🏞｜general-with-images

[solemn topaz]

I see. Have you tried keeping the seed same and playing around with the negative prompt?

[last pawn]

Yes I tried but maybe I did it wrong. I'm lost...I see a LOT of informations around the Ai so maybe I left something. For the moment, I write prompts, then I write Negative prompts, I work with like 30-50 steps, I work with Euler A, PLMS or DDIM, I have the InkPunk.ckpt as a checkpoint, I understood the Seed, the CFG Scale and the Denoise Strengh. I only work with Txt2img, helped by ControlNet.

[last pawn]

I never trained nothing (I don't even know what it is), I don't use LORA (I don't even know what it is)...LOL'

[last pawn]

I have really bad result with Restore Face (I activated the CodeFormer model in the settings)

[wispy nest]

idk if there's a point in using restore face rn, since we have updated vae

vae-ft-mse-840000-ema-pruned
vae-ft-ema-560000-ema-pruned

If you have one of those , you shouldn't have face problems
(unless character somewhere far away from camera)

[last pawn]

Yes I got this:

[last pawn]

working with SD 1.5

[austere pilot]

@tropic shell when I tried adding the add on in blender it doesn't seem to be showing up

[last pawn]

I'm a blender user so check if the addon work on the Blender version you have. You'll need 3 or 5 different version of Blender to be sure all of your addons work! 😅

[last pawn]

Maybe it helped!

[austere pilot]

3.4

I do see Skinify plugin tho create a mesh from bones but idk if that it is

and that is only available in pose mode which idk anymore how I can get there

[last pawn]

Sorry I really don't know what you're looking for but I allow myself to answer you because I'm a Blender user...I have BIG troubles with Stable Diffusion but if I can help with Blender...hehe! So the Pose Mode is like the Object Mode or the Edit Mode. You access to it by pressing Ctrl+Tab

[austere pilot]

I don't see it there

[last pawn]

If you selected the right element (a Bone) you'll be able to access the Pose Mode.

[austere pilot]

how do I add bones?

[last pawn]

that's because you select an object that don't have bones

shift+A...Armature

[tropic shell]

I haven't used Invoke in awhile. I do love what they have to offer! I only recently got Auto to work

[austere pilot]

how do I get Make Human to work?

[tropic shell]

Make Human is its own application! And you can import to model into Blender, then use the Blender SD plugin with it! Just google MakeHuman. You can also search for VRoid

You can export it as a fbx, or dae file

[austere pilot]

it doens't show up in the blender add on...

[tropic shell]

vroid studio: https://vroid.com/en/studio

But you'll need the CATS plugin for Blender to import for Blender

Or again, MakeHuman

Once you have created your character, like so

[austere pilot]

hey!!! I remember now my teacher showed us this

[tropic shell]

you go to the pose/animate tab

Add the skeleton

Default

[austere pilot]

but tbf I should stick with using it as a blender addon cuz I need it for school

[tropic shell]

You can then export this as a fbx

[austere pilot]

isn't there like a standalone thing?

[tropic shell]

Yes, it is standalone

Then you can import it into Blender

Ofc, you can do this with any 3D character, but since MakeHuman is free

[austere pilot]

I should unzip the zip after downloading it?

[austere pilot]

does it have like different hair styles etc as well?

[tropic shell]

Yes, although, I am actually working on some stuff for the community

Rn

To make it easier for ppl to use the plugin

[austere pilot]

woah nice

[tropic shell]

When I get the tutorials out, I'll post links

[austere pilot]

feel free to ping me with them if you remember

[tropic shell]

Anyway, once you have the .dae file, you can import. It will bring in a weird animation. It'll be distorted

Dont worry. Just reset the pose. Also, make sure the bones are connected

[austere pilot]

wait .dae from the downloaded zip? or when I exported the character from MakeHuman

[tropic shell]

(If you import animations)

When you export from makehuman

[austere pilot]

ah so from makehum to blender?

[tropic shell]

yup

The eyes and eyelashes will be a little messed up but it's not a problem

You might also need to sculpt some adjusments

[austere pilot]

these are the files I got how do I open it?

[tropic shell]

Lol, I forgot the eyelashes. You'll need to adjust the alpha settings for the eye texture, etc

Anyway, I have to go now

I have an appointment

[austere pilot]

@tropic shell how do I open?

[tropic shell]

Also, the eye has two layered meshes

Double click the exe

[austere pilot]

that's to install it

the setup

[tropic shell]

Did you make an icon on your desktop? Apparently, that file links to

[austere pilot]

think I found it ye it's on my desktop

[tropic shell]

Lol IDK I click icons, they work--I never went farther than that

Best of luck!

[austere pilot]

euhmmmmm??

[last pawn]

I got this now...It's better but it's really not what I'm looking for...it's rough, it's sloppy...

[austere pilot]

I have put it on alpha hashed but I still don't see the eyelashes?

[wispy nest]

How does it suck for you?

[last pawn]

Because I see stuffs like that on the web...😟

[ripe cedar]

alpaca is out for PCs. The best (30b) version can be run on 32gb RAM (not vram). Anyone know how it helps in prompts?

[dense tapir]

I have 48gb of ram but meh to chatgpt

[ripe cedar]

alpaca was trained in part from chat gpt3 but they're not the same. I just wondered if it could help with SD because that would be cool. It can help in writing projects though

[dense tapir]

well, for prompting I dunno I am on the side of the people that if you can't even think of a prompt that is pretty lazy and/or lame

[grizzled sage]

No better than any other prompt generator really. Since it's trained with a different LLM, it might not align with the image prompt systems as well as we'd hope. Ideally, we want something that is trained with the same clip system to generate prompts, so that it uses tokens and phrases that affect the model the most

[dense tapir]

there is that as well

[grizzled sage]

i think it'd be worth experimenting!

[dense tapir]

with the advent of SDXL everything is on hold for me

flow

see the above pic I posted?

[grizzled sage]

i don't think it's so much as lazy as it is a move towards automation and batching. also, there's the bill gates quote i love. not sure if its real or apocryphal though. "I'd rather hire a lazy programmer because he'll find the easier way to do things and create the better tool"

[dense tapir]

that is some of what I am talking about that would be bad to train with due to the cold temp and glowing objects.

naw, he is too lazy to create anything let alone a tool

[grizzled sage]

cutting corners is the bad kind of lazy, but there's also the good kind of lazy, like being smart enough to have a reaching stick instead of getting up out of your chair ! 😄

[grizzled sage]

i see it

[dense tapir]

DB, and lora/locon will see it and bake it in

real, gen, or painted

then all faces come out distorted

I have no idea how I could change that in say photoshop

[grizzled sage]

would describing the color grading of each pic in the caption help?

[dense tapir]

no, but cool_lighting, or warm, etc... warm_glow does but my lord the time involved for each pic to do that is immense

some I looked at and had no idea what it would be considered

[grizzled sage]

i get that. i'm captioning pic sets of friends lately and it's exhausting. i really need to level up that game. learn some vi skills again

[dense tapir]

when it takes you hours to caption it then half an hour per try then go back and try to find the something you need to add it is hell

[grizzled sage]

i get done captioning 20 pics and i'm just over the whole process

[dense tapir]

yeah, I had 81 and I wanted to commit murder. I was frustrated with some as I had no idea what the ai would consider it.

[grizzled sage]

people training on 10k+ images astound me

[dense tapir]

yeah

[grizzled sage]

base set is just unfathomable

[dense tapir]

75k-300k seems to be a big thing

[grizzled sage]

i'm trying ot get to a point where 'm happy about captioning 1000

thats my spring ML goal

[dense tapir]

See, why isn't there a tool that has AI and says "I see this in this image" now I mean the stuff that it sees when it trains YET we are not told about in the captioning models?

the lighting for instance. If it sees it to train on, and can be captioned away then where is the tools to make that caption for us?

[ripe cedar]

How long until decent auto tagging comes you think?

[dense tapir]

I was thinking soon but now I am not sure

I mean this thing is screwing me over because it has cool lighting

or warm lighting

[ripe cedar]

The interrogate function exists but idk how good that is for training

[dense tapir]

no model I have used even mentions lighting. glow yes but not lighting and its temps

[grizzled sage]

text to image turns out is an easier problem to solve once cracked, than image to text description has been

GPT4 seems to have good capability for oddball instances , but it's trained on openAI's clip model not SD's openclip

[dense tapir]

I can't use interrogate as it is all OOM 😦

[grizzled sage]

blip has stopped working for me in webui

[dense tapir]

the one in webui I can use but is pretty worthless the one in the extension they all downloaded but I just need 12gigs vram to run them

[grizzled sage]

there's an extension? i didn't even look because the one in webui is there

[ripe cedar]

Ideally we could get away from human language altogether. That's why I don't like to use text much. Ai is best without it

[dense tapir]

yes, good models as I know them from the web

15-30 gigs each model

[grizzled sage]

naw. we want to supplement natural language descriptions, not move completely away from it

[dense tapir]

what I wonder is for training I get the sense it really only trains style well if you have 5500Kelvin studio lighting

[grizzled sage]

somethings are just much better described through text. if we start using emojis to describe ideas, then we just made another pictograph version of human language.

