickly struck down by the intellectual property owner). This is a shame that the franchise cannot even allow fanart made in a.i. art form. There are drawings of those heroes on devianart and they were never censored like this.

carmine gale
#

i keep getting "AttributeError: 'Options' object has no attribute 'sd_lora'" when using Lora
how can i fix it?

tulip tangle
#

getting controlnet error like this Loading preprocessor: canny
0%| | 0/20 [00:01<?, ?it/s]
Error completing request

zenith bobcat
#

just found out about invoke

#

seems so convenient

#

woulda saved me a lot of time and effort

vast ingot
#

it's quite a fun tool to play with

#

love its outpainting and inpainting feature

fervent thunder
#

hmm whats invoke?

vast ingot
#

invokeAI

fervent thunder
#

new company?

#

oh web ui

vast ingot
#

no, it's another tool, that isn't as popular as automatic

#

it has some really great things going too though

fervent thunder
#

updated as often?

#

or at least updated frequently

vast ingot
#

less. but when it is, it brings features that rocks. but yeah, that's the big problem of it

fervent thunder
#

cool. ill have to check it out

vast ingot
#

it's very solid, and continues to work, it has a lag behind though

#

more "designer" oriented

#

let me show a "speedrun" using it lol

zenith bobcat
vast ingot
#

really not the same approach

indigo ore
#

Hello

zenith bobcat
#

it does all the suffering of python being unwilling to cooperate in a one click installer

#

quite awesome imo

raw narwhal
#

Does anyone know how a 1070 would perform in SD?

Got loads of GPU choices, want to spend as little as possible lol

fervent thunder
#

yeah i see that auto installer. very nice

zenith bobcat
#

soooo pleasant

fervent thunder
#

gonna check it out when i move over to my rig later

zenith bobcat
#

tho idk how many vram you have

#

i think minimum recommended is like 4 GB

#

it runs on my 1050ti so uhh

raw narwhal
#

1070 is 8GB, the best choice i have around this price range is a 1080TI (11GB) or a Titan X Pascal which would be 12GB

zenith bobcat
#

idk

raw narwhal
#

The question is, how large of a render would this cap out at?

#

i would love to render in 1080p lol

zenith bobcat
#

i dont think theres a size limit on any gpu, its more of a question of how long itd take

#

correct me if im wrong tho

fervent thunder
#

doesnt going larger than 512 just make the model funky? like it starts stretrching more the bigger you go with image size

zenith bobcat
#

also this

#

you can upscale tho which is basically the desired result

fervent thunder
#

yeah

zenith bobcat
# raw narwhal i would love to render in 1080p lol

but yeah basically these models are designed to work in 512x generally and dont really know how to deal with larger resolutions, so your best bet is generating one at 512 or at absolute most 1024 and then upscaling

raw narwhal
zenith bobcat
#

which i think some webuis do automatically with a checkbox

fervent thunder
#

or use a 2.x model you have more grace upscaling since its 768 base

raw narwhal
#

Ah, so instead of natively rendering at 1080p you upscale?

zenith bobcat
#

yeah

raw narwhal
#

Interesting because ive seen some really high end renders that just cannot be possibly upscaled

fervent thunder
#

i can never find a good 2.x model i really like. maybe i just am liking the faster speed still with 1.5 models lol

zenith bobcat
#

generally makes for better results that way

raw narwhal
#

So are we thinking 8GB of VRAM is enough?

#

Got a guy with a 1080 and pretty ready to purchase

zenith bobcat
#

i have never tried anything over 4 tho so uhh

#

im not the best guy to ask

raw narwhal
#

Ive tried 16GB

#

with CPU rendering

#

Atleast it was faster than my 970

zenith bobcat
#

because they uhh

#

removed all nsfw from the dataset i think? which means way less images of humans which means way less knowledge of anatomy which means way worse humans on generated images

#

i heard smth like that

#

like for legal reasons either all nsfw or all images of children had to go, you cant have both in the same AI

#

you can uhh

#

guess why

#

so they removed all nsfw right before 2.0 is what ive heard

#

which is fair on their part tbh

fervent thunder
#

yeah I think thats what happened

#

2.x stuff just seems a struggle to me to get good results

zenith bobcat
#

yeah just use 1.5 thats what everyone does

#

afaik

fervent thunder
#

hopefully the next official model brings the wow factor back

zenith bobcat
#

unlikely but hope dies last

fervent thunder
#

both stable and dalle 2 seemed to be the big news makers last year, then they both just faded to the midjourney updates i think

#

i was looking at dalle2 the other day, first time in months, i assumed theyd have updated it somehow, but it seemed the same

zenith bobcat
#

midjourney more like uhhh mid journey

#

like its mid

#

like its pretty uhh average like its alright but like yknow like

fervent thunder
#

yeah im not paying for it myself, but they at least keep pushing out models that wow the general public

#

and supposedly they have their next update about to release

zenith bobcat
#

main reason im using sd is because you can put all sorts of things in it

fervent thunder
#

yeah for sure

zenith bobcat
#

like theres models and hypernetworks and embeddings and i still have no clue what a lora is but theres that as well i guess

#

thats what makes it so cool for me

fervent thunder
#

lora seems like just another type of embed to me

#

only you can control the weighst easier with a slider

zenith bobcat
#

yeah i mean it probably does what a hypernet does just in a different way

#

anyways its late gn guys tomorrow ill check out invoke and like idk make a big chungus hypernet or smth

oblique star
#

What are decent system requirements to run SD?

vast ingot
#

to run minimum, 4GB
to run comfortably, 8GB
to have the full experience with some training in it, 12GB
to have the mega full experience with the motherload training or very fast/big pictures, anything above

#

depends a lot on your price range

#

if you have access to 16GB and upwards, it gets quite fast too

#

like I make 8 pictures of 640x512 in 10 seconds, on 24GB

oblique star
#

Ok, and you are talking about VRAM or RAM?

vast ingot
#

VRAM

#

you don't do anything on RAM, almost, in SD

#

there are some small benefits of more RAM, but not on the picture generation themselves

oblique star
#

And If I buy a server, can I see how much of a VRAM it has? Because normally there are only RAM shown

vast ingot
#

there are other requirements that I could go on about, like:
processor speed, disk speed, for faster loading of SD. not that useful in picture generation
processor speed and ram for swapping models faster/keeping more things in cache

vast ingot
#

I do it at home so I'm not sure

zenith bobcat
#

would it be possible to use RAM to generate images thatd be funny i think

vast ingot
#

it is possible, RAM + CPU

#

just takes 10+ minutes for a single 512x512

zenith bobcat
#

so awesome

#

bc i got 16 gigs of RAM and only 4 VRAM

#

so uhhhhhhh yeah

vast ingot
#

you should be able to run it on vram though

zenith bobcat
#

i should

#

im gonna try

#

i just wanna train too

vast ingot
#

first by removing from VRAM both process that take the most usually : windows and discord

