🏞｜general-with-images

[dense tapir]

the whole discord thing is them using people as paying to be guinea pigs.

[wispy nest]

I can understand it, well, depends why you're not a fan. I believe the most "off putting" thing about MJ, is how secret they are about their tech, I just want more people to be able to create cool stuff and keeping it hidden isn't making that easier in my eyes :(

I don't mind checking it out, just not exactly now, but I got no issue with trying it when I do get time a little later :)

[dense tapir]

That, plus I can't run it locally, and they charge people and those people who pay are beta testers they then turn around and use for their high paying clients.

[hasty nova]

Well if you insist, here is the page in the models-embeddings forum. https://discord.com/channels/1002292111942635562/1066559052328472596

[dense tapir]

Their entire setup stinks as far as I am concerned.

[wispy nest]

heh, I actually downloaded that one when you uploaded it the other day, or when it was, I thought it looked cool! ;P
But I haven't had the chance of trying it out yet, I can put it on top of my test list tho :D

[hasty nova]

I got into AI art from Dall-E 2, ran out of credits, then eventually found stable diffusion somehow, don't remember exactly. Have been here ever since

[hasty nova]

Oh wow! How cool. Well I would love to see what anyone can make with it, all I have seen are my own gens, and the two people that have left a review for it.

[dense tapir]

I never could figure out MJ on their discord so left to never come back. I tried dreamstudio I think it was and was hooked but as I was playing around figuring it out I used up all my credits and got mad. Came to SD. I mean I could figure out MJ now but I don't like the MJ style. I mean it is supposed to give me art, photorealism, etc... not inject its own style.

[wispy nest]

hehe, I believe a lot of people may download it, but then not give a review, or haven't had the chance to "grow on" it :)

[hasty nova]

Yeah, it has like 600 or 700 downloads now. Pretty surprised tbh, since I really didn't know what I was doing when making it. I must have done something right.

[dense tapir]

I don't like civitai so my embeddings are on huggingface and people are too lazy to figure that thing out and they sit and rot. Meh

[hasty nova]

I think I saw a plugin for blender that hooked into dreamstudio, used up all my credits, and then found out how to locally install.

[dense tapir]

PLUS most people are on 1.5 and I moved on from day one

[hasty nova]

I have a model on huggingface, but they don't really have a reccomendation system that I have found, or an easy way to find things at least.

[hasty nova]

Like me...

Well, I just used 2.1 today to help that one person up there

And it was cool, I must say

[dense tapir]

I don't like recomendation systems, or rating systems but I do like a way to find stuff.

[hasty nova]

Very photorealistic dogs

[dense tapir]

I dropped 1.5 like a hot potato the day 2.0 released. Yes, mad but CLIP in it is superb and I was the first one to find out it could count. OMG, that sold me instantly. 1 knife, two knives, three etc... I was sold. Model is stripped way down due to legal reasons and that sucks.

I had a lot of time invested in 1.5 with DB models I released and some I didn't. I would need to redo them for 2.x but meh

[hasty nova]

Woah it can do that? Is there a specific way you have to prompt that in or is it generally gonna figure it out?

[dense tapir]

I said a knife. It gave me one. I said two knives and damn 2 were there. I then tested to five and it gave me what I wanted perfectly. 1.5 could not count.

I found that out the day it released on this server.

[hasty nova]

so like "2knives" or "two knives"

[dense tapir]

I posted it to have everyone try it and sure enough it counts

[hasty nova]

Hmm

[dense tapir]

I spelled it out as two knives.

[hasty nova]

testing it now..

[dense tapir]

The major selling point of 2.x is 768x768 as the fabric, cloth, leather, fine hair is better than 1.5

[hasty nova]

I noticed the hair was much less grainy in 2.1

not perfect...

But more or less between 1 and 2

[dense tapir]

yep, it does know how to count

1.5 no way. that is the difference in the clips

clip got better model got worse

[hasty nova]

I find it interesting the knives look like guns

[dense tapir]

mine were butter knives

they were, ahem, funky

[hasty nova]

🧈

[dense tapir]

this is what is so bad that due to impending lawsuits they had to strip all the goodness away from the model while giving us a much more powerful clip. Had they not been forced to strip the model goodness everyone that could would have jumped over to 2.0

[hasty nova]

Which means less people making models for it, which would fix the issue

more or less

have you tried the primevalNova model?

I think it makes some pretty good results

[dense tapir]

Now we get a bare basic model and are expected, per emad, to make embeddings, and dreambooth, lora, hypernets to make up for what they had to remove.

if for 1.x no as I am strictly 2.x now

[hasty nova]

I should really start making stuff for 2.1

Just got this with 2.1 and double xposure embedding

[dense tapir]

I am a photorealistic guy so when I saw 2.0 do its magic my jaw dropped as the photo realism is so much better.

[hasty nova]

It is

[dense tapir]

We lost camera settings in 2.x though and that was a huge kick in my ass

lens speed, aperture, etc... just POOF

[hasty nova]

some camera settings embeddings in your future then?

Or would taht not work?

[dense tapir]

can't be done as embeddings can only do one concept

[hasty nova]

Well, make multiple, oner per setting I guess

That would probably be a lot though...

[dense tapir]

it would

only way though

[hasty nova]

Custom model

[dense tapir]

I am not sure why that was removed.

[versed furnace]

lora, maybe.

[dense tapir]

possibly

[versed furnace]

Been having decent luck with it today

[dense tapir]

I know nothing of lora since koyha doesn't have a colab trainer

[hasty nova]

How much VRAM do you have?

It can run on like 8gb

[dense tapir]

6 and it required 40-60mb more

[versed furnace]

I only worked on subject training today with it. Going to be trying effects and styles tomorrow. My knowledge and access to photography is limited so I won't be touching that stuff.

[hasty nova]

oof

[dense tapir]

seems I have to let it rebuild the sd folder and my TI training will come back. Not sure why that is but imma try it. Still not sure what to use as a prompt so all trial now.

Another reason I wish I could train any thing locally but 1-3k USD for a 24 gig card isn't happening. No to used and new is nuts what Nvidia is pulling.

[versed furnace]

yeah. I got a 4090 when they were 'cheap' at 1600. still waaaaaaay too expensive.

[dense tapir]

they only went up and my poor 4060 ti I was saving for is now 200 more than a 3060 with less specs all around even 2 gigs less.

they spit on the ti making it a joke. Seems Nvidia is trying to ditch the gaming market and head strictly for AI/ML and let AMD have the gamers. If they get some gamers cool but that isn't their market it is corporate and AI/ML now. Charge double and sell half as many to equal the same in their minds.

less work for them

[hasty nova]

I got a free RTX 3070 with 8gb for my pc (I also have that in my laptop) so I can train locally, which is why I have been doing so recently. I find training the most fun.

[dense tapir]

For me I train as that is my enjoyment, then releasing it, but just making AI pics isn't as fun to me

[hasty nova]

That's the fun for me as well, which is why I also make video games. I love seeing other people do things with the stuff I have made.

[dense tapir]

makes me so mad so I tried Linux and everything is so broken in it I gave up. days I tried and lost a lot of stuff doing the transition too.

Linux I would have had enough vram but I couldn't get it/this to work in Linux

tbh, the 4060 ti as it currently is speced is a $279 card and no more.

they want 600

anything less than 16 gb is really going to hurt in the future especially when 1024x1024 hits with 3.x

Already 24gb is bare minimum for some things

[tame forge]

My queen

Gonna outpaint this baby to get her whole head in the picture

That's 1.5 btw, still the undisputed king of models

[dense tapir]

@sick remnant

[tidal hawk]

@dense tapir So i was using old version of stable diffusion for months now and updated to newer version and found that UI is completely changed, can't old settings like CLIP, hypernet,etc . Anyways to get them back?