[dense tapir]

devolution is real

[grizzled sage]

we are devo! d e v o!

[dense tapir]

yep

[grizzled sage]

🎶

[dense tapir]

how would I colour gradient that pic above to work with training?

I tried normalizing histogram by hand for 81 pics and it baked in the glow around the gen subject that it saw in the master images from the normalization

[grizzled sage]

colour grading? i'm not sure. in my experience (very little!) i'd assume that pic was fine for training! i wonder if you're running into the same problem that the illuminati diffusion guys trained against. The offset noise situation where it tends to average the luminosity of every generated image

i've got to learn about offset noise yet. i've seen sliders for it in the dreambooth extension

[dense tapir]

no, because I had someone tell me, without even seeing my training data what it was. Sure enough they were spot on. They had 100 images and 2 had a green tint and from then on all people generated looked like they had vomit on their faces. He captioned the green light and POOF, perfect.

[grizzled sage]

"Dark templar lighting" or whatever their protoss dark homeworld is called

[dense tapir]

for instance I have some wearing a halloween mask of a skull. Yepo, sure enough the people gen have a slightly elongated face but put skull_mask in the tags and train again gone to normal faces

[grizzled sage]

"purple lighting" may not do it

[austere pilot]

@last pawn you answered some of my questions before about blender, what's this arrow called between the nodes? it's to create hair. (for those that are out of context, I am creating my SD generated character as a 3D model so I can make her do different poses etc and perhaps even train a mode on it)

[grizzled sage]

from what i understand for captioning, you want to caption everything that you don't want to appear in every image otherwise the training generalizes that image concept as one that fits everything

[dense tapir]

I have heard it both ways but the tagger should see it. It sees most things but what it doesn't see is still a major issue

[grizzled sage]

if i were smart about command line stuff still, i bet there is one command i could type with a regex in it, taht would append lighting tags to a ton of files easy. for that process i would sort everything into folders of similar lighting, then run the command in each folder for each similar tag.

[dense tapir]

oh, I can do that already, sort of, but the problem is orange, warm, blue cold. what about the rest?

it only know warm and cold lighting so you have to figure out which is what

orange glow is warm_glow

red warm etc...

some I have no idea.

[grizzled sage]

yeah. i would call that one "magical dark energy" or something. it's hard to intuitively choose a color grading for that

[dense tapir]

problem then becomes it doesn't work like that for training

be nice if it did

[grizzled sage]

why not? lighting temperature only works because it's natural language and it has references to higher or lower temperature images

color grading is something different from color temperature

[dense tapir]

magical dark energy means exactly what to a blind person?

[grizzled sage]

what would warm or cool lighting mean to a blind person?

[dense tapir]

hot cold

it knows red is hot, blue is cold like water temps

[grizzled sage]

would a blind person intuitively know hot means orange?

[dense tapir]

you'd be surprised

anyway, the point is training magical dark energy will not do as you expect

I lock onto dark and think of no light

[grizzled sage]

they were told if they do. or they had sight once. we're talking intuitively. Anwyays, we're here now. Arguing about the advice i'm being asked to offer. I told you how i feel about that images lighting. I'd call it dark magic or dark templar energy since blizzard and gamesworkshop lean into that style a lot.

[dense tapir]

I didn't consider this arguing I just said it doesn't work like that in the real training world. Try it.

[grizzled sage]

"go do it for me" another cliche people pull after asking for advice. Look, i'm not standing in your way at all brother. you asked how i would describe it. i did. "dark magic" has succesful prompts with that sort of lighting too.

i'm not that experienced, but you came to me. I'm not doing your HW for you now because you're disagreeing and arguing with my idea.

[wispy nest]

What model is it? was looking for realistic model which can do magic
(kinda found one already, but wanna know other options)

[wispy nest]

yea, you can make things like that with right prompt and right model, but that's too big of a change to go from your image, it'll change completely...and well , it needs too...

[smoky oak]

So sorry, I fell asleep early last night

There is a video on how to do it, but it's by AItrepreneur

I know many people dislike him, myself included

[stark vine]

Lmfao? Who on their right mind would dislike aitrepreneur

The guy is a saint. Or should I say the machine

If anyone should be getting hate is that bald fuck olivios l0l

[grizzled sage]

i dont have anything against him personally, but i think he's a victim of youtube's profit cycle and he's been groomed to chase those view counts and click ratios.

[stark vine]

His info is still legit for the most part

[grizzled sage]

rushing content to release with hot takes that fire up the community for example

[stark vine]

Ehh I wish he would focus more on technical aspects but I don't know any better alternatives

SEcourses does videos that are way too long

and others are just flat out annoying with their clickbait

[smoky oak]

Great, so I see Adobe just unveiled their new AI generator

Good thing it looks solidly behind Sd

[grizzled sage]

UX though it's way ahead and tehy're only going to improve the model side

i think they keep theirs as SAAS

[smoky oak]

@stark vine @grizzled sage

My problem with AItrepreneur that I have come to notice is he gives really bad advice for a lot of things. Like his videos on training LoRA's/Textural Inversions are really damn bad. Not even just bad, but legitimately damaging. He spreads really bad information that damages people's understanding of how all of this works

[stark vine]

I did catch him slipping one time I'm not gonna lie, but other than that pretty much everyone gives good feedback on his advice at the very least that's a good place to start

[grizzled sage]

I was watching his lora videos and it was as if he was just reading from the first guide he googled and hadn't verified the methods at all himself. It's a trend of rushing content to release

[smoky oak]

What I'm trying to say is my main problem here is that so many people rush to defend him, especially in his comment sections, to where people who actually know what they're talking about get dogpiled under the guise of hating AI

It's happened to both I and @dense tapir

[grizzled sage]

honeslty i hate video guides since the youtuber makes it all about them generally. i just want a page to read through and study. I can pull in so much information so much faster than listening to a guy blab on about his week for 3min

[smoky oak]

His video on training LoRA's is truthfully terrible. He has no understanding of how it works, and he spreads techniques that actively damage the community

But if you try to comment on that, his comment section eats you alive

[grizzled sage]

yeah a lot of youtube culture is it's fan bases too. A lot of "celebrity" worship there and i gotta wonder why, they're just youtubers

not trying to diminish how hard of work it is to do, but wow. People get defensive and worshippy

[smoky oak]

I only click on his videos long enough to find what he's talking about, so I can research it from people who actually know what they're talking about lol

But yes, in the future as a little PSA, he does not know what he's talking about when it comes to training LoRA's, or TI's. Hell he probably has no idea what dream booths either

Sure, the information from his video is what turned me on to LoRA's, which is cool. But the information in his video is also what took me so damn long to figure out how to do them, because his information is just objectively bad

[grizzled sage]

and usually made very specific to whatever version was out at that time, because it's just rehashing online step by step guides instead of explaining what it is each setting does

[smoky oak]

I say just take everything with a grain of salt from him. He is far from a good source of information, letalone reliable information

[stark vine]

I kind of disagree with that, it was because of him that I learned how to do dreambooth

And no, I saw his Lora videos and TIS and I honestly don't see what's wrong with them. Then again, I don't specialize on those because they just tend to be shittier than dreambooth

[grizzled sage]

i like the longer videos when they're going into explanations about what a learning rate might affect, or why gradient accumulation helps, things like that. Instead of quick "do this step, just set batch and gradient to 1 because lowvram. no explanations here"

[stark vine]

He probably doesn't know, but then again if you ask around here, everyone will give you a different explanation.