#

then colab

#

tbh colab is really good right now

zenith bobcat
#

fair enough thats what i been doing so far

vast ingot
#

it's enough to train or make pictures for 2 hours per day for free

#

I get it's not unlimited

#

but it's quite the offer still

oblique star
#

VRAM is part of RAM, correct?

zenith bobcat
#

VRAM is part of your GPU

vast ingot
#

most people that I see that do big training/have big requirements and do it online (and I mean 100+GB VRAM), do it through notebooks, either on google colab or other online services that rent GPU that way

#

if you intend to buy a server for your home for that though, go more on the A100 road

#

there is a very good comparison pinned

oblique star
dreamy igloo
#

not only do you need a gpu, but go team green (nvidia) who has a monopoly on cuda and rtx. I have not looked at amd for over a decade because of those 2 things being important in games/simulation. I don't like the monopoly but thats how it is now

#

If you just blindly buy a GPU and wanna save a buck and go team red amd, you will shoot yourself in the foot in the world of ai art (so far)

fervent thunder
#

yeah i just go with nvidia for years now

#

no complaints either. all the cards ive had have been great

#

Hey, im new to stable diffusion and dont understand some things. When downloading Citivai models and using a LORA model, there are training images given. Do I need to download another software to use these training images, or is it possible in stable diffusion itself?

brisk bloom
slender quarry
#

Rip karrstuen

#

L stable ai

fervent thunder
brisk bloom
#

select the model and prompt away

#

add the lora to the prompt and it uses that as well

slender quarry
#

@delicate oxide agony i just watched ntts vid

fervent thunder
tawdry forum
#

My new model is up!

cerulean igloo
#

Hi, can someone tell me where are the bots to generate images?

oblique star
#

How to see how much of VRAM does google colab provide? using !nvidia-smi I can't see VRAM

#

When I go to resources it says
GPU RAM
5.8 / 15.0 GB

#

Does it mean that I'm using 5.8 and available is 15 GB?

#

And does this mean that I can run SD on a server with 8 GB of VRAM easily?

fervent thunder
fervent thunder
vestal dew
#

that's why I bought all the stuff for a new system, so I could do more of this on my computer. Net system won't work, of course, so $2K wasted. But I tried.

fleet mantle
#

Hey, so I was watching a video about what Textual inversion is and how it works, and I have an issue : https://i.imgur.com/t6aR44n.png

I don't understand what the 'Text embedding' in this schema represents(I have a little idea, which it is the float numbers representing something about some images, likely rgb/posX/posY)
But more so I don't understand where it comes from. To me I thought this thing was inside a model, and not something from outside plugged onto it

#

(I'm a total newbie at this)

swift sluice
#

How do I make a model that produces 512x768 generations well?
Do I use instance images with those dimensions?

vestal dew
#

My guess is yes

dire sun
#

I don't see anyone talking about the latest samplers, such as uni_pc_bh2

whole drift
#

Hello guys please when trying to render the video output by uncheking "skip video for run all" on Google collab, I got this error "ValueError : max() arg is an empty sequence"

Someone has a solution to this please? Thanks

celest dagger
#

Hello, help please, trying to create a image , and it's works only with CPU but i have a GPU 3060 and this Is not recognized, somebody have the same problem ( l'm working in Ubuntu).

vestal dew
celest dagger
vestal dew
#

I pray it works

peak spoke
#

hi i'm new to SD and wonder if i could get some advice on where i should be placing the models off Civitai example i.e Type Checkpoint merge vs LORA

#

do i put all Lora files in Models > Lora & Checkpoints in Model>Stable-diffusion

warm junco
bleak wolf
#

Hi! I am the developer of Sci-Fi Diffusion. I have been away from the AI and Stable Diffusion scene for a few months now because all of the hate I got made me a bit sad (not from this community obv).

However I want to return back to it. What happened while I was gone? What major tools/software are new? I heard of stuff like ControlNET. Also any major update on finetuning/training models?

forest apex
bleak wolf
#

Anyway gotta get back to this stuff

#

and also Sci-Fi Diffusion 2.0

forest apex
#

the latest video butteryness has a tweet calling it theft with 100k likes, my brain is having a hard time with it too lol

bleak wolf
forest apex
#

but a 13b gpt model that can fit into 8GB that is looking to be able to perform like chatgpt is being worked on, and the open assistant has some preliminary releases, high level gpt bots won't be closed source for very long

bleak wolf
#

Lets go FR

#

Also I love how SD is opening the gap with MJ now LOL

I told people that this being Open Source is a force no corporate funding can cope with

#

Sure MJ is more user friendly

#

but the vastness of SD

#

I love it

warm junco
#

OpenSource is the best!

forest apex
#

yea, I just used the MJ bot recently and it was ok, but just compared to the stable horde bot there are so few options beyond prompting to manipulate your generation

bleak wolf
#

Oh lmao I was one of the earlier workers in stable horde

#

I need to check them aswell

forest apex
#

yea lots of good work being put in, it used to be like 5 options and no models

bleak wolf
#

So I can make my OpenRC project

warm junco
#

OpenRC ?

bleak wolf
#

Open Remote Control

#

rc vehicles

forest apex
#

like openTX?

warm junco
#

Ahhh

bleak wolf
#

Like Voron Design (very cool stuff look it up if interested with 3D Printers)

Basically what they do is make 3D printer designs that are open source, and suprisingly its the best printer out there. Since its open source and not made by a company or anything, any shortcoming or problem is solved really fast and there are billions of mods/customizations

#

I want to apply the same concept, but for making RC vehicles, from toy vehicles for kids to learn to professional rc vehicles

forest apex
#

aww just cars no rovers or drones?

