🏞｜general-with-images

[ripe cedar]

My power supply is 350w and the 2060 12gb I want says it should have 550w from the manufacturer. Do you think I can get away with it? The bottleneck calculator shows the CPU and 2060 don't have bottleneck with each other so I thought maybe the power supply would be okay

[dense tapir]

no

[ripe cedar]

😦

[dense tapir]

absolutely no

you will trigger undervoltage protection, if it has it.

My 1060 uses almost 150-170w when generating and the 2060 is slight more power hungry

[ripe cedar]

Yeah I just saw a YouTube video showing a blowup. Don't what that to happen. There's one on eBay which might go for a good price but no point of I can't use it I suppose

[dense tapir]

Like me if I had the cash for a 4090 I would still need 200-300 more for a 1200w psu

my limit is around a 4070/4080

850w psu

Oh, I account for spikes

[ripe cedar]

I see, it's a bit like upgrading a car; not just the parts but the supporting parts

[dense tapir]

yes

like trying to start a car off a 9v battery

[ripe cedar]

The PC is a MSI which doesn't have much physical space either. Maybe I'll have to just upgrade the whole thing

[dense tapir]

a good rule of thumb is 200w for your system alone.

[ripe cedar]

But I'll check power supplies and dimensions now and maybe I can avoid changing the entire desktop

[dense tapir]

550-650 isn't all that expensive but beware really cheap brands

they simply have no ability for spikes

[ripe cedar]

Gotcha.

[frigid vigil]

Kinda the exact opposite for what I was intending—horror—but it's still neat.

[ripe cedar]

Is it possible to combine the depth info as well as the textures on the 3d objects themselves in image generation?

To make the result closer/more accurate to the intent

[dense tapir]

depth yes, textures not really

at least as I think of textures in a 3d program kinda way

[ripe cedar]

How about the depth+the render?

Rather than the texture itself

I suppose that's image + depth

The 2060 12gb I was looking at went for 190 USD on ebay. Good price but I didn't buy it, ugh

[dense tapir]

190 is a bad price for it because it is used, probably used in crypto mining, and darn near the price it was when new.

maybe half but way too old to still be 190 though thanks to Nvidia the market is far overvalued

[ripe cedar]

Apparently it was not used much according to the description. Who knows...there will always be another deal anyway

I wonder when a comic will come out that uses ControlNet and other new stuff. Time is ripe to make one imo

[dense tapir]

Yes, I think so are darn close

[wispy nest]

messed up, but still kinda cool

[ripe cedar]

I did rigged kitbashing by just getting a bunch of rigged heads and rigged bodies, then just parenting the controls together. Then applying a line-art generator and toon shader. Slapping on an outfit and one-click applying weights. But with SD perhaps it can avoid much of the 2d post processing by applying shadow and cel shaded clothing folds; 3d isn't great at that

[dense tapir]

You can

[ripe cedar]

With that 3d technique and SD, it could make production much quicker than usual

I think 3d artists get a lot of benefits out of AI. 2d artists complain about AI the most.

[dense tapir]

2d takes less work imo

shading and shit

[ripe cedar]

It depends. If the assets are recurrent, 3d takes less time, but more upfront time cost.

2d can make short works faster

[dense tapir]

I mean by hand not computer

[ripe cedar]

O ok yeah

[dense tapir]

getting that 3d look by hand is a pita

time consuming

[ripe cedar]

For comics, typically people don't like the 3d look. But the 2d disadvantage is that it takes forever to animate.

[dense tapir]

Not sure why that is I mean a comic is a comic.

[ripe cedar]

People often want to see the human hand behind the art in comics. They like the consistent imperfections and organic look

[dense tapir]

meh

[ripe cedar]

I'm literally lol now, we're complete opposites. I love the 2d look

[ripe cedar]

Can blender 3d models bodies be highly stylized and then used in ControlNet to make the output highly disproportionate as well? Even if it's in "realistic" style? Ie an inhumanely muscular character

[dense tapir]

I don't know as I never could get with that blender monstrosity as I was more 3dsmax, Maya (both before Autodesk took them over) and Cinema4D.

[ripe cedar]

I only use Maya on the high seas. I didn't like blender either but slowly warming up to it. Some of the tools in it are really powerful. But for toon renders I use Maya. Pencil+4 is a plug-in for Maya which anime studios use, but it's coming to blender this year. Then I might switch to blender completely.

[static tusk]

use the CN depth model to capture body shapes

you can actually use openpose + depth for more accuracy

[dense tapir]

I couldn't stand the workflow of Blender at all no matter at what stage of it, or my life, I tried it.

[ripe cedar]

This is a huge advantage for stylization and consistency. I just wonder how background shots from different angles will be consistent. 🤔

[static tusk]

consistency depends on how well the model or your prompts can reproduce consistent output

it can be random

[dense tapir]

btw, to help get rid of some of that don't use xformers

xformers is nice but an abomination at the same time. I suspect when we get temporal coherence it will have to go bye bye

[ripe cedar]

If an animation is shot at a high fps then perhaps taking out the inconsistent frames while still having material enough to work with, would be a more useable way. Same for sequential art as far as making more generations go and then cherry picking

[dense tapir]

Whale shit, lol

[ripe cedar]

Yeah it would be a lot but oftentimes it's the same areas that give it away. For characters, their hair

But I think when AI gets good, people can make movies like studios of today. But the actual companies will make mind blowing stuff.

[dense tapir]

I don't care much for illuminati 1.1 as 1.0 was much better imo

[ripe cedar]

Makes sense. The illuminati is a regressive organization after all. Can't wait for 3.0, it will probably look like toilet content after a large meal

[dense tapir]

Hollywood is well past its prime as all this woke bullshit shows. They aren't doing very well now.

[ripe cedar]

That's good. The masses should be the ones leading content imo

[dense tapir]

I agree

[static tusk]

ID 1.1 gens follow your prompts more accurately and has more details, I find.

[dense tapir]

Once temporal coherence comes in your bedroom can crank out a full feature film in under 3-4 months. fan 15m films in less than a month.

they make all my prompts look uglier

[ripe cedar]

That's why I want to learn more about AI. It's still early days. Widespread adoption isn't even close. Gotta get the advantage, but being passionate about it is the main thing.

[dense tapir]

Well, we shall see because I firmly believe the governments of the world will clamp down on all this the best they can.

1.0 then 1.1

[static tusk]

yeah thats the issue with females I mentioned the other day, using 1.1

[dense tapir]

I missed it

[static tusk]

stretched and thin

[dense tapir]

yes, like devils

overly saturated and dark

[static tusk]

theyre working on that

[dense tapir]

aye

Not sure why 1.1 gives them pointed witch's chin

the eyes too

all like that I found

[static tusk]

play around with the negative nfixer etc, there are combos that work best

[ripe cedar]

There are restrictions being put in place already for AI, depending on the country. But I think it'll proliferate faster than they can clamp down. Genie out of the bottle and all that

[dense tapir]

can't stop it but once caught it goes real bad for you and I hope this doesn't go underground against the law.

[static tusk]

I guess restrictions will be on the sale of AI art, not so much production

[dense tapir]

1.1 dudes are fine. Seems weird since 1.0 and 1.1 should have been the same dataset just one had the noise contrast fix.

One thing bad in 1.1 for either gender is the eyes

with nfixer

[static tusk]

this was posted in the ID server

[static tusk]

multi diffusion upscaler works pretty good. just upscaled an image to 7k x 5k, but cant get it to work with t2i and controlnet

[smoky oak]

I cannot believe how well my subject LoRA's have been coming out lately

[smoky oak]

looking very nice :p

[lament vessel]

#👥｜roles

[crystal delta]

FAQ: I'm new here! How do I get started and is there a bot?

Welcome! There's currently no bot on the server to generate your images. Start by heading over to #1072220168534642768 to get yourself situated and help find the channels you are looking for! Please make sure you review our #✍🏼｜rules-and-tos and feel free to assign yourself some #👥｜roles as well! Answer any questions your may have at our #1072229020520947753. There are many ways of accessing Stable Diffusion, take a look at #1080946152318443610 to start your journey!

[wispy nest]

Didn't want cat ears there tho 😦

wil probably have to put it into negative prompt too with this model

[smoky oak]

Should have been fairly easy to remove with inpainting or negative prompting

oh, you're not negative prompting?

