🏞｜general-with-images

[sterile temple]

Their code will be closed source? So who's to say they aren't just copying open source code anyway

[dense tapir]

btw, I saw what Microsoft is doing with AI integration into Windows

[smoky oak]

I am confused as to what you mean

[boreal falcon]

oh, true

[smoky oak]

I think it could be great, but knowing microsoft, it won't be lol

[sterile temple]

You're talking about the Adobe Firefly yeah?

[boreal falcon]

hes throwing it in the room that Adobe might hide their code because they are stealing it from others

[smoky oak]

also, I am curious where it runs

[dense tapir]

There is one question that the answer to would either set my mind at ease or send alarm bells ringing. Does the computer I use it on, and/or even Windows itself, require the Internet. In other words can a PC stuck in a corner with no connectivity use any of it? From word, to copilot, etc...?

[smoky oak]

oh, extremely likely, I thought he was meaning those single devs I have seen do it better than all of adobe haha

[boreal falcon]

i doubt they would do that but people are throwing accusations at them for every sh*t lol

[smoky oak]

thats my question too. The demos they were showing aren't just something you can easily run on standard consumer hardware, which means there is a high chance its connection and monetarily fueled

[boreal falcon]

but they might have been inspired by some features of other programs ofc

which is legit

[dense tapir]

BINGO

[boreal falcon]

stealing code is another thing tho

[smoky oak]

Adobe is about as scummy as it comes, with my experience and other peoples first hand

I used to use all Adobe, until they abused TF out of me when I was younger

[dense tapir]

if the pc never needs to be connected to use any of it then fantastic but if it does then there is 100% snooping and telemetry happening

training their model on all our data AND the privacy concerns are too much for me

[smoky oak]

IDK about all that, but if the AI is off system, then that cannot be sustainable for free on microsofts income from windows

[boreal falcon]

i have only good experience with Adobe for now and they are even ready to cut deals with me lol

for Autodesk not enough experience as of now

and some others

or Maxon for example

[smoky oak]

good luck with that, only a matter of time before they gut you

[dense tapir]

Windows is free hence they are not going to allow you to opt out.

[boreal falcon]

gut in what way?

price going up?

[smoky oak]

every way imaginable lol

[dense tapir]

Autodesk is a shit company just not as much shit as Adobe

[boreal falcon]

someone mentioned they might introduce credits system for Firefly but make it still very affordable and available to people

but thats just a rumour

[sterile temple]

adobe used to be so expensive in Australia, someone did an article where they showed it was cheaper to fly first class to the US and buy it there than to buy it here

[smoky oak]

fraudulent transactions, false claims of copyright infringement, badgering into deals with hidden info, changing packages and feature sets just enough to force you into an uncomfortable situation/spending habit

[boreal falcon]

prices are high but if you can afford its easy peasy

[dense tapir]

I suspect the days of running windows is coming to an end for me as I refuse to be forced to be connected to use it. As much as I really do not like GUI Linux I will go to it

[boreal falcon]

luckily Autodesk introduced indie licenses for 3ds Max and Maya

[dense tapir]

I don't care about their programs it is their attitude and the company itself

[smoky oak]

Autodesk is no saint, but I would take them every day of the week over Adobe

[dense tapir]

same

[smoky oak]

maybe thats cause I have heard several friends and myself having hostile interactions with Adobe

[boreal falcon]

its up to people and their views and ideology to choose whom they support and use their products and whom not

im quite the opposite of you xD

[smoky oak]

its only a matter of time

[boreal falcon]

i refuse to support Stability AI while having a deal with Autodesk, Adobe, Maxon and others and you are the opposite

[smoky oak]

Oh right, you're the guy who was saying Stability is less ethical than all of those other companies, right?

[dense tapir]

I rather support open than closed

[smoky oak]

I rather support a company that can't steal money from me again lol

I'd like to think its funny, but Adobe scarred me in that way

[dense tapir]

OSH/OSS over any closed ecosystem

[boreal falcon]

thats what i mean, we are quite the opposite there 😄

ethically as well as in terms of interests

[smoky oak]

You'd rather be given access to something than actually own it

[sterile temple]

And I don't like adblockers.. jk

[smoky oak]

allow them to rip it away from you and monopolize your dependency on it to hike the prices over and over until you notice they are scamming you

its what happened to me lol

[boreal falcon]

we share different views on that one too

[smoky oak]

then I ditched Adobe after they stole from me, and I ended up finding better free alternatives for what I needed

[boreal falcon]

also keep in mind the interests of the customers/users

[smoky oak]

As a customer/user of 4 years, I have them in mind

full Adobe suite

[boreal falcon]

yeah me too

but we go different paths there

we both actually end up using our products despite controversies around their companies

[smoky oak]

your path seems purposely in opposition to the masses and general views towards extremely corrupt and provably problematic companies, so I am not sure what to make of your views. I am open to other views and beliefs when logic and reason are involved, but that still has yet to be part of this equation lol

[dense tapir]

You know I just like being able to see the code, manipulate it, change it, etc... instead of the big corporation way of being treated like a Mushroom (kept in the dark and fed a bunch of shit).

[smoky oak]

I like that analogy haha

[boreal falcon]

the masses arent the oppositional to those

[wispy ether]

wont

[smoky oak]

so you are saying that most people like all of the regulations and limits that come with using Adobe software and other locked down and iron gripped suites?

oh god

[boreal falcon]

dont have to like, but tolerate is already enough

[smoky oak]

well luckily, many people are starting to not tolerate, myself included

[boreal falcon]

just as i tolerated Midjourney for quite some time when using their product

later i ditched it

and SD too

[smoky oak]

I am not allowing myself to be put in a position where a company can steal blindly from me and get away with it. At least not as willingly

[boreal falcon]

although i never played around much with Stable Diffusion tbh

[smoky oak]

it kind of shows, as it does almost everything better than firefly on your own hardware that you own, with code you can see, and communities and individuals you can get to know

[dense tapir]

wait a minute, you ditched Stable Diffusion? Why are you even in here then? No offence, as I would like to know.

[smoky oak]

I find "they haven't screwed me over" in response to people saying they HAVE screwed them over as a very weak argument, personally

[boreal falcon]

well first of all FF is in beta and very young and also other factors play a huge role for me too

[boreal falcon]

to listen, to discuss and also follow whats up

[smoky oak]

I just still find it hilarious that this"beta" they are in is so bad lol

[boreal falcon]

its not like SD remains static in every way possible

also i dont like to be in a bubble

[dense tapir]

but you ditched SD so you data to discuss is out of date.

[sterile temple]

don't forget to post an image every now and then

[smoky oak]

its about as amorphous and constantly growing as it gets

apparently, much faster than Adobe and OpenAI can keep up with lol

[boreal falcon]

thats not what professionals have to say about it

but okay 😄

[smoky oak]

as in the time that firefly's features have been announced, SD has made answers to almost all of them in faster and more efficient workflows

[boreal falcon]

SD faster and more efficient workflow? Hell nah

[smoky oak]

That take actually hurts to read

"But the people with their livelihoods depending on the grace of Adobe don't speak out in objection!"

[dense tapir]

No one company can make improvements, and additions, as fast as an entire army of users can. If all we had was automatic1111 without any extensions that would be a doa product already.

[smoky oak]

right

the reason SD is growing so fast and is actively beating billion dollar corps is cause of the passion in the individuals

[dense tapir]

Will Adobe release a SDK for Firefly for free to all?