#

:>

bleak wolf
#

Anyway, thanks for the chat guys, Good day to all of you! Its nice to be back

forest apex
#

right on 🙂

warm junco
#

Loras are trainable by the Kohya tool.

hybrid marsh
#

How to get stable diffusion to understand a particular color? Lets say i want a certain shade of pink and I want it to understand the color using hexcode... can we do it?

alpine bear
vast ingot
#

definitively in the category of "possible but damn near impossible" right now. The few people that needed that for now did use another tool instead. With the good prompt, you sometime manage to get a flat color, on the most used pantomes, but no consistency

pallid flame
#

hello

vast ingot
#

Hey Ryan !

pallid flame
#

does anyone know why sd does not recognize the Annotator i give to it

#

it understands normal picture poses

#

but doesnt work with open pose editor

vast ingot
#

you'll better ask in #🤝｜tech-support so you can post pictures of the problem, it must be just a wrong setting or something like that ?

pallid flame
#

ok thanks

warm junco
thin tapir
#

does anyone has a good video or documentation about the install and prepare of stable diffusion web ui

warm junco
thin tapir
#

thx mate

warm junco
thin tapir
#

al right

oblique star
young ivy
#

whats the command prompt to rename an entire folder of files to name_numbersequence, Where name is whatever file title I want and numbersequence is generated to go sequentially from 1 to as many files as there are

#

nevermind found an easy solution

alpine bear
#

^that's the perfect code question for chatgpt. Super accurate with basic requests like that. Saves me a lot of time.

fervent thunder
oblique star
fervent thunder
#

yea

oblique star
#

And one question

#

I'm making a bot that generates images based on SD.. And I have a problem, I was testing it on google colab, but seems to be blocking, like I can't generate 3-4 images at the time with different prompts.. Do I have to use something to parallel those things so it can generate more prompts at the same time?

fervent thunder
#

It can generate 8 different prompts at the same time

#

on TPU, not GPU tho

#

I never tested it and idk if I will be able ever, since I'm banned on kaggle now lol

thick bone
oblique star
fervent thunder
#

idk if you can put more into one thing

oblique star
#

Does anybody knows if stable diffusion bot had a limit of number of prompts at the same time?

ripe kernel
#

oi

cursive rain
#

hi, what is the min gpu to run this software, i am using gtx 960

blissful jetty
#

the addon for blender says no gpu required, can it utilize an amd gpu or is it cpu only?

raw narwhal
#

2GB definitely won't render 512x512

#

Does anyone know if automatic1111 supports multi GPU render? (Not sli)

cursive rain
raw narwhal
#

Ideally you would look for something like a 2080ti, or 3070ti. Not the newer stuff.

#

Makes more sense to buy used

regal snow
#

Well yeah you need more ram to actually make the images but the better models will do it faste

#

Toms hardware did a benchmark tests for stable diffusion please don’t mind my voice input

cursive rain
regal snow
#

I'm happy with the 4070ti, although there is a lot of hate for it out there. But it cranks out images very swiftly and I haven't felt like I'm missing out on anything yet due to only 12gb vram

#

My opinion may change on that if I get more involved with model training, or if some vastly superior open source software hits that is more vram hungry

raw narwhal
regal snow
#

While generating images, task manager shows that CPU usage is < 10%

#

the CPU has no bearing on Stable Diffusion

cursive rain
regal snow
#

< 30%

#

and GPU 100% of course 🙂

hybrid marsh
weary aurora
#

Where do I type my prompts, can someone help me please?

regal snow
#

You could probably get similar results inpainting with t2I style transfer plus depth/canny

sleek osprey
#

with controlnet perhaps

cursive rain
lunar wasp
#

I have a gtx 1660 and I was thinking of upgrding to a rtx 3050 is it worth it?

hybrid marsh
regal snow
#

canny tolerance could be set to ignore the wrinkles while maintaining the shape. depth could have a lower weight allowing some variation in that regard

#

...and t2i style transfer

#

I believe would have similar results, but might not be what he did he could have some method thats less convoluted

hybrid marsh
hollow raft
#

Saw this today, don't know if its been discussed, but seems like this is another misguided attempt to 'reverse engineer' source images from an AI tool generated image: https://www.stableattribution.com/

regal snow
#

There's tutorial videos on YT how to use controlnet, and t2i style transfer

sleek osprey
hybrid marsh
regal snow
#

Controlnet will trap the generation, so with that active and you inpaint her chest and arms, the resulting image should remain constrained to that shape

#

Then just include another controlnet with the t2i style transfer.

#

It should force that style to be transfered into the inpainted area

#

the arms would have to be a second round of inpainting

hybrid marsh
regal snow
#

to answer that I'll just be repeating what I already said

#
  1. canny - 2) depth (0.5 strength) - 3) t2i style transfer --- should have similar results
hybrid marsh
sturdy stratus
sleek osprey
midnight sphinx
#

Yes

sturdy stratus
#

There are a few... there's an official one with limited functionality... there's one you can use with a local server that is a bit hard to use... trying to make something that Photoshop users will like

sleek osprey
#

the demonstration is nice

sturdy stratus
#

Thanks

past sentinel
#

how to install xformers on windows?

regal snow
#

Just edit the webui-user.bat and add the cmd argument --xformers

#

well,

#

if not then I believe you need to build xformers from source.

past sentinel
#

will it work with nvidia 3090?

sleek osprey
#

yes

regal snow
#

I was reading that xformers will make generations no longer "deterministic" (perfectly reproducable) - but I haven't experienced that yet. I've only had deterministic results as far as I can tell.

vast ingot
#

it is deterministic

#

it's not hardware agnostic

#

meaning it will give different results on different computers

#

but the same results on the same computer with the same settings

regal snow
#

Ahhhhhhhhhhhhhh very interesting

vast ingot
#

and even the small differences are usually very minor

regal snow
#

the language isn't very clear in the wiki on that

vast ingot
#

like, when I reproduce an example of a sample on civitai, I don't see the differences

#

but yep, I have experienced it with trying to help other users on very specific details sometimes

#

it's also why it's harder to install than other things : it needs to be compiled locally, or a precompiled wheel corresponding to your environment needs to be downloaded

#

I tried both ways, and if you have the wheel (like automatic does) it's a lot easier and faster ^^

buoyant jacinth
#

all I needed for my 750TI was to add --xformers. I was pleasantly surprised.

vast ingot
#

yep, it's quite automated in automatic (lol)

#

I discovered the pain when I was trying to install dreambooth by hand, in the early days

buoyant jacinth
#

apparently, I got into this at just the right time. automatic1111s web ui has great low-mem support, basically everything just works. And thetn unipc is added the other day, cutting my generation time in half.

vast ingot
#

I'm trying the new toy in town right now, comfyUI

#

great potential too

#

node based UI

#

Command line option: --lowvram to make it work on GPUs with less than 3GB vram (enabled automatically on GPUs with low vram)

buoyant jacinth
#

wooo. yeah, looking at it on github now

vast ingot
#

can't really say more, I used it for 10 minutes for now ^^

#

but the video sold me on it

#

just seem to need some adaptation time maybe

quaint tapir
#

Upscaling to 16mp

toxic bay
#

are there models available that are better at creating textures and patterns?

solid kite
#

I m not able to use the dreambots, how to get access in those channel ??

surreal perchBOT
#

@solid kite

FAQ: How do I generate images?