[wispy nest]

it's fine , it's just one image, I have plenty more to check later on, still generating 😄

[wispy nest]

I do, I just didn't expect it to do cat ears

[wispy nest]

new model , didn't know it will do that

[smoky oak]

Ah, alright, I was gonna say lmao

Not using negative is asking the Ai to screw you over haha

[wispy nest]

yea...

[smoky oak]

Like the guy on here who was wondering why his results were bad

his entire prompt was "timber wolf"

no more positives, no negatives

just

timber wolf

Not making fun of him, BTW, I am just meaning that some people need to really add some details haha

[wispy nest]

They'll get it after some attempts, it's fine

[smoky oak]

Yeah, only thing was they were saying the AI was trash cause of it

[wispy nest]

any ideas , how do you call the thingy, when pieces of cloth intertwined with hair?
Like...straight lines of fabric in a hair

[smoky oak]

@celest geyserThanks, after reading your name, I need to pee now lol

[celest geyser]

Don't worry I'll go with ya

Gotta live up to the name after all

[wispy nest]

no memes intended.

[cobalt ether]

Guys can i generate images in this server? And how to

[wispy nest]

#1072220168534642768

We don't have a bot on disc to generate images, you download it locally and generate images on your PC or use other services

[sand pike]

Ribbons?

[wispy nest]

kinda , but that's too wide 😦

Can't even google up what I want lol

[sand pike]

Post an image (any image) of the thing you mean, and I'll try to come up with the right word.

[wispy nest]

Looking at my images I think I've got better idea of just mixing 2 colors , or doing just a little bit of hair in a other color, idk if there's a name for this thing

[sand pike]

That really is a hair style thing... The basic term is a "streak", but you might also try "dreads".

A phrase to Google if you're looking for more elaborate styles (and their associated terms) is "cyber gothic hair".

[wispy nest]

I'll see tomorrow what this will make on SD

looking at google images , yea "hair streak" should be good enough, but mixing black and white hair and not getting monochrome image might be challenge on itself lol

[sand pike]

A different way of incorporating another hair color is "ombre" (horizontal rather than lengthwise).

[strange fox]

https://media.discordapp.net/attachments/981697547171151953/1083305957465002055/FB_IMG_1678067731153.jpg 3d model of a face on a white background 4k details

[sand pike]

Simply sticking "monochrome" or "monochromatic" in the negative prompt usually works for me.

[wispy nest]

yea , I have greyscale, monochrome on my main negative

[spice flower]

Nice, which model ?

[spice flower]

I would like to know the model for this one too please, it look insane 🙂

[wispy nest]

oh...it's been some time...

One of this 3:

chilloutmix_NiPrunedFp32Fix
realdosmix_.safetensors
ddosmix_V2.safetensors

Should be easy to find in https://civitai.com/ by start of the name

[spice flower]

Thx !

[wispy nest]

nvm my deleted messages, discord sent me to wrong thing 😄

[glacial token]

heres another dragon for you all

[green plover]

First time I have printed anything out of SD

[smoky oak]

New Mr.Wolf LoRA is awesome!

works with described expressions as well!

[stoic pivot]

A quick Kilo test:
The prompt is, "anime_render, cute bunny robot, forest background" and then I added 8K or 4K to the end, this do not affect the quality of the image at all, but 8K altered the bunny, but same thing did happen when I used 8M, L8 and 42069k.
My conclusion is that this is useless tags that does not do what folk hope or want the tags to do, the reason is that you can not have 4K that is 512x512, 4K and 8K and so on referee to pixel density and not quality.

Model used: Stable Diffusion v1.5

[stoic pivot]

AI is strange and every word we use is translated into a number that in some way can generate an image, so we can insert what is total nonsense and still gat a image.
Prompt: "hyper realistic, Hasbro cartoon, cinematic light, 8K, blender render, best quality, 4k, RAW photo, drawing made by a master of art, super sharp, trending on artstation, cute, kawaii, extrem detailed, photorealistic anime, pixar style, Unreal 5, Pruff and Krutt."

[stoic pivot]

There is AI creators out there that can use AI to create great images and some of them even have an idea what they want to create and then spend time on creating and edit the image to be superb, I see that as creativity. When I make fun of make "art" with silly prompts, that is more aimed to me and how I used AI and I later found out what work and not. I still use all kind or words to experiment, but I also love to make AI give me random images that can be very fun, creative and a source for inspiration.

For me the AI technology is new and we shall test, and when I say that things like 8K and "trending on artstation" do not work, that may not be true in the future due to how fast AI change, I often just add random letters and words just to see what is happening. So keep experimenting and be happy with the images you get and like.

The silly promt used for the image above was: "hyper realistic, Hasbro cartoon, cinematic light, 8K, wide, blender render, best quality, 4k, RAW photo, furry gimp, 50 mm, masterpiece photos drawing made by a master of art, super sharp, trending on artstation, cute, kawaii, extreme detailed, photorealistic anime, pixar style, Unreal 5, by Salvador Piasso".
And it gave result that I thought was fun and creative.

[ripe cedar]

Has anyone tried using a close up high-poly head sculpt + depth to generate images?

[smoky oak]

@stoic pivotI continue to feel that your experiments are far too narrow/non diverse to draw any conclusion on the effectiveness of specific prompt words/phrases. I get what you are trying to do, and its cool for personal use, but I don't see any benefit in people adopting/abandoning words you have tested, just based off of the lack of consistency between variables

sadly, its just something far too diverse to run on a large scale and come out to any meaningful outcome, and it will always be massively circumstantial

I appreciate the effort and commitment, really. Its something i occasionally test on small scales as well, but I don't want you to continue barking up the wrong tree if you feel this information is widely applicable, cause there are just too many variables to adjust for sadly

as you can see, some words don't have too much of an affect, but it is still noticable. You can see how both "realistic" ones, have a sunset and more contrasty fur, where as both "detail" ones have a more drab sky, and softer fur, and the last 2 have a very big effect, completely changing the image, and looking different from each other

[stoic pivot]

My point is often that some words are to vague to have any real impact on the image, folk spent time on "quality" rather than content, if some want a "sharp" image it is better to point creators to postprocessing in drawing apps rather than spend half a prompt to get a "sharp" image.

I am a bit off today, I may misunderstand and so, so I may not try to debate or test today and just be silly.

Master prompt = "Gurka the mazurka".

[smoky oak]

you can even see here that worst/best/high/low quality do have some affect. Specifically the best and high look a little more punchy to my eye

and interestingly, worst is the most off and has a completely different background

[stoic pivot]

Try it also with "nice quality", "Coffee quality", "My quality" and "Master quality".

[smoky oak]

another interesting thing is that best and high have signatures, while the other 2 don't

you can even see it a bit here, where high and best have white details and more dramatic lighting, but worst and low are flat

(different model)

[stoic pivot]

And what is quality, that look like a medium quality bunny on low quality snow 😄

[smoky oak]

so that 2 models that have shown consistently better results with high/best

albeit, slight

but a lot of slight changes can make a huge one

get 10 words that add minor differences, and you get a considerable one, which is why I don't see the point in focusing on edgecases, especially when they are not consistent between AI's

oh, see, this model latches on hard with hyper realistic

very interesting

let me try the other one

this one has a very different response, latching on to hyper realism instead

[stoic pivot]

How do you do those images, it is the XYZ script

[smoky oak]

yes

I put the word realistic, then use XYZ with S/R

it finds realistic and replaces with what is next

the point I was trying to make is that every model is trained differently, so what if you told somebody that Hype Realistic added nothing, and they missed out on a wildly different result because of it. Its all very circumstantial

I actually really like abstract realism here

[stoic pivot]

"Hyper realistic" is a word that work, but folk often use it in the wrong way, it is not the same as "good realism" but more towards "exaggerated realism".

It is often called "hyper realism" when movies have a "up and down! in space and things fly like aircrafts here on earth but in space, like in Star Wars, that is not in say a bad thing, but it is more than it would be in reality.

[smoky oak]

ok, so abstract realism really does make a massive difference. Very interesting

[stoic pivot]

Those are nice test images.

[smoky oak]

WOAH

sampler makes a huge difference!

