🏞｜general-with-images

[glossy herald]

https://comfyanonymous.github.io/ComfyUI_examples/area_composition/

this is the technique I would recommend for such things now

https://comfyanonymous.github.io/ComfyUI_examples/noisy_latent_composition/ this one can work too, even if a little more complicated

[smoky oak]

I'm still gonna be waiting for ComfyUI to continue to improve before I switch over to developing workflows for it. Lord knows I have my work cut out for me with my guide in a1111

[glossy herald]

yeah, I can't go back anymore... I still do some fun things in AUTO, but it's become my side software now

since I plugged the LLM inside comfy too, hard to think of AUTO as competent anymore (the software, not the person)

[smoky oak]

For me, there are just too many creature comforts in auto, like its still consistently lower VRAM usage

not all of us have 24GB VRAM to just throw away 😅

[young solar]

I'm not the person who originally asked, but I am wondering if there is a good way to generate a consistent character using just SD. The usual way would be to use a LoRA like you say, but then I would have to use preexisting images. Probably someone else's character. Is there a way to get a consistent character using SD to then train a LoRA? Even if it means trashing 9/10 images in the initial work.

[smoky oak]

simple controlnet can do it if you prompt right

[ripe cedar]

I have created a sketched out short with panels and dialogue, that part is done.

LoRa is needed, got it. I'll assume not to use any 3d characters, just AI generated ones. If I don't use a Lora, I'll try out the unique character name and generate batches to get close. (I can do some 2d art too, so cel-shaded will be easier I use that style, rather than a gradient style).

For backgrounds, how should I keep them consistent? I like how using SD and ControlNet gives a good stylized look, but the color consistency and composition is off. Ie even if the structure is the same, it will make doors and windows on walls in different generations. But with using img2img for backgrounds, it will not be as pure stylized, but should remain more consistent ie basically a filter on top. I was wondering if I could have my cake and eat it too ie get both.

[glossy herald]

for sure. you can totally prepare your dataset on your character yourself
It requires a good base model and/or lora, a nice prompt that is very detailed, such as most results look like the same character, quite some luck and seeds to run, and cherry picking the best pictures that feature the same character with the same details almost. Doing so with different backgrounds for your character, different poses, ... will land you a dataset good enough to then make your own TI, LoRA or model trained on it (TI seems enough here by the way, LoRA seems better and faster, and DB seems overkill but best possible quality)

[smoky oak]

I trained my own LoRA on the tigers from Zootopia in order to get my character looking right. I spent a ton of time on it

stupid clyde is blocking the image

[young solar]

Thanks. I am going to remember that. I think trying to make consistent characters is the next main challenge for myself with SD. Either that or learning controlnet to maintain a consistent background across multiple images.

[smoky oak]

There are better tools for doing things like consistent backgrounds, if you want to give it a try

like zoned prompting

[ripe cedar]

Yes, there is a trick by using prompt blending. So basically you use Famous Name:Other Name to combine them with a weight for each one. Famous names will be consistent but they are too recognizable so when you combine them, it will still be consistent but look different.

Another way is using a random name generator with different last names according to ethnicity. It only works on some names, but you can make it like that.

Also you can make it with famous people but then change ethnicity and gender. This will look like a completely different person but still give you a consistent character. And, you can make an embedding of it too. So similar to a LoRa without actually making one.

[glossy herald]

brb sorry

[smoky oak]

really hoping SD3.0 will drop sometime soon so we can finally have decently accurate color prompting

just trying to get a ginger woman with green eyes instead of blue is practically impossible

[young solar]

Thanks. That is also something to try. I have not attempted using names, generally just physical description. It works pretty well, but there are small variations. Consistent face shape is the biggest problem I run into.

[ripe cedar]

Ah yeah I use both ComfyUI and Auto

[smoky oak]

I basically never generate women, but man SD 1.5's limitations start to show when going for specific eye colors

[young solar]

What kind of tools? Zoned prompting is a feature in ComfyUI, or is it also in A1111? What else tends to be good? For reference, what I am trying to do is put different characters onto the same background image or change a character without changing the background.

[smoky oak]

I don't know enough about it to be confident enough to answer personally. I just know that in A1111 you can divide images into different prompts, including backgrounds with zoned characters on top. Never messed with it myself

[ripe cedar]

There was a link I have for GitHub but it is 404 now 😦 but it shows the specific process that the guy used and also he had several free TIs for characters that he made using this method that anyone could use

[young solar]

If you can find it, ping me. I will be quite interested.

[ripe cedar]

https://github.com/BelieveDiffusion/tutorials/tree/main/consistent_character_embedding#readme this was the link, maybe you can trace it down somewhere

[young solar]

I will see what I can find, thanks.

[ripe cedar]

https://civitai.com/user/BelieveDiffusion it's gone from his GitHub but I found it on his civitai

[young solar]

Ahh, thank you. Bookmarked.

[smoky oak]

Damn, I am not sure how I should continue on this ultimate upscale guide. Things seem like they suddenly changed half way through demoing

[ripe cedar]

A1111 changed?

[smoky oak]

no, but my results seem to have changed a bit

like the things I had cemented down seem to have shifted

[ripe cedar]

Hmm, consistency changed?

[smoky oak]

the process seems to have. I am not sure how to explain it

My process is several different steps and nuanced parameter changes. But I had found options that should never be changed and now they are... not behaving that way

Looks like I need to sit down and crunch some gens to redraw these conclusions. I really don't want to keep delaying this guide

So many people seem to want it, but I need to make sure I am not putting out bad stuff

[dense tapir]

@smoky oakWell, to upgrade to a new Nvidia card I have only two choices. FE only, or retire my old full size ATX case for one that can handle 14-16 inch long cards.

[smoky oak]

as somebody with an FE 3060ti, it has been the best GPU I have ever owned

sleek, compact, lower cost, great temps, very very quiet

[dense tapir]

The problem is I have to buy a FE only from Best Buy then they send it to my local brick and mortar and I have to go pick it up when it arrives.

[ripe cedar]

I jumped on a deal on a craigslist desktop which had a good GPU, but still trying to sell the old one since that time. Not so easy to get rid of...

[dense tapir]

takes about 4 days it said

[smoky oak]

So are you really considering dropping $2k on an NVIDIA GPU?

[dense tapir]

Aye. My case can handle 12inch long cards, or 13inches if I work it but these cards are super long 14+ inches

[dense tapir]

No, but I am getting one, yes

[smoky oak]

what? I am confused

You are not considering it, but you are doing it?

[dense tapir]

Not cosidering 2k, hell no

[smoky oak]

I am extra confused now. Did you not say a 4090?

[dense tapir]

4080/4090 same length :/

[smoky oak]

I could justify a 4090 before a 4080

[dense tapir]

Always sold out. I mean always due to Nvidia severely limiting supply

This is how they are trying to get us to buy their lesser cards. Jensen mentioned limiting supply.

tbh, Nvidia are ultra dicks now and I am very unsure why.