[tidal hawk]

(cant even see the what model is loaded like this)

[dense tapir]

show me what you are seeing

[dense tapir]

pm?

[tidal hawk]

i had no idea UI changed this much since i last updated SD lol

[tidal hawk]

okay

[distant monolith]

I did apply it but why can't I see dreambooth tab?

[grizzled sage]

@tidal hawk

[tidal hawk]

yup found it

[acoustic eagle]

lol what model made that

[fervent rune]

One I am working on called SmolOnes, should be ready sometime tonight or early tomorrow 😄

[wispy nest]

yesterday I couldn't generate 768x768 images, today I can 832x1152

on 6gb vram

[acoustic eagle]

why

what changed

[wispy nest]

nothing

I was so mad yesterday

today it just works

very good

do people generate 2k x 2k images?

Or the results are too repeating?

I'm considering buying rtx 3060 laptop and I'm not sure If it's that necessary

[versed furnace]

with the hires.fix, or img2img upscaling, for the most part, or with images without people in them. things get weird at resolutions far beyond what the models was trained at.

[wispy nest]

I wonder if the new version will be trained on big images.

Is there info somewhere about that?

[versed furnace]

3.0 is rumored to be training on 1024. not sure how reliable the info is. seems reasonable, though.

[versed furnace]

biggest I usually render without upscaling is like mid-900's. larger resolutions work if you're doing patterns and things, but I stay pretty close to the training res for people.

[hallow sparrow]

How do I stop the buildings from being cropped? I have this in negative already: ugly, broken, distorted, hazy, (cropped), cartoon, zoomed-in, (close-up:1.1)

[dry crow]

@hallow sparrow
Try adding clear sky

[untold plinth]

Make a caricature #1002292387051212860 kid having fun at the computer

[strange jungle]

Easy Diffusion UI has option to take an image through 25 more iterations. Is there a way to do that in Auto1111 AI?

[versed furnace]

whoa. https://www.reddit.com/r/StableDiffusion/comments/10jjatu/wednesday_addams_dance_edit_experimented_with/?utm_source=share&utm_medium=web2x&context=3

[stoic pivot]

Are you for long prompt to get exactly what you want, or short ones to get random images for inspiration? I am the later and here is "four funny feminine furry fursuit friends, fishing fine fish frantically for fun".

[heavy quarry]

Incredibly detailed long black war astroship with tiny green light inside in the dark depths of a black ocean, HQ, 4k

[orchid wagon]

how can i stop img2img from modifying it so heavily

i want it to still look as if its the same city

[stoic pivot]

Try use low CFG Scale and as low Denoise as possible. I use Automatic1111 and can set (in settings) so it use colors from the original image.

[orchid wagon]

okay thanks ❤️

[stoic pivot]

And try Euler rather than Eulre a.

[orchid wagon]

ill try it out now after the next gen, running on cpu cause im not trusting these random ass ckpt files lol

okay bet

[stoic pivot]

But img2img can be a bit random, I some time get images that are easy to work with, then some days images that put up a fight.

[orchid wagon]

only amd tut i found that includes img2img had the download on some shit like "aiinfra.visualstudio.com"

and ah okay i see

thank you for the help 😄

would you happen to know

if this is a safe place to install from? its for stable diffusion except compatible for amd.... link sus af

[orchid wagon]

@still bobcat

[still bobcat]

coool

[stoic pivot]

Found the golden long banana.

[orchid wagon]

lol

[still bobcat]

Lol

anyway thank you for the help

Thank you so much for your help. I really appreciate it.

[orchid wagon]

np

[sharp merlin]

as they should

[sharp bronze]

Ive used Dalle for image creations and like the look, this "surreal", painted look I always get. How can I get this style( mostly for concepts) in Stable diffusion?

Ive tried so many different things but cant get it work. Would appreciate some tips or a hint. Thank you!

[wispy citrus]

Hi, another newbie here looking for help! I generated the first image using F222, and I'm reasonably happy with it, except for the faces. OK, move to inpaint, mask the faces, load up the inpaint 1.5 model, run a DDIM at 10 steps and 10 batches and a random seed as a rough copy for face replacements, but nothing I try seems to give me anything near realistic. I'm using the same prompts for inpaint as for the original generation, namely: (Pos: two aristocratic women, beautiful, bored, looking at camera, sitting on antique couch, victorian, indoors, well dressed, casual pose, photograph, expansive, color, symmetrical, ((saul leiter)), (((radoslaw pujan))), (pier paolo pasolini); Neg: 3d, cartoon, sketch, blurry, disfigured, deformed, old, ugly, poorly drawn, extra limbs, missing limbs).

I'm using CodeFormer face restoration both in the original and the inpaints.

Is there anything I can do here to get normal looking faces?

[steel yarrow]

huh... ok

[fresh hound]

anyone here tried nmkd SD gui?

just installing it rn

[still flax]

prompt/model? That looks like dreamlike-diffusion tho...

[sterile kiln]

With Photoshop plugin (Stable.Art); Generated an picture with New York, took a T-Rex toy picture, and used a layer and img2img in Photoshop to generate a new different one, works well, need to do an example with more work :

the result :

The trick :

[grizzled sage]

i noticed discord had no towel emoji, so i went to SD and said to it "SD, give to me a towel emoticon of the highest quality" and it gave

[cosmic sky]

Question for you all. Whenever I do a high res image with a resolution that isnt 512x512, my subject always gets duplicated and weirdly stretched out. Almost as if it is trying to stretch it to fill the resolution. How do I make the ai make one subject no matter what resolution I do. Or how to center the subjects in a certain portion of the image so they arent just floating around. THANKS!

[mighty flame]

can someone give me some advice to find a model that looks like midjourney and can generate pictures like this?

or just help me do something similar

[outer plaza]

it's just lexica and the prompt was pirate, centered, symmetry, painted, intricate, volumetric lighting, beautiful, rich deep colors masterpiece, sharp focus, ultra detailed, in the style of dan mumford and marc simonetti, astrophotography

[calm burrow]

crate a gamer controller futurist

[round kelp]

seems to be a 512x512 generation, what are you using? I mean if you are using something like automatic1111 you can try a bigger resolution like 768x768 and upscaling.

[wispy citrus]

Yes, it is 512x512. I’ll try a larger res, thanks for the tip 👍🏻

[round kelp]

No prob, it happened something similar to me when playing with some models. And trying larger resolutions was the answer.

[round kelp]

Are you using the default model?

[cosmic sky]

watcha mean by default

[round kelp]

Stable Diffusion 1.4, 1.5 or 2.1

[grizzled sage]

it occured to me that maybe my random cuda crashes might be due to the nvidia overlay, whiich always just randomly pops up and disturbs me as a desktop user. it might be doing that to the training process too

turned that off. don't know why i didn't sooner tbh

[cosmic sky]

oh 2.1

protogen v2.2

[round kelp]

mmmmm strange, what resolution values are you using?, normally they should be resolutions whose values are divisible by 256, or at least that is how I reduced that kind of problem

iirc values divisible by 128 also works

[cosmic sky]

I trained it using 512

[grizzled sage]

its cause 512 trained models create images in 512 sized blocks. its pushing the same seed into every 512 block it looks at, and diffusing what content is already there out some more

[round kelp]

yeah, but the image with the orangutans seems to be 388x721

[grizzled sage]

thats what the hires fix is good for. you do a small close to 512 image, then img2img it bigger with a denoise setting that gives a good rub to it

the attention isn''t fixed to a 512 grid. it's attention can be at any 512 block from any origin

so if the seed has a strong face to it, it'll want to create that face anywhere it can fit it

[wispy nest]

Daddy government said it's here to help me.

[fervent rune]

https://discord.com/channels/1002292111942635562/1067710900187963504
new model, please check it out and show me the results 🙂

[still flax]

what dataset did you use for that?