Now completely different images stand out

how weird

[stoic pivot]

But try a thing, use same as above but add an unrealistic element to it, "robot bunny" and see how it handle to make something unrealistic to realistic.

[smoky oak]

@stoic pivot

it looks like basically every word had different outcomes, and both samplers had different outcomes too

@stoic pivotHere is an extreme test haha

[stoic pivot]

Thank you for the tips, I will use that in some of my test, this one was to be a bit silly and troll the AI by asking for realism in a prompt that is aimed to be unrealistic.

Prompt: "anime_render, cute robot bunny, normal"

[smoky oak]

Here is a fun one@stoic pivot

[keen kindle]

geneate a donation poster, the donation theme is a one-to-one aid education activity.

testing

[stoic pivot]

Ty.

[smoky oak]

No worries haha

You can post it with it censored, I'm sure

[keen kindle]

A painting of a cute goldendoodle in the sky, wearing a suit, natural light, with bright colors, by Studio Ghibli

[stoic pivot]

I hope this one is okay.

[dry crow]

Haha xD

[stoic pivot]

Clyde blocked the image that was censored.

[smoky oak]

Wow, but not the uncensored one lmak

[sterile kiln]

Looks "The Stand" (1995 mini serie)

[glossy herald]

true ! it was the luck of the draw, just a part in my PoW work, but you're right, it gives that feel !

[stoic pivot]

I had to test after I got the tips that "Anything V3" was trained on "best quality" for the aesthetic, I will do two images that do several different "quality" so we can see if "best quality" in some way is better or useful.

First prompt "a hand, ", Second prompt "Anime_render, cute robot bunny, forest background, ".

[ancient trout]

Smart baby, cartoon, profile pic

#1072220168534642768

[steady matrix]

Hi

[wispy nest]

Looks nice , perspective \ proportions look off tho 😦

That triggers my Alice in the wonderlands syndrome, lol

[dense solstice]

trending on PH, never thougth about that one 🤣

[stoic pivot]

Internet broke me 🙂

[wispy nest]

But seems like SD actually knows what is it, interesting

or maybe it just took first part

[dense tapir]

Larger than the car, lol. I'd ride both if given the chance.

[dense tapir]

I better hurry up and jump on this right now!!!

[wispy nest]

-imagin

hello ¯_(ツ)_/¯

[sterile kiln]

Yes x) you guys are right, I was so concerned about the quality of the image that the proportions didn't realize it, haha

[dry crow]

Insurgency, Total War Three Kingdoms, Civilization V,

[smoky oak]

assuming you are outside the US?

[smoky oak]

Glad to see you abusing the xyz plots now haha

[steep shale]

care of scammer

[soft quiver]

https://giphy.com/gifs/2013-jared-leto-12ANYkldS5ZHsA

[pearl sleet]

Hello, recently uploaded my first Stable Diffusion automatic1111 extension titled "Corridor Crawler Outpainting". It takes an image from the user, shrinks and pastes it into a prefab image, runs it through depth-to-image, and then repeats the process for as many steps as you desire. This is my first serious programming project and there's a lot of features I'd like to add, but it currently works and is public here: https://github.com/brick2face/corridor-crawler-outpainting

[dense tapir]

That was tongue in cheek and notice it is in USD. What a fucked up, utter damn shit price and damn all who would pay that for it to help force all the damn video cards to go up in price. As long as people will pay it will never stop. I am pretty burnt out on all this anyway and the stupidity of the current pricing didn't help. Probably heading to Intel with their next gen ARC and to hell with this AI stuff. YMMV.

[smoky oak]

I thought it was Canadian dollars, which use the same symbol. Also, if it was canadian dollars, that would have been a reasonable price, so thats why I asked.

that would have been $1615 USD, which is a very reasonable price USD (compared to what NVIDIA values their GPU's as, at least)

[smoky oak]

This is really cool! I have no idea what i would use it for, but I appreciate the work!

[dense tapir]

Naw, even in CA dollars it would still be crap. Honestly, all of their cards are 25-60% over priced and the 4060 is utterly a damn sick joke. I really have a strong dislike for Nvidia and their CEO and just glad he, and I, would never be in the same room together.

[smoky oak]

I agree entirely. I wish that Ampere (3000) cards wouldn't have been so VRAM cucked, and then to go and do lovelace even worse in terms of VRAM to performance is insane IMO

[dense tapir]

I have nothing but utter rage at what they are doing, AMD is taking advantageous too, sadly.

[smoky oak]

I don't like NVIDIA either, but sadly they are the only really good option for AI, as AMD has trash drivers/support, and intels GPU's are a joke

[dense tapir]

Yeah, since when was a 3060 12gb then magically all after it were cucked?

[smoky oak]

it was because of the VRAM buss

[dense tapir]

Intel is promising. AMD drivers have always been shite

[smoky oak]

they could either do 6GB, or 12GB, so they went with 12 to make them look better than AMD's offerings

[pearl sleet]

Thanks! Some ways to use it in game development would be for an endless runner game, a dungeon-crawler RPG, or even a rail-shooter. Eventually I'd like to add a script to it that can make loops like this video, but I lack the math knowledge to do it myself. Video created using the prompt "Dungeon crawler hallway" here: https://replicate.com/arielreplicate/stable_diffusion_infinite_zoom

[dense tapir]

3060 12gb, 3060ti/70 8gb. wth?

[smoky oak]

and it sucks, cause not only does the 3060 have more VRAM than the 3060ti (stupid as hell), but its also WAY slower for things like AI and rendering, which are the only things you would need that much VRAM for, as that GPU is far too weak for high res gaming

[dense tapir]

tbh I see where this is headed and <16gb is a bad buy unless you get it for super cheap

[smoky oak]

agreed

my uncles 3080ti Laptop has 16GB VRAM, yet the desktop 3080ti only has 12-

[dense tapir]

I looked at the pricing for last gen cards and they are almost as expensive as ADA cards.

[smoky oak]

I got my 3060ti for $400 just a couple days after it launched, was an insanely good deal lol. It was worth it IMO

[dense tapir]

It is the same as only having one car brand like it was in the old Soviet Union. Now I am not kidding you paid for it and had to wait up to 10 years for it to be delivered.

An old joke Ronald Reagan told about that was something like "A man went to purchase a new car in the Soviet Union. He paid the money and the dealership said it will be 10 years for delivery where he then asked on which day to be sure his schedule wasn't busy that day".

[smoky oak]

but now I am cucked with a trash CPU lol

[smoky oak]

lol

[dense tapir]

That is where this is headed and you can bet Nvidia has their eyes, and ears, out there knowing AI/ML is the next crypto boom for them.

5090 5k 32GB

sad thing is it would still sell 500k-1m units

[smoky oak]

yeah

if there was a better option, I would switch, but they have a monopoly right now sadly

[dense tapir]

I saw people (saudi princes and stuff) buying videocard for 20-50k during crypto boom

I saw that I knew this will never come back to reality then crypto went boom for mining and here this arrives.

they lowered the cost of the 4060 before release it is such a sick joke and even still it is MSRP of 500

what I have read is that they are trying to herd consumers into the 4080.

Not working so far.

Even going so far as limiting 4090 but still not working, lol.

[smoky oak]

hey, at least there is only one 4080 now :/

[dense tapir]

As I said I would rather not have than get bent over but that is how I always have been.

for now the ti is due

I know the last 16% of the die is probably going to the 90ti

[smoky oak]

I'm happy with my 3060ti right now

its crushing my Ai gens haha

[dense tapir]

If all I did was gen I wouldn't mind I could grab anything but fun for me is in the training.

Heck, even AMD does gens. I am not sure about ARC though.

[smoky oak]

I know that even 6 months after launch, arc is still a joke

[dense tapir]

Now they are cheap and where all these should be at in pricing, but the moment it gains marketshare BOOM. You know it.