[smoky oak]

0 chance

[dense tapir]

precisely

[smoky oak]

Adobe has 0 reason to ever do anything for the creative community that isn't laced in underlying predation or monetary gain

[smoky oak]

oo, I like

[dense tapir]

See, if I had a product like PS, and Firefly I would gladly release a free SDK for it because the community would design stuff to up my sales potential and make existing customers even happier

[smoky oak]

I am just still trying to figure out how Adobe is gonna make fat stacks off firefly

[wispy ether]

i made a teeth embedding with 1 training time and 6 images bruh

without it

[smoky oak]

Yeah, but unfortunately Adobe is the Nintendo of creative work lol

[boreal falcon]

Adobe doesnt plan to rely on Firefly to make the biggest buck for them

[dense tapir]

LOL, yeah.

The Sméagol of the real world

[smoky oak]

oh of course not, they make money off of stealing from people when they aren't looking and while they are breaking their banks for 100's if not 1000's of dollars a year to use mediocre software

[boreal falcon]

Firefly is supposed to end up being a tool for artists who are using Photoshop and other CC apps primarily

and its also not suited for many of the prompters here or AI art community

well depends on some things ofc

[smoky oak]

still, the point is any functionality that firefly has, SD will do better in no time

SD can already do images to SVG in the click of a button

[boreal falcon]

mhm thats a speculation but lets see

[dense tapir]

Adobe

[smoky oak]

abobe

[boreal falcon]

also dunno why you cry about "stealing" when you know the controversy around Stability AI and its partially notoriously toxic community

[smoky oak]

Again with the same weak defense :/

[dense tapir]

o.O

[smoky oak]

And I mean STEALING as in they STOLE from my PERSONAL bank account

[dense tapir]

Well, we are toxic?

[smoky oak]

apparently lol

[boreal falcon]

nice tweaking but that doesnt help either

[smoky oak]

tweaking?

[smoky oak]

Adobe stole from me directly and told me to fight them in court to get the money back, thats what happened lmao

[frozen socket]

My "hot take" for the afternnon (well it is here).

The 1080Ti was the card Nvidia made a mistake with. Even 6 years after it was launched it can still hold its own against newer generation cards such as the 3060

[dense tapir]

One they never repeated with any RTX card.

[smoky oak]

the 1080ti is a very solid card, tho its performance is getting pretty cramped now days, and it lacks all AI accel and RTX/DLSS/Tensor benefits

but it is still a solid card for just gaming

[dense tapir]

Yep, PASCAL is pretty dead, or dying on its last days

[boreal falcon]

with that you mean you forgot to cancel the subscription or you ignored the TOS and just didnt know that you will pay a fee if you cancel the subscription too soon

[smoky oak]

I could go back to a 1080ti and miss almost nothing in my gaming experience

[willow widget]

Greggs in the stone age

[dense tapir]

Oh, boy here we go it isn't them it is you. You did something wrong not them. Get that all the time.

[wispy ether]

what is greggs

[boreal falcon]

well thats often the case

[frozen socket]

I use one for SD /Auto 1111

Single 768 x 768 image generation is around 19 Seconds (which I know isn't speedy but I'm not after instant gratificatiion ) with a 512x512 clocking in at around 8 seconds

[dense tapir]

I am still over here trying to figure out how we are toxic.

[smoky oak]

no, what I mean is I went out of my way to put my accoutn on hiatus as we went on vacation, they billed me for the time I was supposed to be off, and then gave me that many months back as "reimbersement", to which I then immediately cancelled my account.

then 7 months later I found they were taking 2x my monthly payment out of my account, for a total of nearly $300 I didn't have,a ndI got hit for overdraw fees.

I called to check in, to which they said I had no account so the transactions could not be reversed.

they then offered me all of those months back as reimbersement again, but with the caveat that I would have to start a new annual contract with them.

To which I told them to kindly go off themselves, and they offered to settle this dispute in the court of law, if I felt so inclined

[frozen socket]

Plus it cost me less than £200 with a EK Waterblock fitted. Idels at 24c and maxes out at around 37c on my daily non manic settings

[dense tapir]

I am wanting to go to an open loop with my new case I am waiting to order but damn, they have went so high at 700-900 USD now.

[smoky oak]

So yeah, they stole from me 2 times, didn't honor their reimbursement the first time, and then threatened to take me to court over it

mind you, I was 16 at the time, and I didn't just have $300 laying around

[dense tapir]

Not even including the block for a GPU

[frozen socket]

aye water cooling can be expensive, but damns its so nice & quiet 🙂

[dense tapir]

the quiet is what I after

[smoky oak]

honestly better than I exected, that puts it at about 1/9th the speed on a 3080, which genuinely is not that bad considering the age

[dense tapir]

the old kits of 359 from EK are now 679

[smoky oak]

only thing cheap in PC's right now is RAM lol

get that crap for a few schmeckles right now lol

[dense tapir]

@tired peak What it/s or s/it are you getting at that rez?

[smoky oak]

I can test too

tho, I am on a power limited card

[dense tapir]

I do about 1.6-5s/it

[dense tapir]

my card is O/C too

though it doesn't affect SD much at all

[frozen socket]

My Daily settings are to run the water pump at around 3k RPM (maxes at 4250) , Radiator fans (2 x 360 external rads) at around 700RPM (max 1800) and case fans at around 700 rpm (max 1600)

[boreal falcon]

so you paid the fees for the cancellation?

[dense tapir]

My fans are about 1k but this case has one dead spot, now the new case has 3 fans in the front and tons of air flow

[smoky oak]

512x512 for me is 15.2it/s (1.1s for 15 steps)

and 768x768 is about 6.8it/s (3.1 seconds for 15 steps)

[smoky oak]

no, they waved them cause I was a minor and the contract I signed was not legally binding cause I was not an adult

that was an oversight on my end. genuinely had no idea when I was 12 that I was signing that sort of a deal as an adult

[dense tapir]

7900XTX is about 17 and 25 in Shark for 512x512. I can't get the nums for 768x768

[smoky oak]

but they still made out with about $300 I did not agree to give them

[smoky oak]

that sounds like about my card when I am not power/perf limited

I peaked at 18.3 it/s at 512x512 when my GPU was properly tuned and setup

[dense tapir]

yeah, without ROCm it is about the SD speed of a 3080

[wispy ether]

https://imgsli.com/MTgxOTQ5

is it just me or does fp16 look better

[smoky oak]

@dense tapirAlso, side note, I can honestly say that Zovya's photo real model could be a 768x768 model based off of how well it does 768x, and sometimes even 1024x without artifacts

[dense tapir]

if ROCm 5.6.0 can get 25-30it/s as anticipated I would be elated

I had one on 1.5 when I went back for the week that was doing 1024x1024 and my jaw dropped

[wispy ether]

||me here with my 7it/s||

[boreal falcon]

you signed that with 12? O_O

[smoky oak]

yeahhh, not one of my better moments. I did technically sign a document as an adult when I was not one, cause I didn't realize at the time

that did at least get us out of the cancellation fees, of which I also didn't know would be a thing at the time

[dense tapir]

I lay odds when the 300MI is released is when the ROCm 5.6.0 gets released which is August for the MI300.

[boreal falcon]

at least you didnt have to pay them

[dense tapir]

remember ROCm is for their business side we just get to tag along.

[smoky oak]

but yeah, they still blatantly stole from me and then tried to get me to try them in court, knowing there was a 0% chance I would do it

[boreal falcon]

dunno how high the fees are right now

[smoky oak]

native 1024x1024 gen in photo real with no high res fix

pretty impressive to me, IMO

[boreal falcon]

i mean obviously the sooner you cancel the higher the fees

[dense tapir]

@smoky oak Realize this the 7k cards are the first ones with real AI on them not the 6k, or before. AI / wmma core if they utilize all those it will get about 30-35it/s

[dense tapir]

I'd be happy with that know 768 or 1024 would be slower but yeah

[smoky oak]

Just know that situations like that with adobe are not an uncommon as you would like to think.