Like we consumers killed his pet dog/cat or something.

[smoky oak]

Its because NVIDIA doesn't need us now

we are kinda just a side thing for them. They don't rely on us for their big money anymore

[dense tapir]

Yeah, agreed.

btw, all three card makers now have contracts with TSMC for their stuff.

Battlemage in 24 and the next in 26 so I am so hoping Intel can figure this out. I think they can.

[smoky oak]

SD is being a bit of a pain for me all of a sudden

and I am not sure why

it keeps super zooming in some of my gens

like yeah, this is definitely what I wanted :/

or this

and this is with close and close up weighted at 1.5 in neg

[dense tapir]

Yeah, having that happen lately too as if it is ignoring my neg

[vocal salmon]

#🏞｜general-with-images cow

[static tusk]

too much emphasis in the positive prompt for details will gen a closeup I found

[smoky oak]

No crap...

I just removed all the details, and they are good now

wow

Guess we never stop learning lol

much better

[sturdy tree]

Premium Quality Face Swap for Videos and Images

[dense tapir]

Yep, just measured it again and 13inches is it. If I move my drives I might could fit the FE in there with not a single mm to spare.

Can't "Dremel" anything either

[young solar]

The size of those cards is obnoxious. I had to go water cooled because it was the only one that would fit other than FE. And FE is impossible to get.

Which reminds me that I need to flip the fans. They aren't blowing the right direction.

[dense tapir]

My case was made before WC was a thing so there is no where to put the radiator. I despise all these windowed cases too.

You know they are now so big and heavy they come with a stand to put under them so they don't sag and bust something.

[young solar]

Radiator was a pain. Super ugly placement, but whatever. They aren't necessarily that hard to place, luckily, but it barely fit in my case.

Though my case is weird. Chimney style. Good for airflow, certainly, but routing wires is awful.

[dense tapir]

My case - https://www.newegg.com/black-silverstone-kublai-series-atx-full-tower/p/N82E16811163102

eATX sized and costs a leg back then with no window as I chose the non window version. Never thought cards would get so damn long

[smoky oak]

Ok, my upscale process is kinda pissing me off

its not being consistent now

[dense tapir]

@young solarWhich card did/model did you get?

[static tusk]

also measure the distance from the motherboard to the side of the case. they are larger in that dimension also, + consider the cable

[young solar]

Going to have to check the exact model in the morning, but it was the Gigabyte closed loop one. There is also another closed loop one that ended up being out of stock, think it was Asus but not certain.

[smoky oak]

Alright, I don't know what to do

my upscale process isn't working properly now.

I am not sure what changed, but this is not a good thing.

[dense tapir]

171mm is my max these cards say 140mm

[static tusk]

thats fine. no worries there

[smoky oak]

sighhhh

Alright, I might have to scrap this whole guide. Damn it

[dense tapir]

even if so they have these neat things I saw on Jays2cents where they will 90 or even 180 degree

[static tusk]

yeah Cablemod make them

[dense tapir]

yeah, I would seriously consider a 90

helps with strain relief

[static tusk]

you need to break the process down into each step and refer to older gens for the metadata

without knowing what your doing its hard to help

[smoky oak]

I would explain the whole process, but its one hell of a journey

[dense tapir]

Yeah, Gigabyte, and I, went our separate ways a few years ago.

[smoky oak]

and now it doesn't seem to work the same sighhh

[static tusk]

you compared metadata?

[dense tapir]

This one too https://www.bestbuy.com/site/msi-nvidia-geforce-rtx-4090-suprim-liquid-x-24g-24gb-ddr6x-pci-express-4-0-graphics-card/6522334.p?skuId=6522334 What is your case?

[smoky oak]

There is nothing to gather from that

[static tusk]

use the same images that worked and go from there

[smoky oak]

maybe if I get desperate I will do that

Right now I personally don't wanna dedicate several hours to that

[dense tapir]

Sheesh, 40 USD for the 90

[smoky oak]

Ok, so I process is working better now that I restarted it? This makes 0 sense

[static tusk]

if it works, it may help to take a screenshot of the window

[smoky oak]

its working properly now, and I have no idea why

I did see that USDU updated before I restarted, but its the same process

[young solar]

Now that I have no idea. I got it on clearance four years ago.

[dense tapir]

:/

I just think I am fucked and will have to buy a new case

far too costly to have to do all of this

[dense tapir]

The real number

4080 sizes

[static tusk]

yup plenty of room. your main issue is length

[dense tapir]

The length, and size, of these cards are like 1980s boomboxes in size

[smoky oak]

Might wanna look into a sag bracket

[dense tapir]

They come with one

[smoky oak]

good, good

[dense tapir]

that is how bad they know, lol

[summer viper]

wowow

very good

as if it were real

[smoky oak]

@dense tapirWeird new issue discovered with only the anime upscaler

this one

it has done it through several images in weird ways

[smoky oak]

Ok, I reworked the process a bit now

big difference

(didn't take the time to fix the eyes)

[dense tapir]

What happened?

[smoky oak]

No idea, it just seems to have a tendency to randomly do that kind of stuff

[dense tapir]

Did something recently change?

[smoky oak]

nope, no updates

weird

[dense tapir]

That is very strange

[smoky oak]

I found a better workflow that seems to be giving even better results, believe it or not

[dense tapir]

wow

Oh, I think I might have found a case

I have been dreaded this but it was coming as mine doesn't even have USB 3 front ports. Always use it for something just not this something. Still, mad I have to upgrade the case

Nice thing is this case as a shit load of room for water cooling in various places

*has a

I really do want to go watercooling as I am tired of fan noise and WC always seems to be so much quieter

Phanteks Enthoo Pro is the one I was looking at. Sad thing is to get some of the cards in it I need to remove the HDD cage though it appears to have a shit load of cages so I need to see if I have to remove all or just some.

It is a 9 year old case so amazing it is this future proof and I saw on parts picker a 4090 strix in it

[smoky oak]

interestingly enough, the new workflow uses switching between different upscalers at different steps

turns out they all have different strengths at different multiplications

[dense tapir]

yes, this is a fact

[smoky oak]

Experimenting with less "identical" results, but very inspired ones

[smoky oak]

this inspirational generation technique is yielding some cool results :>

[dense tapir]

Interesting

I watched a few vids of reviews and that case is nice until it is time to clean all the filters.

[smoky oak]

@dense tapirI may have just found a massive hack for upscaling. Its producing crisp results

It takes even more steps, but it is showing massive improvements

Like, these are the same resolution. left is without the new trick, right is with

way sharper and crisper with far more pixel level detail

Same res gens, without and with

only tested with one image, but lets see how well this works

[rancid quiver]

Perpetual Diffusion 🙂

[glossy herald]

wow ! that is an awesome one !

I haven't manage anything that good with it yet ^^

[rancid quiver]

Yes, very happy by this one.
raw output, no inpainting. It's incredible!