[smoky oak]

drivers so bad they make AMD look good lol

no DX10 or 11 support

[dense tapir]

Actually, not so from the reviews I have seen. They are upping their driver game

[smoky oak]

They can only do so good, and it will never be as good as AMD or Nvidia on their current GPU's, cause they are missing physical hardware

[rustic magnet]

Gid

Привет 👋

[dense tapir]

well, don't be like one of those 1.5 vs 2.0 types as this is really advanced for Intel as nothing before was worth a damn. Give them a couple of more gens and watch. I mean FFS they are Intel so if they can put the chips in cars, and into pcs I would think they could GPUs

[smoky oak]

or my bad, Arc has no direct x 9 support

[dense tapir]

yes, 9 is emulated

[smoky oak]

yeah, i don't doubt they can do it in the future, but this gen of GPU's is objectively flawed, and is very weak for the price

[dense tapir]

They will surely learn from this or fold

[smoky oak]

I dislike intel as much as NVIDIA personally because all of their extremely misleading claims that make AMD and NVIDIA look like saints lol

[dense tapir]

Considering AMD works hand in hand with Nvidia to screw us too I have no faith in either so we need a third manufacturer

[smoky oak]

they make completely false claims all the damn time, and never catch flack for it

[dense tapir]

I don't give two fucks if Satan rose up to offer it as long as it worked and was decently priced. I hate Intel but I am not brand loyal at all.

Now AMD's driver division sucks. Sucked when it was ATI and never stopped sucking so they are now getting into the AI/ML stuff. I doubt they can pull it off.

Great hardware but god their drivers

[smoky oak]

yeah, it is a huge shame

[dense tapir]

I remember about 12 years ago games would be shaded differently than Nvidia gamers and crashes to desktop and all kinds of weird shit. I know cause I was one among those gamers.

[smoky oak]

I am dreading giving my money to both Nvidia and Intel when I buy my laptop

[dense tapir]

The two biggest devils, yes

you know we can squarely blame all the crypto miners for this because they had it so good they just can't let the gravy train go.

[smoky oak]

sadly for what i need/want, they are the option to go with. AMD just falls flat for both criteria for me, as much as I would prefer them

[dense tapir]

I am a Ryzen guy since 2017 so I have no want of an Intel system these days, although I just had a kick in my privates see the new mobos. FFS a budget AM5 board is now 200+ USD

I am hoping that when I upgrade in 2026 AM5 boards will be settled and go back to reasonable.

[smoky oak]

The two biggest things I need are ray tracing/Ai performance, and the fastest single core possible, of which NVIDIA and Intel are the options

[dense tapir]

budget cards are 500-600, budget boards are 200-250. I mean come on.

what do you need single core speed for? I know Intel has always had that.

[ripe cedar]

I found a desktop PC for $600 with a 12gb 3060, i7-10th gen, and 16gb RAM. I'll upgrade the RAM but I think it's a decent deal

[dense tapir]

grab it

the card alone new is 400

[smoky oak]

I would pay $250 max for a 3060, IMO

[dense tapir]

just saying what they go for

[smoky oak]

but for that whole deal, thats nice

[ripe cedar]

Yep I'm going to meet the guy tomorrow. I'm glad the 2060 gpu I was looking at yesterday didn't pan out; this PC is upgraded I'm other ways over my dinosaur so I'm excited

[dense tapir]

For me I am tired of having to train on colab it is absolutely death by a thousand cuts

[smoky oak]

lord knows I need a CPU upgrade lol

but I am done with desktops

[smoky oak]

yeahhh, I bet that is no fun at all

[dense tapir]

Never ever a laptop for me for 3 decades now.

[smoky oak]

audio work/sound design/music production

[dense tapir]

yes

I hate laptops and tablets make me want to skeet shoot with them

[smoky oak]

So I need my raytracing and AI for my renders and AI I do, and I need single core for audio work and music production. So yeah, NVIDIA and Intel are really my only options sadly

[ripe cedar]

I use a 2014 Samsung Galaxy 10.1 for reading in portrait mode, that's it. I only needed it for the resolution which is the same as modern ones.

[dense tapir]

I am looking at the new threadripper due within 1 year now. 96 cores 182 threads

*192

[smoky oak]

I actually just recently started using my samsung tablet as an extra display for my desktop lol

dedicated spotify monitor

and the touch screen works, which is fun

[static tusk]

got a spare 5k for the CPU alone?

[smoky oak]

^

lol

probably considerably more

[dense tapir]

Sure, let me just grab it out of my asshole

[static tusk]

threadripper sure is exp, even the current gneration

[dense tapir]

fact is a lot of the consumer grade Ryzen chips are about to go up in c/t counts I read

[smoky oak]

Yeah, sadly TR is kinda meh as of now

in some cases you are better off buying Epycs

[dense tapir]

I hadn't paid attenting and I thing there is 24/48 now or it was 16/32 I forget

[smoky oak]

there is already 16/32

there has been for 3 generations now

[dense tapir]

people forget more cores beyond a certain amount for we consumers means jack

must have been a 24/48 then

[smoky oak]

windows chokes past 64 cores, and a lot of programs die after 16

you need special programs to use more in a lot of cases

blender can untilize more

[dense tapir]

yep, I can see that happening

[smoky oak]

cinebench has a 64 core max

[dense tapir]

I think days are numbered for windows for me anyway

not much on here do I use that is a windows exclusive and I can always dual boot or vm it

if I had to

W11 pretty much cinched its fate with me

[static tusk]

windows dual boot is a pita, win messes with everything

[dense tapir]

Naw, I have dual boot right now I just don't use the linux mint that much

what killed me was that I could not get xformers to install with it and I spent days. get it all working then it refused or get it and something else broke

all this shit is built on quicksand as a house of cards

Still, some quirks in the modern Linux I do not like

[static tusk]

its the reason I havent moved away from windows. many things in linux are like that

[dense tapir]

they added some sort of security so this audio equalizer pak something or other could not startup at launch of Linux and no way around it. Screw that I don't like having to manually start things like that.

another was the gpu fan (like afterburner). Can't do it at startup either

I thought Linux was supposed to be the OS to allow you total freedom to mess it all up or not?

Used to be.

@smoky oak One thing that kills me about colab is it uses old Python (finally, yesterday, updated to 3.9.16) it only gives me 2-2.5h per day and when I run a training program there goes 10-15m so count on 2h. T4 is so slow too so a 3090 10m training is 34m for me.

[smoky oak]

jeez

my 1000 step trainings are like 5-10 minutes on my 3060ti

[dense tapir]

I am always in a rush on it and can't just enjoy anything.

[smoky oak]

yeaahhh, I could imagine

[dense tapir]

training in what?

[smoky oak]

i enjoy LoRA's cause I can do them at my own speed

[smoky oak]

kohya ss, personally

But I learned a new wayyyy better way to train LoRA's, at least different from my old way, and it needs wayyyy less steps

[dense tapir]

well, lora with locon 1800 steps is 30m for colab.

[smoky oak]

and the results are extremely good

[dense tapir]

a new lora type just emerged, btw

[smoky oak]

old Na'vi LoRA< 49 minutes training with 8 epochs

vs my new one that took 6 minutes and 2 epochs

safe to say i was doing them wrong before lol

this toothless LORA took 5 minutes

mr wolf took 7

[dense tapir]

training on 2.1 is crap and honestly SAI should be ashamed to have ever released the turd.

[smoky oak]

whats this new LoRA you are talking about?

[dense tapir]

for generating it rocks for training it sucks

[smoky oak]

I am glued to 1.5 right now. Too easy too use, and the results are too good for me to go through a whole new headache

Here is my Scar LoRA

actual picture of him

Old LoRA (41 minutes)

New LoRA, 4 minutes

[dense tapir]

1.5 is very easy to train. I tried it 2.1 vs 1.5 where I resized to 512x512 and done.

[smoky oak]

and the way I made my LoRA's they play VERY well together with other LoRA's, or other base concepts

like here, where I mixed scar with a Na'vi

[ripe cedar]

There's this new training thing out that trains in like half a second, compared to dreambooths half an hour and such. I'll try to remember what the name was. I remember it uses way more vram than the others though

[ripe cedar]

Do you have to train characters naked because that way they have their body markings and injuries, as opposed to their clothes which would make the thing think the clothes are part of them?

[dense tapir]

https://imgur.com/f4KgCBb

This was the control net dev shaming SAI for 2.x

I wouldn't have been as polite.

[smoky oak]

I trained scar as wearing clothes, and the Na'vi as being shirtless

So I used it as Na'vi mixed with scar, and then weighted the Na'vi LoRA heavier, and specified shirtless, and it took scars face and applied the Na'vi affect on it, and then used the body

[smoky oak]

yeah, SAI sucks

[dense tapir]

new is called loha

it seems to overfit quicker and most are able to shave 1 epoch off their training

[smoky oak]

is that the 2.x one, or is that another new one?