[grizzled sage]

another misconception that comes out of youtube hype culture. "TI is shit", because they're trying to hype their DB videos for views.

they push these extreme hyperbolic positions and their audiences worship and adore them

[stark vine]

He was also the first one to publicly call out what was wrong with dreambooth when it broke and how to get it fixed. I literally saw zero people give advice on how to get it fixed. So he gets credit for that

[grizzled sage]

hyperboles become gospel and its just a wildfire of toxicity in the community. started because a youtuber wanted more views

[smoky oak]

Textural Inversions are less diverse than LoRA's, that is a fact, but TI's have their use

[grizzled sage]

It's another tool and it has uses for sure. Hyperbole leaves zero room for nuance

[smoky oak]

Yeah, for sure

It's a tool with it's limits

[grizzled sage]

I think the deluge of merges that peaked for a bit, where every new model were teh same 5 merged a different way with a few differences, was because of youtube hype on merge tutorials

model merging is another great tool, but it got REALLY popular there for a minute

neat ot watch this all go down though. Barely 6 month old field and all this stuff flares up often. I just wish there was some more "responsible journalism" surrounding coverage

( i hate alluding to that responsibility since gamer gate toxicity tainted the idea )

[stark vine]

It got popular because it's literally the best way to generate images…

I would say anyone who's not using merges is a noob

[grizzled sage]

we're all noobs, that's the thing.

how can you be an elitist when youv'e been at this a matter of months?

[smoky oak]

That's why I try to share as much as I can, cause I know how hard I work to forward my understanding

Oh, I saw that midjourney made a magazine

God I hate midjourney

Of course a scummy company like midjourney is everybody's first exposure to AI inage generation

[stark vine]

I yearn for the day someone leaks their model for the public

Their entire business model would collapse

[smoky oak]

I don't even want to use their shit, I just want them gone

[stark vine]

R0fl

[smoky oak]

Greedy ass company

[grizzled sage]

they're still very niche. soon it will be adobe

[smoky oak]

Luckily at least for now, adobes image generator is very far behind

[grizzled sage]

What's that wallpaper app for cellphones ? it has ai generation now too

most in my local circle bring that up to me

zedge

[smoky oak]

Oh great, there goes my plans of putting my content on there

[grizzled sage]

heh

[smoky oak]

God it annoys me

[austere pilot]

@tropic shell you know a faster way of doing the hair I want?

[grizzled sage]

i dont blame them but yeah, it's annoying a lot fo companies want to do what i'm wanting todo and monetize this

getting in my way

[smoky oak]

All these shitty half assed implementations of what can only be a watered down diet lite 0 calorie version of stable diffusion

[austere pilot]

hope you remember what I am talking about

[smoky oak]

At least I can confidently say that as of right now, stable diffusion is still the best AI image generation platform, and it's also free, which is a huge accomplishment we should be proud of, cause it's not SD that made it great, it's this community that did

Stable diffusion needs to learn that the only chance they have of continuing to stand out in this world of AI generation is to embrace their extremely generous community, rather than fight them every step of the way like they seem to want to

[austere pilot]

it's better then just free

it's open-source

[smoky oak]

For now, yeah

It's important to not smoke screen ourselves with what's now. Cause the truth is, a lot of people including myself feel that stability AI are just abusing this community to learn for free until they can steal it all back and monopolize it

[austere pilot]

wdym for now? Emad is big fan of open source and is CEO of Stability.AI he will never allow a paywall

or lack of transparancy

hence it's the reason he hates OpenAI

[smoky oak]

Don't be so sure of that

With how shitty they have been to the community, I wouldn't wager too much on them as a company caring about us

[hardy tulip]

flower, vibrant, vivid, intense, (masterpiece:1.2), colorful, vivid, (sharp details:1.2), extremely detailed, vibrant lighting, (highly detailed:1.3), (sharp focus:1.2), illustration, beautiful, (trending on artstation:1.2), cinematic, (8k:1.3)
Negative: (low quality, worst quality:1.4)
OpenJourney v4

[austere pilot]

look even if they go ahead and put it back to closed source what does it matter all the code has been released to the public we can just build further even when it's back closed sourced, so who cares if Emad would decide to put a paywall behind it XDDDDDDDDDDDD

[stark vine]

Stable diffusion is definitely the best platform but definitely not the one with the best model.

[austere pilot]

Mj is SD but with heavy filters

[smoky oak]

That is not true

[stark vine]

That's the indictment of capitalism. Those who have the most money get to hog the best product

Im sorry but v5 is insane

Meanwhile, SD gets no funding so who's going to spend all the resources in making a good competitor for V5?

I've been begging for a 1024 model for ages and I'm starting to think it's not gonna happen especially after the lawsuits started flying off

But since MJ gets to raking the money from all the subscribers each month, they can do whatever they want… It's a bummer

[smoky oak]

I really wanna see midjourney burn man

[austere pilot]

it is tho, MJ uses SD as base model

proof me wrong with sources

[smoky oak]

It's a lot more than that man

I can't deny the amount of work they've put into it

Midjourney is SD if stability actually gave a shot

*shit

Midjourney took the trash out of SD, but then limited it massively in other eays

[stark vine]

No controlnet, no, Boobies no pussy, a really half assed inpainting

[smoky oak]

I'm so passionate about this, I just want stable diffusion to do good. But instead they choose to fight the community every step of the way, and they're making their shit worse because of it

[stark vine]

Not to mention a discord-only interface lmAfo

Mj should only be used as training wheels for AI enthusiasts

[smoky oak]

If stability would swallow their pride in just admit that they are not the driving force of stable diffusion, but rather the community is, and they would come into the community and actively embrace feedback and resources from the rest of the community, stable diffusion would rapidly dominate

[stark vine]

It's not that they don't care about the community. It's just that they don't wanna get sued out of existence.

[smoky oak]

It's the reason why the multi-billion dollar dalle 2 looks like dog shit compared to a standard stable diffusion model.

[stark vine]

Blame the social justice warriors and all the snowflakes from twitter

[smoky oak]

No it's both of that, they actively choose to ignore the community

[dense tapir]

I sub him, but I am not being mean here but he is an idiot. He was telling people, flat out nonsense, and wasted so many people's time. I sort of went on a crusade to help them and I had to break it to them that his "special" numbers and saves were 100% BS. When I would bring anything up to him he refused to answer. Fact is I have yet to see him actually talk to anyone in his comment section. If he does it will probably be from someone licking his feet or something else.

[stark vine]

That's ironic giving that what you're doing right now is giving fake news.

He most certainly response to all of my comments. In fact, he was on the on unstable diffusion discord not too long ago, answering questions, and talking to his fans.

I don't wanna stan him but also examples would be nice to see what he's doing wrong because I would say he's like 85% accurate on the stuff that he teaches

[smoky oak]

You would be about 60% inaccurate lol

[dense tapir]

I said I had not seen it. Fake news. Go fuck yourself you ...

[dense tapir]

I use his vids now just to see what is popping but far more creators out there, like koiboi, that go into GREAT detail about subjects not clicky shit. Sad thing is in the age of the tiktoker the clicky bait wins and the ones with vary good info are "boring" and get little subs. Idiocracy.

[smoky oak]

I agree

@stark vine I would encourage you to go out and find information on these topics on your own, so you could see just how bad his info is, but you don't seem like the type of person to do your own research

Please, try and use his tutorials for TI's or LoRA's and let us know how it goes over, cause I can assure you it won't be well lol

He's not a bad person, I don't think, but he's far from accurate

[austere pilot]

wdym no controlnet and half assed inpainting??? and NSFW can be trained very easily with your own models with SD as base

[stark vine]

I'm talking about MJ lol

[austere pilot]

ah XD

[stark vine]

literally the only thing they have going for them is that badass model

I would suck dick to dreambooth that shit

[austere pilot]

just use SD model that is trained similarly as MJ and you're of

[austere pilot]

the base for that model is SD

[stark vine]

nah G v5 is like base 768 or 1024 at the very least

also has more params

see some of their gens, the amount of detail cannot be replicated with any sd model sadly

[austere pilot]

the base is SD but of course they added more to it

but they didn't built from scratch

[smoky oak]

Not with that attitude at least lol. I'm pretty confident I could match it

[stark vine]

let me grab some examples

[stark vine]

just off the top of their showcase

the tiny details are really hard to emulate in sd

the style yes but the stuff like the tiny creases in her gloves are too intricate for SD for the most part

You can do inpainting n stuff but you cant get it done in one go

[smoky oak]

Maybe it's just me, but that doesn't look that impressive to me, IDK

Feel like I've seen plenty of SD subreddit s that look better

[grizzled sage]

That looks like 100 other examples on civit

[smoky oak]

Right

Like sure, it's cool, but it still has issues all over the place lol

[stark vine]

another example of tiny details

those are pretty damn hard to get in SD

not to mention the custom resolution

[grizzled sage]

midjourney level images have been possible since before 2.x dropped. people just prompt poorly then blame the entire system

[smoky oak]

So by tiny details... You mean grain

[stark vine]

Nah look at the bubbles and the water detail

[smoky oak]

Bro, if you think SD can't do that, then you're high or not doing things right lmfao

[grizzled sage]

embeddings are total shit though so they won't help achieve these results right? even though midjourney is certainly using extra networks on their backend too

[stark vine]

nah embeddings are mostly a waste of time

I go dreambooth or go home

[grizzled sage]

of course! they wont help get that detail at all! 🙄

[smoky oak]

There is your problem lol. Too limited, too slow, too clunky, and you inherit all issues of the base models

[grizzled sage]

how do you even help someone who insists they know everything already?