[fervent rune]

it's included in the civit post but I scraped the midjourney images on the same topic from tophmagoph_ on instagram with his permission

66 of them to be exact

[still flax]

I see... yeah I was thinking something like generate related images then train. Its always nice to learn new workflows, thx

[fervent rune]

I have a lot of TI's with midjourney images and a few models, it works pretty well for helping SD catch up or pass MJ

[still flax]

I gotta try that! Never thought of using MJ to train SD...

[mighty flame]

So no one can help me?

[strange jungle]

seems to be the case

[fervent rune]

I can help you make one, if you provide a few more pictures

[strange jungle]

Hurrah!

[fervent rune]

Training SD models tends to focus in on a specific area, so that style would be doable in a model or TI but would be the limit of the model

[fervent rune]

Honestly, what was the MJ prompt, I can find the similar pictures myself even

[undone furnace]

Maybe this model could help https://civitai.com/models/1274/dreamlike-diffusion-10. You can find some prompts examples below in the images section

[dry crow]

Openjourney is trained on MJ Pics, try the 2gb model variant:
https://huggingface.co/prompthero/openjourney-v2/tree/main

[mighty flame]

Vivid green and purple tones fierce bear

[mighty flame]

I already tried this, it looks awful

[mighty flame]

Can you train a model or tell me where I can learn that?

[fervent rune]

I want to give it a try but here are two good videos for learning such a thing using automatic1111
https://www.youtube.com/watch?v=dNOpWt-epdQ&t for TI/embeddings
https://www.youtube.com/watch?v=Bdl-jWR3Ukc&t for DB

I will let you know my results 😄

[mighty flame]

Do you want more material for it?

[fervent rune]

I was going to search MJ but if you have more that would be helpful

Just finished a Lora model I was working on so I can run another training session now

[pliant field]

Hi. Has anyone had issues with black edges on the output, when using video input in Disco Diffusion? Whatever settings I try, the first image output is fine, but the rest get progressively worsening warped black borders. any thoughts? Many thanks

It seems to have something to do with physical camera panning/movement - like it wants you to use a locked-off camera shot

[sharp merlin]

yes, listen to daddy or otherwise you will kill yourself accidentally under mysterious circumstances one day

[vague oasis]

has anyone seen this? Fiverr is probably one of the few big sites that allows AI art

[errant leaf]

Protogen is not SD2, it's a merge of SD1.5 models

[sterile kiln]

manual additions of prompts using the lasso tools and the Stable.Art (SD) plugin on Photoshop, incredible tool

[wispy nest]

ONG

[gloomy elbow]

One of these cats... is an imposter!

so, here's a fun game - prompt for something that gives you AI text, enter AI text as next prompt using seed of previous generation.

[waxen bramble]

Like at quick glance, everything seem rather normal. However... Longer you look, the more you find... Just stuff wrong.

[gloomy elbow]

You may not like it, but this is what peak performance looks like.

[steel yarrow]

oy! a fiver

[frail isle]

how do i fix it?

[wispy nest]

add that "flag" --reinstall-xformers to your user bat and rerun the webui once, then remove it afterwards

[wispy nest]

Well, maybe I should have started with the question: which webui are you using?

[frail isle]

automatic1111 ig

[wispy nest]

I can't remember for sure, but are you using the webui-user.bat to launch it?

[frail isle]

yes

[wispy nest]

you wrote --reinstall-xformer, it should be: --reinstall-xformers

[frail isle]

Already up to date.
venv "D:\ai\SUPER SD 2.0\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.9 (tags/v3.10.9:1dd9be6, Dec 6 2022, 20:01:21) [MSC v.1934 64 bit (AMD64)]
Commit hash: 6cff4401824299a983c8e13424018efc347b4a2b
Installing requirements for Web UI
Launching Web UI with arguments: --medvram
No module 'xformers'. Proceeding without it.
LatentDiffusion: Running in eps-prediction mode
DiffusionWrapper has 859.52 M params.
Loading weights [9ab3546b0f] from D:\ai\SUPER SD 2.0\stable-diffusion-webui\models\Stable-diffusion\AnightingV4.5+Yuzu3.5.ckpt
Loading VAE weights specified in settings: D:\ai\SUPER SD 2.0\stable-diffusion-webui\models\Stable-diffusion\anything-v4.0.vae.pt
Applying cross attention optimization (Doggettx).
Textual inversion embeddings loaded(0):
Model loaded in 47.1s (1.4s create model, 45.4s load weights).
Running on local URL: http://127.0.0.1:7860

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

[wispy nest]

remove the --reinstall-xformers and add instead --xformers, it needs to be one of the args to be used in the first place

[wispy nest]

sorry, but I can't read that error as it's not in a language I understand. It's a lot of checking to see how you installed it, and I don't got that time sadly. You can remove the --xformers if it is causing the webui to not launch.
Otherwise, check the git page and see if other people have the same errors. Try and follow what the window tells you to do. Good luck :)

[frail isle]

okok thank you

[hasty nova]

Try removing the OpenOutpaint extension (if you have it) and delete the folder called "vae" in the webui directory. I have seen that window pop up before and after a lot of searching on the internet I found that fix somewhere. At least I think it's that error, I can't read it either 😆

[frail isle]

dont need,somone on anime help me ahha

but thanks

[dry crow]

You can delete boath xformers folder in SD and then run the bat again with --xformers. It will install the new version

[scenic iris]

did today's daily as my first experiment with SD online - it's really cool to see the difference between the releases and how they use prompts

[cyan anvil]

What happened to all the hot manga girls from SD 1.4? Look at the difference: prompt: "manga girl with red hair and a gun". And you usually just get a head shot with 2.1. LAME.

:/

[dry crow]

For 2.1 u need to use negative prompts

Like blurry, deformed, lowres

[cyan anvil]

I see. But I don't understand why the cartoon women are less hot 😂

So your negative prompt has to be the opposite to your positive prompt? So "red haired manga girl" the opposite would be "blue haired disney prince"

[dry crow]

Yea but dont make it to specific

[cyan anvil]

idk, i'm not convinced by this latest version yet. But I'll try to get better at prompting

[dry crow]

For positive Prompts try masterpiece, highres, highly detailed

[worthy stag]

https://media.discordapp.net/attachments/868444231293886534/1037197902784761876/caption.gif

[sick remnant]

anyone ever just mess with the base SD model for scifi shit? it looks legitimately epic af

[pseudo aspen]

A friend asked me to generate Saul Goodman and this is the result

(very short prompt)

[quiet flame]

https://media.discordapp.net/attachments/802788240335437896/1068077444046790677/image.png?width=1440&height=666

good lord i dont have to deal with indepth prompts

this plugin uses gpt3

i just command like human instruction like u give to chatgpt

https://cdn.discordapp.com/attachments/1024723587317366874/1068065177368543302/image.png

[sterile kiln]

what plugin? awesome

[dry crow]

Pix2pix on github

[sterile kiln]

This one?
https://github.com/phillipi/pix2pix

[vague oasis]

https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/pix2pix

is not available in auto111 yet?

[dry crow]

I think this:
https://github.com/timothybrooks/instruct-pix2pix

[empty temple]

this is where all the cool kids hang out

[smoky goblet]

I am getting this error whenever I click on dream button

[empty temple]

oh I never used what ever that is

I cant help

sorry

do you have anything in the prompt?

[smoky goblet]

Yes I do

[empty temple]

then I dont know

[smoky goblet]

https://beta.dreamstudio.ai/

[smoky goblet]

Here is what it looks like

[empty temple]

I got it

so a positive prompt is like

Table:1

Example: A cat doing a backflip. 4k, Best Quality, Ultra Detail, Table:1

[smoky goblet]

Wow, thank you so much. What effect does it have with creating the image?