[dense tapir]

a new one

[ripe cedar]

I read that Stable Diffusion took out 80 million images from training of SD3. I don't suppose that job could be manually crowdsourced...

[dense tapir]

I wish there was an exclusive 2.x trainer as it might actually be better

[smoky oak]

I am happy with 1.5 as of now

I wills tick with it until there is an objectively better option

however, I would really appreciate if SD would get a less awful text encoder, so it could actually give you what you ask for

[dense tapir]

I fear with the horrible clip for training we have and 80 million even less images ffs what a turd this will be. Hope I am wrong.

[ripe cedar]

I remember now, it was called ELITE. https://www.reddit.com/r/StableDiffusion/comments/11m5a9e/elite_new_finetuning_technique_that_can_be/

Supposedly a thing to train extremely quickly, at the cost of high vram.

[dense tapir]

with the current clip I love it for generating over 1.x BUT it needs new tools as the old ones were made for the old clip and that isn't working as well with the modern 2.x

[static tusk]

as long as the 80m are all low quality jpeg, it wont be a real loss

[dense tapir]

per emad I don't think that is what they ditched

[smoky oak]

oh I see, it won't run on any consumer level GPU, what a shame

personally, I find I spend way more time captioning/tagging than training, since my LORA's take single digit times haha

[dense tapir]

well, finetuning is the keyword

[smoky oak]

Time to use my external tablet display for my PC to rip images from netflix to train on lol

I can understand netflix not allowing recording, but come on, screenshots?

so stupid IMO

[dense tapir]

Netflix I have no want of since about 2018

[smoky oak]

I can assume, just based off of how anit literally every company you are lol

[dense tapir]

anti woke

to the nth degree

[smoky oak]

ok, now that is just a stupid statement lol

[dense tapir]

no, it isn't.

[smoky oak]

mkay lol

[dense tapir]

The last thing they had I cared to see was The Ballad of Buster Scruggs. Then they went sick and lost me.

[smoky oak]

I mean, you do you, I guess

or is the "you do you" sentiment too woke as well lol

[dense tapir]

Cuties is what tore it for most people and why Netflix began its downward trend

[smoky oak]

Oh well I mean, basically everybody hated that lol. That has nothing to do with a specific agenda, other than just being stupid lol

I can agree with that sentiment, at least

[dense tapir]

I am socially moderate so go too far left, or too far right I don't like it.

[smoky oak]

god, I wish so bad that netflix had a frame skip key, rather than a huge 10 second skip

[dense tapir]

yeah, same with YT instead of 5s

[smoky oak]

youtube does have frame skipping tho

I just wish netflix had it, so I could get my screenshots easier

[static tusk]

on YT its , and . for frame by frame

[dense tapir]

oh? damn I never knew

[smoky oak]

yeah, I use it allll the time

[dense tapir]

never even saw those mentioned, shit, lol

damn, it works, lol

thanks

[smoky oak]

no worries haha

Ok, here is hoping i got enough images at a high enough quality

[dense tapir]

https://www.youtube.com/watch?v=wL-s3P4KzyM

[static tusk]

quick, download all the models before theyre locked behind paywalls

[dense tapir]

yeah, ain't that the shit 😦

I don't even see how that is legal

[static tusk]

as Olivio said in the video many of those models are built using other models, so its hard to justify them being pay to use

[dense tapir]

Yep

I saw this mentioned a few days ago and didn't understand so that made it clear they are pond scum suckers

[static tusk]

the battlegrounds begin it seems

[dense tapir]

yes

I remember this same kinda the same thing with thingiverse where their lawyers made it that models uploaded were theirs.

a lot of creators left and protested

[static tusk]

yeah that was low

[dense tapir]

I left what I had on there but just stop producing

[sterile kiln]

Guys, what is the best name to define the magic staff of a magician like Gandalf or Saruman (in english)? A magic stick? A scepter?

[static tusk]

staff would best suit those I think

[sterile kiln]

ty !

[grizzled sage]

thing with any content host. lawyers will always come in and say "you're going to need to get rights to copy and reproduce and modify all this content or else people uploading stuff to you will be able to sue you easy."

because if you put a copy of your work on their servers, but don't give them any of those rights, they can't touch it

it's sort of a catch 22

so what happens is always, these sweeping changes in TOS are made to protect the company, because intellectual property law literally works that way, and then somene sees those changes and declares "THEY ARE TRYING TO SAY THEY OWN YOUR WORK" and collectively, everyone inside the biz sighs and says "here we go again"

just another thursday really

[static tusk]

I was just looking at the list of models. most are hentai and NSFW, so IDK about those

[grizzled sage]

they're all NAI merged at some point more than likely

[static tusk]

I have realisticvision , deliberate, dreamshaper

[grizzled sage]

dreamshaper i know has NAI in it because i regualrly rely on it's high quality anime tag training, plus the magic tokens like animeful, best quality, masterpiece all work

[static tusk]

yeah they have to be blends of other models

[grizzled sage]

probably caught it from anything v3

reason i bring NAI up is because nobody can legally use it for anything. No one has any rights to it cept NAI. It was never legally put onto any content host. it was stolen from their servers

[static tusk]

oh really?

[grizzled sage]

well i guess NAI's partners too

[static tusk]

didnt know that

[grizzled sage]

Yeah it was a huge scene! I had just started using this stuff right after it happened. i sort of walked into the shit after the storm

https://tenor.com/view/the-darkest-timeline-community-gif-18212973

[static tusk]

Ive only been in this since the beginning of this year

[grizzled sage]

i got in october

i've never been able to find any reliable information on it though. its a whole shit show that people jsut don't want to confront

[dense tapir]

Yes, NAI is actually using stolen property

[grizzled sage]

Yeah, then it became N AI are the ones who stole stuff first

[dense tapir]

rather using NAI I meant

[grizzled sage]

it's a bit of both

[dense tapir]

yes, then

all a hot mess

[grizzled sage]

like i said, total shit show

Emad was condeming A111 i heard, and he was like "nah uh". subreddit was taken over then given back. wow did i come in at a weird time

[dense tapir]

yes, he was and rightfully so

FINALLY A1111 has a licence so all good now

[hasty nova]

I think I came in around the end of 1.4 and the start of 1.5, whenever that was. I didn't keep up with all the news at that time, not there really was any I don't think.

[grizzled sage]

that was an issue but i get it for the boot strapping phase of a project. no need to rush into one always. But it was an issue.

[grizzled sage]

you got in after the fireworks just like i did it sounds like

[hasty nova]

Yep

[dense tapir]

I still think A1111 never asked any dev that worked on the project, or gave any work to it for their consent so the licence is actually invalid from what emad was saying

[hasty nova]

I actually might still have my first image made locally, which used inpainting that I was messing around with. If you want to see it I can post it

[dense tapir]

must be in writing

this is why you set it at the start before the devs submit

[grizzled sage]

lol i shouldve kept my first, but it didn't make it through the many great deletenings since i've started this

[dense tapir]

he has so many extensions and so many pull requests, etc... I just know he didn't ask any of them

[hasty nova]

I don't think it's my absolute first, but it's the earliest I can find.

it's so terrible

I thought it looked amazing at the time

the pegasus was inpainted there, in place of a tree I thought looked out of place

[grizzled sage]

lol look how far we've all come. still long ways to go. i'm spreadsheeting lately. for a couple months i was studying art history and styles. wtf.

[hasty nova]

So I started doing things with sd locally around Saturday, ‎October ‎22, ‎2022, ‏‎5:15:31 PM

[grizzled sage]

now i'm posing dolls and saving those pictures and then putting those pictures into generations. I'm a guy who takes pictures of posed dolls now! WTF

what have i become

[hasty nova]

Actually, the first things I did were earlier, and used a blender plugin that connected to dreamstudio. I ran out of credits with that really fast.

[dense tapir]

I can't go back to 1.5 due to rez but I highly suspect SAI's next move is to just rebadge 2.1 as 3.0. They can't think that will fly.