Currently there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or running Stable Diffusion locally using your own system resources! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using the website or #🤝｜tech-support for any problems you encounter while installing it locally!

oblique star
wild steppe
oblique star
# wild steppe Sure, although I'm not the right person to ask since I didn't work on it myself

So I'm trying to make a bot similar to dreambot.. And I have a question about it.. When I use prompt it usually takes 35-45 seconds to generate an art.. But when I use 2-3 different prompts at the same time then I can see that GPU is working for each prompt a bit of a time, so that ends up around 120 seconds to generate each art.. Can that be managed to finish it sooner (back to originally 35-45 seconds) or not? (I'm using google colab with 6GB of VRAM)

buoyant jacinth
#

Now I'm waiting for a webui for an txt2ai that will take a text prompt to set up the nodes in comfyui to generate an image.

wild steppe
vast ingot
vast ingot
#

I'm sure gonna make myself a library of useful modules

oblique star
wild steppe
oblique star
uneven heath
astral goblet
#

As soon as i saw that comfy ui was nodebased, i got excited by that but, my immediate thought was " A lot of people will not be comfortable with this"

#

Another UI paradigm i thought could benefit these workflows a lot is the editor for Winamp's Advanced Visualization Studio

weak sundial
#

What's the maximum number of words/tokens accepted by SD?

alpine bear
fervent thunder
#

in which channel to use the AI?

surreal perchBOT
#

@fervent thunder

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

alpine bear
weak sundial
alpine bear
steep star
#

What should I do after downloading this?

vestal dew
#

Ditch it. Download 1.5

#

Do you have SD installed?

astral goblet
#

With A1's ui there's also the BREAK keyword that isn't commonly known about. it deals with that batching you're talking about. Instead of automatically cutting the batch at the 75 token marker, you can tell it where to end one batch and start the next, giving you better control over the attention of each batch before it's averaged together.

steep star
steep star
steep star
vestal dew
#

yes

#

it goes in the models/stablediffusion dir

#

Automatic1111\stable-diffusion-webui\models\Stable-diffusion On mine, automatic1111 is the main dir

steep star
#

Ok sir

vestal dew
#

go to files and versions

#

download one of those

steep star
#

Ok
Thank you

vestal dew
#

I don't know what emaonly is for that model

steep star
#

Which ckpt file should I download?

#

Pruned or Emaonly pruned?

steep star
vestal dew
#

I just read emaonly is for every day

#

use the other for training

steep star
#

Should I follow this?

vestal dew
#

ckpt and safetensors are the same saved in a different way

#

not sure if that explains adding models

#

that's about updating a1111

steep star
#

Ok

green warren
#

Who wants separate channels for contest stuff and actual product announcements?

brittle marsh
#

so you are from the stable diffusion server!

fervent thunder
shrewd belfry
fervent thunder
#

yes

shrewd belfry
#

Models > VAE

buoyant jacinth
fervent thunder
#

ok thanks

vast ingot
#

but soon

#

I'm getting to the end of my day

fervent thunder
#

it doesnt change tho output images i didnt understand in the settings is at automatic

hearty dune
#

wuick question, using controlnet, are there models (depth, canny, etc) which are smaller in size? all i see is several gb size for them

brisk bloom
sinful pewter
#

I am looking for someone who can show me how to take real life photos and make them into visual novel style backgrounds with controlnet or without. This is a paid gig.

vast ingot
#

and even if you said it

#

it had me surprised

#

lol is it like this the whole tutoriel ?

#

are there hidden feature to be unlocked by seducing the teacher or something ?

#

this is even just very instructive to explain how Stable Diffusion works

hearty dune
brisk bloom
#

yes, CN models folder

#

safetensor is safer to use than ckpt, but essentially the same thing in use

buoyant jacinth
errant orbit
#

can someone help

#

I dont know how to fix this

warm junco
errant orbit
#

ok sorry'

warm junco
fervent thunder
#

Hi

ebon tartan
#

hey guys

#

is better stable difusion 1.5 or 2.1?

brittle frost
#

so now i can't generate images from discord?

alpine bear
past moss
#

I watched a tutorial that said the weighted value of LoRA additional networks shouldn't go over 1 even if it's the sum of multiple LoRAs. Why? I tried it and I didn't see much of an issue, I know that if that weight value goes too high pictures start to look weird

vestal dew
rapid crescent
#

whats comfy ui?

kind cave
#

anyone know if its possible to get SD to work with tiff files?

#

or dpx

#

any professional image format

fervent thunder
#

But with it you can generate different things on different areas of image

fervent thunder
waxen pasture
#

Ok am I going crazy or was the "member" role that makes everyone's name green darkened?

open cipher
#

when writing captions for a subject in dreambooth, do you describe it using the name it will be referred to (ex. ukj) or an actual description of the character? (ex. a young man)

marble nebula
#

Hi I'm new

ionic sand
silk skiff
#

typing...

broken gull
static lynx
#

Deliberate is an EPIC model. any more like this?

knotty turtle
# open cipher when writing captions for a subject in dreambooth, do you describe it using the ...

I am by far an expert, and have a lot of unanswered questions myself, but from what I understand, the captions (as in the side-car text files) should include very detailed descriptions of everything in the image that you want the ability to change through prompting. In other words, what doesn't get described, becomes un-editable through prompting. So for a single subject, you might consider describing everything in an image (background, lighting, camera angle, surrounding details, etc) except key features of the subjects standard look, since you want those to come through in every generation.

Fam, please correct me off I'm off base here.

solid wren
#

Hey guys, I'm not sure what terms to include in a google search for this... how do I add a picture of someone to use as a reference for a character, and have the engine reference that picture everytime I use the name, say, Goliath, in my prompt? Sorry if this is a very newbie question.

knotty turtle
solid wren
#

I do

knotty turtle
# solid wren I do

I highly recommend this guys channel. He covers almost everything you need to know to get started. For training, I'd recommend starting with a Textual Inversion first. They are fast to train and might help build your foundation of knowledge for how training works. After that, you might consider training a LoRa or DREAMBOOTH, which are a little more advanced and time consuming.

https://youtu.be/2ityl_dNRNw

solid wren
#

ok thank you I will check that video!

#

@knotty turtle Do you know if it would be accurate or not if I gave 4 headshots in one image and then prompted "Group of men standing together"

#

accurate in terms of representing those people accurately

peak spoke
#

Can someone point me to where it explains prompt value I.E masterpiece:1.x Still learning and not sure what this is actually called

brisk bloom
peak spoke
#

Thank you soo much I will read now

vast ingot
#

man

#

I already gave you warning for that yesterday

alpine bear
vast ingot
#

that's a timeout

knotty turtle
#

😬

vast ingot
#

he's welcome to come back but he didn't answer to my warning yesterday, and just posts the same thing today with more data to contact him

#

at least interact with the community and go on about things related to AI

#

this had no link

#

like, linkedin exists...

knotty turtle
#

Oh, for sure. I block people who do that. It's like, c'mon

alpine bear
#

@vast ingot it's a throwaway / spam account. I can p much guarantee that pfp is fake and the guy is just looking to scam some kids.