Same BF turned friend in Hungary got locked out of his Adobe for months in Uni and had to pirate it at the time cause he was gonna fail if not. Their only recommendation was to make a new accoutn with a new year long contract (which would have put him in 2)

[dense tapir]

wmma is AMD's Tensor cores

[smoky oak]

oh, and he wouldn't have gotten school prices either, cause his educational email was already in use on the other

[dense tapir]

Sounds like Adobe to me

[smoky oak]

yeah, a lot of people don't know that RDNA 3 is Ai accelerated

[dense tapir]

does AV1 encoding and decoding too

[smoky oak]

Yeah, tho I have heard that AMD is not too great with that 😅

best for AV1 actually seems to be Intel

[dense tapir]

no, it is actually but RT is forget it

[smoky oak]

They had it first, and their method still seems to be fastest and most power efficient

[dense tapir]

yes

my card I force H.264 so I can get H/W decode on yt

[smoky oak]

thats also weird, cause I have heard and seen the opposite, with RDNA 3 being extremely good at RT, but bad at AV1, as it sues much more power and GPU load than it does on NVIDIA of Intel

[smoky oak]

I always forget, do you have a 1060, or a 970?

I have a friend with the opposite

[dense tapir]

well, know this 350w is what the card is rated at

1060

[smoky oak]

and I get you confused with

ok, so you ARE 1060, right, cause the 6GB VRAM

[wispy nest]

yo sytan

[dense tapir]

yes 😦

[wispy nest]

what do you think about making a lora for optical illusions

[smoky oak]

I have a friend with a 970 who I always get confused for you

[dense tapir]

if the Nitro+ fell to 900 I would buy it instantly

[wispy nest]

should be interesting

[smoky oak]

could be cool, but I am taking a break from LoRA's at the moment. Reallyyy trying to get a form of stable income from Ai

[wispy nest]

plis

[dense tapir]

I wish I could get that from SD but nope

[smoky oak]

If etsy could get their thumb out of their ass and fix my account, I would be "Elated", in the words of yee Awareness, the General

[dense tapir]

hehehehe

[smoky oak]

I don't know why, that word just tickled me lol

probably cause I use ecstatic instead of elated lol

its more... fresh to me

[dense tapir]

I saw a vid on how to sell your shit from SD/MJ on etsy and I rolled my eyes and went to another vid

Why pay for the milk when you can get the cow for free?

be a hard sell iow

[smoky oak]

You'd be surprised

expertise can be sought much more than versatility

[dense tapir]

sounds like NFT skullduggery to me.

[smoky oak]

Thats why I have a return customer for LoRA's

she only buys from me, cause I am the only one that does what she wants, and I do it damn good too

[dense tapir]

Ahem

[smoky oak]

Niches are a powerful thing

[dense tapir]

sorry, my mind just ended up in the gutter

[wispy nest]

i wish i knew how to train loras

i watched a video tutorial but shit was too much for my brain

[smoky oak]

I do hot men LoRA's lol

Something she said almost nobody was willing to do, or wasn;t good enough to do

[smoky oak]

most tuts are trash anyways. FAAFO, best way to get good at LoRA's

[wispy nest]

wtf is this

[dense tapir]

Well, hot men isn't my thing, nor hot women. I am more into the landscape and horror shit. Death, blood, gore, depravation, hunger, etc...

[smoky oak]

understandable haha

but I fill some biggg niches

hot men
gay stuff
and most of all
Furries

and my GOD do furries pay out the ass

like, comically so lmao

[dense tapir]

Stay away from the AIEntrapeneur he is....

[smoky oak]

trash

unreliable

pompous

egotistical

self righteous

snarky

[dense tapir]

did you see my furries? I didn't mean to make them. I tamed it down to make cute animals instead.

[smoky oak]

full of himself

need I say more?

[dense tapir]

I unsubbed him a few weeks back as he went over the cliff

[smoky oak]

WRONG 99.9% OF THE TIME

there we go, big text moment

to be used sparingly lol

[smoky oak]

I did! Nice stuff!

I have a very special model that is ungodly good at anthro stuff. Makes realism models look like absolute pathetic jokes lol

[dense tapir]

LOL

[smoky oak]

hands down best trained and most advanced model I have ever used. Nothing comes even an 8th close

let me get the list

just a sec

[dense tapir]

I didn't mean to do it but I liked the result

I showed this on the other channel

[smoky oak]

It can RELIABLY do:
50 hue consistent colors
851 species/categorical characters
1972 art styles
554 articles of clothing/outfits/armor
368 predefined jobs and statuses
332 personal descriptors
40 quality alters
102 setting prefixes
82 concept suffixes
26 defined surface materials
268 predefined structures
146 predefined countries
26 fantasy worlds
180 direct locations
111 nature biotopes/biomes
195 "themeatic" settings
405 general/generic locations
11 body poses
30 limb poses
8 look poses
21 general poses
62 emoji triggered emotions
31 text triggered emotions
120 action styles
6 image crop adjusters
103 accessories/addons
181 object/character adjectives
335 feral animal species
502 types of food
16 nature objects
51 singular/misc objects
439 general group/misc objects
28 shape modifiers
19 art themes
106 vehicles
17 ambiance enhancers
27 weather ambiance
35 camera modifiers
10 camera brands
21 camera compositions
76 camera models
59 color pairs

[wispy nest]

whats the name of it

asking for a friend

[smoky oak]

can't share here, name is too NSFW lol

but my GOD is it good

you can drive values to 1.6 with no artifacts, and even use CFG up to 18 with no artifacts

its strictly SFW unless you say otherwise

its diverse, supports multiple characters semi reliably

its just jawdropping, and makes me cringe going back to realism models

[dense tapir]

yeah, but 1.5, right?

[smoky oak]

perfect faces, hands, feet, mature stuff, EVERYTHING

I can't even detail all the mature stuff it can do here lol

just know its trained on 2.9 million images from E621, with 77k individual tags for... content lol

[smoky oak]

yeah, but my god, it gives SDXL and Midjourney a huge run for their money in terms of how damn good it listens

[dense tapir]

I don't do 1.5 because all my loras and special sauces are for 2.1. I go to 1.5 and become naked.

[wispy nest]

i abandoned 2.1 because it doesn't do nudity

prudes smh

[dense tapir]

SDXL will not either

need models from the community for it

not my wave so not even sure any have been trained or not

[sterile temple]

men's clothing is hard to get perfect, because it often has to be perfect. shirts, polos, trousers, belts. so simple, but so easy to mess up

g'nite

[oak osprey]

https://images-ext-1.discordapp.net/external/JGECEjqLOdrqRXmxeXbrPMCPtmH9MdZqhJG-CQkb18c/https/tripleback.net/public/discord/1685189557.3441982f4b9420a6bd42e006fee64c457750db8.png?width=844&height=844

[oak osprey]

uhm it absolutely does nudity

that was my first fine-tuning project.

[neon crystal]

Thank you this was purely accidental. I fine tuned the my usual 2.1 prompt salad for horror on 1.5 and just for laughs tried it in 2.1 and got this stunning result.