[hasty nova]

These were what I made with that extension

I literally thought those super blurry landscape images were amazing

[grizzled sage]

Nice! I was doing some early stuff with craiyon an when it was also dalle mini

[hasty nova]

I actually might be able to find my even earlier gens in my discord server, I was trying to make some concept art for a video game I am making.

[grizzled sage]

i've been watching all sorts of ML breakthroughs for years. i've always loved computer vision an graphics tech so this all really caught my attention hard

[hasty nova]

I found them! I have some really old dreamstudio gens.

And Dalle

And now I do things like this:

Crazy

[grizzled sage]

Those are all high quality. Minigolf generations look fun

[hasty nova]

yeah but they are pretty bad considering what we can make now. Now I wanna go back and try another mini golf gen

[grizzled sage]

with noise offset

gee god now that i've seen what illuminati can do it's so obvious to me that EVERY image gen before it was "mid"

i think the researching being done is just getting started and there'll be spectcularly cool new dynamics in all models coming up

[hasty nova]

Im currently trying to work on a game development specific model, it's just a lot of work and I barely even have like 1/5 or 1/10 of the dataset done. Captioning is hard, finding good images is also hard.

Im thinking it's gonna be for 2.1, but I don't know how to really go about training it, like if I should make it a small lora or a whole ckpt, and if my 3070 can do it quick enough.

It still isn't great at mini golf. Maybe I should make a mini golf embedding.

[grizzled sage]

my goal is to create a dwarf fortress model and all of the captioned images will be straight descriptions from dwarf fortress raw texts. so hopefully, i'll get something that when you plug an item description from DF into it, it'll show

[hasty nova]

I WANTED TO DO THAT TOO

[grizzled sage]

creatures are a goal too but i'll start with crafted items

[hasty nova]

Omg

I just wasn't sure how to get the training images since it's all pixel art which doesn't make for a great model

[grizzled sage]

haha yeah it's something a lot of people are all converging on and i look forward to seeing what comes

i'm not gearing for pixel style. fully capable model geared towards DF descriptions

[hasty nova]

I mean, it would be a lot of work, but you could img2img hires reaslistic versions of the items and use those for training

or even controlnet them, since that's a thing now

Gosh it all moves so fast

Omg someone should train a language model on dwarf fortress

Then feed that to a dwarf fortress sd model

[grizzled sage]

Yeah for source imagery my intention is to get a controlnet of a dwarven hammer. do 12 vector outlines of them. Maybe 20. Maybe 50. Vary it up. Then blast those through controlnet and get 1000s of variations of hammers of different materials

[hasty nova]

Man I need to play more df

[grizzled sage]

eventually i'll have this spreadsheet set up so that i'll just do what the next field wants that day. should take me a couple weeks of synthing up an image set of 100k

We all need to play more DF but like, man, it's hard

[hasty nova]

Im just trying to do 250-500 for my model, is that too few (game asset creation model)

[hasty nova]

I have yet to play long enough to be attacked by an outside group of enemies

[grizzled sage]

No i don't thiink so. Dreamboothing is fun. I just want to go big since i saw a few examples of how great those results can be

Dreambooth results work right?

[hasty nova]

I keep starting new fortresses before my previous one fails

[grizzled sage]

i think a good think about that is it leverages the power of the base model you're refining

[hasty nova]

Usually, I am planning on doing lora dreambooth extension since I only have 8gb of vram

Though I might wait until I get a new case for my pc, since it has no airflow

[hasty nova]

Yeah I don't know what model to train on. I was thinking base 2.1, since that's all there really is.

[grizzled sage]

i've been playing DF for a long long while. I still have barely touched what can be done in it though. Got some pretty deep forts. 100 year forts. many lost to slowdown cause of processor optimizations

so much has been improved in the steam release

[hasty nova]

I have only played the steam release

[grizzled sage]

i love that it has official graphics now xD

lol mine would be in the 10k hours range by now i'm sure

[hasty nova]

dang

[grizzled sage]

steam release i bought day of though and have barely scratched lol

[hasty nova]

I've barely scratched the surface of the game. I streamed it like twice then kept playing on my own.

lol

[grizzled sage]

make sure to grow plump helms or something easy to brew into drink for them

they need their alcohol

[hasty nova]

Yeah I have watched a bunch of tutorials, I know the basics.

[grizzled sage]

or else they get pissed and kill everything

[hasty nova]

I'm tempted to start in like a scary zone just to see how long I last

[grizzled sage]

hahaha oh yeah, do explore all the regions. i figured that if you made it as far as having a fort, you had to have known a lot about it already

[hasty nova]

Very basic forts, but forts nonetheless

[grizzled sage]

one game i started on the ocean shore and giant whale men attacked me

[hasty nova]

Woah

There is a lot of content in df

[grizzled sage]

https://www.moma.org/collection/works/164920 literally the museum of modern arts

[hasty nova]

crazy

[hasty nova]

man I love blip sometimes
"a sci - fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi"

well thats 100 images cropped and hand captioned, time to repeat that 4-9 more times!

[grizzled sage]

its such a huge task. theres gotta be ways to automate captioning and having a way to easily flip through files for edits

[grizzled sage]

that's why blip isn't this way lol. i wouldn't use it cause you end up having to hand edit it all anyways

[west sundial]

anyone have any advice for fixing the arms?

[green plover]

1. make sure you have a good negative prompt
2. If the AI gives you errors, but you otherwise like the picture, it's time to send to inpaint

[west sundial]

i mean, i like to think my negative prompt is good, but i sort of stole it off of hugging face without understanding and adding my own stuff when its not giving me what i want, im just looking for it to redraw the lower arms but its not changing anything despite the sliders i change, i am already using inpaint and im not getting any errors

[green plover]

"Errors" meaning if the image has mistakes

[west sundial]

ahh my bad

[green plover]

what is your denioising strength set to in inpaint and I know this is kind of a stupid quesstion, but did you mask off what you want to change?

I usuallly make that mistake thinking I'm in img2img only to find I'm in inpaint and no mask, no change

[west sundial]

denoising is set to .7-0, but i have tried higher, and its just deleting the arms even though i have it set to fill

unless thats not the right setting

[green plover]

Then you need to lower your denoising. The higher the denoising strength, the more it will change your picture

and make sure you only mask the parts of the arms you want to change. also you may need to change your prompt to describe what you want it to do with the arms

I very rarely change my prompt when I'm inpainting, but when it just won't cooporate, I have to say "no do this"

[west sundial]

do i delete the rest of my prompt when inpainting?

[green plover]

if you have to change it then yes

[west sundial]

or just add on to my current one

[green plover]

I would copy and paste it into notepad so you don't loose it though

another thing that I find works is to just inpaint a small section at a time

more work, but more control

[west sundial]

more like this?

[green plover]

if it works then yes. I have even had to go so far as one arm at a time

when it comes down to corrections, patience is the name of the game

[west sundial]

alright

let me try

fill is the right option i should be using yes?

[green plover]

depends on how drastic a change you want. When I want it to be similar, I use original when I want a major change, I use latent noise

I don't think I have ever used fill or nothing

[west sundial]

getting better

gonna try one arm at a time

[green plover]

sometimes it takes a few tries. make sure you do more than a single generation to give yourself some options

[west sundial]

yea i do 8 at a time

[green plover]

ok

[west sundial]

1 batch of 8 i should say

im running on a hard drive or this would be faster

OMG think i figured out one of the main issues, because i pulled the data from the image i was working on last night, it pulled the seed as well

[green plover]

yah. I forget to clear my seed all the time

[green plover]

I see dead people

[smoky oak]

Anybody know how one would train an offset noise infused LoRA?

[dense tapir]

they all have that as an option now

[teal kelp]

Hello,
I have a question,
I am trying out the rest api
and can't seem to figure out what SD is looking in this param?
Plz help thxs

[teal kelp]

It will also complain if set to none("")

[stoic pivot]

"hyper realistic, Hasbro cartoon, cinematic light, 8K, blender render, best quality, 4k, RAW photo, drawing made by a master of art, super sharp, trending on artstation, cute, kawaii, extrem detailed, photorealistic anime, pixar style, Unreal 5, (Pruff unt Krutt)"

"hyper realistic, monster cartoon, cinematic light, 8K, octane render, worst quality, 4k, RAW photo, creepy drawing made by a master of horror, super sharp, trending on artstation, scary, kawaii, extrem detailed, intricate_details photorealistic anime, pixar style, Unreal 5, (Grunt eats Fromling)"

This is BS prompts, that is that they have no meaning and make no sense, it is like give most of the creativity to the 'AI' in combination with the checkpoint/model. After a stroke that affected my creativity I saw this as very helpful and it gave me many new ideas.