[stark vine]

r0fl. So ur saying TIs > DB

[smoky oak]

Dreambooth only makes sense for huge data sets

[smoky oak]

For specific things absolutely lmfao

[grizzled sage]

embeddings are a useful tool in a lot of cases

[smoky oak]

Just sit and watch as they fail lol

[grizzled sage]

Those youtubers who told you TI is dead because long live DB, well, they were just hyping you up for click bait

they lied bro

[smoky oak]

LoRA's too lol

[stark vine]

I might have outdated info but embeddings only worked well on 2.x models

on 1.5 they were pretty poopy

[smoky oak]

What lmfao

[stark vine]

and I only do 1.5 so... it's pooh

yeh I ain't bullshittin

[smoky oak]

You are a work of art man. And I mean like first version of SD work of art

first you think pixel upscalers are better than diffusive upscalers, which they aren't

Then you start spouting off nonsense about MJ v5 being insane, when its still really unimpressive, and now you are shilling dreambooths and saying embeddings and other networks are worse than dreambooth... Come on man

Even if you were a broken clock, you would have been right at least once by now

[wispy nest]

You can do that with SD on higher resolutions and better.

but yea, MJ*knows more words and easier to prompt things

[stark vine]

I guess it depends on the case usage

Since I pretty much only do people, no Lora no TI nothing has even come close to me than the power of dreambooth

Dreambooth recreates people better than anything out there, I don't even bother combining db with a TI just because db gets it so right the first time

not to mention db has gotten buffed over and over again. I can train 2 subjects now with like 30 mins of processing

[wispy nest]

no idea about dreambooth

[stark vine]

ur missing out

[wispy nest]

on what?

[smoky oak]

I don't even want to waste anymore time with somebody so closed minded

[stark vine]

SD without db is like peanut butter without jelly

[wispy nest]

Something more exact?

[smoky oak]

Please don't fall for their over hyping of dreambooth

[stark vine]

essentially training ur own subjects lol

nah dont listen to the haters

dreambooth is flat out magic

[smoky oak]

no, its not lmfao

[stark vine]

idc if its 4gb a pop lol

worth the space

just imo tho. I have like 30 models I can vouch for

[wispy nest]

I don't do training anyway, my gpu won't train shit, so whatever

[smoky oak]

Dreambooth is only reasonable for huge datasets

[ripe cedar]

I think simple stylized stuff is really good for AI regardless, especially when making in quantity. Less detail, way easier to edit in PS/CSP, etc.

[stark vine]

do u have any loras or tis of specific subjects

[smoky oak]

Yeah, tons lmao

[stark vine]

cuz if u wanna put ur money where ur mouth is I'm game

[smoky oak]

wanna see my WIP Na'vi LoRA?

[stark vine]

whats that

share it anyways

[smoky oak]

Blue people from avatar

[stark vine]

ooh

[stark vine]

yeh share

[smoky oak]

oh wow, whats this?

a fantastic result with just a LoRA

😱

[stark vine]

rofl? An image doesnt say anything

Share the model

[smoky oak]

no, I am not sharing anything with you

You are not worth reaping my time and efforts when you are so misguided and obliviously wrong about all of this

do something for yourself for once

[stark vine]

I can literally prompt avatar high resolution and get that picture

let me see some custom shit

[smoky oak]

do it then, I'd love to see you try lmao

[stark vine]

no, I am not sharing anything with you

You are not worth reaping my time and efforts when you are so misguided and obliviously wrong about all of this

[smoky oak]

exactly, now keep your trap shut

[stark vine]

😆

[smoky oak]

I don't think any of us want you here

[glossy herald]

maybe some cows, people like cows

[smoky oak]

please, just stop feeding people false info man

[stark vine]

how come u have a cat then?

I love cows. I want a cat sized cow.

[glossy herald]

I have multiple cats, I am a cat, but I can still love cows. I don't see any incompatibilty there

[smoky oak]

There should be a rule or something in this server against people spouting straight up nonsense, ISTG

[stark vine]

everyone here has different opinions

I think there is no one real answer to anything

[smoky oak]

then we wouldn't have to deal with people who mooch off others while feeding people misinformation

I am fine with people being wrong on their own cause its funny, but when it draws back others, it sucks

[glossy herald]

yeah I'll be honest there

[smoky oak]

So many of us spend dozens if not hundreds of hours to work on this shit, just to have a troll tell people its bad/useless, while also asking is for help? lmao

Making it really hard to want to give back to the community rather than gatekeep

[glossy herald]

I don't like false info, and I'm not sure we do have a rule about that specifically, mostly sure not.
But I also didn't look up the whole chat and I frankly don't want to start and say who's right and wrong

why not just stop there when you disagree that much and feel the other person isn't genuine ? just block them and move on ?
don't harm yourself in the process I would say mostly.

Like the famous cow story says :

Once upon a time, there was a false fact that claimed cows could predict the future. According to this false story, cows would line up in a specific order before a big storm, and their positioning would indicate the severity of the coming weather.

People started to believe in this false fact and would consult cows before making important decisions or planning outdoor events. However, as time passed, it became clear that the cow predictions were not always accurate, and people began to question the validity of the cow oracle.

Despite this, some people felt obligated to continue discussing the cow predictions, even though they didn't really believe in them. They would argue over the accuracy of the predictions and become increasingly heated and argumentative.

Eventually, the discussions turned into full-blown arguments, with people berating each other and refusing to listen to any opposing viewpoints. It became clear that the false cow prediction had caused more harm than good, and people decided to put an end to it once and for all.

and if you feel there is really dangerous info going around in there, I'm all for checking it on our side too, and in those case, using the ⚠️ reaction or "report to staff" application let us do that exactly

[wispy nest]

kill da cow, kill da cow

[smoky oak]

Because I would rather people receive help over hindrance. I spend HOURS working on this stuff, just to be questioned by somebody whos not even willing to work for their own stuff. Its depressing, and it brings down the collective of the server.

I feel very strongly that people who routinely give out bad/false information should receive some sort of role that indicates that they are not exactly reliable as a source of information

[grizzled sage]

i'm more of a cat person than a dog person. dogs are needy and need that constant "what next boss?" where as cat's just bug you when they need something and figure the rest of their business out on their own

also, sometimes, i'll use dreamboothed models AND textual inversions at the same time 😮

[smoky oak]

how are we supposed to give back and support this community when people are actively trying to spread misinformation about important topics. Its frustrating and demoralizing

I'm not even talking about anybody specific anymore, I should clarify

I just mean in general now

Sorry if I get/got cross. I am passionate about helping people not hindering them

[glossy herald]

that's a very logical and well positioned argument.
About the first part, I reiterate, don't let yourself feel berated/diminished when you know your stuff. but yeah, easier said than done.
I'm not sure we should tag bad advice givers, but I get your point though...

and yeah, I feel you on the misinformation being so depressing. I'm fighting against scams and trolls half of the day, so I can see how much we get too.
I'm also quite tired and I can't say I'm in the best position to analyze correctly the complete situation, so I'll take the time I need on that and chat with the mod team too, I get that this is a real problem that does impact you and others.

and no, don't feel sorry for spitting your guts. but do preserve yourself, call on us if you feel you are getting in a position of danger too.

[grizzled sage]

i'm reminded of how i was having problems and someone insisted i swap to use an older python for many reasons. 3.10.6 i think it was. insisted like they had good reason that it would cause instabilities with pytorch.

after i solved that initial problem for other reasons, i kept the older python and couldn't ever figure out why i had so many issues with NAN errors. a month later i updated python to a newer 3.10 and all my woes faded. once i started taking the initiative and finally learning more about python instead of just doing what i could based on advice, i came to find that there is no reason to recommend the older one other than it being in a wiki on one unupdated page. all the "it causes so many errors" was just made up to sound expert i feel.

i've come to stop trusting advice that lacks explanation and beat my own path since then.

[smoky oak]

Thank you for remaining level headed, I appreciate it and I will try to as well

Its hard for me to keep my cool around people who so confidently spew out nonsense like its a game. Somebody who complains about why their stuff isn't good/doesn't work, demands other people help them, shits on their work/advice cause they don't listen to it, and then from there on preaches gospel that its all bad.

Like why would you ask how to do better and then write everything off as bad before shilling the things that are making things worse for you.

I give so much back to this community, and I just want it to be a place where people can learn, and ask stupid questions without being fed misinformation or be made fun of.