[cyan snow]

woooooooo

[oak osprey]

yeah there's models that make young women by accident. and then there's that one

[cyan snow]

eh, it's in the prompt

no problem with incorrect generation

[wispy nest]

let the bodies hit the floor lol

[cyan snow]

wait, there are two men holding hands in the background

wtf

[oak osprey]

looks like a fist bump

her left arm is deranged

[cyan snow]

nah, that's just old people arms

[oak osprey]

thats the cryptkeeper arm

[neon crystal]

Extraordinary claims require extraordinary evidence. Verifying it does ALL these 1000s of things RELIABLY would take a ton of time. Is there a large community that is testing this model? I’m assuming this is a list of tags / attributes in the dataset they trained on. It could be amazing ( based on your personal experience ) but I’m a bit skeptical about how much better it is compared to other models out there. I could be wrong of course. And that bit about 1972 art styles makes it sound like bunkum unless of course that is meant to say art styles of 1972 artists that are present in the dataset it is trained on. My apologies if I’ve come off as rude btw I’d be happy to be proven wrong.

[neon crystal]

Referring to this 👆

[oak osprey]

i think he meant the year 1972

that said, the claims behind that model are due to the large training corpus but it remains to be seen what the training data is, how it's labelled, how it was acquired and so on

doing 2.9 million images through SD 1.5 sounds really, really, expensive. SD 1.5 takes more steps to train than SD 2.1 does.

[smoky oak]

They actually have documentation where they show off around a thousand of the styles and such, as well as examples for every species as well

I would send the examples, but they are quite NSFW

[oak osprey]

nsfw species??

that sounds fucking broken

[smoky oak]

It's a furry model

[oak osprey]

then it doesn't really count 😛

[smoky oak]

What's that supposed to mean?

[oak osprey]

if you can't get what you want out of it without it going nsfw and furry, it's broken

[neon crystal]

Wow that’s incredible. Would love to get hold of it.

[oak osprey]

try putting 'nsfw' and 'drawn' or 'animation' in your negatives

[frozen socket]

Oh it does do nudity, yiou just have to watch out dor 2.1 nipple syndrome

[oak osprey]

2.1 knows nipples but you have to teach it areolas

and anything below the waist comes back as like a smooth plastic doll

[frozen socket]

ahhh maybe that why th enipples end up like Yorkshire puddings lol

[oak osprey]

i don't get the reference

[frozen socket]

I also find starting the prompt with and you have to force it, I generally start the prompt with (NSFW,nudity,naked,erotica) helps and adding thinkgs like (lingerie, knickers, panties,bra) to the negative helps

[oak osprey]

that looks like neither pudding nor yorkshire

[frozen socket]

these are Yorkshire Puddings

[oak osprey]

😩

[smoky oak]

I mean it's a hyper trained model, it does one thing, and one thing specifically, and that thing is furries and furry art. SFW and NSFW

[oak osprey]

hyper trained, over trained, under deliver, it's all the same thing

[smoky oak]

It's not overtrained

It is hands down the best trained model I have ever used

Makes all the realism ones look like jokes

[oak osprey]

"it does one thing, and one thing specifically" -> it lost all of the baseline variation capabilities

[smoky oak]

Well yeah, it traded baseline blegh results for unrivaled quality over a massive amount of species and characters for a specific use. It's a fine-tune, so what?

[oak osprey]

i'm just saying, you don't have to be upset that it was overtrained. i've said it many times, it's about what your goals are. i would consider that a very poor and broken result. but i appreciate the wide range of the baseline models.

[dense tapir]

Naked as this is 1.5

[smoky oak]

You would consider a model being exceptionally good at over 70k tags as "very poor?"

If so, idk what to say

[oak osprey]

i have never tried that model but the baseline is good at like more than 2 billion tags so yeah

[smoky oak]

Nobody else in the community has come even close to a 10th as good as this model for reliability and diversity

[oak osprey]

then again you're probably only thinking of 1.5 and i've never gotten stuff i'm happy with from baseline 1.5. but 2.1 is a different story

[wispy nest]

quality is in the eyes of the beholder. For one person something might be perfect, but for someone else it might not. That's why I try and not mention something is better than something else. But I'm not perfect...yet :P

When it comes to art, there's such a wide gap between models and while some are better than other. It's hard to say which is without actually seeing any results :P

[smoky oak]

Your fixation with 2.1 is stifling

[oak osprey]

because it's far superior, lmao

[smoky oak]

I'm done here, anyways

[smoky oak]

@neon crystal if you want the model, I can link in DM's. Name is too NSFW here even tho it's a SFW model first

[dense tapir]

I gotta say this 1.5 model is not obeying my prompt but damn, the hands, wow

[oak osprey]

[neon crystal]

Any way I can apply for access to this ? Can DM you if it’s alright. Would love to get my hands on it and play around with it to see how it follows prompts and other fun stuff

[smoky oak]

Friend request sent

I'll meet you in dms

[dense tapir]

Sytan, look at the above, why is there no skin texture?

Is there something I have to do with 1.5?

[wispy nest]

and if you want, you can send the name of it to me as well. I'm kinda a collector of models :)

[smoky oak]

My assessment of this model is off of it being able to do exactly what you ask for, as well as having a huge range of tags that do not interfere with each other, while also having individual LoRA's for dozens of aspects of it to be appended to other models, as well as the ability to do complex multi character scenes reliably

[wispy nest]

don't worry, I don't assume it's a perfect model as everyone who gets good results aren't just using one thing for it :)

[smoky oak]

It's not perfect by any means, but for a model looking to achieve something, it's by far the most capable of any specialized models I have ever used

[oak osprey]

it's easy to overfit keywords but continuing the amazing prompt comprehension is hard

[dense tapir]

this is that model.

[oak osprey]

like you can teach it a new concept, but can you get that to seamlessly blend with other concepts it knew. for example, you teach it what Bigfoot looks like. now you want to make Bigfoot go surfing. it might not work.

apparently i deleted all my 1.5 models on the 9th and haven't felt like using one since

[smoky oak]

It can do that exceptionally well

I don't know how to describe it without getting into NSFW

[oak osprey]

[dense tapir]

damn

[oak osprey]

i gave a pretty good example

i wish the wizard would stay this beautiful forever

is that jared leto

[hasty nova]

also yes this tree was mostly ai generated, I just did some minor touchups

[oak osprey]

https://images-ext-1.discordapp.net/external/EiEk8Yo6DLeHQoD9D9KqmqPkJ2y-zvfflo_pTFa_vVs/https/tripleback.net/public/discord/1685202230.218285132dfd96bac4a6661b6a8c20f9fc67a2.png?width=1757&height=1004

[astral flax]

Now make her wink 🦭

[oak osprey]

the textures in those images look whack

[oak osprey]

me: "you cant do bike chains. figure it out, dummy"
AI: "I gotchyu fam."

[wispy nest]

[oak osprey]

oh

that's Jeremy

the crow

[oak osprey]

from NiMH

[wispy nest]

dunno

i just made a crow

was supposed to be an anthro crow but it was so good i kept it

[wispy ether]

Kinker

Latex suit and obvious boobs

[oak osprey]

well this is the first time i'm seeing a damn truck show up in the alien prompt during training

[odd tangle]

I love this style lol (generated using f222 I think, might be anything v3)

ignore the jacket

[dense tapir]

RuntimeError: Expected all tensors to be on the same device

[wispy ether]

i dont have a single style in my folder

[wispy ether]

if you run 2 instances at once

[dense tapir]

No

[oak osprey]

it's just a code error

probably have to restart because the model CPU offload probably screwed up

https://images-ext-1.discordapp.net/external/96ZZOoFuI3O2Bp7XQiy9k4cQjPjJaYgkySPDzjr4jeA/https/tripleback.net/public/discord/1685207557.958769e2207f7a987e04169bff4fe58bbe2ac1.png?width=1267&height=844

a bitcoin moon lmao

https://images-ext-1.discordapp.net/external/u40x4-WYOxClwzCojAbPPyniNb4RL0tTj9sjqOBmj-k/https/tripleback.net/public/discord/1685207678.456774716a9f53504830411a2db33cafed7685c.png?width=1267&height=844

ooooh i like that

[wispy ether]

id call that a successful generation

since its fake

[odd tangle]

LMAO what model?