[empty temple]

It adds Intensity

if you made a house and said Grass:1.5 then there would be more grass than if you just said Grass

not sure what Negative prompting would do

[smoky goblet]

I think negative prompting makes sure it doesnt include what you specify

[empty temple]

oh I guess its different in this one

since there isnt a seperated box for neg prompts like stable diffusion (local)

[smoky goblet]

Oh okay, thanks for helping me out, really appreciate

[tranquil garnet]

Yes, is in the available extensions list in auto1111. You have to download the model manually. Here's the extension link.
https://github.com/Klace/stable-diffusion-webui-instruct-pix2pix.git

[quiet flame]

r there specfic settings for not getting blurred final images or distorted portraits

in this plugin

[sterile kiln]

Guys, do you know a good prompt to have the "Aerial perspective" effect?

[strange jungle]

What does it do? What is it used for? I don't understand.

[vague oasis]

you can change specific parts of an image in natural language

[strange jungle]

AH

[vague oasis]

heres a demo: https://huggingface.co/spaces/timbrooks/instruct-pix2pix

[strange jungle]

I take it it will not run with my GTX 970

[vague oasis]

neither in my GTX 1050, but its really cool

[strange jungle]

I am trying the demo you linked. Thanks

[obsidian shadow]

would it work with other .ckpt s, or do you absolutely need the pix2pix one?

[vague oasis]

only the pix2pix model works but the embeddings made for SD 1.5 are compatible

with the pix2pix model

[hard wagon]

Glitch embedding + coming RetroFutur v2

(The latter by @errant leaf)

[vague oasis]

we get it bro stop lmao

[hardy anvil]

Faces I generate looks similar even with different seeds. Any idea on how to change

[strange jungle]

use a different model

use img2txt

use names of known people at different weights

describe facial features

[vague oasis]

are you using a custom model?

[hardy anvil]

NO,its dreamlikediffusion photoreal

[vague oasis]

thats a custom model

its a common issue cause the dataset used to train that custom model is not broad enough

protogen also have that issue

[tranquil garnet]

auto1111 emphasis is supported
so if, for example, the prompt "make it blonde" changes the image too much, you can try "make it (blonde:0.7)" or something like that

[steel yarrow]

drone shot

[sterile kiln]

You don't understand, Aerial perspective is supposed to be this effect

[steel yarrow]

Yes, I don't understand

[steel yarrow]

have you tried what I told you ?

[sterile kiln]

it is this effect that makes the colors paler in the distance, on the most distant layers, an optical effect due to the atmosphere

[steel yarrow]

that's atmospheric distortion, called distant Sfumato in painting

that has nothing to do with height of the camera

[sterile kiln]

Ah, thanks, english is not my native language, so

[steel yarrow]

You're french ?

[sterile kiln]

yep

[steel yarrow]

I figured that part

[sterile kiln]

x)

[steel yarrow]

go try it, tell us how it went

[sterile kiln]

works well ! 🙂

[steel yarrow]

nice

[sterile kiln]

i need to use it in Photoshop plugin on a creature, but hard associate with something i think

[steel yarrow]

Why don't you do composition like the "stelfie" guy does it ?

[sterile kiln]

working on this, generated the left on SD, and working with Satble.Art and drawing with Photoshop

[steel yarrow]

is stable.art working ok ? been meaning to try it

[sterile kiln]

WOrking perfectly

[steel yarrow]

nice to know, have to try it out soon

Didn't you post that on le reddit ?

[sterile kiln]

Nop, WIP

[steel yarrow]

weird, someone posted something very similar to that

[sterile kiln]

???

[steel yarrow]

it had a godzilla on the right side though

[sterile kiln]

I was writing a pavasse to prove that this is my project but a video is much better xD

yeah dino on right side

[sterile kiln]

https://youtu.be/jtsGTZp14D4

If you have a link of this post on reddit please, have to talk with this guy

[steel yarrow]

we knew you like sfumato, so we put some smufato in your sfumato dawg!

[sterile kiln]

inceptsmufato?

[desert jolt]

https://www.youtube.com/watch?v=-Y5PJLdjySQ

[quiet flame]

lol i thought it would be like chagpt prompting

[wispy nest]

Guys, anyone doing blender or low poly renderings here? Trying to render something like this. What model and sampling methods etc would you recommend?

Just as you can see, soft, sort of blender style.

[wispy nest]

what's the best upscaler option?

[wispy nest]

got your nose!

[wispy nest]

the boring and probably what most would answer: it depends on what you want to create :P

I've not even named half of the ones I liked. Their default names tells me nothing other than that they have no idea what words mean! That, or I have no idea what it's supposed to mean ;P

[static tusk]

I usually use Remacri or Ultrasharp, depending on the image https://upscale.wiki/wiki/Model_Database#Universal_Models

[wispy nest]

I got almost every upscaler from that site, it's such a good addition :D

[static tusk]

there are too many to te4st for each image so I only use a few. even then its a little fiddly

[wispy nest]

yeah, some does almost nothing as they are are made for super specific things or areas of the image, but the horder in me wouldn't let me just get the ones I needed! ;P

[dry crow]

@wispy nest here for example. model and vae from the same creator

[wispy nest]

ooooh ok i'll keep an eye out for those

also "Oh" for realizing this is the chat where i wanna be, i figured it was for chatting while posting SD images

[dry crow]

its general so yea we could post created images also xD

[wispy nest]

thanks again! time to play around with it!

[deft dune]

Been hatching a phoenix fox 🙂 Love Stable Diffusion, really addictive stuff

[dry crow]

wow really nice! which model did you used?

[deft dune]

latest one, hatched from an image

https://www.youtube.com/watch?v=CoTP4wzU0NE

[dry crow]

official SD 2.1?

[deft dune]

yeah

[dry crow]

ahh

[deft dune]

Ive been creating a pick a path story using google sites, gtp chat and stable diffusion. Only been doing it for a few weeks, but starting to get the hang of it now

[dry crow]

very nice 😄

yea SD makes addictive

[deft dune]

100%, its crazy fun when you strike a good seed, almost like a win on the pokies lol

[dry crow]

haha 😄

wait until you found out about upscaler highres fix, that changed a lot for me

[deft dune]

How do you do it? I kind of follow the composition. I see a composition in an image and then create a prompt to try and hatch it, some timjes just upping the steps, other times getting the new image and starting from 15 steps and 4 images to try and improve it

[deft dune]

What's that?

[dry crow]

so you mostly use img2img ?

[deft dune]

I start with prompts sometimes

but usually I get better results from other iomages

using tehir colours and composition

[dry crow]

ah okay

[dry crow]

that makes you image resolution higher while creating, it creates and image of 512x512 and then upscale it automaticly by 2x and you get 1024x1024 as result

[deft dune]

Oh awesome, yeah I knew that was out there but haven't done much with it yet. I'm trying to get a heap of images so I can pick the top ones to do that for

would be a lot better for the pick a path pictures, thanks, I should get it into my workflow now

[dry crow]

okay have fun 😄

[deft dune]

Cheers!

[analog goblet]

woo, so i think i will finally switch back to 1.5 from 2.1 🙂

so whats new here ? 🙂 haven't checked the 1.5 stuff since 2.0 came out

[wispy citrus]

What's the best way to reduce (or eliminate) the glow/halo you often get around faces after you inpaint them?

[fallen echo]

anyone got any idea how to "force" CodeFormer or GFPGAN to use the whole picture?

it's not even able to fix/upscale the neck

[dusky hamlet]

Working on a hyppernetwork to fake plastic miniature , happy with the result 🙂

[pallid ruin]

I wanted "A cake with the text stable diffusion" and I got a SUS cake lmao

[pallid ruin]

You can´t do that on codeformers/gfpgan as far as I remember, you have to use another AI like srgan/bsrgan/remacri/etc. You can also try using img2img in stable diffusion with a low denoising strength

[wispy citrus]

Groly lol

[waxen bramble]

Ok since you like girls... Way more interesting in my opinion than what you have there.