[rancid quiver]

Have you tried my Prompt Generator? 🙂

[glossy herald]

I still need to try your prompt generator, yes

not yet

time flies by :p and great idea to include such thing

that feels like a big release to come 🙂

[rancid quiver]

I may open some slot today maybe

[wispy nest]

i know it's a blasphemy but i find it cool

[smoky oak]

@glossy heraldCan I pester you for a moment?

[glossy herald]

sure

[smoky oak]

I just wanted to flex my newly found muscles, cause I added an unorthodox step to my process and DAMN are the results paying off

I am sure you saw the cat example above

those are the same resolution image, but the right one has my new ✨ F A N C Y ✨ step

[glossy herald]

I saw it but didn't detail it, just checking for NSFW in channels. i'm training and preparing 3 model pages right now

nice ! what's the magic sauce ?

[smoky oak]

all good. You can just look and see the detail improvement at the same settings

[smoky oak]

Its a secret :p ||I'll tell you in DM's||

Secret has been shared

SO

I have been testing it

Base image

upscaled

I have achieved pixel perfect detail

its been a struggle of my process, but it has been solved

[glossy herald]

ENHANCE

[smoky oak]

and now I have a process that not only works much faster for the same detail, but also has higher detail at the same res

[smoky oak]

I am gonna upscale 2x again to see how far I can push lol

this is gonna be 59MPX

[glossy herald]

nah,just enhance the face now

and then do a zoom video :p

[smoky oak]

Oh, that could be awesome haha

[glossy herald]

with infinite details, ending in the eye

[smoky oak]

just zoom into the atoms haha

[glossy herald]

yeah why not

you may need to change the prompt a little

[smoky oak]

Already worked on that. After a certain res, having the original prompt becomes detrimental as the tiles get so zoomed in that they start to hallucinate detail

[glossy herald]

ok I need to go back to training now, later 🙂

[smoky oak]

sounds good. This upscale isn't gonna be particularly fast at 59MPX lol

[dense tapir]

Now that is nice.

[glossy herald]

it arrived, and it feels glorious ! thanks for the tip !

[ebon gate]

sooooo, what's the secret?

[glossy herald]

they are writing a guide on it from what I understand, and don't seem to want to "spoil" the info before

[coarse pendant]

Why does the image i generated look so unpolished?

[glossy herald]

512x512 makes it hard for small details like the windows on the castle to really work. You would need to upscale to fix details like that.
The rest seems cooked enough, I don't think adding more steps would help. Just higher resolution for more details/sharpness

[coarse pendant]

oh I thought upscaling is really just to increase the resolution. haven't thought that i can also add features

[glossy herald]

#🏞｜general-with-images message

check those examples higher

upping resolution lets you achieve a lot better end result, the AI having a lot more work on each part of the picture

[coarse pendant]

I see thanks good to know

[rancid quiver]

I'm cooking SAM as new controlnet

[pliant oxide]

Ukiyo-e painting of dog on unprompt.ai

[honest flume]

some random thoughts

[grizzled sage]

i'm glad to hear another great rave review on that device. Ive been wanting to pull the trigger on buying it for a long while but i got a pile of controllers that won't let me justify it

i've demod one of them twice and both times they feel amazing

[glossy herald]

I'll say more once I really tested it

for now, the only bother is that the buttons that are where I'm used for Pause and menu, those are for other overlay features, and can lock you out of the game still playing for a few seconds. Just got to change habbits

The material is really good

nice feel

[grizzled sage]

i hate game overlays hahaa. even on my modern system they seem to get in my way and jank up my game . i might have an interupt problem regarding many storage drives though, but also, i just hate overlays

[stray comet]

guys which are the must have models and utilityes? i started yesterday, using Deliberate at the moment

[coarse pendant]

Why part of the tree is blurry and part is not after the upscaling?

This is what i'm using

[dry crow]

Dreamshaper

[fresh hound]

LEAKED - TRUMP MET UP WITH TRUMP

[fiery carbon]

I am cheating 🤙

[coarse pendant]

Anyone can help me to fix the shoes?

[fiery carbon]

Inpainting?

[coarse pendant]

I tried not sure which checkpoint I should use

it doesn't really give me normal size shoes

always distorted

[fiery carbon]

It depends on the image size a lot. Best results i had with InvokeAI for such matter.

[coarse pendant]

so should I upscale it first then try to inpaint?

[fiery carbon]

Yes, upscale first sounds right to me.

[coarse pendant]

will try it out thanks

[fiery carbon]

You're welcome 🙂

[faint locust]

for example i promted: white, caucasian, man, short brown hair, short beard, looking at camera, 0.5 meters away, 81.2 degrees fov, 2.7mm focal lengt

got this:

i believe i need some extra package

im running SD 2.1

[vivid forge]

i think "0.5 meters away, 81.2 degrees fov, 2.7mm focal length" may not work, have not seen prompts like that working

also it looks like not enough steps

[faint locust]

ok can u send me a promt that u already prooved

[vivid forge]

https://civitai.com/images/422820?modelVersionId=38255&prioritizedUserIds=7146&period=AllTime&sort=Most+Reactions&limit=20

[vivid forge]

adjust # steps and CFG settings, if that does not work, i would try another ckpt model

[dry crow]

Sd 2.1 x768 ?

[dry crow]

Ah then your resolution is wrong

You need a 768x768 resolution for this model

[faint locust]

okey i changed it, lets see

better i guess

[dry crow]

Looks okay 👍

Make sure you add some Quality tags like high quality, 4k, masterpiece, in the prompt
And blurry, low resolution, deformed,
in the negative tags

That can help a lot

[smoky oak]

Just a little reminder for whoever didn't see it last night.

I made a monumental leap in the quality of my upskill guide by introducing an unorthodox step

*upscale

The difference is pretty damn amazing IMO

No in painting, no cherry picking, just pure results out

This new technique can have 2x the detail at half the res, and I know that sounds like some random arbitrary value, but it's not

These are the same resolution images, without and with the new step

[dry crow]

Ouh very nice and that with the same res

[fiery carbon]

What exactly is the step for your upscale render?

[smoky oak]

It's a lottt of steps, but now that I have finally found the holy Grail step that allows for pixel level detail, I'd say the guide should be finalized within a week

Okay, it's not a lot of steps, but it's a lot of variables that I have to thoroughly explain so people can adapt the process to suit their needs

[white pivot]

Parseq + ControlNet

[smoky oak]

Looool

[smoky oak]

Thank you! Here are the full view images without and with

You can See how big of a difference it makes on improving texture, as well as how much more it looks like it was painted, which was the goal

Used it to make this image last night as well

(working on the crunchy whiskers problem)

[fiery carbon]

Is it better than 4XUltraSharp upscaler?

[dry crow]

Wow epic shot

[smoky oak]

Oh yeah, way better.