I wish I had some better form of way to share my/other peoples genuinely amazing information on here for people to see, rather than an echo chamber of people being hostile and downing each other, which I realize I am now a part of because of my actions, but it comes from a place of caring for this tool, community, and the betterment of AI as a whole

Sorry to bombard you with my great wall of text, I am just really feeling tired of people trying to make things actively worse for others while telling them to not listen to the people who are actually dedicating massive amounts of time to the betterment of this tool as a whole

Also, felt the need to post this in response to some false claims about MJ V5 outdoing SD in terms of detail. no inpainting, no editing, straight out of SD (cropped cause she got a little nippy lower down lol)

I think its pretty clear that stable diffusion is still winning

[grizzled sage]

very very high quality images. but because of ||nitpicked subjective taste issues|| it's not as good

[glossy herald]

I'll make party cat come out

[smoky oak]

Party cat?

[glossy herald]

but yeah, hard to say A is better than B imo.

[smoky oak]

Its a mix of skill issue, lack of experience, and various other things

[smoky oak]

I can agree with colors, but the argument was made that there is no way SD could match the detail level of MJ V5, which is just... clearly false lol

i do agree the colors on MJV5 are better, but you can color grade the SD output as well

[grizzled sage]

i certainly have some metaphors to describe this aspect of internet culture. It's nothing new of course. disrupting discourse has been something people done for fun since usenet

[wispy nest]

color grade is different from model to model too, idk if mj is better there

[smoky oak]

I don't care for people in this community trying to put down... well, this community

People in here trying to discourage people saying competitors are doin better, and discouraging them from even trying

[grizzled sage]

"you can't"

[glossy herald]

it's not a fair fight to said "SD" of "MJ" already. like what SD? the base brick, or the extended feature set with custom model and style randomization ?
what MJ ? the black box that we imagine is just a model, or a more detailed pipeline with models, prompt interpreter, hypernetworks or embeddings ?

And also for what "better" result ? Can I get my very simple photo of a cat in MJ or will I be obligated to have lots of artsy artefact around it even if I don't want ?

Here take another party cat

[smoky oak]

I put maybe 5 minutes of work into that example. I could easily make something WAY better

[smoky oak]

haha, this one is dope lol

[grizzled sage]

I think some people do want MJ though. A black box you just yell at and it does stuff. But then the best MJ users get into advanced API stuff with it instead. So it's a mixed bag still

there's just a lot of "you can't" all the time

[pseudo aspen]

One more Chester. Enhanced the prompt a little

[wispy nest]

we need more neon lights

[smoky oak]

I agree, I am not saying MJ isn't good, cause it is, and it has its own use case and user base that it serves very well. But to actively put down a community you are a part of is just so toxic, OMG

[grizzled sage]

thinking MJ is better is cool but i do think there's a bit of skill level difference with more advanced tools and a magical black box with presets

[smoky oak]

If you wanna just have a fast, simple, and fairly good looking output with no hassle or additional work, by all means, use MJ

[glossy herald]

for sure, lots and lots of people do want MJ. it's fun, it's good, it's effective.
Lots of people need the more versatile version that is SD though.
Like any professional that would need to train its product in their model.

[smoky oak]

exactly for sure. A lot of us here are high level/advanced SD users, myself included. Its not unrealistic to expect that MJ will get better results for the average person over SD. But the part I hate is the idea that "SD could never"

when it quite literally does lol

[pseudo aspen]

Imagine if Stability.ai and midjourney worked together

[grizzled sage]

Could buy that mall kiosk pocket knife or get a proper tool that Mick Dundee would carry

[glossy herald]

(this is #🏞｜general-with-images people, don't forget your pictures if you want to make an argument, temporary rule)

[grizzled sage]

mall knives are cool and useful too

[smoky oak]

Midjourney is too money hungry to ever consider doing something for the betterment of people around them lmao

[grizzled sage]

maybe crocodile dundee is a bad analogy

[pseudo aspen]

Unfortuanetly

[grizzled sage]

Stability acquires MJ, now they will

[pseudo aspen]

Still think SD is better thanks to community who makes extensions, models and modifications for it.

[grizzled sage]

that party cat is hype

[glossy herald]

high as a cat mostly yeah :p

[sterile kiln]

rip

[smoky oak]

Stable diffusion is infinitely more powerful / diverse than mid journey is, simply off the fact that mid-Journey is so closed down and censored, and they share absolutely nothing for the betterment of AI for the masses

[grizzled sage]

inpainting. boom. done. i've always said that.

they automate some common remix tricks too like masking out a subject, rising or mirroring it, etc..

probably have a script that randomizes a few subject removal and positioning tricks

i suspect they have other tricks for the way they blend two images that might be more advanced and possibly novel

i'm thinking something about the way it feeds both images into the noise alternating. its a neat looking mode. not sure what they're doing with it tbh

here's the big example google pulls up

[wispy nest]

too lazy to gen more at higher resolution or fixing hands, but that's my contribution on that 😄
Light prompt too

[grizzled sage]

i think now a days with controlnet and some jiggering between clipvision and color inputs, you could blend two images. i haven't played around on those fronts yet

[smoky oak]

See! Its not hard to mimic/surpass the detail level of SD haha

[wispy nest]

His image is 900x2k or something, I mean...what kind of detail you expect

[grizzled sage]

MJ just isn't worth what you pay for it. do the math and buy a SD capable machine instead of a subscription

[wispy nest]

my thing in 768x768

[sterile kiln]

(https://www.youtube.com/watch?v=YLHpvjrFpe0)

[wispy nest]

That's weird upscale, noisy

[wispy nest]

I asked before...what model are you using for it?
Was looking for realistic model which can do magic, already have one, but another option would be nice too

[dense tapir]

I use a couple but I mostly use the one @smoky oak uses the realistic 1.4

[sterile kiln]

the image is not optimal because of the lyrics, probably. But I've seen/done much worse.

[wispy nest]

lyrics?

[sterile kiln]

see EXIF on pict, and link on top of image

[pseudo aspen]

One of my favourites. ❤️‍🔥

[wispy nest]

oh...I mean video upscale is weird, I can see white\ gray dots everywhere on 4k

[sterile kiln]

ah, have lossless album on my comp, don't need listen it on YT, but yeah

[wispy nest]

They are very clear on white backgrounds

I thought video is whole point first 😄

[sterile kiln]

it can only be an upscaling, making videos in 1080p at that time was a dream

[wispy nest]

yea

[smoky oak]

@wispy nestI use Realistic vision 1.3/1.4 99.9% of the time

Still so glad the creator got off his high horse and gave back to the community

[pseudo aspen]

RealisticVision for the win

[wispy nest]

underwater water drops lol
Spongebob kind of thing

[smoky oak]

only now that he made the model available

[smoky oak]

I vibe haha

[dense tapir]

50 steps instead of 20

[dense tapir]

heh, the foot is backwards

the whole bottom leg too

[wispy nest]

yea

[dense tapir]

SD doesn't know right from left

[wispy nest]

hands and feet are rarely good

especially on higher resolution gens

[sonic crane]

jus a lil gif

[dense tapir]

Dang, the nose 😦

[smoky oak]

theno ose
o o
o o

[dense tapir]

Yeah, wth? lol

Otherwise it was great

Well

Maybe whale

[smoky oak]

oh, the double horizon mirroring with the cloud bottom looks reallyyy nice

I have been meaning to see if I can train my first style LoRA

I would share on what style, but some people here

like to steal ideas before people can make them

So I will keep you all updated with progress, assuming all things go well

[grizzled sage]

thats totally me with the DF model i want to work towards. sorry. i'll get it done first though really! i swear

[fresh sparrow]

A stunningly beautiful girl. Cinematic in nature, this hyper-detailed scene is filled with insane details and beautifully color graded using Unreal Engine. The use of DOF, Super-Resolution, Megapixel, Cinematic Lightning, Anti-Aliasing, FKAA, TXAA, RTX, SSAO, Post Processing, Post Production, Tone Mapping, CGI, VFX, and SFX has created an insanely detailed and intricate world. The hyper maximalist approach and hyper realistic Volumetric and Photorealistic rendering bring out the ultra photoreal and ultra-detailed aspects of the scene. With 8K and super detailed visuals, this scene bursts with full color and Volumetric lightning, using HDR to create a realistic and breathtaking environment. Powered by Unreal Engine and rendered in 16K with sharp focus, the intricate details of this scene are truly mesmerizing.

[smoky oak]

@green plover @dense tapir Ok, so adobe firefly is awful lmfao

[sterile kiln]

I take it as a challenge!