Also I didn't need a massive set of prompts and wall of negatives. Can you even prompt with your using greggy?

IN case you are in to milf or gilfs

Not judging!

[tame forge]

they look like oil paintings

[waxen bramble]

They are supposed to

[tame forge]

I dont dig the painting look

[waxen bramble]

well not really...

[tame forge]

Thats why I always put greg rutkowski in the negatives

[waxen bramble]

They are supposed to be in the realm of old film reel and painting.

Kinda like old school matte

That is what i am going for

[tame forge]

thats kool. I just dont do that style. I do modern photography

Well if u like 2.1 just keep using it. I just dont like it cuz it doesnt do what I want it to

especially with celebs

[waxen bramble]

Well you could at least try modern photography techniques...

Like art photography is all sorts of interesting

They try and do all sorts of shit

Like IR and UV filters. Polarisation...

Hell lot of them have gona back to film since it allows you to do things digital doesn't. Maiinly due to how the exposure works.

[wispy citrus]

That's a pretty cool style. That's one thing I really like about this AI art is that you can actually develop a style with it, just like with painting or photography

[waxen bramble]

Well yeah. I take my best generations of some style. Make a style embedding - not always as easy as you'd think. And then keep iterating that

[wispy citrus]

Style embedding....that's on my list to look into. I only just discovered SD this past weekend so still lots to learn. Happily consuming info at a voracious rate though lol

[kindred willow]

Instruct Pix2Pix is very nice

Thoughts on this by the way?

Based on the Creation of Adam by Michelangelo, of course

[potent hornet]

just found this prompt on civitai, and, well, maybe if we ask with this level of kindness, SD will generate good images? hahahahaha

[dry crow]

hahaha nice one xD saw also prompts with very very very good eyes or five fingers!!!!!!!

[potent hornet]

i have to admit that I already used the "five fingers" one, hahaha

did not work very well

[dry crow]

hehe 😄

[potent hornet]

if there is one thing I learned about hands, is that... when you put "biomechanical claws", there is no right or wrong, so it will always look cool

[dry crow]

yea with my cyborg prompts its also very good xD

[potent hornet]

hahaha

[dry crow]

give me sec

[analog goblet]

hi 1.5 guys 🙂

[potent hornet]

that looks great!

[dry crow]

hey 1.4 guy 😛

[analog goblet]

😉

[dry crow]

thx i discovered better results with model merges

[dry crow]

ouh very nice

it looks realy cold

[analog goblet]

yeah, cold touch 🙂

[dry crow]

which model is this? 1.5 ?

[analog goblet]

yes

[dry crow]

frozen cherry trees

[dry crow]

looks like an old photo

[analog goblet]

sometimes stuff goes into ... unexpected directions

[potent hornet]

model merges? you merge them by yourself in webui?

[dry crow]

yes i merge a lot and test with them

[potent hornet]

amazing colors and composition! I would easily frame this to put on a wall

[analog goblet]

🙂 thanks

[dry crow]

@analog goblet do you use the 1.5 VAE ? im not sure

[potent hornet]

sorry to ask but, is there a tutorial that you would recommend to watch and learn this? I just have no idea, haha

[analog goblet]

this is v1-5-pruned-emaonly.safetensors

i dont know if this answers your question

i downloaded that VAE file, but i have no clue how to apply it

[dry crow]

yes thx i would suggest to use the vae to get rid of the tiny pixelation

[analog goblet]

what graphics card do you have @potent hornet ?

how to do that ?

[dry crow]

you need the vae.pt file of 1.5, then rename it to match the model file, the vae goes into the models folder

the webui will then be apply it automaticly if its set so in the settings

like here:

[analog goblet]

so i rename it to ... ok

its a safetensor right now

so i rename that, too, yes ?

[dry crow]

no

the boath files need the same name but not the .filetype

[analog goblet]

v1-5-pruned-emaonly.safetensors is my mondel

*model

v1-5-pruned-emaonly.vae.safetensors would be the name then

[dry crow]

v1-5-pruned-emaonly.vae.pt would be the vae name

[analog goblet]

ok, so do rename to .pt

[dry crow]

no

you need the vae file

[analog goblet]

i downloded it somewhere i think, but it was name .safetensors

[dry crow]

how big is it ?

[analog goblet]

can you link me that pt file ?

326mb

[dry crow]

ahhh okay

then yes

rename it to .vae.pt

at the end

[analog goblet]

then restart the webui thing ?

[dry crow]

yes

[dry crow]

i can give you a short explanation, idk if there a tuts

[potent hornet]

NVIDIA GeForce GTX 1050TI, 4 gb VRAM

[potent hornet]

I would appreciate, hehe

[analog goblet]

that's very borderline slow

it might maybe not be too much fun

[potent hornet]

thats right.. too sad that here in Brazil the GPU prices are insane

I can never make high resolution pictures

[dry crow]

you should use --xformers and --medvram in the webui-user.bat

[analog goblet]

maybe pay for a monthly midjourney ?

[potent hornet]

what exactly does it do?

[analog goblet]

seriously a 4gb 1050 ?

makes it a little less slow hopefully, but still slow

[dry crow]

it then uses less vram of your 4gb and you will be able to generate better

[potent hornet]

nice

ill sure try this

[potent hornet]

yep, thats it, hehe

[dry crow]

@potent hornet your webui-user.bat should look like this:

[analog goblet]

i think there's even a --lowvram options, isnt there ?

[dry crow]

yea but with 4gb you should use medvram

[analog goblet]

ok

[dry crow]

did the vae work malicor ?

[analog goblet]

no

but i think it did after i renamed it to .vae.safetensors

[dry crow]

make sure in the webui settings under Stable Diffusion the vae is set to Automatic

[analog goblet]

it is automatic, yes

[potent hornet]

just did this

autolaunch is for...?

[dry crow]

autolaunch is starting the webui in browser after starting the bat

[wispy nest]

I found that switching vae's manually is easier than naming them after the model, less headache for me anyway :P

[dry crow]

yea if you have many vaes or models then yes

[wispy nest]

What's the most appropriate channel for prompt / model assistance. Trying to replicate an artstyle I use a lot on midjourney.

[dry crow]

#📝｜prompting-help

[wispy nest]

and for some reason I don't trust myself with changing the web-user.bat, instead I just made my own bat @echo off cd /D E:\stable-diffusion-webui REM git pull OR git checkout master OR (before highres fix change: git checkout fd4461d44c7256d56889f5b5ed9fb660a859172f) REM --precision full --no-half --api --xformers --disable-safe-unpickle --medvram --reinstall-torch --reinstall-xformers set COMMANDLINE_ARGS=--xformers --api start "opens a new incognito chrome browser" /MAX "chrome.exe" --incognito call webui.bat :P

[dry crow]

nice xD with the commands to remember

[potent hornet]

wow

[analog goblet]

well, i think i somehow bricked my system with that vae thing

[wispy nest]

yes, no way in hell I'd remember then, my memory is so bad that I don't even remember it all!

[dry crow]

what okay yea never saw a .vae.safetensor file xD

[potent hornet]

since i'm no dev, this is crazy for me, i'm really struggling to learn all this thing, hahaha

[analog goblet]

when i try to train on a embedding i get "KeyError: malior_wild3"

i got malicor_wild3.pt in my embeddings/ directory though ... i was training it before the vae thing

[potent hornet]

hey @dry crow , just tested the --medvram thing, and wow this is a gamechanger

[analog goblet]

i now cannot pick ANY embedding to train anymore

[dry crow]

also with xformers ? glad it worked

[potent hornet]

yes

[dry crow]

top

[potent hornet]

thank you so much

[dry crow]

np i help where i can, you mostly see me in #🤝｜tech-support xD

[analog goblet]

any idea whats wrong with my installation now ?