I think folk shall not be afraid to use more of the AI and let the AI spit out crazy ideas like a bunch of drunk friends can do when they try to come up with a game idea, most things they come up with will just be odd and maybe dumb, but from time to time this give an idea that you normally would not thought about.

[stoic pivot]

"underwater realistic cute pink|green kavain monster swimming in cinematic light, funny render made by a master of silliness, intricate_details (H.R. Giger) style"

It is the "cute" part the AI struggle with just now 😄

[tropic shell]

If you need help, please check out #🤝｜tech-support

[teal kelp]

Ok thxs

[tropic shell]

Np!

[raven agate]

prompt

[stoic pivot]

This was first like a joke, to make a test of just "trending on ArtStation", but it become more interesting than I thought, for this prompt do not make an image that would trend on that site but how images on the "tending" page look like. That is why we often get footers and text and so on, but by dragging it through a "XYZ Plot" script we can see things like "DDIM" pick up text better while PLMS create solid footers. This is not "bad" but show what those are good at. Euler a and DPM++2S a did keep it simple with fever details (can be good for Anime and Cartoons.

DPM++ SDE and DPM++ SDE Karras went a diferent creative way, and DPM fast just made it good but differently (I like that), and also DPM adaptive was a bit different in a cool way.
I will create one more image like this to see if it follow same pattern.

[pure gust]

A Cat smiling

[sterile kiln]

when I try to cross a portrait of D. Trump, with a chameleon (ControlNet Clip_Vision), strangely the face of Trump does not change.
It's a conspiracy of the reptilian men

D.Trump + a Cameleon... ==============================

[stoic pivot]

Can you mix the silly president with the "not the bees!" guy?

[sterile kiln]

Nicolas Cage??

[stoic pivot]

That was his name.

[sterile kiln]

hahaha, can't remeber this movie, maybe saw long time ago

[stoic pivot]

Oh, it is Nicolaus (bees in a) Cage.

[sterile kiln]

Legolas Cage

I could do this mix, but i don't think the result will be good, because Clip_vision is for concepts visual, not representation

i could use a depth of a "Not the bees"

[stoic pivot]

It can just be Trump in a bee cage... Trump in Cage... it may be NSFW, best skip this idea.

[wispy nest]

i havent visited the site “thispersondoesnotexist.com” and now it doesnt feel like the same (off topic)

i like the one where the ai faces just appear randomly

nvm i found it im dumb

[sterile kiln]

🤣

[stoic pivot]

Omg, that is epic.

[grizzled sage]

i dont get it

is it funny because it's a racist eating bees?

[sterile kiln]

You can be paranoid and see evil everywhere, that way yes.

[proper kelp]

heheh

Name ‘em all? Should be easy .. save one perhaps

[wispy nest]

Oh yeah it's the jokster

[ripe cedar]

I wonder if there are political cartoons made with SD. Haven't found a model for it yet

[wispy nest]

He eats protein shakes too

[grizzled sage]

Not everywhere, but with him, certainly. All that birther nonsense happened and there's no coming back from that

[grizzled sage]

Maybe is a little wishful thinking. Anyways, blocked cause you're clearly trying to get at me, but it's weak.

[wispy nest]

Oof poor little sensitive critter getting triggered at the sight of massive gains. Many such cases. SAD

[grizzled sage]

Like I said. Obvious

[sullen surge]

Hello all,
I have created a character where i found the face very cool and want to use that particular face on different images. Is that possible based only one image?

[pastel mango]

yeah tis pseudo anime, fear it

[sterile kiln]

today I went to see my garden if everything was fine

[left pond]

Just a reminder that Stable Diffusion is free and fucking rocks. 😉

[hasty nova]

I just use it to generate the text files

[dense tapir]

I went to wd1.4 as git/blip/clip not for me.

https://www.youtube.com/watch?v=rGsnkkzV2_o

[stoic junco]

nice, which model is this?

[stoic junco]

also very cool. which model are you using?

[sudden iron]

hi, can I use a photo of a person to put her in a new scene and make other changes while keeping her face the same?

[sterile kiln]

RealisticVision1.3, ControlNet (clipvision) and the prompt on the image of trump above, with the image of cameleon in CV.

[crystal bloom]

inspired in elden ring

[static tusk]

fantasyai took down Aitrepreneur's latest video apparenty https://i.imgur.com/Fa4aN13.png

[dense tapir]

illuminati

[dense tapir]

ffs

[static tusk]

cant even criticize them

getting a bit over the top

[dense tapir]

Seen it before.

Though, to be honest he was asked and caved. Not cool imo.

[static tusk]

Im betting he was threatened with legal action, not simply asked

[dense tapir]

I mean never cave if you are right and have truth on yourside.

[dense tapir]

I would just tell them to bring it.

Although I don't live in his country so no idea how law works there

[static tusk]

I would say fantasy ai has some big money behind it, so a youtuber would have no chance in court

[dense tapir]

do you not see something seriously wrong with that? I mean it is 100% but shows how far gone this world has gone.

rich win all others suck it

[static tusk]

yup the best lawyers win, money talks

[dense tapir]

injustice reigns

bad look for fantasy even more though, but I have to say if they did threaten me in his case I would have to take it down and I would tell everyone they had threatened legal action.

[static tusk]

quite a number of people follow him so maybe a boycott of sorts can get rolling

[dense tapir]

Many youtubers before him did the same and explained they were threatened legally.

I hope so as it shows where there is smoke there is fire and they are just scum

[static tusk]

exactly

[dense tapir]

civitai is now doing the same shit I heard

[static tusk]

doing takedowns?

[stoic junco]

for what?

[dense tapir]

Even CivitAI is now licensing models...they licensed Chillout Mix and Sunshine Mix! You legally can't, and the owner had to withdraw all the info on how he mixed it!

more pond scum rising to the surface will come for sure.

this is why I stuck with huggingface

[static tusk]

all the hentai and NSFW shit attracts bad actors

[dense tapir]

precisely

I mean when I go to sign up I am hit with porn that would make pornhub blush something just isn't right.

[static tusk]

if only huggingface had a better description panel of sorts like civitai, which is quite useful

[dense tapir]

well, as sinister said they aren't really setup for this and why it is the way that it is, and at any time they may come sweeping through it and take all the non project files down.

they are more for researchers

[static tusk]

I know they were working on icons for models hosted there, which does help a little

[dense tapir]

if you are a researcher at a uni they will give you free resources

man, I hate their file stuff cause if I make a mistake I can't delete one or even rename it I have to start over which I did countless times due to that.

[static tusk]

yup not use friendly at all

[dense tapir]

a few months since I released anything as I don't play that game any longer so they may have changed it, but I doubt it.

[ripe cedar]

Picked up my new PC from craigslist. i7-10700k, 3060 12gb, and 16gb RAM. I want to upgrade the RAM to 32x2=64. My question is

Is it worth going to DDR5, or DDR4 is good enough? Also, the speed of the RAM is more important than the brand that makes the RAM sticks?

Or maybe even 32gbx1 DDR5 over 32x2 DDR4 if it's better (I don't think it is?)

[dense tapir]

stick with ddr4

[woven dawn]

erm...

[wheat flume]

Round medal, silver medal, Himitsu lettering on the rim of the medal, child's profile portrait, super detailed, excellent quality, 3D--ar2500

[woven dawn]

ddr5 is only compatible with 12/13th gen intel & 7k series AMD processors

You literally can't use ddr5 with that rig

:V

[ripe cedar]

Okay. I see 3600 mhz ones on eBay, which are better than the 2600

[ripe cedar]

Okay good that makes things easier

[dense tapir]

yeah, ddr5 is brand new stuff so no need right now.

[woven dawn]

You need to make sure the motherboard supports 32gb of memory on a single dimm. If it doesn't support a max of 128gb total, then you can not use 32gb dimms, and need to stick with 16gb

This also depends on the number of memory slots

[ripe cedar]

Yes, I checked that part, it can take 2 sticks of 32

[woven dawn]

If I were you I'd aim for 64gb of ddr4 4ghz

but

do you really need 64gb?