Ultimate stable diffusion upscale utilizes diffusion technology in order to generate new information that pertains to the underlying image, whereas pixel upscalers really just sharpen the information that's already there

[fiery carbon]

Well, i want the guide for that method, thanks 😉

[dry crow]

I wasnt sure if i see celebretys until 0:55 xD

[smoky oak]

When I do post it, it will be on Reddit, and hopefully pinned in the #1003034183716835418 channel

[fiery carbon]

Okay, i will look out for it.

[smoky oak]

I'm working on a full step by step guide with images, 20 example images, recommended prompts, FAQs for problem testing, and a full explanation on the variables that should change for different use cases. It's going to be a very big guide lol

[white pivot]

first one is David Bowie, then Jared Leto, than Dwayne Johnson 😄

it is weird because celebrities have a weird face sometimes

like Will Ferell as Louis XIV

[dry crow]

Haha right but i didnt thought everyone is someone xD

[white pivot]

If you recognize Brad Pitt and Tom Cruise, Im happy 😄

Even Matt Demon as Marat wasnt easy to get...

[dry crow]

Yea Brad pitt was easy

Still looking for tom

Francois ler ?

[white pivot]

yes ;D

lol I know

I dont know why some celebrities are so well made and some others not. I even tried Leonardo Dicaprio.

But it was just a random blond guy

so I changed for Chris Hemsworth

but still look like a random blond guy

Even Anne Hataway...

https://tenor.com/view/looking-nick-cage-nicolas-cage-the-unbearable-weight-of-massive-talent-staring-gif-25135498

hope people get Nicolas Cage face 😄

it was my favorite

[dry crow]

Haha yes 😄

Awesome done

[white pivot]

l,ol thanks

[muted garnet]

hello

[stoic pivot]

What do "negative" prompts actually do? I want to test if suggested negative prompt actually work or are effective. I made a prompt "kawaii neko" and then tested "worst quality", "bad hytentot" and just "bad", but to test that I first need to generate a image with "worst quality". So one question can be; Have you at any point made a prompt that made so bad images that you had to add "worst quality" as a negative prompt to get it to look okay?

Fogot the examples

[muted garnet]

i aint readin allat

[stoic pivot]

For lazy, "Have you at any point made a prompt that made so bad images that you had to add "worst quality" as a negative prompt"?

[muted garnet]

no

[stoic pivot]

Nor have I.

[tropic shell]

Though somewhat outdated, esp as the bot is no longer running, you can see my negative prompt guide here: https://docs.google.com/document/d/1aHJ9RBt_vlCwJQBVUUsb7VghKB-wynv7WGTxm9ozL1k/edit?usp=drivesdk

[smoky oak]

This has to be one of my favorite gens I have done. Love the slight cryptid look mixed with a slight cuteness IMO

[stoic pivot]

I know. Got that liken so many times. But my question simplified "any suggestions how I can test how effective different negative tags are".

[tropic shell]

Your kitties are very cute

[calm canopy]

can simply generate a picture (or a bunch) of each: 1 with no negative tags and 1 with the usual negative tags used with ur model, then compare

[stoic pivot]

That is what I do now, but how can I know if "ugly" work as a negative tag if I can not generate an image that become ugly without trying to make a ugly image?

[tropic shell]

Test the positive relationship first. Like, "Cat, best." Then, "cat, best quality." Then "cat neg: best" , " cat neg: best quality." What you are trying to do us understand how cat and best are related, then how best and quality are related. Then you want to test them as negatives on both sides of the equation. So, "cat, best neg: best" "cat, best quality neg: best quality" and so forth

Finally, go "cat neg: best" and "cat neg: best quality."

If you generate like 20-30 images, you can get a pretty good idea of what key words you might need to use. Over time, you'll learn what ones do what

[snow basin]

guys, what's I'm doing wrong? ControlNet just doesn't work

[tropic shell]

I suggest checking out #🤝｜tech-support

[calm canopy]

have you enabled controlnet?

you're gonna want to enable it here if not

disregard, i missed that second image

[stoic pivot]

Activete Model OpenPose too?

[calm canopy]

oh you found it

yeah thats the issue

[snow basin]

you talk about this?

[stoic pivot]

As I said, I saw it was empty, I think there is a model for OpenPose, but I have self never tested it. I often use Canny and I need both to be set to Canny to get that to work, so I thought it may be same for that.

[calm canopy]

yes that model needs to be set to openpose

the preprocessor is only ever turned on if you are uploading an image that you want the extension to extract the pose from

you already have some that you made so set the preprocessor to none and set the model to openpose

[snow basin]

There's still 'None'

[calm canopy]

you want it to look like this

[snow basin]

¯_(ツ)_/¯

Maybe I should install something?

[dry crow]

Your missing the Controlnet models it seems

[calm canopy]

ah it sounds like you havent downloaded the openpose model

[smoky oak]

@dense tapir Ohhhh, the 4070 has 12GB. That makes me feel WAY better about the whole GPU

[calm canopy]

my mistake, i didn't understand your message about it only having none but it makes sense now

https://huggingface.co/webui/ControlNet-modules-safetensors/tree/main

you can find the controlnet modules here

[snow basin]

Oh thanks, I thought it was enough just to download the extension from the 'Extension' list

[smoky oak]

There are even some other models on places like Civit AI, but not many cause control net training is insanely intense

@dense tapir I didn't realize just how good AMD has gotten recently. They released drivers semi recently that unlocked a huge performance gain on their GPU's

$400 less, and only 290 watts

also, 4070 vs the $520 6800XT

identical

also :/

ok, I didn't realize the 4070 was hyper efficient. This changes things again. Its a 200 watt card that can beat a 3080 in compute, and also has 12GB VRAM... I guess this card is not as bad as I thought

[grizzled sage]

amd drivers have always sucked. its why i left them after nearly 20 years . finally just fed up enough to make the jump.

i'm not surprised that a random update pushed huge gains for no apparent reason

[smoky oak]

when the 3080 is using 310 watts, the 4070 uses 170 for the same perf. Very interesting

[grizzled sage]

someone probably left some dumb debugging code on in the production release

[smoky oak]

It sucks cause AMD's hardware perf, efficiency, and price kicks NVIDIA in the groin, but their GPU's are useless for anything else (except some very niche workloads where they make NVIDIA look like pocket watches lol)

[grizzled sage]

i'm a 20 year AMD user, and shit, i don't know my way around the driver panel there today. After 20 years. Shit changes so often. ZERO consistency. Hell, i made the jump to linux cause i was always figuring it out on the windows side anyways. The driver situation for AMD is infinitely better on that side.