[odd tangle]

reminds me of BB from death stranding

[wispy ether]

realisticvision

[wispy ether]

the fuck

[odd tangle]

art

[odd tangle]

evil baby

[dense tapir]

why?

It's not evil only misunderstood, lol

[wispy ether]

ai created a new dog

[oak osprey]

https://images-ext-1.discordapp.net/external/9Xr3KHwqELJRQ-iiFS-0DmeSfh_iA8fGIWUJ2-oMD6M/https/tripleback.net/public/discord/1685207989.78826075e85bd1878be7c6e55215ae9eccee2ba.png?width=1267&height=844

it's the best when it makes the kids start flying with the kite in the park

this needs like 12,000 - 14,000 steps of training to start happening lmao

[wispy nest]

hello mega picture

[oak osprey]

i like that one a lot

[oak osprey]

https://images-ext-1.discordapp.net/external/EMZao-efWgxU0cIccV78SwS2syLEgmtVSBU_O3DGn68/https/tripleback.net/public/discord/1685216974.761840611a8156da1b8d38e3a0405cf54b7a9ae.png?width=1267&height=844

[oak osprey]

https://images-ext-2.discordapp.net/external/S-1Jfn32Hr8hX-GrA8gLJUFKiOI3Orr9LlVsM_xrFjQ/https/tripleback.net/public/discord/1685216974.7814262dab7f22dcd691eb9910d95b7fc9ddac2.png?width=1267&height=844

[oak osprey]

https://images-ext-2.discordapp.net/external/NjdJKZSO1dAUF9WZeb1HDnrjJhz2dPks3YJYuOCGeeo/https/tripleback.net/public/discord/1685217475.2908689adcc4020c6e10891b106969bbe92533.png?width=1267&height=844

https://images-ext-1.discordapp.net/external/G2pSiSLlGotwcHUuNp3ZXIbLEgi2WqyU0S3Ypq-XIPE/https/tripleback.net/public/discord/1685217714.50829277a00731eadb1469afa169b6b52c34b72.png?width=1267&height=844

[fickle gorge]

I'm trying to upscale an image and, when using ESRGAN_4x, the result isn't blurry but I notice it adds lines all over the image. When using Lanczos the image is blurry, it's nearly like doing it on GIMP, but no strange lines are added.
I'm using Ultimate SD Upscaler with 0,04 denoising strength and upscaling 4x from the original size, varied step counts as I'm still trying to find something that works.
The image I'm trying to upscale is this:

I think the image looks not bad, but getting it upscaled is giving me trouble

[hard wagon]

Just released a new multi-style model called Aether Lux

https://civitai.com/models/76276/aether-lux

[oak osprey]

my model's output for that first prompt

[hard wagon]

Ok 🙂

[hard wagon]

Another image made with that prompt on Aether Lux 🙂

[pastel mango]

Aether Lux, Aeon Flux

[hard wagon]

Really cool reference 😀 totally had forgotten about that movie

[oak osprey]

came out around the time Shrek did

i think they're about 40-something years old now

[pastel mango]

I feel old. That movie is as old as many horror classics were when I was a kid

[oak osprey]

Friday the 13th

sleepless nights followed for me

that one where Jason comes out of the lake

[pastel mango]

that's every one

except the first

or well no

the first too I think

in a dream

[smoky oak]

@hard wagon that looks dope for a 2.1 model. Great work man

It's a pain in the ass to get 2.1 fine-tunes that look good

It def does still have some of the 2.x artifacts, but man that had me guessing for a bit

[oak osprey]

it's easy for a limited subject matter but hard to make it generally applicable

[smoky oak]

probably the cleanest 2.1 I have seen so far

[oak osprey]

i found there's a way to adaptively freeze layers once they converge in the text encoder so that it only trains layers that still have errors remaining to fix

[smoky oak]

That could be cool. I am just happy to see 2.1 models that don't look so nasty in the fine details. Thats the first barrier to entry for me

[oak osprey]

i wish the leopard gecko would come back

you see, i noticed the prompt for the gecko images is a leopard gecko, stalking a cricket and it's never put a cricket into the output but it, at one point, morphed the gecko into one

[wispy nest]

might be the ai interpreting your prompt as "a leopard colored gecko stalking like a cricket" ;P

[smoky oak]

Cricker geckit

[wispy nest]

maybe add a conjunction or two to help the ai on what you want? :P

[oak osprey]

it's a sanity check prompt to look at changes during training

[wispy nest]

aren't sanity checks for when you want to ensure the correct values are being added? Like inserting a string when it should be an integer and stopping it from being used? :O

[hard wagon]

Thanks! I am not claiming that this is a perfect model (as if there are any). I hope it will be of used to some anyway. There sure are few custom 2.1 models out there.

[smoky oak]

Yeah, from all of the dozens I have seen, it looks the best in terms of the 2.1 artifacts

2.1 is messy AF

[wispy nest]

I think my largest issue with 2.1 is that so much of it look so, I don't know how to explain it, smudgy maybe?
Blurry as well as even when the image is focusing on something it still look off to me :(

[hard wagon]

Ok. Cool! Didn’t consider that here.

[oak osprey]

i love how sytan talks about artifacts when it's coming through as a JPEG on Discord all the time

worst compression algorithm for images if you want to preserve detail

in other words, the images probably look fine locally

aw man, leopard gecko now. i've seen this before

catastrophic loss

[spring kelp]

was this with auto1111? i can only achieve this sort of creativiy with comfyui

[oak osprey]

no, this is with a small python script

[fallow thorn]

Found this on another server

[oak osprey]

lol

[smoky oak]

That's not what I mean, it's not discord compression, it's compression right out of SD

[oak osprey]

its compression in your brain

[spring kelp]

i like the results, very nice

[fallow thorn]

The amount of filters people use these days and the quality of generated images

[oak osprey]

i just wish i could reliably see what the heck artifacting you mean

[fallow thorn]

That's basically no difference lol

[oak osprey]

@fallow thorn using a ND filter on a camera makes it look like that too

[fallow thorn]

Hmm it's insane

[smoky oak]

it is quite prominent in most 2.1 gens, let me get some examples

[sterile temple]

morning

[smoky oak]

@oak ospreyAlright, here we go

these images were generated at 2.1 res

Image seeds were decided off of best of 16 for the 2.1 model, 1.5 model uses the seed selected from the 2.1 model

First prompt was a Japanese man in a Kimono

You can see the 2.1 model is just very compressed and has a lot of noise and artifacts

That is the main thing that keeps me from using 2.1

[sterile kiln]

(me)

E. Macron

[smoky oak]

Next is a bonsai tree on a desk

the 2.1 image has a ton of noise and compression artifacts in this one. One of the worst I have seen

Severe compression, bit crushing, over sharpening, and stair stepping effects

last one is supposed to be a cozy autmn forest sunset with snowy mountains in the background

2.1 did not listen to this prompt well, but I do like the color processing and aesthetic more, even if it is once again, full of tons of issues and compression artifacts

i.5 vs 2.1 tree close up

2.1 is always just so blobby and messy, and that is what I am trying to get away from with 2.1

don't even get me started on the muddiness in the darker areas

What is even happening here

just dirty, nasty, over processed artifacts everywhere in the shadows

so that, @oak osprey , is what I am referring to

[wispy nest]

yeah, I recognize those things. It can be somewhat fixed with the use of X, but it's just so time consuming and exhausting for me than just using something from 1.5.