I guess its pretty cheap nowadays

[ripe cedar]

I think 32 would probably be enough, but I do Blender too and it's only $50 to upgrade

Does the brand matter too much? I think the RAM market has matured to a point where it doesn't matter much

[woven dawn]

Not really. Personally I prefer g.skill & corsair kits - but it doesn't matter too much.

[ripe cedar]

The motherboard doesn't care what speed the RAM is as long as both cards run at the same speed right?

[woven dawn]

different boards support different speeds. Just because you get a 4ghz kit doesn't mean you will be able to run it at 4ghz.

both sticks should be identical

[ripe cedar]

Okay, I'll check it out

Ah, it's a Z590I Aorus Ultra. It can take up to 4.5ghz according to PCPartPicker. Not that I'll be getting that much though.

[floral flicker]

I started doing some 1024x786 images today on a 2.1 model, but there’s like a 50/50 chance it errors because I’m short <0.5gb vram. Is there a simple solution in this case?

[celest fulcrum]

i think u can use your cpu

it takes a long time but it doesnt crash

[floral flicker]

🤔 I also realized I haven’t tried xformers yet

[wispy nest]

xformers, medvram

If you're really desperate - lowvram

[smoky oak]

That will benefit you massively. You should be able to do wayyy higher resolution generations, as well as way higher resolution upscales

my 3060ti went from being able to do 768x1024 max without it, to being able to easily do above 1080p

[novel ermine]

realism is getting scarry nowerdays

[sacred turret]

Majoras mask moon

[smoky oak]

Here is my first successful attempt at getting 2 extremely different characters trained into one LORA

[dense tapir]

That is pretty cool

[smoky oak]

So, based off of this, My plans to make a single LoRA that can do both species and sexes of Na'vi is confirmed to be possible

[rain schooner]

I would like to ask why my RENDER will be blue?

I would like to ask why my RENDER will be blue?

[green plover]

Try putting "blue" in your negative prompt and see if that fixes it

[dry crow]

Try 50 steps and CFG 7 and disable Face Restoration, but looks like a missing vae

Can you link the model you used?

[dry crow]

Try an other model

[rain schooner]

Thank you, there is no problem at present.

[dry crow]

Seems like the other model is missing a vae or only knows blue dogs

You could also try cute Brown dog

[green plover]

and maybe put blue in the negative prompt

[smoky oak]

I had no idea that AI image upscalers like gigapixel could get rid of compression so damn well

I am gonna be doing this from now on

That difference is insane

[glossy herald]

gigapixel has some quite nice presets

[smoky oak]

Like damn, this is gonna make so many previously unusable images just fine for training

[tropic shell]

That's an adorabib blue dog though, ngl

[cyan shoal]

uniPC with 1 step (2.x model but just showing you guys how good it looks at such low step count)

this will help those low end people lmao

[dry crow]

@cyan shoal interesting, in my tests it was worse then DPM++ SDE at 5 steps

like shown here:

[cyan shoal]

1 step

it could be because of a different model architecture, maybe cause it's anime? 🤷‍♂️

[dry crow]

yea that could be it

[cyan shoal]

9 seconds

in these cases it doesn't look as good of course, but at least it's not distorted

10 seconds with 5 steps each

[stoic pivot]

use "hug" in a prompt to test models and samples, first of I see that this word is treated very differently in models, second is that it often merge characters in very strange way and therefor good to see how the samplers manage that. Just "man hug woman" is often good.

[small meteor]

Stable Diffusion

[stoic pivot]

I testing the new "UniPC" sampler and can not see any larger difference between that and Euler.
UniPc = one cup, euler = two cups.

[crystal bloom]

Nice. how did you design the characters?. ControlNet scribble?.

[river mauve]

I got it ya, ai art is the junk food of art.

[dense tapir]

All text to image as I don't use control net.

[fresh hound]

I'm a pirate mother fucker

[crystal bloom]

what you did in order to get the cartoony style of the character?

[stoic pivot]

I used a checkpoint / model that can do furry characters that I found on https://civitai.com/ and I could get it to look the same for I locked the seed and did not change the text prompt

[main coral]

on my 750TI, UniPC is giving me decent results at 15 steps. Dropping image generation from 2m50s (30 steps dpm++ 2m karras) to 1m12s. So, real damn spiffy

[stoic pivot]

Then popcorn isn't junk-food unless you pour yucky fat on them as they do in the cinema, I can not understand your comment/post, so a long rant will follow. (and do not take it to seriously, we have fun).

For now AI is new and folk are afraid and unsure and when folk is afraid they tend to be rather skeptical and hateful as racists.
But keep in mind that we are in the forefront and it is we that dictate the rules and whatever AI is junk or not, my goal is that AI shall make art and creation more democratic and equal over all classes, that it will help small creators to be able to do something that look what the big media companies do to day, it is a tool to create art and art is never junk, have some created it and had fun doing so it is good art.
I often say that we have a problem when it come to the "creative" part that is needed to get copyright on what we create, but long before copyright existed there was artists that could live on what they where doing, so junk or not, be proud of what you do whatever folk call it. 🙂

We are making are, we are making the future.

[wispy nest]

I mean...you will often see similar results using different samplers , it doesn't mean anything.
IN some situations you might like one sampler's result - in another - another sampler's result more...

It's just a method how your pc math basically.
All samplers can do all tasks, it's just how they work "inside".

But...on more complex side some of them changes more and has bigger changes over higher amount of steps, while other - almost don't change after 20-25 steps

and well, some of them faster

[stoic pivot]

When it come to create same characters and style that can be hard just now, but soon it will easier with ControlNet, for now we can make custom embeddings with our own style.

Images here are rejected result, I am working on a thing called "Popcorn Propaganda" and I just had a bit fun with the prompt, but I can get a character that look almost the same, but I often need to do some post work in Affinity Painter.

You have boring party with no popcorn, come join our "party" and you get popcorn.

[vague oasis]

unipc looks prettier

[glacial token]

my best dragon yet with dreambooth

[vague oasis]

i tested unipc and seems to be better

[stoic pivot]

Comrade, this is 150 step of UniPC to pop this party image.

Step test with UniPC sampler.

Same as above but with smaller steps around 20 where it acted a bit odd with more variations.

[pseudo aspen]

My Barry Allen LoRA. I am proud of myself

[stoic pivot]

Some post.

[timid cove]

a photograph of circus people living in the 70's, during their show, detailed costumes, flame flower, Kodak portra pro, detailed pictures, realistic skins, reflection, realism, professional color grading, clean sharp focus

[stoic pivot]

Crazy prompt "lineartgrayscaleinkwhiteteddybeardrawingpopcorn warpostercutebear", Negative prompt "horn, human, nsfw, nude"

[glossy herald]

love that line art style !

[plush hamlet]

#🏞｜general-with-images car cyberpunk

Cyberpunk car

[crystal delta]

@plush hamlet

FAQ: I'm new here! How do I get started and is there a bot?

Welcome! There's currently no bot on the server to generate your images. Start by heading over to #1072220168534642768 to get yourself situated and help find the channels you are looking for! Please make sure you review our #✍🏼｜rules-and-tos and feel free to assign yourself some #👥｜roles as well! Answer any questions your may have at our #1072229020520947753. There are many ways of accessing Stable Diffusion, take a look at #1080946152318443610 to start your journey!

[glossy herald]

@timid cove 👆

[dense tapir]

Than ye. 🙂

[smoky oak]

popagand-corn

[pastel mango]

pew

[pastel mango]

you'd better make sure that Johnny Mnemonic fights Neo

[subtle saddle]

close enough?

[ripe cedar]

Windows 10 is better than Windows 11 for Stable Diffusion? Mine came with 11 but I think I'll go to 10 if it's better

[cyan narwhal]

when you try to generate anything except anime girls or cat pictures

it has no idea what is freaking catapult is

[stoic pivot]

I think I getting closer to finding my style, now I need to get back in form to draw and edit with my Wacom so I can correct and adjust to images I want and maybe need. I will also need to learn ControlNet to get more exact pose for my need.

[stoic pivot]

Nice details and pixel sharpness.

[celest geyser]

Thanks!

[smoky oak]

They look like they are about to fight in the sheets lol