[potent hornet]

yea, you already saved me sometimes there, hahaha

[dry crow]

an update could broke it. or you can rename the vae back like it was maybe it works

[analog goblet]

no, doesnt work at all now anymore

[dry crow]

the training or the whole webui?

[analog goblet]

training

which sucks, since 98% of what i do is training

[dry crow]

i dont think the vae has something to do with it

[wispy nest]

did you try a fresh new training embedding without continuing from another one?

[analog goblet]

i think the git pull updated 2 files

[dry crow]

try to search for updates in the extensions tab

[analog goblet]

i m not sure how to see what happened

i cant see the 2 files anymore

(closed and restarted the cmd.exe )

can i somehow with git revert all changes from the last hour ?

[dry crow]

yes there is a tutorial

its for pix2pix but in the first minutes he explain the whole git update and revert stuff:
https://www.youtube.com/watch?v=0fkGd9wIhrA

[analog goblet]

should maybe be git reset --hard master@{1.hours.ago}

i m not sure if i'm at master, but i suppose so

[dry crow]

you can try idk but yes you should be on master

[tranquil kindle]

AI is not that good at generating the "text" you wanted it to do.

[dry crow]

@potent hornet here is a quick explanation of the Merge tab:
If you have 2 Models (A, B) always use Weighted Sum: It will merge them with this formula where
M is the Slider Scale: A * (1 - M) + B * M
It will merge the two with the Percentage of the Slider (0.5 = 50%).
If you have 3 Models (A, B, C) always use Add Difference:
It will merge them with the formula where Slider is M: A + (b - C) * M
In Fact it will erase all of the C Model of your final Merge.

Interpolation is only when you have 1 model (its for making ckpt to safetnesor, baking in a vae or just renaming).
ckpt and safetensor are the output files you can choose for the model.
Choose dont copy config if your models are based on 1.5
also dont check fp16 if you dont know what it does.

[waxen bramble]

I have yet to find a model that show significant difference in inteference qualities between 32 and 16.

But what it does - for those that don't know -ä just means you can store 23 fractions instead 10.

Basically every number is twice as big.

[dry crow]

i heard that gtx cards can only use fp32

[analog goblet]

C:\Programmet\AI\stable-diffusion-webui>git reset --hard master@{2.hours.ago}
HEAD is now at 9beb794 clarify the option to disable NaN check.

any idea how to fix this ?

i dont get anything to work right now :/

looks like now my webui-user.bat is broken, too

its empty, i dont have the --xformers and i dont know what else i had in there

[dry crow]

oh crap, --xformers and --autolaunch is all you need if you have much vram (8gb+)

[analog goblet]

got 24, yes

it asked me to reinstall torch and xformers

no idea why, but i tried to let it by doing

set COMMANDLINE_ARGS=--reinstall-torch --reinstall-xformers

[dry crow]

because its outdated

they pushed the version check a few days ago

[analog goblet]

Successfully uninstalled torchvision-0.13.1+cu113
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
xformers 0.0.14.dev0 requires pyre-extensions==0.0.23, which is not installed.
open-clip-torch 2.7.0 requires protobuf==3.20.0, but you have protobuf 3.19.6 which is incompatible.
Successfully installed torch-1.13.1+cu117 torchvision-0.14.1+cu117

how to fix that one ?

[dry crow]

you need to delete boath xformers folder and only use --xformers. it will then download xformers again

[analog goblet]

both xformers folders where ?

[dry crow]

in the webui folder search for xformers

with windows explorer

should find folders

[analog goblet]

this will take ages

[dry crow]

found them in stable-diffusion-webui\venv\Lib\site-packages

[analog goblet]

and there's a second directory i need to delete, too ?

[dry crow]

nope just two xformers folders

not the directory

[analog goblet]

whats the difference between folder and directory ?

in venv/lib/site-packages i have a directory xformers and a xformers-0.0.14.dev0.dist-info

delete those two ?

[dry crow]

yes

[analog goblet]

then restart webui-user.bat

[dry crow]

these two and then run the webui only with the --xformers --autolaunch

[analog goblet]

ok

what does --autolaunch do ?

[dry crow]

it just opens the webui in your browser after you launch the bat

so you dont have to type the ip or copy it

[analog goblet]

okay, it says "installing xformers" now

i still cant choose -any- embedding under train

[dry crow]

hmm do you get an error? then post it in #🤝｜tech-support maybe someone knows it

[analog goblet]

no error, just an empty dropdown list

uh, on startup it said TI embeddings loaded: 0, and skipps(229)

so it skipped all

but why?

hmm, maybe because it was the 2nd time it tried to do that and it had them already

i try restart

[dry crow]

make sure you select a model that is compatible with the embeds

then make an image with it and then restart so it will load the compatible checkpoint

[analog goblet]

what do you mean with "make an image" ?

[dry crow]

txt2img

[analog goblet]

ah, just a picture ?

doesnt help

i switched back to the v2.1.ckpt

and created a picture

it even said it loaded all embeddings

[crimson flower]

So Shutterstock has their own image generator now.

[analog goblet]

okay, after a lot of back and forth and restarting i can now choose the embeddings again in the 2.1 model

i m pretty sure the 2.1 embeddings worked on the 1.5 model, too though

[dry crow]

no thats impossible

embeddings work only for model base they are trained on

[analog goblet]

hmm, then this is a bug probably

[dry crow]

is it free?

[analog goblet]

so if you do:

1. kill the cmd.exe
2. start it again with a 2.1 model active
3. create an embedding there
4. train that embeddig 5 steps
5. switch to 1.5model

-> the 2.1 embeddings are still available under "train"

[dry crow]

ahh okay seems like a bug then

[crimson flower]

Free to generate previews, paid to download.

[dry crow]

paid to download xD wow

[analog goblet]

i could have sworn that i trained those landscape images on a 2.1 embedding while using the 1.5

but i cannot prove it in any way unfortunately

[crimson flower]

Cat generated by the ! token.

[dry crow]

in shutterstock?

[crimson flower]

Yes.

[analog goblet]

on the other hand, being confused is basically my normal state in the stable-diffusion-game 🙂

[crimson flower]

Seems like it's treated similarly to regular stock photos. Subscription or on-demand. Expensive either way.

[analog goblet]

first 1.5 image, kinda 🙂 i think 🙂

[dry crow]

oh you always used 2.1 before?

[analog goblet]

i have no idea

well, i m pretty sure no actually

[dry crow]

or did you meand your first image of SD ?

[analog goblet]

because i tried around with some nsfw stuff and that worked out and wouldnt have at 2.1

but i think the embedding training i did was probably 2.1 only

i was just confused that all embeddings seemed gone while i thought i used them on the 1.5 model

but maybe i just -thought- i used them, while they had no real effect to the ai

[dry crow]

xD that can be possible

[analog goblet]

i m for sure confused, even more so than usually 🙂

kinda nice colors 🙂

[dry crow]

yea with constant updates new features and the community with extensions xD its a lot to test and watch

[dry crow]

wow nice also the texture looks like a real painting

[dry crow]

yea, i love that we can use SD local for free

[crimson flower]

Shutterstock AI logic:
Anime? Ok.
Woman? Ok.
Girl? Ok.
Anime woman? Ok.
Anime girl? Hol' up.

[dry crow]

hahaha xD

[analog goblet]

woo, 1.5 is again a whole new world 🙂

[crimson flower]

And "anime woman" looks more like 3d game characters than anime. You have to use "2d anime woman" for something decent.