[smoky oak]

I am happy to see that Adobe Firefly is nowhere near as good as SD or Midjourney, and is in fact quite horrible lmao

[green plover]

And if it says Adobe on the box it probably comes with a price tag. Probably a monthly one

[smoky oak]

yeah

have you seen the un-cherry picked results? lmao

its fucking dreadful lol

[errant isle]

excuse me, how do i use my seed to recreate the same photo
i dont seem be able to enter the seed anywhere

[smoky oak]

Somebody just released a comparison of firefly vs SD, and its fucking HILARIOUS

[green plover]

Directly below CFG scale

[green plover]

@smoky oak you see what I've been doing with SD and iClone?

[errant isle]

am i using a wrong version?

[smoky oak]

oh, this is not normal SD, best of luck

[green plover]

What interface is that?

[errant isle]

https://beta.dreamstudio.ai/dream

[smoky oak]

yeah

[green plover]

ask in #1025467151206854736

[errant isle]

thank you!

[errant isle]

which url do you generate your prompts?

[green plover]

I use automatic1111. it's a local install

[errant isle]

Is there a link for me to get it

[green plover]

one sec...

[errant isle]

Thank you DF

[green plover]

https://github.com/AUTOMATIC1111/stable-diffusion-webui

Sorry for the delay. I'm baking a couple normal maps. Slows down my system

[errant isle]

Thank you!

[green plover]

NP

[wispy nest]

This prompt breaks the record of max meaningless for AI words I've seen so far 😄

hm...
idk what happenned to gravestones there, but I still like that image

[smoky oak]

LMFAO

[wispy nest]

oh...I think I know what happened lol...
I put "grave stone" into prompt as 2 seperate words

although other gen was fine...hm

kinda, it added just "stones" too lol

[stark vine]

Besides the classic photosh00p does anyone know an easy way to inpaint a specific object? I wanted a rusty wooden cross on the blank space to the left but inpainting even at max denoise doesnt work. I remember there being an easy way but I already forgor

Also, after further experimentation I can definitely vouch for the ultimate upscaler but only for far away subjects like my girl. I would say it is slightly superior to controlnetting the original image and regenerating at a higher resolution

Then inpaint the artifacts and badaboom, no more warped faces on far subjects, no need for facefix

For closeups, the artifacts become too much of a pain to sh00p out, classic upscaling is preferable until we get a dope 1024 model. That's my two cents on the extension

[wispy nest]

I didn't understand anything

We have high res models, both 768x768 and 1024x1024

[stark vine]

Which 1024

[wispy nest]

well...not high res, but higher then base 1.5

I don't remember , there was few

[stark vine]

This one? https://civitai.com/models/20842/char

People have recommended it but it seems to be pretty inflexible

[wispy nest]

I'm currently genning with this one on 1024x768, but idk if it's actually trained on that high

[wispy nest]

never seen this one

[stark vine]

did you get it from civitai?

I havent even tried newer models since RV1.3 lol

Just because the blend I made works so well, and db models arent transferrable between models yet sadly

[wispy nest]

yea

[stark vine]

also does anyone know if training a lora with 1024x1024 would work on 24gb vram?

[wispy nest]

There was other models which was saying they are trained on 768 or 1024, I just don't remember even all models I have...

[smoky oak]

It would work, but it would probably be useless

[stark vine]

well he says there are 1024 models out there

so it should in theory work with those

[smoky oak]

Unless they are all 100% 1024, then you're gonna have artifacts and distortions all over the place sadly

[stark vine]

Yeh I figured, thats why I havent messed with em yet

[smoky oak]

That's why the 768 SDv2 sucks ass

It was trained on 512 most of the way, and 768 at the end

[stark vine]

wut

[smoky oak]

It led to some very weird issues, like cars being super stretched and flat, and people having abnormally long necks

[stark vine]

so it isnt even native 768?

so its even worse than I thought lamo

[smoky oak]

Parts of it are, but not consistently

[grizzled sage]

768 is the 512 v2 model, further refined with 768

[stark vine]

yeh idk why so many ppl advocate that crap model

[smoky oak]

There is a full 768 prune that somebody did, however the results out of it are terrible, just like all stable diffusion base models

[stark vine]

even their custom models I dont touch with a 10 foot pole

[grizzled sage]

better clip knowledge

[smoky oak]

The base models aren't meant to be good, they're just meant to be diverse. They give us the base models to train into several focused high quality models

All of the base models are basically raw data with no proper direction

[grizzled sage]

i've seen some 1.5 models that were refined on 768 too. while it's not an ideal 768 generation result, these extended models do see improvements

[smoky oak]

Jack of all trades, Master of none as everybody likes to say

[stark vine]

except 2.x isnt diverse at all

it got nerfed to oblivion on so many things

[smoky oak]

2.x is very diverse, however it uses a completely different text encoder, which makes it extremely hard to get specific ethnicities or poses out of it

[stark vine]

if you cant make b👀bies with it it's garbage

[grizzled sage]

yeah that philosophy is important. people keep treating the base model like it should be hyper stylized midjourney

[smoky oak]

Stable diffusion V2 should be significantly better than stable diffusion v 1.5, but sadly they introduced an absolutely abysmal text encoder that makes it nearly impossible to train, and then on top of that they decided to withhold information from the community

That is why you can get significantly better results out of well-trained stable diffusion 1.5 models

[stark vine]

they also explicitly said they removed nsfw because of kids being in the model

so I know for a fact it's adulterated

[smoky oak]

They went back on that and stable diffusion 2.1 I believe

I'm pretty sure 2.0 had less celebrities and less NSFW, and then for 2.1 they brought it back after backlash

[stark vine]

the question is, who will be the brave adventurer to train a native 1024 from scratch with as much variation as 1.5?

[smoky oak]

Don't quote me on that, I know a lot of things, but I'm not certain on that one lol

[stark vine]

nah both stink

[smoky oak]

That would cost millions of dollars

[grizzled sage]

2.1 i thought was extended off 2.0 and not a fresh start

[smoky oak]

However, let's just say that @green plover And I have some extremely lofty goals of retraining stable diffusion 1.5 all over again using completely free open source and non-intringing photographs, while also improving stylistic understanding and consistency

*non infringing

I don't want to explain the exact process here, as we are both slowly working out the details in private, and we don't want somebody else to mooch off of our work

[pastel mango]

any ideas why my auto1111 suddenly looks like this? I tried a full clean reinstall and all

[stark vine]

sounds like a bad idea

[smoky oak]

Oh Jesus Christ

[stark vine]

2.x from what a lot of people say is already trained on copyright-free image

[smoky oak]

That may win for the most curse a1111 installation I've already seen

[stark vine]

and errubody knows what a steaming pile of pooh that turned out

[pastel mango]

but its not an installation yet since I can't get it to actually run

[smoky oak]

The problem with stable diffusion 2.x is not in its data set, it's in its text encoder

[pastel mango]

I have before, worked fine

many many times

suddenly now I have this UI crap

[smoky oak]

Stable diffusion 2.x should theoretically be better than stable diffusion 1.5 in every single way, however it's using a significantly worse text encoder, which makes it impossible to unlock the potential

And to put that into perspective, a lot of people like to bring up how much better mid journey is at delivering on what you ask for then stable diffusion. That all comes down to their text encoder. It's better than stable diffusions

[pastel mango]

2.0 should be bit isnt, in their effort to make it PG13 they screwed something up

[smoky oak]

It has absolutely nothing to do with them making a PG-13

[stark vine]

it does

its contested at least

no one has verified it

[wispy nest]

Just for experiment, have you tried other browser?

[smoky oak]

The failure of stable diffusion 2.x has nothing to do with the data set. It's entirely the encoder

This is a factual statement, that is the exact reason why control knit doesn't work on it

[pastel mango]

yeaeh I also cleared cache and cookies

[smoky oak]

In fact the developers of control net themselves quite literally told stability AI that they need to get their head out of their ass and fix it

[wispy nest]

yea, that was my thought...

[fiery lichen]

inpaint that hole in the middle of the suit into heart shape please?

[wispy nest]

I should try to restore my kaggle account...my google collab doesn't work anymore, genning on my gpu kinda sucks

[pastel mango]

I managed to add a custom ui after strugling and I think it partly fixed it

very partly

no actually not at all

[wispy nest]

maybe it's one of extentions which adds something to UI?

[fiery lichen]

which colab does not work? lastbens'?

[smoky oak]

What GPU? I may be able to speed it up drastically depending

[pastel mango]

but I said I did a clean install

[smoky oak]

Define your clean install, please

[pastel mango]

nothing there

vanilla

[wispy nest]

I don't think anyone but me using it here lol
it's sagiodev's , at least it was broken few days ago, idk , maybe fixed now

[smoky oak]

as in what? The entire folder, the venv, the original cached files?