[smoky oak]

This will never not make me cackle

[grizzled sage]

after 20 years, go back to Nvidia, i know my way around. They haven't changed shit for 25+ years

[smoky oak]

AMD wins so hard in SNX-04 and solidworks that people who main those softwares buy a separate AMD GPU just to run them cause they are so disgustingly good at them

or well, I suppose the gap is SNX-04, not solid works

[grizzled sage]

nvidia gimps their consumer end cards for cad work. If you're doing cad work you'd be an idiot to buy a gaming gpu. naturally.

cackle all you want. that's not the market focus for those cards

[smoky oak]

No like, AMD still slaps the hell out of even the Quadro cards for specific programs, SNX-04 being one of them

Its cause AMD has some types of hardware acceleration that NVIDIA does not

[grizzled sage]

i'm so over amd. beyond. the way they done my vega64 so dirty on windows is indictive of how they'll treat current rdna 2 cards going forwards

new architecture means old gets the b coding team

[smoky oak]

Oh, the vega64 cards were trash from the start. I kinda blame the person who bought them 😅

[grizzled sage]

Great card actually. Until driver updates nerfed them in favor of memory speed. Killing all the bandwidth capabilities of hbm2 and making it a pointless integration

Maybe stop being a fan boy for a second instead of being like "that's your fault"

amd's driver team sucks. It's just a fact. people who are hyped about them are just caught up in marketing. i'm a 20 year vet there

[smoky oak]

I am not being a fan boy lmao

[grizzled sage]

you literally just blamed me though

wakeup a bit sytan

be reasonable

[smoky oak]

It was a bad card from the start, you bought a bad card, whos fault is that?

But yeah, AMD's driver team was deplorable before the rework, which is why we now see these massive perf gains on drivers

[grizzled sage]

sigh. It was what AMD offered then. The other option was the 1070ti or a 1080. Way out of my price range. To prove a point you're passing the responsibility for bad driver code off on me. it's just dumb. i'm walking away from this

[smoky oak]

I would have to say, from my experience and the experience of other people big in the community, that AMD's drivers are about as good as NVIDIA's now, but they lack the feature support NVIDIA has

[grizzled sage]

meh. marketing convinces people so easily. it's always been that way. "no the drivers are good this time"

[smoky oak]

"I bought a car that sputtered when it started, and now it doesn't work. I blame the manufacturers"

You should have seen the warning signs when basically everybody was saying not to get them because of their niche use cases, lack of most modern forms of videogame hardware acceleration, and already extremely shaky stability in the tasks they were made for.

Sure, it was their fault that they got worse, but you weren't just blindsided by them being bad. They always wre

Vega 64 is a mess, just like it was when it launched

[grizzled sage]

stop being literally dumb. if it was a used car why would i? if it was a brand new one i sure would. easy. literally dumb analogy.

[smoky oak]

however, RDNA and RDNA 2 have been seriously good

hell, even the RX4xx and RX5xx cards are still quite good for performance and efficiency

[smoky oak]

At least I made an attempt, rather than calling somebody who is... quite literally against both companies a "fanboy"

You're floundering to try and it make it seem like you didn't have a bad idea purchase that was met with an equally bad outcome. Both parties are to blame. The company that put out a blatantly bad product, and the consumer that bought a blatantly bad product from a company that at the time was known for not being good, reliable, or timely with drivers...

Anyways, moving on as this is a useless conversation, and your money invested in that card that is still gonna be useless after all this anyways, so yeah

[sand copper]

What is it?

[smoky oak]

I'd have to explain the whole process first, which is what the guide is for

[sand copper]

oh I didn't see you mention that before sorry 😅

[smoky oak]

No worries! I'm working on a whole huge guide and breakdown of my process

[dense tapir]

Yep, and I can even train Dreambooth with them. 7900XTX is more akin to a 4080 with 24gb

[dense tapir]

I have settled, not happy but settled, on a new case.

Windowed and LEDs which I hate but I think I can turn the LEDs off via my motherboard RGB headers

What I can't turn off is the leds of the video card and that stuff is bright

Remove the middle drive bay and it can handle up to 392mm long video cards.

@smoky oakOn the AMD what you do lose with AMD is xformers.

[crimson dawn]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations has sdp now though. not sure what it means but its faster than xformers

[dense tapir]

Not sure I know on the 4090 it matters on a 4080 not nearly as much and on 3060 not at all, so not sure.

@crimson dawnMy fun is not in the generating it is in the training so I need something that can train, and train well. If it can train at the speed of a 3090ti then I don't care who makes the card.

[crimson dawn]

yeah. it works on nvidia too. i think it has gains on 40 cards because it needs pytorch 2

[dense tapir]

Yep. I know the pytorch.org guys said it needs to be programmed for it and the real gains would be with 4090 cards until then. Once it is fully programmed for, which it is not yet, we will all see performance gains even my little 1060 though barely for me.

@crimson dawnFor now the 4090 people saw 3it/s faster using --opt-sdp-attention on average while the 3060 people saw nothing.

[dense tapir]

RDNA3 was a huge step up as is DLSS3 vs 2 of each. A lot of real AMD fanboys are waiting on RDNA4 and even the reviewers are saying with as much of an advancement 3 was over 2 4 may actually be the one. Not hopium/copium just based upon what advances has transpired. Now with AMD getting into ML with their laptops via a chip for ML it might just get interesting because they know this is the future.

[crimson dawn]

rdna3 dgpus just came out. everything about 4 is hopium

i even saw what jim keller said about it. still hopium if he's just looking at engineering refs. same old game

[wooden silo]

hello

Can you tell me how to fix it ?

[tired peak]

what kind of creature does this look like?

fish, dog, chimera... hm

some kind of elemental

[feral pebble]

Is there a controlnet model that will still accept color?

like, I want it to draw that muscle there still.. since it loses color, the edge detection and stuff just kinda ends up merging it with the skeleton and whatnot

and stable diffusion doesn't know how to accurately draw psoas muscles itself 🙂

[storm badger]

The dog is generated with Img2img but with segmentation selection of Facebook's Sam model.Working pretty good.Compared to control-net,you have much more power and control over your selection.Yes you can in-paint, mask manually if you want,but this is smarter when it comes to selection as the models were trained specifically for this.

[feral pebble]

@storm badger is the outline from a deliberate backlighting or an artifact of segmentation and limitations of SD handling it?

[storm badger]

The outline was a side effect of img2img.it has left some traces like ghost and img2img tried to fill inside that area. If it was filled with %100 accuracy that should not be the case.Maybe in the future it can be done better,but this is pretty new atm. Artifacts or unwanted spots can be cleaned in the post production if it is intended for a professional delivery etc,I believe...

[feral pebble]

@storm badger how do you use fb sam?

plain img2img destroys my pic (all defaults) :}

better with low denoising (.27) :}

[storm badger]

@feral pebble https://github.com/continue-revolution/sd-webui-segment-anything ,use it with automatic's webui...manual installation and model downloads are required.also after downloading the models put them inside the extensions own model directory without renaming. You can use it on img2img tab.

[feral pebble]

thanks! 🙂

[storm badger]

Np,have fun 😉

[feral pebble]

'tis for love

[dense tapir]

Everything is hopium with all of these fuckers from hope that Nvidia stops fucking us all over to hope that Intel can finally do it right to AMD finally able to pull AI/ML off to compete. All is hope right now because all of this situation for all of us fucking sucks.