I really want to like and use 2.1 as it got better possibilities, but it's just too much at the moment to get good results that I want from it :(

[fallow thorn]

I mean even real mud is dirty so

[smoky oak]

You have to put so much more work to try and get a result thats as good

[smoky oak]

different use of dirty

and this is at native 2.1 res, if i did 1.5 res, it would be even worse

[wispy nest]

yeah, I just want to get good results I want, and it might be a me thing as well, but I have trouble in almost every step of using 2.1.
There's so little documentation on it, so few people using it. So many more watermarks from my experience. And I found out that the best prompts for 2.1 went something like embedding of embedding a embedding embedding embedding embedding embedding :P

[sterile temple]

Is there a config file for this?

[smoky oak]

stepped down to 1.5 base res, and now the 2.1 model is just unusable

So 1.5 can do 2.1 base res just fine usually, but 2.1 can't do 1.5 base res

[oak osprey]

@smoky oak can you give me some prompts to try

[smoky oak]

sure, I can share in DM's

[sterile temple]

does 2.1 have a bias towards copper, sepia, brown tones?

I'm not familar with negative prompting for 2.1

[wispy nest]

depends on the prompt, some words have a tendency to give a more yellow/orange color. Like something pretty might be a gold colored ring. And overall I'd say that 2.1 has more "warm" and bright colors

[ripe cedar]

Any decent realistic upscaler model anyone uses?

[sterile temple]

after adding some more negative prompts

[smoky oak]

I would not recommend asking illegal things in the server lol

[sterile temple]

pretty close

[smoky oak]

Chang of Hart

Just makes me think of a guy named "Chang Hart" lol

Kevin Hart's long lost Pacific Asian half brother lol

[dense tapir]

I wonder what dead people, or near dead people, SD was trained on when I always get this kind of results?

clavicles are always so jacked

[wispy nest]

I feel that the ai see no difference between dead and something like bored as I'm not sure there are many "realistic" dead images that it trained on. I think it trained more on "dead ghoul people from horror movies" and those can vary a lot :P

[dense tapir]

training on the undead, lol. I bet you are right which really fucks up images of the living.

[wispy nest]

I learned to trick the ai into making what I want more than telling it what I want :P

[dense tapir]

I don't know that trick, and I also see this with armour sometimes too.

[wispy nest]

I don't know how well it works using 2.1 stuff, but using a movie title and stuff like "wearing corpse mask/makeup" might trick it into creating a human with a human mask made of corpse skin.
Very weird sentence, but that's what I learned works very well :D

the ai know more about stuff like "makeup" than "dead stuff" so it might not know where to apply the "dead stuff" but if it was makeup, then maybe it would know, but yeah. It's a experience thing that gets easier over time to "think about."

Or I know nothing and I'm just taking bulls— :P

might just be a bias as well hehe

[oak osprey]

https://images-ext-1.discordapp.net/external/ORhmXi9Bj2fqTtuEo8EPAJdkfkcmqvFO3oYiOo7S-qA/https/tripleback.net/public/discord/1685237852.288849d2a0d8ae91f1374357280e724ea962f0.png?width=1267&height=844

[wispy nest]

not wanna show too dark stuff but here's pure horror! :P

[dense tapir]

2.1's clip is a bit funky with such stuff

[pastel mango]

testing like two dozen models now with different prompts to find strengths and weaknesses. Occasionally you get images you just gotta love

[wispy nest]

this is probably a safe horror image I made some time ago :P

[dense tapir]

I am not even sure how I would fix it in post either

[wispy nest]

I hope these aren't too dark, but this is the limit of what I think I can show on this discord channel when it comes to horror, if not, and it's too much, then I can remove them. Or if someone else does before that :)

[sterile temple]

time for lunch. nom noms!

[wispy nest]

haha

[smoky oak]

Pacific islanders, one of my many weaknesses haha

I guess at some point, I have to stop saying specific races/ethnicities/types of guys are my weakness, and just say men in general are my weakness lmao

[smoky oak]

Love this kind of stuff

been hoping to get some good visceral satanic imagery out of a model sometime soon

[smoky oak]

FOUT BMUAIY

[wispy nest]

depends on how "authentic vs movie coolness" you want, but it's easier than one think to get those type of images :D

[smoky oak]

hot demon men are in my sights

Wonder if I could do that lol

I am trying to sleep with the devil lmao

[wispy nest]

is he a snorer? :P

[smoky oak]

I am gonna stop before I say something horrid lmao

Unlike some people here, I know when to stop with mature things lmao

[sterile temple]

Face peel

[wispy nest]

here's also a 'safer' one :P

[sterile temple]

Margery, how could you??

[dense tapir]

Hehehe

[sterile temple]

la croix, tesla, bottled water, wifi, macbook

[neon crystal]

More alien delicacies.

[kind quartz]

Volha?

[kind quartz]

yes nice light of scene

[dense tapir]

Thanks

[sterile temple]

cabaret

[marble wasp]

https://tenor.com/view/master-master-chang-big-chief-grand-chef-senior-chang-gif-25145635

[marble wasp]

i really like these cars and the textures.

[dense tapir]

Yeah, they do look nice and classical YET, not real.

[marble wasp]

i did a similar fictitious vehicle

[dense tapir]

Nice one

Anyone use Kohya_ss and explain why when I just reinstalled it I run it to get "Torch not compiled with CUDA enabled"?

[kind quartz]

this remind some hybrid between porsche and quak quak citroen 🙂

[marble wasp]

This is Rojelio, made with SD and epiCRealness modeler from Civitai.

[dense tapir]

Not bad

[marble wasp]

This was a different version of Rojelio, as a vagabond.

[dense tapir]

I like the texture of the skin

[kind quartz]

ghost of titanic in streets. I like how it changed its paint to white 😄
Also it somewhere lost one chimney 🙂

[marble wasp]

not to over do it, but I did a series of Scarlett Johansson with SD and the same modeler but fine tuning prompts using "draw" and "indraw" functions.

[dense tapir]

that painting was an indraw

[hard wagon]

Config file?

[dense tapir]

.yaml I think he meant

[sterile temple]

it's working without one, all good

[wispy nest]

i think i just made my new favorite reaction image.

EDIT: actually think i like it better without the grass/shadow effect.

[kind quartz]

do you outpainting or inpainting? @dense tapir

[dense tapir]

Never got it to work for me.

[dark flare]

Change this one into a boy and it's my son! He'd need extra demon-glow in the eyes though 😉

[dense tapir]

LOL

Looks to be its mate

[sterile temple]

that's a spicy meatball

[dense tapir]

Heh

[neon crystal]

There's this different kind of chemical system, 2000 Flushes. Keeps the water crystal clear.

[dense tapir]

Not sure what world it is colliding with but not going to be much left

[marble wasp]

Hahahah, ok, so I was working on a series of images with an elderly asian male with tattoos and one rabbit tattoo, and I wanted it to be a minimalist portrait. One unexpected blooper:

[dense tapir]

Oh, look at this one

[marble wasp]

cyber Thomas

[dense tapir]

LOL

[sterile temple]

choolorean

[marble wasp]

My eyes are up here

[sterile temple]

what's the scented candle for?

[dense tapir]

Mood

lol

[marble wasp]

^

[sterile temple]

now try to get a darker image 😅

[marble wasp]

which?

[sterile temple]

short train

[dense tapir]

with the energy I am seeing it is too long

[sterile temple]

we were struggling to get low lighting images yesterday of a woman wearing headphones

[dense tapir]

That is just the engine

[marble wasp]

ahh i didnt see it, let me try

[smoky oak]

Today I learned that windows has built in text to speech

or sorry, speech to text

[dense tapir]

for decades

[smoky oak]

I have a custom model mix that could probably do that

[dense tapir]

reason Dragon speech went kaput

[smoky oak]

had no idea! lmao

[smoky oak]

I remember that lol

[dense tapir]

yep

[smoky oak]

my grandmother used to have that

[dense tapir]

we went from dragon speak to Siri in less than a decade. Amazing really.