#

Whatsapp is extremely sus, just another red flag

vast ingot
#

most likely yes. problem is, it's not really possible to be sure in all cases. I try to keep everyone to the same rules, and it's rare we jump the gun on a ban, but since it's 2 times the same in 2 days, I'll check if I upgrade to ban with the team

alpine bear
#

A ban is an upgrade? I want one anyasmug

vast ingot
#

plus this gives me work to do man !

#

lol upgrading the sentence :p

alpine bear
#

You... you mispelled the word sentence, in a sentence about upgrading a sentence, and then upgraded your sentence by correcting the spelling of sentence.

vast ingot
#

that's a sentenception

knotty turtle
# solid wren <@108690328109723648> Do you know if it would be accurate or not if I gave 4 hea...

You may have already tried this while I was away, but I don't recommend trying to train multiple individuals in one model/embedding (if that's what you were meaning). It is possible to generate an image with multiple people, but is much more advanced and involves training each person as a LoRa on each of the people and then using a multi layered approach with a plugin called ControlNet and the open pose editor

pastel fern
knotty turtle
fervent thunder
#

h3ll0 :3

vast ingot
# knotty turtle Oh, you can?! I had no idea

lora can function like dreambooth. you can finetune as many concepts as you'd like at once. you'll just need to adjust setting for the quality not to go down too much, test a lot your datasets to find the right mix for every concept to be train and learned at the same time.

#

to be fair though, if you go to those extends in efforts, you should do a dreambooth instead imo

pastel fern
#

Dreambooth is good if u aren't planning to share ur shit right

#

But if u are then Lora should be better I feel

vast ingot
#

depends on the bandwidth and for what
a good dreambooth is always nice for a future mix

#

mix mixed with more mix tends to lower the quality accross the board, starting from trained checkpoints a new mix can be best

#

but yeah, LoRA have their merits for sure

#

like, in Auto, you cannot use multiple checkpoints at once

#

but you can compose a prompt with multiple LoRA

#

comfyUI solves that though, you can make pictures from multiple models at once now

pastel fern
#

I always felt like training Lora = training dreambooth

#

Since u can just merge loras into ur model anyway

vast ingot
#

it is very close but no cookie for me. I find some difference in quality, and I can't stand that, I prefer to go dreambooth, but yeah, it's because I can with lots of disk space and bandwidth

pastel fern
#

What I like about loras is u can mix on the fly

#

I usually do like 3-4 loras at 0.25 str each

vast ingot
#

they do train quite fast too

#

they are the current best compromise in quality/speed/size for sure, and quite versatile

#

but still, I would advice mostly 1 thing to every trainer out there (no pokemon)

#

keep your datasets, keep them in the highest possible quality, and tag there.
Do crop and resize, but keep the originals.
You never know what training tomorrow will ask

pastel fern
#

Btw stupid question

#

Can we finetune loras lol

vast ingot
#

what do you mean ? full finetune in lora ?

#

full finetune is just a nice word for dreambooth with captions on a large dataset

#

so if you put enough layers on your lora to retain enough data

#

it should work

pastel fern
#

Like can I resume training someone else's lora

vast ingot
#

I'm not sure I see any function for that in the official repository

#

but

#

like you said, you can merge a lora into a checkpoint

#

and train on that checkpoint as if you were continuing the training the person did

#

as long as you merge the LoRA on the same model it was trained on

#

but I may have just missed the option

#

plus there are 3 different LoRAs implementations currently at least

knotty turtle
#

When y'all are captaining for an individual subject, do you describe the key features of that individual or leave them out?

I've seen YT'ers describe doing it both ways, and am not sure which is best.

It's my understanding that leaving those key features out will make them more malleable in your outputs, but maybe that's not ideal if you're trying to create a recognizable character.

For example, I'm trying to make a model for myself. I'm bald and have a beard. Should I mention bald and beard in every caption, or leave that out

zealous narwhal
#

Anyone played with the MidJourney 4 model from Hugginface into SD 1.5

pastel fern
west raft
#

Is dreamstudio Stable Diffusion v1.5 the same as runwayml/stable-diffusion-v1-5 from hugging face?

wheat sand
molten lodge
#

Hello world

warm junco
stark grail
buoyant jacinth
#

@vast ingot my main impression of comfy so far: it is more efficient in both RAM and VRAM usage (than auto1111's). I can generate images and use other 3D apps at the same time with comfy on my wee 750TI

vast ingot
knotty turtle
#

What's your guys' workflow for captioning large datasets?
Do you use a specific app or interface?
I'm brute forcing this like a pleb opening each file individually in a text editor and a Windows image previewer and ... maaan, it's a slog

buoyant jacinth
#

My workflow is to download someone elses and click the ❤️ button (:

vast ingot
knotty turtle
# vast ingot First pass using an auto caption tool, but not always. Then i use bulk rename ut...

Thank you. Are you using this utility to somehow create/edit captions in your txt files? Or just for renaming files? Or am I just misunderstanding how captioning works? Haha.

I've been using BLIP and deepbooru to get things started and then opening each text file to finetune descriptions.

Was hoping to find a utility for streamlining the process of opening each text and image file for comparison and editing

vast ingot
#

Not sure how to mass edit text files easily

#

I know in notepad++ you can search and replace text in all files in a folder though

#

Using regular expression and what not if needed

buoyant jacinth
#

I was thinking notepad++ or something like atom

knotty turtle
# vast ingot It depends on your training tool and options. Mine uses the files names, but the...

Interesting. Okay. I wasn't trying to mass edit the text files, just hoping for a utility that would look at both the training image and it's associated text file at the same time to save me a few seconds from opening and closing hundreds of files and risking me accidentally opening the wrong two files and miss-captioning the wrong image (as I've already accidentally done - heh). But maybe captioning just the filename would work better/faster than the text file way. I wasn't aware I could do that.

Do you use A1111 to do that? I don't see the option to caption inside the file name

vast ingot
knotty turtle
#

For me, the main tab is "Train", then "Preprocess images"

vast ingot
#

there is some training built in automatic yes

#

text inversion and hypernetworks

#

the preprocess is about the same, let me start it, I'll check

knotty turtle
#

I was actually just looking at the every dream tool. Someone on a different server told me about it, but I have absolutely no clue how to install or use it. I'm doing some research on that now. I'm not a programmer so this stuff is a bit out of my depth

vast ingot
#

windows/nvidia ?

knotty turtle
#

Yes. Both

vast ingot
#

1/ you need to have python 3.10.X in your path

#

check that by typing "python --version" in any console

#

2/ git clone https://github.com/victorchall/EveryDream2trainer cd EveryDream2trainer windows_setup.cmd

#

that's about it

#

the doc is quite complete

steep star
#

I have python at this location but it still says

vast ingot
knotty turtle
vast ingot
knotty turtle
vast ingot
#

it stays very superficial, it's an introduction into looking more in the method you choose

#

but I also try to define most concepts in the terminology

knotty turtle
# vast ingot it stays very superficial, it's an introduction into looking more in the method ...