[dense tapir]

@smoky oak I just watched a review by this content creator I trust a bit more than the big names and I am surprised at the results for gaming. If I were a gamer the results shows me which to buy. 6900XT vs 4070/ti beat those cards in almost every single test and really the only thing they have going for them is DLSS3 which AMD doesn't have. Turn DLSS3 on and the results are slightly (very slightly) better than the 6900xt's pure rasterization. That is a last gen card vs brand new ones. I gotta say I am impressed, and depressed at Nvidia. These tests were at 1440p.

[smoky oak]

As I've been saying for a while now, AMD has become a serious force to reckon with when it comes to gaming performance, but the biggest reason why most people won't buy their cards even though they are higher performance, less power draw, smaller, and they cost less, is because they don't really support anything other than gaming

[dense tapir]

Thing is look at this to get sad again, but there is hope.

for strictly SD there is really no reason to go to a 4070ti vs a 4070.

not for 200 more

[static tusk]

AMD cards are just far too limiting

[dense tapir]

for now, but they are getting into the ML

CEO said so

I wish I remembered the laptop ML chip she said it begins with

[static tusk]

they are yet to release the mid range and low end cards for the 7000. ML is going to be a while off

[dense tapir]

I am brand agnostic I just want realistic pricing again pre Crypto mania nonsensical pricing that can give me AI and I don't give two shits who it will be.

Personally, I loathe Jensen with a passion and would probably end up in jail if I met him in real life for how he turned out. Such a bright star 12 years ago to turn into this thing we currently have.

Damn

Funny they did not show the flagship models from each but I expect the same curves

No matter if I grab an AMD or a Nvidia I have to buy a new case which blows

[static tusk]

yep 3D rendering was the exact reason why I bought the 3080, 2 years ago. AMD still are terrible in that space

[dense tapir]

I sort of have settled on it being a Lian-Li as either of the two cases allows water cooling in the future after I recover the financial loss of having to upgrade everything.

216 vs III as I am unsure which to get.

I see in the III more room but a lot of shit to go wrong over time as there are too many moving parts and clicky parts.

GN mentioned that in their review too.

I am leaning towards the 99 USD 216 without led fans.

[rocky charm]

Hay

[dense tapir]

I think its max is 395mm long GPU while the III does 423-425mm

Funny thing is if I buy this case I am sort of wanting to get the smaller more expensive water cooled 4090.

[rocky charm]

I need a photo showing qnc1te PARTY in holiday style

[dense tapir]

Gigabyte has one so that goes out the window

[static tusk]

Its so funny. Im watching a couple of 4070 reviews on YT and my initial response was to laugh at how small it is. Ive been conditioned to expect large cards....

[dense tapir]

That, and its power consumption get a BIG thumbs up to but that is it.

1 of 2 things, or both, that card needed that would have knocked it out of the ball park. 16gb and/or 499

They sort of shot themselves in the foot with the 12gb 4070ti

4060ti at 8gb and 499 is DOA imo

if they end up doing 10gb for it I could give it a pass and say alright for a 1080p budget card.

[smoky oak]

Damn, they need to get somebody who can actually benchmark SD, cause those results are trash lmfao

13.84 images at 768x768 on a 4090 lmao

[dense tapir]

Seems fine to me as I am lucky to get 1 per minute

no 4090 shown

not 7900xtx

[smoky oak]

sorry, 4080

its really small in the thumbnail view

thought it was 4090 lol

[dense tapir]

As I mentioned they did not show the flagship models from either company for some strange reason

honestly 768x768 is slow af

so a 4080 able to push almost 14 per minute makes me drool

[static tusk]

yeah looking at it, its images per minute. thats off

[smoky oak]

I am confident I could break 12 images per minute at 768 on my 3060ti no problem lmao

[smoky oak]

even that seems off, cause my 3060ti can go 512x512 sub 2.5 seconds

maybe just cause I actually know what I am doing

[static tusk]

how many steps are they using?

[smoky oak]

probably 50 or something like people who have no idea what they are doing lol

[dense tapir]

None of this matters if the same metrics were used for all posted. YMMV for various reasons and no mention of steps, sampler, etc...

[smoky oak]

yeah, see, I just did 16x 768x images on my 3060ti in 70 seconds

[smoky oak]

only problem is that its extremely misleading to people already in SD

if I looked at that graph and didn't know better, I'd think the 4080 is a small upgrade over my 3060ti in speed lol

considering mine got 12 images per minute at 768 lol

[dense tapir]

Then you are looking at it wrong

[smoky oak]

it does at least show the relative performance between those GPU's, but that performance is way under what it would be if they were testing right

[dense tapir]

if your card is not compared in the results, it isn't, then you have no idea

They do not tell us what the step count was or the sampler etc...

[smoky oak]

if thats the case, then that graph is completely useless to know if its worth it for me to upgrade lol

they should at least have the common sense to include basic settings

sampler and steps alone would be a massive insight

[dense tapir]

that matters if you are chasing after raw numbers but what they are showing is card A vs card b

I agree

they should have included the 4090 as they did prior

[smoky oak]

yeah, I guess. Just an extremely niche graph that is poorly made for literally any other useful interpretation of it

[static tusk]

in the January benchmarks they used 100 steps and Euler. My guess is the same settings here

[dense tapir]

you are missing the point as they are not selling you a card this is just saying this graph is to show how each of these cards did against each other. No other information is intended.

Yes, pure is right as I remembered 100

Jan showed 4090 too I believe

[static tusk]

https://www.tomshardware.com/news/stable-diffusion-gpu-benchmarks

[dense tapir]

Of course if I had a 4090 I would 100 steps all the time

[smoky oak]

I personally wouldn't increase the steps and waste time and energy for no increase in output quality

[dense tapir]

I think what this testing was for, after reading the comments, is to show just how much the 4070 sucks ass juice. Nvidia is charging more for, at best, the same performance to worse performance than last gen 3080

They have seemingly hit a wall which is probably why Jensen said they are starting over from scratch for the 5k series and the 5090 is 2-2.9x faster than the 4090 but the pricing I can only nightmare over.

3nm too

See how they hit this wall? There is a theoretical limit to how low they can go and 2nm was about it. 4 to 3nm and probably 2nm then stuck because the traces are so close the electron jumps connections etc... really messy.

your trace separations are smaller than the electrons themselves.

iow we are about to hit a nasty wall with a head on collision I am unsure how we work around it. Maybe 3d tech or something.

[static tusk]

more cores, larger cache, faster vram, will be the temporary answer

[dense tapir]

bigger dies

no way around it the dies will have to get larger

[lilac pine]

Hlo

[dense tapir]

this is why I think AMD went to 3d tech so the die grows upward and it saves in latency I think they said as well.

be sides they can afford to grow up more than out by much more.