[analog goblet]

i m not really sure if it's a nice world yet 🙂

[dry crow]

1.5 is really nice and support the most stuff

[dry crow]

wow, did they trained their model on shutterstock photos or what xD

[analog goblet]

i wonder if embedding training works about the same as in 2.1

but we're about to find out 🙂

[dry crow]

it should, but yea let us know ^^

[analog goblet]

i m not convinced yet 🙂

1.5 seems to be more into the longneck / foureye / twoheads stuff

[dry crow]

ohh xD

[waxen bramble]

They work the same way as you can navigate London with map of Paris. Both are maps of a city, and might lead you to somewhere, but sure as hell not to the right place.

Also due to the fact that 1.5 uses CLIP VIT-L that was pretrained, 2.x OpenCLIP vit-h which they trained themselves. Totally different models and structures.

So you navigating London with map of Paris written in German.

If you have trained a model, like with Dreambooth and touched the text encoder, the embeddings are most definitely absolutely porked and must be retrained for that model. If you only touched Unet, then... chance that they do something approximately correct however no quarantees on that.

[analog goblet]

makes sense 🙂

[analog goblet]

this is pretty hard

embeddings are stubborn in 1.5

getting weird stuff 🙂

this is pretty nice actually 🙂

[analog goblet]

it looks to me though has if doesnt give an F about what you give it as training dataset directory when you train an embedding in v1.5

[wispy nest]

"Save a copy of image before doing face restoration."
Where is it saved? I thought it would be in Log/images but nope

[kind orbit]

Hi, if you have an image of an dog on a blank background, how do you inpaint a background (like some flowers) behind the dog? I kept getting almost nothing in the background.

I'm using a mask of the dog during img2img inpainting

[sacred spruce]

You will need to either use a denoising strength of like 1 with (maybe latent noise?) or you could add color paint to where you want the flowers to be.

[analog goblet]

say, is textual inversion embeddings completely broken for 1.5 ?

it seems that it compleeeetely ignores the dataset directory training images

[sterile kiln]

Tell me if i'm wrong but, to use pix2pix locally, 18 GB of Vram is needed? Right?
I only have 8 GB (2070 SUPER), i suppose there is no way?

[dry crow]

wrong, it works with 8gb vram for me xD but you should have xformers

[outer moss]

anybody know how to use wd1.4 e2 on automatic 1111?and then this happens

i have the model and the yaml it wont work

tried visiting the link i dont understand it

[dry crow]

you have to name the .yaml the same as the model ckpt:

Example.ckpt
Example.yaml

[outer moss]

same file right?

here?

[dry crow]

yes

[outer moss]

still same thing happens

oh nvm

my goofy aaaah also coppied the file format

[dry crow]

xD

[analog goblet]

hi @dry crow 🙂

say, do you know if textual inversion in 1.5 is simply completely broken ?

[dry crow]

hey, i dont think so, but who knows xD

[analog goblet]

i was doing about the same as in 2.1 and from my tests it completely to 100% ignores the training dataset pictures

[outer moss]

whats the best webui in yall opinions?

[dry crow]

hmm thats very strange

[analog goblet]

try it out, its easy to see

[dry crow]

My fav is Automatic1111 webui, it has the most Features and also support Extensions.
For the Design the win goes to InvokeAI. Very pretty interface

[analog goblet]

put "create embedding -> initialization text" to *, and then give it 10 pictures of dogs

[dry crow]

@analog goblet maybe watch this tut based on 1.5: Maybe there are some steps you need to change?
https://www.youtube.com/watch?v=2ityl_dNRNw

[outer moss]

isnt there one that lets you see prompts that would work? like if you type a part of the prompt it completes it i found it when i used a colab but i cant tell what its called but there was a big "nocrypt." watermark on it

[analog goblet]

thats about the idea, but try it out, doesnt seem to work

[dry crow]

@outer moss i use the extension Booru tag autocompletion. That gives me autocorrect of anime tags

[outer moss]

how do i install it?

[dry crow]

in the webui there is a tab "Extensions" then click on Available, then on Load from, then search it and install, then apply and reload the webui

[outer moss]

is it confirmed to be safe tho ?

[dry crow]

its on github, so you can check the code. but i use it for like weeks and its working good:
https://github.com/DominikDoom/a1111-sd-webui-tagcomplete

[outer moss]

the closet ive been to coding is editing game code and minecraft command blocks

[dry crow]

same xD and a bit c++

[analog goblet]

got a few min to try the textual inversion thing @dry crow ? help would be really nice here, i have no clue whats going on

[dry crow]

if you give me the instructions sure

[analog goblet]

train -> create embedding

name: cs1o_dogs

[outer moss]

idk man seems pretty sussy i might get coronavirus on my computer

[analog goblet]

init text: *

[dry crow]

its up to you, like i said its working for me and friends and other community members

[outer moss]

how many people have used it?

[analog goblet]

then -> create embedding

[outer moss]

approx

[dry crow]

i think like the whole anime com

[analog goblet]

then tab : train

[outer moss]

seems safe enough that a entire website uses it

[dry crow]

yea its just scrapping tag data of danbooru, nothing special

[outer moss]

do i just paste the gethub link in it?

[dry crow]

no its listed there under extensions when you click on load from

[dry crow]

next step?

[analog goblet]

embedding: pick the one you just made

emb learning rate: 0.01

dataset directory: c:\somewhere\dogs

[dry crow]

i have no dogs xD

[analog goblet]

make a directory c:\somewhere\dogs and put 10 dog pictures in there

[dry crow]

can i use random stuff

[analog goblet]

just an example

10 of one-of-a-kind

[dry crow]

okay

[analog goblet]

not one dog, one cat, one house

[dry crow]

ok give me a min

[analog goblet]

Save an image to log directory every N steps, 0 to disable : 1

[dry crow]

okay

done

[analog goblet]

you set the save-image to 1, yes ?

how many pictures did you put in the dataset dir ?

[dry crow]

14

[analog goblet]

how much vram ?

[dry crow]

i have a gtx1080 with 8gb

i dont want to take this an hour xD

[analog goblet]

delete 2 images

do batch size 4 and gradient accumulation steps 3

[dry crow]

done

[analog goblet]

(or 3 and 4, depends what the system can handle)

then "train"

"train embedding" that is

[outer moss]

dont understand how to install the extension still put it in cave man terms

[analog goblet]

chaos: step 1: click on "extensions"

acknowledge with "check".

[dry crow]

@outer moss

then search for Booru tag autocompletion its near the middle bottom

[outer moss]

this and click it?

[analog goblet]

yes

extension index url put: https://raw.githubusercontent.com/wiki/AUTOMATIC1111/stable-diffusion-webui/Extensions-index.md

then press "load from"

[dry crow]

no dont type it there

just click the button

[outer moss]

what do i check here?

[analog goblet]

put https://raw.githubusercontent.com/wiki/AUTOMATIC1111/stable-diffusion-webui/Extensions-index.md as url

[dry crow]

check, localazation, and ads (these then get not listed)

sort by a-z

then look for Booru tag autocompletion

@analog goblet how long does the training goes?

[outer moss]

ples send a video or something i still cant find it out lmao

[dry crow]

did you clicked the button load from?

without any input next to it

[outer moss]

no i pasted the link on it

[dry crow]

okay good

then click load from

[analog goblet]

training goes forever @dry crow

but it shows you the generation pictures (epochs)

and those should convert towards your training dataset

[dry crow]

so what to do now? wait a bit?

[analog goblet]

yes

[outer moss]

worked now which one do i install?