Sweet. Understood. I've done a considerable amount of research on the different methods and am pretty set on using dreambooth for optimal quality. And I've done some trainings with basic settings and automated captioning. The results have been pretty good, but I'd like to dig in deeper to understand the best way to train through proper captioning. I'm still very new to this process. Right now I'm just trying to develop a workflow for speedily captioning large data sets

vast ingot
#

it's the full suite, not the trainer itself

#

it lets you prepare the datasets

sharp robin
#

If some are interested, 2 stability.ai employees working on DeepFloyd are participating in wandb fully connected today, talking about WandB and DeepFloyd + txt2Img

fervent thunder
#

Hi, i am new to stable diffusion. I have questions about text2img and controlnet image generation. Is there people can help me? I am Turkish so if i contact with a Turkish person, i express myself better. Can those who can help write me from PM? Thanks in advance.

vast ingot
#

I can try and answer your questions in simple english here if you want

#

but I can't talk turkish

fervent thunder
#

Thanks. then I try to explain my thoughts simply.

#

I want to replace the characters in a illustration I found on a site like Deviantart with other characters I want. But which extensions are good for this, which models would be better? I don't know exactly what settings I need to make.

#

Anime characters

vast ingot
#

ok

#

so you want to keep your background

#

and change the characters

#

only the faces, or the whole person ?

fervent thunder
#

Yes actually i keep the pose of the characters, and change the characters

#

Whole person i think

vast ingot
#

and is the background important to keep ?

#

or is it mainly the poses ?

fervent thunder
#

No it is not important

#

Mainly the poses

vast ingot
#

controlnet is your solution here

#

I will need to show some pictures

#

and we cannot in this channel

fervent thunder
#

Ok

buoyant jacinth
#

comfy can probably do that

knotty turtle
vast ingot
#

there is not a lot of reasons to use EveryDreamTrainer anymore imo

knotty turtle
vast ingot
#

it's a complex thing ^^ I've been in it quite a lot, you can also ask if you'd like

knotty turtle
dull hare
#

hello guys , i am learning more about stable diffusion and i was asking where can i find the PNGInfo script ?

#

inside stable diffusion files

wild steppe
waxen pasture
silver bronze
#

is 59.46s/it good?

vast ingot
#

Depends on your situation but no, not so much if just making a picture

silver bronze
#

how to make it better?

vast ingot
#

I'm not sure without more context

silver bronze
#

mk

grizzled harness
#

59.46 it/s?!

#

my current image is generating at like 4

#

granted, it's a big size 😄

vast ingot
#

if it's s/it it's bad, but it/s would be good. you said it both ways

grizzled harness
#

oh LOL

#

touché

clever venture
#

yoo

#

guys

#

do you know if there's any good fnaf model?

thorn hare
#

is there a wiki where i can read what exactly is the difference between hypernetworks, embeddings and LoRA ?

solid wren
#

If I'm training an embedding and I close stable diffusion then come back, will it continue where it left off?

lime violet
#

Anyone using AUTO1111, I'm trying to find a away to get around the 2000px limit on img2img

boreal oxide
#

I’ve got a request for anyone here, if they’re willing to help me with something for a minute-

I’m trying to get the img2img alternative script running with the 2.1 model, but it seems to fail halfway through.

Can anyone else try running a 2.1 model with the img2img alternative script? I’d love to know if the errors are just isolated to me or if the script is broken for everyone

solid wren
#

You can't see the error in your CMD and look it up on google?

boreal oxide
#

first the image gets re-noised into latent noise, and that part seems to work, and then it doesn't de-noise it. It seems to fail on the first iteration so I get returned an image that's one big wash of color

vast ingot
#

and the console stops and closes at that time, or goes back like everything is all right ?

boreal oxide
#

no errors

#

everything's running

vast ingot
#

just all ok according to it

boreal oxide
#

yep

#

it thinks it succeeded

vast ingot
#

can you show us the settings in the UI ?

boreal oxide
#

yeah

vast ingot
#

it may be something in there that prevents steps to happen for any reason

#

I didn't use that script yet tbh

boreal oxide
#

are pictures allowed in this channel?

vast ingot
boreal oxide
#

hopping over there!

astral goblet
#

Sometimes i wish i was still on top of the coding game and was sharp still. I'd do stuff to a1's to make it use controlnet for hires fix. i should start migrating to comfyui where the nodes are

crystal compass
#

hi

placid bane
#

So my wife was scrolling thru Tiktok and saw these really cool semi-animated images that at first sight seem to partly been made using Stable Diffusion. I'm sure the creator used some kind of extensions to further enhance the images. I could be totally wrong, but if anyone here understands the creation of these kinds of semi-animated pics, I would love an overview of how to create them. The user who posted this on TT is: at vblank_digital . Thks

astral goblet
#

animations are typically polished out in another app like blender or da vinci. SD just generates all the detail needed to create them. Depthmaps are combined with the image to produce interesting results

#

a lot of it is still artistic hand and skill, rather than star trek "computer make me an animation!" level yet

placid bane
#

This is good to know! I appreciate u sharing your insights.

wild steppe
winged hatch
#

could anyone send me a tutorial on how to install a model from huggingface

solid wren
#

search what you want

#

download the safetenser/chkpt file

#

put it in webui/models/stable-diffusion folder

winged hatch
#

ty

fervent thunder
ripe escarp
#

Hello everyone! I asked my friend what AI he uses and he gave me a link to this server. please explain what and how is done here.

solid wren
#

lmao this Zephyr dude is a chatgpt bot

ripe escarp
#

I can't speak English well, so I'm using translator.

astral goblet
# fervent thunder Do you make embeds?

i have but i'm no expert. i've just wung it everytime. used automatic1111''s mostly and once stable tuner. training i've mostly been reading about lately. might abandon embeds for now while i learn loras and locons

solid wren
#

Does anyone have experience training embeds? I'm doing it locally on stable diffusion UI and it's... A) Taking a long long time, B) not giving me the most accurate portrayals of the person I'm training it on.