Just think NYC hit this problem over 150 years ago, lol.

[static tusk]

look at what Intel are developing for Granite Rapids server chips. Thats a big boy CPU socket

[dense tapir]

OUCH

how big is the motherboard?

[static tusk]

https://wccftech.com/intel-super-massive-lga-7529-socket-pictured-ready-for-next-gen-xeon-cpus-with-up-to-132-core-p-cores-512-e-cores/

[dense tapir]

Latency begins, yes

I agree

through hole I understand but through hole water cooling?

all I know is I lay odds the 5, or 6k will be the last gen before the brick wall smashes Jensen in the face and they will be forced into a radical change.

Ahh, I thought that is what you were saying

even 3d can only go so far before issues begin per Samsung

I think on their drives they use 3d tech

possibly but the latency involved

I am just not sure how we are going to dig ourselves out of the upcoming hole

you can't go any denser or higher or further out. You have pushed the silicon as far as it can go.

[static tusk]

something other than silicon possibly

[dense tapir]

What happened to the graphene that was all the tech rage?

[static tusk]

died off

[dense tapir]

yeah, strange

died off because graphene is actually hard to make and dangerous too. I was watching this scientist show how it is made and how to make it and man I can see why as silicon is so easy.

in comparison

I bet when we have no other choice is when they will revive it.

[static tusk]

there was also a hint at man made diamonds as a thermal solution

also died

[dense tapir]

I remember that

not sure why that one died so fast. Maybe because it would possibly dilute the heavily controlled diamond business? I dunno honestly.

[static tusk]

yup the price of natural diamonds would plummet

[dense tapir]

Thought so. Screw the advancement of mankind cause muh wealth.

I am pretty Vulcan about that. Needs of the many and all that.

[static tusk]

Im sure there are plenty of inventions and ideas bought out and sitting on shelves till its time to make more profit

[dense tapir]

Sadly

[dense tapir]

@smoky oak https://www.reddit.com/r/StableDiffusion/comments/106zzm5/4070_ti_if_you_have_one_care_to_do_a_couple_tests/

[smoky oak]

Lets just see if its not trash like their last 3 generation of rapids CPU's lol

[smoky oak]

👁️

As it is now, intels CPU's are pretty damn pathetic in almost all fronts

They cost more, their boards cost more, they are way less performant for that cost, they use way more power for less performance

just a mess lol

Epyc Genoa makes all of intels server side CPU's seem like cheap jokes

[dense tapir]

What is HRO?

High res fix?

[smoky oak]

no idea, didn't look at the post for long

[dense tapir]

Not sure what that is and the link doesn't say

Good to see some tests for 70ti

[smoky oak]

Those results also seem bad lmao

I am so confused

maybe its early adopter drivers

[dense tapir]

what is funny is the 768x768 on the 70ti is what I used to get on my 1060 at 512x512

[smoky oak]

I know my 3060ti was a problem for like 4 months when I bought it

let me run their test and see how I do, just a sec

[dense tapir]

Not sure but those are what people are getting even today

I am not sure what HRO is as that may mean something

[smoky oak]

blegh, have to download 1.5 base

[dense tapir]

he did the shit in that thread on github and 1/3 less time

45s to 15s

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/2449

[smoky oak]

That would be 2/3rds faster

not 1/3rd

[dense tapir]

2/3 less time

[smoky oak]

yeah

[dense tapir]

What I meant.

Basically, it is just changing a few files and the 4070ti got faster

Still

[smoky oak]

thats it?

[dense tapir]

your 3060ti is almost as fast? what is yours at?

[smoky oak]

My 3060ti if I really crank can breach 11.5it/s

[dense tapir]

20 step euler a is even fast for me

[smoky oak]

its not euler a, its euler

[dense tapir]

his above is euler a

[smoky oak]

so yeah, my 20 step test was 2.94s

so considerably faster than this persons results with a 4070ti

let me try 50

5.47s

[dense tapir]

alright so 1.73s vs your 5.47s

[smoky oak]

what?

His was 7.66s

[smoky oak]

oh, with that one, yeah

wait no

its 1.73s vs my 2.94s

my 20 step was 2.94

[dense tapir]

you said that was euler not euler a though

[smoky oak]

my 50 was 5.47

oh sorry, let me change to euler a

[static tusk]

512

[dense tapir]

which card?

[smoky oak]

Euler A was 5.39s

for 50 steps

[dense tapir]

20 steps

[smoky oak]

let me try 20

[dense tapir]

I am trying to compare

[smoky oak]

2.45s

[dense tapir]

sheesh the 70ti sucks

that means the 70 will be even worse

[static tusk]

1024

[smoky oak]

let me try 768x768

5.50s

let me try 1024x1024

[dense tapir]

seriously, this speaks so bad of the 4070 anything

[smoky oak]

11.56s

[dense tapir]

it got ya there

[smoky oak]

it did max my VRAM, so I figure that contributed some slow down

yeah

[dense tapir]

yes, obvious

12g vs 8

[smoky oak]

which is very weird, no other models max my VRAM at this res

I usually do upscale at 1200x1200 with 96% usage

[dense tapir]

upscale buckets

little chunks

that was a full blown 1024x1024

[smoky oak]

no, i mean ultimate upscale, it does full res gens then stitches them together. Its teh exact same speed as a single img2img

[dense tapir]

anyway, this shows the 70ti something is wrong

[smoky oak]

high res fix does bucket

let me try my preferred settings

on my preffered settings, 512x took 2.1s

[dense tapir]

I am just not sold on these new 4k series cards for this stuff.

better but barely

[smoky oak]

switch was DDIM on 15 steps (which I have shown is petty good quality)

[dense tapir]

I mean the 4k series

4k over 3k is barely a step up for SD

[smoky oak]

its not like I am just dropping the quality super low for speed, the quality is still there

15 steps ddim

vs 20

trying 768

4.71s on my preffered

cool result heh

[dense tapir]

btw, today is the official FE release day for the 4070 and tomorrow is for the AIB cards

[smoky oak]

1024x is 9.6s on pref settings

fun result

the fact that the 3070 and 4070 have the same core count shows that NVIDIA is not after gains, but rather ease of manufacturing

as LTT showed, its around 26% better on average, for about 21% more money :/

its die is also stupid small, which means it costs way less to manufacture as a single silicon wafer can fit wayyyy more, and also the single failures account for far less of a disks lost yield

153664mm vs 87025mm

just slightly under 50% smaller

so that means if you took the die area of 100 3070ti's, and then used the same area for the 4070ti, it would be able to make 176 die in the area it took to make 100 last gen

76% more die to wafer efficiency is monumental

[dense tapir]

yep, but none of that given to the people as the cost should have been less.

[smoky oak]

the only difference is its a much smaller process, which means the yields are naturally lower, and the wafer cost is naturall higher, but still. That should be at max on par

[dense tapir]

https://youtu.be/cZ6ODLbEmjk?t=197

your card against the 4070ti

[supple ocean]

Hi!