[smoky oak]

and now we have significantly better tools than even siri. Our progression in tech is truly inspiring, I think

[dense tapir]

I 100% agree

if the govt gets out of the way we can advance if they do what they are trying to get away with we will head into a technological dark ages.

[smoky oak]

No more hot furry boys and big tiddi goth anime girls

lmao

[dense tapir]

Well, that one is hot in more ways than one

Well, damn

[wispy nest]

What model is it?

[dense tapir]

A mix I did

[wispy nest]

gimme dat, demon girls is my thing

[dense tapir]

oh, shit

honestly, demons aren't in it so surprised

[wispy nest]

hm...which sampler is it?

[marble wasp]

a little darker:

[smoky oak]

We are two sides on the same coin lol

Demon boys are my jam lol

[dense tapir]

++ SD Karras

[smoky oak]

I have a custom string noise offset semi realism model I mixed myself, and it can do dark images like this very good

[dense tapir]

Yeah, this model I pinched from is so heavily woman weighted as I said demon it gives me women.

[smoky oak]

Sick and tired of that, personally

[dense tapir]

so am I

[marble wasp]

nice, i was just reading an article on this, didn't fuly understand

[dense tapir]

I had one model I put Samurai and out pops a chick

The female Ninja was hot, but not what I was going for

[wispy nest]

lol

[sterile temple]

look, if you want dudes. you just gotta put (woman lady female girl:999) in the negative 😄

[dense tapir]

Well, I added men to it and it helped

[wispy nest]

overweight negitive doesn't overcook?

[dense tapir]

was a merge

Once I get my card I can finally go back to training

I love the background it is giving these

[marble wasp]

https://www.crosslabs.org/blog/diffusion-with-offset-noise

[smoky oak]

noise offset shouldn't need to be a thing for much longer

[smoky oak]

SNR solves it with no downsides

[dense tapir]

yep

sad it only works going forward all old models can't use it

[smoky oak]

Here we go, got the semi realism custom mix I made working

I am rusty with it

its also based off a pretty poor realism model by now lol

real vis 1.3, so it was never that great to begin with lol

[dense tapir]

What I hate is everything in 1.5 seems to be mixes of orange whatever. If it said orange it is in almost all mixes.

[smoky oak]

man, every time I got back to realism from that furry model, I am just disappointed lol

much better image using a better modern realism model

[wispy nest]

SNR?

[smoky oak]

I don't know how to explain it

its noise offset, but as a fundamental part of the model with no need for specific triggers

[wispy nest]

Is it something that SDXL will have by default?

[smoky oak]

its already in SDXL now

yeah

[wispy nest]

okay

[smoky oak]

they added it right when the paper came out

it makes a huge difference

[dense tapir]

SDXL, lol

[smoky oak]

@wispy nestThis is what it looks like in SDXL at 40% training

it just has a wayyy better way of rendering colors, and it can handle extreme contrast and such

Also should make VAE's pointless IIRC

[dense tapir]

Yep, real black

[wispy nest]

I think vae's still will have usecases, if you want to apply very specific style \ color schemes

[smoky oak]

for color filtering, yeah, but you shouldn't need them to fix improper base colors

like, it should get rid of over crushed or over flat images IIRC

also god, I wanna get my hands on SDXL soooo desperately lmao

[smoky oak]

It is so outrageously good, and its only at 50% training. I can only image how insane the final full release will be

[dense tapir]

If people can't train it as easily as 1.5 people will give up and go back to 1.5

[sterile temple]

I thought you could use SDXL through the dreamstudio site?

[smoky oak]

I know I won't lol

[smoky oak]

you can, but its like a really cut down experience

[dense tapir]

If I can't train it I know that I will come back to 2.1

[smoky oak]

when you look at what actual people who have access have been able to do, its insane

[smoky oak]

I just personally couldn't leave such a vastly superior model for something like that, personally

[dense tapir]

I like the personal touch not just a conveyor belt of what they will give you. That is MJ

[smoky oak]

@sterile temple@wispy nestHave you guys seen the SDXL 50% checkpoint demo images?

[wispy nest]

nope

[smoky oak]

They are... Something magical lmao

just a sec

These are all straight out of SDXL

50% trained

[smoky oak]

This 50% trained base model makes even the best 1.5/2.1 finetunes look like jokes lol

[dense tapir]

new model

[smoky oak]

It can do everything, and its all consistently better than previous models

that water color one floors me

[dense tapir]

This reminds me of an animated movie I saw. Same style.

[smoky oak]

Some more from their demos

it can do so many different styles, and it can do them DAMN well too

[smoky oak]

gets me hyped for whats just right around the corner for AI

[smoky oak]

Here is the link to the full album, if anybody wants to look

Its 49 images IIRChttps://imgur.com/a/jsi4sAM#0RDvTQq

[wispy nest]

That's alot of variety, but idk, I want to see that myself, how it generates \ how it reacts to different prompts , how many words it knows and shit

it's just...cherry picked hard to judge with no prompt info

[smoky oak]

I can say just from using the completely raw non improved SDXL, There were few things it couldn't do

Several colors on a single character, diverse and complex poses, and it has a natural language model interface, so you just talk to it like MJ

I did like 200 SDXL tests the first few days it dropped, before any finetuning, and it was like magic

[smoky oak]

@wispy nestHere, I generated these at base 512x512, I know they do not look good, but they show the control

Statue of liberty:

From behind, from below, from a helicopter, far away at night, being demolished in black and white

and that was just with the trash dataset they trained 2.1 on

[smoky oak]

Its so easy to prompt

Rainbow water color style geometric art of a symmetrical wolf face

And

Michelangelo painting of jesus getting down in a roman night club lol

no negatives

[kind quartz]

It could be nice hobby 🙂

[smoky oak]

was even able to do advanced things like this

Photograph of a white woman with red white and blue tie-dye hair in braids

it also had no weapon filter

not sure if it still doesn't

A photograph of a native American chief sitting in the sand of a desert next to a campfire while smoking at twilight

[smoky oak]

Same, gay disembodied floating water color wolf head, same

I relate 100%

[sterile temple]

getting handsy

[neon crystal]

This has gotta be one of the best I’d seen so far damn there’s so much going on!

[sterile temple]

the lord works in mysterious ways

[neon crystal]

I don’t think we’ll witness something like that again! STOP TRYING! lol

[smoky oak]

I actually almost gave me a stroke reading

[wispy nest]

yea...same here

[sterile temple]

word soup

[marble wasp]

Portrait of a Young Boryokudan in Safe House, circa 2023. AI portrait made with Stable Diffusion, epiCRealism modeler on Civitai, and SD++ Karas.

[smoky oak]

@dense tapir have you seen the new talks about sub nanometer CPU fabrication?

IMEC laid out their roadmap, and it includes a 0.7nm fabrication, of which they are already researching.

The "A" is 0. Number followed NM

So A7 is 0.7nm, and so on

Very very curious to see how they plan to do it

[sterile temple]

Jesus heads to the club

we've all been there...

[smoky oak]

Sounds like a British person with a thick accent saying "A, are we - we - we OK?"

Lmao

I could picture it being a song lol

Oh shit, cluster headache

[sterile temple]

I just put random song lyrics I'm listening into and you get funny results

[smoky oak]

I love we we we gock the O

My jam

[dark flare]

Are you sure that isn't an actual song?