Is there a better/easier way?

fervent thunder
#

Trying to decide if you should name the actual images with a prompt or just all the same name like 'chaz sitting in a cafe' and 'chaz standing in front of a cathedral' or just 'chaz(1), chaz(2), chaz(3) etc'

solid wren
#

I am using 5

#

I tried to run it overnight... but it seems to have just stopped running by itself

#

at image #138~ instead of 3000

fervent thunder
#

i found if my text file and image names didnt match up it could stop but idk

solid wren
#

They match

#

instead of putting the name I'm iterating on gCro I put a desciptor of the image though

For example, instead of "gCro sitting on a chair", I put, "a man sitting on a chair"

fervent thunder
#

Send me anything you want

zenith bobcat
#

browsing loras on civitai

#

wow this artists style looks cool let me check their actual real art proceeds to witness the most spine chillingly detailed and realistic genshin impact porn humanity has by far managed to concieve

#

i hate it here

acoustic thicket
#

Hello everyone. Can anyone help me extract an "image url" from the stability_api objects generated from the following call? How can I get the URL that contains the generated image?

answers = stability_api.generate(prompt="expansive landscape rolling greens with blue daisies and weeping willow trees under a blue alien sky, artstation, masterful, ghibli",steps=30,cfg_scale=8.0,width=512,height=512,samples=1,sampler=generation.SAMPLER_K_DPMPP_2M)

vestal dew
#

url?

#

there's no URL in there

thin widget
#

There's a community of lucky ⚡ 4090 ⚡ owners hidden away on discord where we talk about all things SD and AI. If you'd like to be part of it feel free to reach out 😄

fervent thunder
#

hey all, Ive been messing around with controlnet for the last couple of days, using the sd21 models from last week... So i think this is awesome, but im having some issues weighting the controlnet and find im consistently having to push models I like up to way higher CFG scales to get the result I want, does this seem correct? If I have a model and config that generates images I really like, or consistently cool stuff, and then try and use a controlnet on it, it stops doing what I liked about it, and I have to fiddle with config scale values (like going from 7-8.5 usually to 11-15) and samplers to find something that works. Anyway, I find when googling this it seems to be lacking any real discussion or have yet to find others talking about it, so it makes me think maybe something is weird my side. Any thoughts/knowledge appreciated, or if you've seen this addressed/discussed anywhere id love to know, otherwise ill assume im doing something incorrect 🙂

fervent thunder
#

has anyone tried comfyUI?

#

its really amazing

fervent thunder
zenith bobcat
#

is there a way to use loras on invoke btw?

#

cant seem to find the button on the interface

fervent thunder
fervent thunder
fervent thunder
weary jackal
#

set COMMANDLINE_ARGS=--vae-path "models\Stable-diffusion\vae-ft-mse-840000-ema-pruned.ckpt" i ve this in .bat file but i get this error Couldn't find VAE named vae-ft-mse-840000-ema-pruned; using None instead file is in place it is 326.851 kb , why i got this error ?

fervent thunder
#

damn. I guess Midjourney 5 dropped

#

I'm so desperate for some new models I may open the wallet lol

dreamy igloo
#

when v3 went to v4, it was a wow moment.

#

SD will always have the nsfw edge i guess lol

fervent thunder
dreamy igloo
#

I am seeing the rating party images here, but nothing that the current crop of SD models can't do. especially the newer 2.1 stuff

fervent thunder
#

too pricey for me still, Im just anxious for a new model like I said lol

dreamy igloo
#

maybe its just easier prompting yet again...as the article said

#

it is already pretty easy

#

that steam punk woman on the above link look less gaunt/edgy than the usual midjourney v4 retro-fashion model look that i usually see. more in line with the stuff i get on like rmada's stuff on 2.1

#

one thing i hope sd can solve is multiple faces in a scene. Unless you inpaint to fix each one, there are always messed up faces in a crowd. at least midjourney v4 show recognizable faces, even though they usually look like a bunch of creeps to me for some reason lol.

#

here is a v5 portrait by someone. Looks like something that 2.1 classic neg and other analog model for 2.1 can already do comfortably

regal snow
#

hey guys, is it possible to leverage the power of 2 separate computers for SD generations?

dreamy igloo
#

yea, you can generate more images with 2 separate computers...but if you are thinking of doing an SLI type thing, I don't know if it is possible.

unreal mesa
#

Hello

regal snow
#

Oh nvm yeah I was thinking SLI type of thing

cerulean merlin
tiny hamlet
regal snow
#

Should work but it’ll be slow, and you’ll be restricted to minimum resolutions

magic citrus
tame notch
#

anyone know how long it’ll take to upload anything v4.5 to google colab?

coarse blade
#

trying it out now, this is definitely not ready for prime time

#

on a random image, 6 HOURS

#

and I have a pretty beefy pc

potent spire
#

6 hours?

#

someone else did his in less than 60 mins with high settings

#

i didnt try it myself yet

coarse blade
#

the important part is the render quality

#

in terms of the glazing, the best option was medium

#

but even then, for images with white it becomes VERY obvious up until the lowest options

potent spire
#

true

#

but the artwork from Ortiz was glazed beautifully

#

they merged so nice together

coarse blade
potent spire
#

^^

#

another artwork i looked at being glazed was almost looking like someone ate too much mushrooms

#

the background was white tho

finite frost
#

really hoping for StabilityAI to release a competing model to GPT3/4 sometime soon

potent spire
#

they dont have anything with such features tho

#

as of now

#

plus i doubt they will have any strong enough benefits to convince people to switch from OpenAI to them

#

somehow like Krita vs Photoshop

#

although, Stability AI has ofc much more ressources than Krita

finite frost
#

a big one would be having a lot more transparency about what its trained on

#

people would value that

#

having unrestricted and actually open (🙃) AI would be an instant sell to a lot of people, myself included

potent spire
#

Im not a fan of iz

#

it

#

But people have different approaches, opinions and co

finite frost
#

sure. but it is a way to differentiate themselves

#

and its the whole reason why stable diffusion became so popular and displaced DALL-E

#

you can prompt whatever you want. not what we say you can prompt

#

for as much as Emad talks about the problems with GPT3/4 I would be surprised if he had no plans on doing anything about it

#

theres huge potential there for stability AI

vestal dew
#

SD is psychic

#

earlier today in the midst of generating totally different images, it made an image of a crashed motorcycle

odd mural
#

hi

vestal dew
#

a few hours later, neighbor's son was working on his motorcycle, riding it to test, crashed it

#

ambulance came

vestal dew
#

people are paying for it just to goof around trying to get it to write silly things

covert kettle
#

Hi

copper umbra
#

Can you find any error messages throughout?

forest apex
vestal dew
#

your right

forest apex
#

only the few should make our decisions, it just feels right

vestal dew
#

its better for it to be controlled they're

#

they're on the paid web site

fervent thunder
#

Are you on Windows?

#

How are you running it?