[dry crow]

@outer moss Booru tag autocompletion

[outer moss]

you know how the dreambooth extension effect the webui or the program itself since im running on low vram

[dry crow]

it dont work xD

with less then 12gb vram

[outer moss]

it says optimized for low vram

[dry crow]

wow xD with my 8gb i cant use it too

@analog goblet hey how long should it run? how can we test if it works?

its at 50 steps now

[analog goblet]

50 steps should be enough

do the images look remotely like your training dataset ?

[dry crow]

i used a blue parrot and every image is a blue birdlike thing

[analog goblet]

did you put an init text or set * ?

[analog goblet]

and which ckpt ?

[dry crow]

what wait

[dry crow]

where

[analog goblet]

at "create embedding"

[dry crow]

the text is *

i did not stoped the training yet

i dont know how xD

[analog goblet]

can you send me 1 of the dataset and 1 of the ones you got from training ?

you scroll down and click on "interrupt"

[dry crow]

ahh thx

okay what now

ill try an txt2img with my embed? but how to trigger it

is it * or cs1o_dogs

[analog goblet]

now you go to c:....\stable-diffusion-webui\textual_inversion\2023-01-28\nameofyourembedding

[dry crow]

okay

[analog goblet]

there you find those training images, right ?

[dry crow]

yes

a lot of blue parrots

[analog goblet]

send me #1, #20 and #40 pls

[dry crow]

it starts with 2

[analog goblet]

sounds fine 🙂

[dry crow]

also added 60

wait there are not sorted right

[analog goblet]

no problem

[dry crow]

ok

[analog goblet]

ok, now take 12 pictures of some-other-thing

like dog / cat / house / whatever

make another dataset directory

and do you use the v1-5-pruned-emaonly.safetensors ? exactly that one ?

[dry crow]

need they 512x512?

[analog goblet]

not for this test

[dry crow]

no where do i have to select it for training? or do you mean for generate?

[analog goblet]

its just topleft stable diffusion checkpoint

[dry crow]

no there was a custom merge model selected

[analog goblet]

its all so weird

can you send me that ckpt you use ?

[dry crow]

i can try train again with 1.5

the model was based on protogen and an anime model xD

[analog goblet]

oh, it wasnt a 1.5 base ?

that would explain everything

[dry crow]

boath are based on 1.5

[analog goblet]

i m training on portraits like this:

and these are some of the outputs i get:

[dry crow]

lol

[analog goblet]

(i really like the last one, but still ...)

[sterile kiln]

? It's used in the Automatic1111 graphical interface, I didn't quite understand how and where to install it, isn't that right?
https://github.com/timothybrooks/instruct-pix2pix#other-ways-of-using-instructpix2pix

[dry crow]

@sterile kiln watch this 😄
https://www.youtube.com/watch?v=0fkGd9wIhrA

@analog goblet so should i train again? i have 10 anime pics of the same person

but udk if 1.5 ema will work that good with it xD

[analog goblet]

can you try with the 1.5 ema ?

[dry crow]

sure

[analog goblet]

for the next test pls make a new embedding

again with init text *

rest the same

by the way, the pix2pix thing doesnt work remotely as good as promised in that video

[dry crow]

it works very good for me, changing hair colors and stuff, but not for big changes like weather. its limited by only changing the surface of an image

[analog goblet]

it cant make a portrait photography to a cartoon for example, but thats a whole different thing 🙂

i d be happy if i could fix this embedding problem

[dry crow]

thats only possible with img2img

[analog goblet]

curious if it works with the emaonly 1.5

[dry crow]

its saying preparing dataset and it loads ages

[analog goblet]

how many images in the dataset dir ?

[dry crow]

12

but 512x768

[analog goblet]

ah, that may have just used too much vram then maybe

[dry crow]

yes its seems so

[analog goblet]

you can try interrupt

[dry crow]

100% gpu, 7,8 of 8

[analog goblet]

might work, might not work, maybe you have to kill the cmd.exe

if you want to go with this, reduce batchsize (and up the grad.accumulation thing)

like to 1 batch and 12 grad-acc

instead of 3 and 4

[wispy nest]

How do you generate stockphotos that actually look like real photos?

I can always recognize that they're not photos

I've tried prompting with "photo, realistic, real person, 16k" etc.

[dry crow]

change the resolution to 512x768 and tag portrait

[dry crow]

so i interrupt and start again?

[analog goblet]

yes

(technically you interrupt and continue)

[dry crow]

okay

now with 1 / 12

[analog goblet]

and ? 🙂

[dry crow]

30 steps

images look similar but not very good

[analog goblet]

does it look close to anything ? 🙂

[dry crow]

but very close to the original

i gave it an anime girl with the same clothes and hairstyle

the colors always are used but randomly

so i guess its working

[analog goblet]

weird

i wonder why its not working for me

[dry crow]

trained on pics like this:

[analog goblet]

maybe not the nicest output but defintively works

do you have the 4gb or 7gb ckpt ?

[dry crow]

i used the pruned emaonly

i have boath

[analog goblet]

well, then i dont know 😦

[dry crow]

hmm im uptodate with the webui,

[outer moss]

anyone know how to fix this? i alr use low vram and this still happen?

[dry crow]

wrong channel, pls #🤝｜tech-support

[outer moss]

thanks

[reef merlin]

does anyone know how to get DreamShaper to work? i'm not sure what im doing wrong, but im getting results like tihs

posting this in case it helps

[dry crow]

You need a minimum resolution of 512x512 also use the tag dreamshaper and dreamlike

[reef merlin]

i turned the resolution down cause it gave me an out of memory error

[dry crow]

How much vram?

[reef merlin]

i think 4?

i have a 1650

it says RuntimeError: Not enough memory, use lower resolution (max approx. 384x384). Need: 0.0GB free, Have:0.0GB free

[dry crow]

Do you use xformers and medvram ?

[reef merlin]

i have xformers but most of the time it says it's starting without them for some reason

i don't think i have medvram on tho

[dry crow]

Try edit your Webui-user.bat and behind Commandline_ARGs=--xformers --medvram

[reef merlin]

like this?

[dry crow]

Yes

[reef merlin]

i'll test it out, thank you :)

[dry crow]

Np hope it helps 👍

[sterile kiln]

Xformers works with RTX, not GTX

[dry crow]

Wrong

[sterile kiln]

ah

[dry crow]

It works with boath

[reef merlin]

is "set PYTHON=" where you can specify a python version? it's trying to launch in 3.9.12 and xformers isn't supported by that

[dry crow]

Yes there you can set the path

3.10.9 is the highest Version that works. But 3.10.6 is recommended

[sterile kiln]

Guys I'm going crazy, I'm pretty sure that a lot of people manage to create an Embedding with 8 GB of Vram, personally I can't do it and it's driving me crazy, someone nice who would have 8 GB and who would have managed to give me a hand?

[dry crow]

If you scroll a bit up, @analog goblet gave me an short instruct how to train an embed

[sterile kiln]

CUDA crash, each time

[dry crow]

Lower the batch size

[sterile kiln]

yeah, i just did what's in the video

[dry crow]

Ah okay

[sterile kiln]

I have 9 images

[dry crow]

How much vram?

[sterile kiln]

8 GB

2070 SUper

[dry crow]

With 10 images you could try 5 / 2

[sterile kiln]

5 / w?

5 batch 2 gradient?

[dry crow]

Yes

[sterile kiln]

CUDA crash

[dry crow]

Make sure you vram is cleared

Also the images to train should be 512x512

[sterile kiln]

it is

[dry crow]

Okay

[dry crow]

Okay should work

[sterile kiln]

What ARGs are better to do this, some peoples says --medvram create problems

I use this :
--no-half --medvram --xformers --api

[dry crow]

dont use --no-half

its not needed for 8gb vram

also dont use --medvram for 8gb xD

your capping yourself

[sterile kiln]

Yeah but without it, i had black pictures

i will try to remove it

Yeah, it now works without --nohalf 🙂

(txt2img)