I did this yesterday, thought it was pretty sweet... the prompt was:
asdl;jas dkljasdkla jailogf jhaoig8s0j h80gh 24389ghnw8 vh sd89fh38ghef8wghe09ghids njdc

[umbral lava]

Lexica + stable diffusion

[sterile temple]

https://youtu.be/l1uMxrN4ePU?t=12

sometimes you get funny, sometimes you just get nice pics

[dark flare]

Did you find a way to download Lexica's model(s)?

[smoky oak]

shittttt, this cluster headache is coming on soooo strong

losing vision in my right eye rapidly

[dark flare]

Well maybe ....... just silly talk here .... step away from the computer 😅

[smoky oak]

Thats not how clusters work, and I have been off my PC for well over an hour

haven't had one in months

hopefully I took my pills soon enough, or its no sleep for me tonight

[smoky oak]

I-

wow lmao

[dark flare]

ah! Sorry 🙂 shoulda googled it 😇

[smoky oak]

There is nothing I can do to combat it, its just pain until the pills kick in

weird thing is, its on my non prone side

[sterile temple]

I read the other day that doing handstands, or swallowing with your stomach above your head can help with heart burn and gerd, for those out there who have that

[smoky oak]

I do have that, but luckily I do have some eating habit changes and omeprazole that combat that for me

I have a ton of medical issues, just how its always been, but its been a very long time since I had a full on cluster

gonna have to keep pills on me, cause I am probably entering a new activity period if thats the case

[dense tapir]

TSMC is already switching over to AI because they could barely get 3nm. Honestly, getting that low the electricity will jump the gap with ease so just because you can do it doesn't mean it actually will work in the real world.

[smoky oak]

We'll have to see

IMEC is the best in the world when it comes to chip fabrication, so I would tend to take what they say pretty seriously 😅

They already have world first 13.5nm light emission sources from Tin plasma that cost some 40 million dollars to get a working prototype

Which is what is doing all of the fab available to the public right now

They supply basically all chip fab equipment

well, all that matters

[dense tapir]

As I said there comes a point the gap between the traces is too small and the shit will jump over as if there was no gap

Where that gap is I don't know but we are approaching it. Might need some sort of new insulator between the traces.

[smoky oak]

yeah, but I will take the word of the best in the world who know this better than anybody else, personally heh

They supply TSMC with their machines, so if TSCM can do 3NM, imagine how many several steps lower are being tested by them already

They are also already talking about a new way of handling smaller chips

[dense tapir]

I take no one at their word I just know science. Be like trying to contain an atom which would require sub atomic structures

[smoky oak]

GAA NSFET

[dense tapir]

the other thing is if they pull it off can we even afford the damn thing?

[smoky oak]

now that I have no idea about

but they did say that they are researching 3D matrix based chips with several magnitude faster connections than we have right now

[dense tapir]

already the price for 3nm is through the roof and 2nm will be almost 100k USD per wafer

[smoky oak]

I believe they said at 1.4nm, they are moving to 3D matrix's, which will require a huge rework of computing chips in general

[dense tapir]

yeah

I see them going up rather than out. which helps that gap jump

ing

[dense tapir]

AMD already has a rudimentary 3d chips

[smoky oak]

They commented on that saying that their implementation can't be compared to real 3D chip fab

GAAFET is already a proven process, so it will be interesting to see how that goes

[dense tapir]

Well, I am done getting hyped over anything in science because at the last minute it fails, or goes quiet or gets out and fails. I now take a stand back to wait and see. Cool if it works

[smoky oak]

@dense tapirok, so apparently, GAA NSFET, FSFET, and CFET are already real process that can achieve those sub NM transistor sizes

Only limit is making them more practical and more reliable

[smoky oak]

FREE MEN!?!?!?!

WHERE!?!?

[sterile temple]

Thought you'd like that 😄

[dense tapir]

there is the rub the reliability part

[smoky oak]

FREEEE MEMM

[sterile temple]

oyo!

[smoky oak]

OWO

but

with a Y

O Y O
V

[smoky oak]

ahhh, GAA NSFET are all side insulated

Looks like they all are after FinFET

[sterile temple]

please

[smoky oak]

FREE ME DOM

Dom, you need to free him

[sterile temple]

free him please

[smoky oak]

MORE FREE MEN!?!?!?

A DREAM COME TRUE!

[sterile temple]

it's so freaking exciting!

[dense tapir]

Athlete's foot?

[sterile temple]

a tee hee ~ 🤭

[dense tapir]

I see controlnet is on version 200 now

[wispy ether]

https://tenor.com/view/kek-zoil-gif-22981488

me trying to train lora and got no hope

[dense tapir]

Notice a full moon YET no light on the ground as a real full moon would be, lol

I hope SNR takes care of that

[sterile temple]

A camera would struggle to pick up the light off the ground when the moon is so bright?

[dense tapir]

Nope, not unless directly pointed at the moon as I have seen the photos out in the desert on moonless, and full moon, nights. The desert is so lit up it is like daylight.

What really sucked is when SAI took away the other clip model because it had all kinds of photography settings it knew and it knew them DAMN well too. That I miss greatly with 2.x

[sterile temple]

dagnabbit, looks like the dynamic prompt extension is broken again after the A1111 update

[tired peak]

great!

[dense tapir]

Been broke and I reported it. Finally was told to roll back. Screw it

Problem is how they get random numbers changed so my wildcards went bye byye and I use them to check with

[sterile temple]

I have to reload A1111 if I want create a new wildcard file

[dense tapir]

it sucks

I reported it 4 or 5 days ago

[dense tapir]

Now try that natively because I see the contrast fix there

[sterile temple]

I'm not using the lora

[dense tapir]

nfixer is the same but as a TI embedding

[sterile temple]

no embedding either

[dense tapir]

even emad admitted it can't be done with stock 2.x

a lot of models now train with noise offset which is exactly what I see in your above

[sterile temple]

maybe that then yeah

[dense tapir]

a lot of models are already starting to be trained with SNR

Joe Penna added it a week or two ago

[sterile temple]

It's a mix I made, but I got the photorealism by subtracting an anime model from a mix. Can't exactly remember how I did it because I was just playing around

[dense tapir]

As much as I have read on mixing I don't get great results with faces as it always tears faces up

[sterile temple]

I have tried to improve this model I mixed. Must have tried 100 different things, but I keep going back to this one

I have the one that I subtracted the diff between BRA and realistic vision, but that's better at doing traditional art stuff, not so much photos

[dense tapir]

I don't get how others do it so easily

[sterile temple]

my latest wildcard, band names

[kind quartz]

it is one band?

[sterile temple]

all different

[kind quartz]

is there iron maiden or led zeppelin or deep purple? Just asking not expert on that sort of music.

[sterile temple]

yes 🙂

[sterile temple]

cradle to grave?

[dense tapir]

yes

Now undead things

[sterile temple]

get in ma belleh!

[dense tapir]

PIE > CAKE

[sterile temple]

apricot, apple?

[dense tapir]

Rhubarb

Peach is good too then Apple

[sterile temple]

don't think I've ever had a peach pie

[dense tapir]

Peach cobbler is good too which is just a broken up pie, lol

[sterile temple]

nom nom

[dense tapir]

Cheesecake, oh my

[kind quartz]

never tasted this one 🙂

[sterile temple]

what is it?

[kind quartz]

Upside down cake

english specialist

[sterile temple]

don't think I've had that, unless it's the one with pineapple?

[kind quartz]

realy dont know. Know it only from some TVs things.

[dense tapir]

I have pineapple upside down cake for my bday each year.

Made my late wife a peach upside down cake for hers once.

[wispy ether]

i had a stroke reading that

[sterile temple]

g'nite