#

extension

#

you shouldn't need to, you can just load them in the webui

#

I'm making a comic but its not marvel style

peak spoke
#

@fervent thunder your working on comic I was think of doing the same

#

Have you made any before?

fervent thunder
#

Never

#

First time, I posted an image in the general with images

#

It will probably be bad, but its more of an experiment, the whole thing is done with A.I

vestal dew
#

you didn't do the words with AI

fervent thunder
#

Yes I did

vestal dew
#

how do you get words?

fervent thunder
#

Not image but text

vestal dew
#

oh

fervent thunder
#

I used gimp for the speech bubble and words

vestal dew
#

also, the characters look different in each panel

fervent thunder
#

Ye that’s to be expected

#

They will look slightly diff I will try to improve it later

peak spoke
#

I am thinking of trying to use Lora but question becomes can’t it do

#

I have

fervent thunder
#

I made a text embed and I tried dreambooth a while back haven’t made a Lora yet

#

Doing dream booth in the webui was terrible last time I checked I ended up using python

uncut plaza
#

DAVID.G9

lament solstice
#

just downloaded stable diffusion, took me an hour to set it up 💀 my lord i have a lot to learn

fervent thunder
lament solstice
#

Automatic1111

#

tomorrow I’m planning on learning how to use new models n all

fervent thunder
#

There's nothing really to learn, just download them (you can do it there - https://civitai.com/)
And drop into \stable-diffusion-webui\models\Stable-diffusion folder

#

oh yea and change model inside webui

lament solstice
#

Ic thank you

late saffron
#

/chat

elder cradle
#

Damn, I really want glaze to be open sourced...

quiet solstice
#

hello

vast ingot
#

hey there and welcome

quiet solstice
#

I have some questions about stable diffusion...

vast ingot
#

well, this seems like the good server 🙂

#

do tell us 🙂

#

it's not really rush hour yet though ^^ quite calm for now

quiet solstice
#

so, I became really interested about it and already downloaded it couple of times. I have an amd card so I thought it would be a good ideas to download it on linux

#

it was arch linux

vast ingot
#

I think it's how you got to do it on amd yes

quiet solstice
#

when I tried to download it on Ubuntu, it didn't really work for some reason

#

and then I started searching in the internet

vast ingot
#

damn, too bad it's the simplest to set up

quiet solstice
#

I know, I really tried everything

vast ingot
#

🫂

quiet solstice
#

I will try again today

#

but

#

my question is

#

what'S the difference between stable diffusion and webui?

#

both have good ratings

#

what's better for what usecases?

vast ingot
#

Stable diffusion is the brick, the ciment that makes the pictures.
WebUI is one of its implementations

#

there are other UIs for it too

#

like invokeAI or comfyUI

quiet solstice
#

oh

#

can I send links here?

#

github

vast ingot
#

for links and images

fervent thunder
#

Comfyui is awesome, I prefer the workflow

broken gull
#

how do you call those hair styles for ladies with like a bulb from their hair in english?

vast ingot
#

not sure, ask chatgpt ? x)

thorny sorrel
#

hey all, is it generally better to use underscore in prompts instead of a regular blank space?

broken gull
vast ingot
thorny sorrel
#

i thought that most models used the Danbooru tagging pattern, is that not the case?

vast ingot
#

depends really on what you call most

#

I don't use anime model personnally

#

so never

#

x)

thorny sorrel
#

oh its mostly for anime models then? i see

#

ty

vast ingot
#

it could be just replaced by space in captionning to conform with the current naming patern

thorny sorrel
#

not even trying to do anime though, its just that i found the tagging so complex that it shocked me

#

there are tags for bangs over one single eye and other crazy specific things like that

vast ingot
#

the main models trained by S.AI were on some part of LAIONB

#

so not danbooru

#

yes those are quite crazy

#

it's one of the reasons there is so many anime models out there

#

the tagging system was already in place and going full force

#

so it's quite easy to make datasets

thorny sorrel
#

when merging models, do the tags remain the same?

#

like, can i take a realism model and merge 10% of anime into it to get the crazy tags?

vast ingot
#

Like I said, you should ask anime on this. I think they use the crazy tags in quite a lot of models but not with underscores

#

And when merging, it merges the same tokens

#

So if merging things trained completely differently, it can have adversary effects

thorny sorrel
#

oof

#

ok ty for the help, imma go ask anime

whole nacelle
#

Goodmorning everyone!
I'm surely not the first person to come here and say that I'm developing a new Stable-Diffusion based software with a team of developers, but I was wondering if any of you would be interested in sharing with me (both here or on private messages) their own workflow in order to produce AI generated pictures in order to help us to improve it and make everything easier.
Since I'm the lead designer for the project and I'm also a traditional illustrator and ex-Adobe certified trainer I might have some ideas on the most efficient workflow but I might be biased towards a specific approach to the software!
(the beta will be released in the next months)

broken gull
#

new AI software is always excited

#

the unified canvas from InvokeAI is the future imo

main grotto
#

where do I create images

whole nacelle
# broken gull the unified canvas from InvokeAI is the future imo

InvokeAI looks great but I'm not sure that "unified canvas" is a necessary approach!
Right now we're working hard on a whole different aspect for the UI. Let's say it will be a lot more focused on the "creative process" with a zone focused on brainstorming ideas and a different area just for evolving those ideas with tools to fine-tune the "promoted" results.

surreal perchBOT
broken gull
vast ingot
#

(mainly look at #1080946152318443610 on the FAQ I just linked, there are ways to make pictures on there)

whole nacelle
broken gull
#

InvokeAI from what I understand also works for the professional sector, and they give payed plans for companies but for the casual user it is free

whole nacelle
broken gull
#

great

whole nacelle
#

Also we want to offer custom model and split the "pro" subscription (for those who wants to pay that) with the creators of models and plug-in active on their software / we're still thinking about how to do that - but we want to give a big economic boost to those who wants to work for this community :3

#

And we want to give the users the possibility to save their progress and keep working the next day from every step they made along the journey towards the end result - which will give a lot more value to the artistic interventions of the user (like: The AI didn't do this art on its own, I did this whole thing because I made the whole journey to this point)

broken gull
#

kind of like adobe but different companies offering their products in one unified app

whole nacelle
exotic acorn
#

@vast ingot

#

Nope wont work here to

vast ingot
#

I don't think there are any other than those x)

exotic acorn
#

For real

#

Haha

bleak matrix
#

Good morning, everyone! How are we today?

whole nacelle
bleak matrix
# whole nacelle Great, thank you, what about you? ^_^

Trying to get my hands warm, haha! I'm freezing and after my morning pills, I only found one one of my arthritis gloves. 🤣 I think the other took a trip behind something. I'm just going to hide under these blankets and pretend I found it

whole nacelle
bleak matrix
#

So, what's everyone up to today?