[wispy nest]

Doubt it, it's 1050ti lol

[sterile kiln]

Haha, i could

[fiery lichen]

did you update python recently?

[pastel mango]

no

[smoky oak]

oh god, you can even get it to start with that?

that is impressive lmao

2gb VRAM

wow

I am genuinely amazed it even starts!

[pastel mango]

unless it was through a venv but that would be localized

[wispy nest]

yea ,I have no problems with it, it's just super slow

[wispy nest]

it's 4gb vram

[smoky oak]

That is insane. I guess Xformers is working magic

[smoky oak]

wait really?

just a sec

well noooo shit

it is

damn

[pastel mango]

Imma erase venv from this auto build and rerun it and see what it does

[smoky oak]

uno-momento

[smoky oak]

best of luck

you may need to delete the cached files as well

[pastel mango]

thank you

where are those at?

[smoky oak]

that way it direct downloads all of the components, incase one of the cached files is messed up

[smoky oak]

I am not sure sadly, though I might know who can help

[wispy nest]

yea I'm out of ideas on that one...except fixing css manually and saving it somewhere or using plugins which adds custom css to specific sites

[smoky oak]

I am very new to all of this venv stuff

[pastel mango]

Im just gonna do a whole new reinstall

[smoky oak]

@dense tapir Are you here by chance? If so, do you know how to delete the cached files used to make a venv?

[pastel mango]

but still need to know cache

[smoky oak]

Let me see if I can find that

ok, ok

so it looks like you can add --no-cache-dir

DO NOT QUOTE ME

that should stop it from checking for previously cached installations

[wispy nest]

why do I like weird images sometimes more than actually decent ones

[pastel mango]

I still want to delete them though cause they'd be using up a shitload of space

[smoky oak]

Its a common thing. I tend to like some AI wonkiness haha

[smoky oak]

alright, let me see if I can find info

[pastel mango]

I'm trying too

thank you

[smoky oak]

no problem. Love learning new things, and I hope I can help too haha

@pastel mangoDo you know what version of pip you are on?

[pastel mango]

pip 23.0.1 from G:\Programs\Python\lib\site-packages\pip (python 3.10)

[smoky oak]

ok, so it looks like this command should work

pip cache purge

it should get rid of all cached/wheel files

https://stackoverflow.com/questions/9510474/pip-uses-incorrect-cached-package-version-instead-of-the-user-specified-version/61762308#61762308

[green plover]

I love it how people have to chime in with whether or not someone else's project is a good idea. When you take on a project, you can decide on the viability of it

[smoky oak]

have a look for yourself if you would like

[smoky oak]

right

[pastel mango]

ERROR: pip cache commands can not function since cache is disabled. apparently

[smoky oak]

hmmm

I am in uncharted territory sadly

I know 0 about pip or python. I am just a very fast learner and I am good with google haha

so when you make a new VENV, does it say its using cached files?

[pastel mango]

I dont know

gonna run an update on my gpu drivers first, before I do anything else

safer to do it first

[wispy nest]

%LocalAppData%\pip\Cache

• pip cache location on windows

[pastel mango]

ah thank y ou

yeah there's 9gb there

[smoky oak]

yeah

[pastel mango]

doing an install with studio drivers for nvidia for more vram optimization

[smoky oak]

the command I gave you was supposed to clear that file directory

hmmm... weird

[pastel mango]

yeh

im gonna try things from this point and see what happens

[smoky oak]

let us know please

hope it works

[pastel mango]

meanwhile I will make a comedy conversation by people who don't exist, with people who don't exist

[static tusk]

that pip purge command worked for me 👍

had 7gb, now 2kb

[smoky oak]

yay :>

glad to hear!

[smoky oak]

relatable

[static tusk]

weird how it doesnt have auto cleanup

[smoky oak]

its actually quite nice for some things

[wispy nest]

oh, that's nice

[pastel mango]

ah got most down now

[errant isle]

do i need to download anything else to be able to use this?

[pastel mango]

The UI is normal again @smoky oak

[smoky oak]

Glad to hear :>

Nothing worse than a mistake you can't fix

[pastel mango]

well there wasnt really a mistake, it just sorta worked when I went to bed, then it didnt

it might have had to do with an extension or a python update

[green plover]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs Here's the install instructions

[errant isle]

oh dear... looks complicated for a non-tech person.. and ill probably need a good hardware

these kind of things can be run on laptops? or not powerful enough. im not going to do too crazy things

[pastel mango]

damn I lost what extensions I was using though

some toolkit, some additional network, and the rest I think I got handled

actually got additional now too

ok got model toolkit too, yay

[green plover]

Depends on what GPU you have.

[simple fossil]

👍

[smoky oak]

I'm sorry, I had to share

My mom and I (mostly her) spent a ton of time working on her spice collection, and I wanted to show the fruits of our effort

These are all of the ones that can fit inside the drawer. The rest are stashed in our garage pantry

I'm just glad that she uses them all. My mom cooks authentic Chinese, Japanese, Korean, Vietnamese, Cajun, southern soul, Indian, Mexican, Italian, you name it, she cooks it lol

[young solar]

I find your lack of cloves disturbing.

[smoky oak]

She goes to every extreme length she can to go to the most authentic places, and talk to people of native culture in order to get recipe straight from the source

[smoky oak]

Not all of the spices are in yet, we had to throw out our clothes because the original bottle broke in moving

All of the empty ones ended up shattering and we had to get replacements

[young solar]

Fair enough. Otherwise quite cool.

Though I would suggest grating/grinding nutmeg fresh. It is easy to do and has a much better flavor than pre-ground.

[smoky oak]

All the extra ones she doesn't normally use are either here or in the garage pantry lol

[smoky oak]

She typically does that as well, most of our herbs are freshly grown as well

She just has all the stuff for all the reasons lol

She takes pride in her collection haha. Most people try to give her shit for having tasteless food just because she has light skin lmao

[young solar]

Fresh herbs are the best. I need to set up an indoor garden for them.

[smoky oak]

She is actually out buying a whole new harvest of plants for the season. Now that we are in our new house we are going to be building our own corrugated metal freestanding planters with a water irrigation system in our backyard. We live in the middle of the desert, but this section of the desert is shaded quite well by a large mountain behind us

I also do aquarium keeping and aquarium plant growth, so all of my residual fast-growing plants can be decomposed into fertilizer for her plants. It really is quite self-sufficient

[young solar]

It would be pretty crazy to set up, but you should look at aquaponics. I saw a home setup a few months ago, and it was quite the thing. Especially if you like eating fish.

I love that herbs are easy to grow, though. Managed to keep some basil in just an apartment window and was having to figure out recipes to use it because it grew so fast.

[stark vine]

BRO NO FUCKING WAAAAAAAAAAAAAAAAAY

@fiery lichen I just came up with the most insane method of producing high detailed images with the ultimate upscaler

with minimal photoshop involved. this is insane

check out the examples. sorry that I had to censor them lamo. Left is before, right is after.

Check out the before and check out the after in open browser. Night and day. No artifacts. Not just the face look at the right hand

Video TuT coming soon? 😏

[smoky oak]

my aquariums are all nano tanks (sub 20 gallons) so there is no fish eating for us, and probably not enough nutrients to grow plants. But my floating plants grow like weeds (I may or may not have 11 species of floating plants that I shouldn't lol)

And by like weeds, I mean more than double in volume every 5 days. So I can grow an insane amount and use them as compost to fertilize our garden. thats the goal at least

[stark vine]

I'm officially addicted to this shit.

[smoky oak]

Good. Glad you are experimenting for yourself

Glad you decided to give it a shot after all

[static tusk]

I think you owe @smoky oak a thank you

[fiery lichen]

good work. a lil write up would be good, vid would be great.

[smoky oak]

[stark vine]

Yeh I need to play with it even more but yes I'm glad you introduced me to this tool cuz I initially thought it was shit

If I can make it work even better than I did with that example ima drop a tut

it just takes a lot of gpu power and sadly the extension is glitched, it doesnt generate batches

[smoky oak]

I am assuming you did multiple upscales on the same image back to back?

[stark vine]

Yeh but it only saves one 😦

and boy did I waste time figuring that out

[smoky oak]

As in you are batch upscaling, or what? I may be able to fix it for you @stark vine

[magic raptor]

give logo of id:home unknown name

[stark vine]

No I just need several different upscales of the same image

If you can fix the code that would be sweet

otherwise I have a workaround. I might also dig in there to see if its an easy fix

I might even get greedy and find a way to use this with canny I aint even bullshittin dawg

[smoky oak]

I have used it with canny before

I used it with depth as well

[stark vine]

nah it cant work with canny cuz it doesnt split the mask

if it did tho, it would work with higher denoising

cuz it would stop it from generating garbage

it may seems like it works at lower denoising because it doesnt change much

[smoky oak]

@stark vineCanny can work with it, I can prove it right now

you just have to do it differently

you have to save the canny image from the annator and then feed it back in

You can see I saved this canny map

and you can see the hair in the water

it does work. Only issue is they need to be the same base res I believe, cause you can see mine is not lined up. Which shouldn't be too hard for the same image

gonna see if I can get a better result with it on the base image itself

aha, I see

so its hard cause the canny map res needs to be the same res as the final output

which I might be able to do, just a sec

[stark vine]

wait wut

what am I supposed to be seeing here

[smoky oak]

ok, I was able to get canny to work is what I am saying

[stark vine]

the upscaler works with the image in chunks

[smoky oak]

I used the canny of one image while generating a different one

[stark vine]

yeah but that doesnt make sense

[smoky oak]

yes, and it chops up the canny map as well

[stark vine]

lol let me try it

[smoky oak]

but thats how it works

[stark vine]

im sure it doesnt

[smoky oak]

look above

or well rather, no it doesn't

sorry, I mispoke

you need the canny map to be the same res as the final output, then it lines up

[stark vine]

let me try

[smoky oak]

ok, I was able to get it working and properly lined up with a high res output. I am trying it on a higher scale now

an actually good image, i mean

[stark vine]

how do you get the canny map to be the upscaled resolution

[smoky oak]

you have to do math and trick the canny exporter

[stark vine]

just regular upscale it as an image?

[smoky oak]

you figure out the res of the image you are upscaling it, and then multiply the smallest value by the amount you are upscaling

then you take that result and make that the res of the annotator

then you preview it, save it, and feed it in with no pre process

I am about to do it

ok, it seems like it may be detrimental to the process

I am about to see if thats the case

@stark vineCanny makes the results wayyy worse, OMG

[stark vine]

no shit lol

its cuz it doesnt work

but let me try

I want to be wrong here lol

[smoky oak]

it does work, just not beneficially

Low res base

normal 2x upscale

2x upscale with canny

Here is the canny map

every other setting is 100% identical

same seed, everything

so canny does clearly work here, but it is not beneficial to the process

the issue is with the current implementation, you have to do a canny process on an image that is 4x lower res than the canny map itself, so I am sure that makes a lot of noise

[stark vine]

o shit

its working !?

[smoky oak]

yes, it is

I already told you that haha

[stark vine]

let me see what the results are

[smoky oak]

you just can't let it process the canny live

I... just showed them lmao

they are above

it makes it look worse

[stark vine]

but that doesnt make sense, why would it

all the canny does is tell it not to color outside the lines l0l

damn still getting mad artifacts

7 min till the final thing is complete tho

I'll be the judge to see if there's an improvement

[smoky oak]

the reason its not working is its having to draw a canny map at 4x the res of the actual image its looking at

so its probably very noisy (I could also have guidance too high, we are in uncharted waters after all)

I have used depth with ultimate upscale as well

[stark vine]

yeah I can already tell from the previews its way worse

well sadly I was right lol. sort of. I mean it can process it it just doesnt do what we expect it to do

[smoky oak]

Yeah, we were both right

it does "work" but it doesn't benefit

so it works, but it doesn't... work lol

[stark vine]

did you try setting it at 2 weight

I have it at 1

[smoky oak]

its at 1.5 weight in mine

I am proving that it does "work" by putting one of the other canny maps ontop

[stark vine]

it just still doesnt do what its supposed to

is the map supposed to be black lines with white bg or viceversa

[smoky oak]

yes

ok, its very hard to see here, but there is a different canny map ontop of this image

this is the canny on top

you can see the hair waves on her left shoulder

[stark vine]

wait thats black bg white lines

[smoky oak]

here, so it is clearly applying, but its now working how we wanted

[stark vine]

I gotta try this at 2 weight, I just got greedy with a 4x upscale lolol

ugh I wanna see the final result already

[smoky oak]

oh, did you set the annotator res to 4x the base image res? if not, its not gonna line up

[stark vine]

yeh I upscaled the canny map and set the rez to the same proportion

but Im still confused why it shows black bg white lines

[smoky oak]

oh, you upscaled it. Curious to see how that works

[stark vine]

on the preview yet the output its inverse

[smoky oak]

oh waiitttt

[stark vine]

and I have no way of knowing if its reading it correctly cuz it looks like ass as it usually does

[smoky oak]

let me try it with inverted colors to see if it works

[stark vine]

yeh I have no idea which colors its supposed to be

[smoky oak]

trying inverted, weight 2x

[stark vine]

yaaas

[smoky oak]

I am seeing both online, so IDK

only one way to find out lol

fuck around and find out time!

My favorite heh

[static tusk]

canny is white lines, scribble is black lines

[smoky oak]

ok, so canny IS supposed to be black with white lines?

[stark vine]

ooh

im using the opposite

[smoky oak]

well fuck it, I am trying it backwards lol

I already have one

we also need to try low guidance. It could be too strong

[stark vine]

no the issue is that its not guiding it well enough

[smoky oak]

and that could be messing things up. Maybe it just needs a subtle push

[stark vine]

if it did we wouldnt have bs artifacts

[smoky oak]

OH OH!!!

INVERTED WORKED WAY BETTER

[stark vine]

aaand the final result is dogshite.

so u did black lines white bg?

[smoky oak]

bro, inverted worked way better

[smoky oak]

yeah

just a sec

[stark vine]

I gotta try that

[smoky oak]

its still not the bestest, but its a lottt better

I think the weight is too high now

not inverted

inverted

she has teeth again, and her eyes are less fucked

alright, trying some weight variations

.5 first

[static tusk]

its better because it had no annotator data to use

it was inverted

[smoky oak]

oh man, if thats the case lmao

that could be the case lmao

however, the disabled version looks way different (identical settings) than the inverted one

this is the disabled one

[stark vine]

nah its still ass for me

[smoky oak]

first is off, second is on

off looks better, but inverted is still doing something

[stark vine]

hmmmmmmmmmmm

let me try with my method but I aint optimistic cuz right off the bat I saw artifacts

[smoky oak]

ok, this is off vs 0.5

[stark vine]

o shit hold up

it kinda does work?

I wont say anything till the final image

[smoky oak]

kinda sorta, I think

off vs 0.5 on

actually, they look really close

only difference seems to be in the hair

[stark vine]

ok this is strange.

I can see the white lines on the output

looks like absolute dogshite of course, but its weird that those lines are there

let me try lower weight actually

[stark vine]

nah still a bunch of horseshit artifacts. what a disappointment.

I guess if you think about it, even if we can get the proper map it still wouldn't work, let's say it's just working on a chunk of hair

[smoky oak]

what in the heeeeeeeeeeell happened here

[stark vine]

the prompt still says "photo of xyz high definition face"

[smoky oak]

mine look worse, but not that bad, jeez

[stark vine]

its gonna try to apply that prompt to the hair alone whether canny is there or not which is gonna look like bullshit lol

nah the solution is native 1024 model trolol

or my method 😏

I'll try to do a good prompt tomorrow with my method, if u can fix the code that would be pretty neato mosquito

im too lazy to try to fix it myself cuz i suspect its not an easy fix

[static tusk]

when you upscale dont use any descriptive words for the subject. Use hi-res, masterpiece, etc

[stark vine]

well im trying to replicate the exact same image

So it looks as similar to the original as possible

with more detail of course

[smoky oak]

weird, I tested that for a while and found it yielded way worse results

I do suppose I have a better setup now. I could try it

[stark vine]

you'll waste your time

if the image doesnt match then piecing it together wont work either

[smoky oak]

I have time to waste in the pursuit of better SD stuff. How do you think I found all the tips I did to get SD upscale good enough to do a 6x upscale with no detail loss? lol

I am testing!

[stark vine]

insanely better results with lower weight on canny

actually hold on

let me do without it next

[smoky oak]

that seems to be because its just paying less attention to it 😅

[stark vine]

yeh l0l

thankfully I can test the theory.

[smoky oak]

it seems to do way better when there is no canny, from what I can see lol

[static tusk]

gen your image, regular dimensions. Use hi-res fix to make it a bit larger within vram you have. This fixes most problems. Then ultimate upscale.

[smoky oak]

Thats how I already do it

I am testing some things right now

[stark vine]

ultimate upscale already uses the upscaler u select there to upsize it

I assume thats what that option is there for but its just my guess

dude

your theory was wrong