[smoky oak]

so basically he is showing my 3060ti is the same speed? lmao

[dense tapir]

if the 3060ti had 12g too it would be faster across the board. wtf?

[smoky oak]

but the VRAM makes the difference

lmao, wow

There has to be a problem with drivers

[dense tapir]

has to be

that was 2 weeks ago too

although the bus width

[smoky oak]

however, I will say, this is a null issue when using processes like my upscale, as my upscale doesn't ever run into VRAM overflow #SelfPlug

no VRAM slow downs, even at 8k+

[dense tapir]

what we are seeing, I firmly believe, is your 256bit bus vs the 4070ti 192bit. It does matter.

[smoky oak]

THE 4070ti IS ONLY 192 BIT!?!?!?

[dense tapir]

regardless the copium from Nvidia

[smoky oak]

BRRRRRRRRRRRRRRROOOOOOOOOOO

VRAM buss makes a MASSIVE difference

its why the 3060 is so much slower at tons of things compared to the 3060ti

in blender, the 3060 can be less than half the speed of a 3060ti in ray traced rendering cause of the VRAM speed

[dense tapir]

I know and Jensen says no it no longer does because the buffer chip is 32mbs now. NO, it does matter as we are seeing it.

[smoky oak]

hell, in Etherum mining, the best way to increase perf is to max out your VRAM overclock as much as possible. Blender too, and SD too lol

[dense tapir]

my 1060 6gb has the same bus as that 4070/ti

[smoky oak]

my GPU OC is 150+/900+

[dense tapir]

they seriously took a HUGE step backwards

[smoky oak]

when I used to Ether mine, it was -200/+1350

VRAM matters so much for so many things, like what lmao

[dense tapir]

one of the big things of the 4090 is its 384bit bus

[smoky oak]

its only 384 bit

[dense tapir]

even the best intels have a hard time not bottlenecking it

[smoky oak]

the 3090 is 512 bit ._.

sorry, sorry

3090 is the same

384 bit

[dense tapir]

Yep

I will bottleneck it but once in its memory for training it can work it. personally I would prefer a 4080 if it were 24gb

4080 has a 256bit bus like your 3060ti

I just feel every single card not a 4090 is a con job from Jensen

[woven dawn]

^

That's why I broke my wallet and got a 4090

[smoky oak]

Thats exactly what they wanted you to do too

[woven dawn]

I know

[dense tapir]

Yep

[woven dawn]

Everything else is ass value

*by comparison

[smoky oak]

yup

[dense tapir]

Funny thing is you can grow into the card as the cpu bottlenecks it

[smoky oak]

in some ways, yeah

[woven dawn]

In the bright side, 48it/s

[dense tapir]

yeah

which model did you get?

[smoky oak]

4x as fast as my 3060ti basically

for like 6x the price

[dense tapir]

apples to oranges

[woven dawn]

I got the zotac amp extreme - the only one that microcenter had when I went

[smoky oak]

what did that run you?

$2600?

[woven dawn]

1800

Usd

[dense tapir]

I have to buy a new case for any of them 😦

[smoky oak]

oh, thats not that bad

[woven dawn]

[smoky oak]

$1800 is a lot more reasonable than I expected

[woven dawn]

Microcenter actually sells stuff at ~msrp, so I'm glad I live near one

[dense tapir]

These new cards are far too long

I wanted one now for over a decade but no dice.

[smoky oak]

I am kinda glad that I have an 8GB GPU now, cause its leading to me pioneering a method of upscale that is not only way better quality than what we have now, but also can run over 8k resolutions on 8GB cards

[woven dawn]

Waow

[smoky oak]

Just to flex it again really fast, here

Base res vs my process

[dense tapir]

If my case allowed water cooling I would get a 4090 water cooled one but nope so it looks like a lancool III or a lancool 216

[smoky oak]

512x512 vs 3840x3840

[dense tapir]

I am extremely not happy at being forced to change my case

[smoky oak]

and yeah, it shows lol

[woven dawn]

Where's your outline posted?

[smoky oak]

Still in the works. Whole process got a rework last night as I discovered something massive in the process

results dramatically improved

[woven dawn]

[smoky oak]

For example, this is a screenshot close up of the same image with both processes. Old is left, new is right.

Same res image, believe it or not

shows how big of a difference it makes

with my new process i can guarantee pixel level detail at 4k, and I am sure I can adapt it for pixel level detail at 8k as well

512x512 vs 3840x3840

[coarse pendant]

Very nice 👍

How long does it take to upscale it to 3840?

[smoky oak]

on my 3060ti, about 100 seconds?

give or take

so 1:40 or so

[coarse pendant]

That’s pretty fast

[smoky oak]

yeah, considering its not possible using things like high res fix 😅

And its also way better quality and coherence than high res fix can offer either

[pastel mango]

meow apocalypse

[violet salmon]

Is there a way to apply a "Pixel art texture" to an "isometric reference model"?

Here is the model

and I would like to apply one of the "textures" here to the walls:

could img2img with controlnet work? Or should I look for image interpolation solutions instead?

[pastel mango]

why not make a lora?

with very specific trigger words for each image in the dataset

[stoic pivot]

I am testing negative prompts and for me "beige" was good. It look like many interior images are based on armature photos that have a beige tint due to auto white-balance, when "beige" was added to the negative there was more variation in the images.

[violet salmon]

What would be the best way to proceed? A different lora for each "tile"?

[pastel mango]

nah one lora that you teach the style for and then describe each tile in the prompt text file for it. Describing it very simple and accurate with a specific triggerword for each

that way you could also alter the looks via prompt, if you do it successfully

I dunno if it would be better to have an embed though

[violet salmon]

oh, maybe with a lora for style + controlnet for each piece of wall then?

[dire current]

hello, im new here

[dire current]

how do you do that? is it a prompt or a additional api?

[glossy herald]

Little reminder, the AMA with Emad with start <t:1681401600:R> in https://discord.com/channels/1002292111942635562/1005545148211527800
You can already come and ask your questions in #1029055412764422214 before the event starts ! (sorry for the messages :p but fun time to come)

[oak siren]

how is progress with hands?

[zealous stag]

https://cdn.discordapp.com/attachments/895353923978821692/1096105206086897765/image.png

simple prompting works too

[lusty jasper]

I'm surprised at how good it can do futurist art

[rugged pendant]

Anyone tried SDX? Is it better than custom models

[vague oasis]

its def an improvement from sd 1.x

i actually like it a lot

[ruby swift]

this is great. what is your process?

[runic zephyr]

Is it something with new code or you've just found a way to use existing addons to achieve this?

[slender needle]

These are awesome results. Are you using auto1111 or comfyui or something else? And where/how are you planning to share your guide?

[smoky oak]

@slender needle @runic zephyr @ruby swift

They are using ultimate stable diffusion upscale, I am working on a detailed guide as it has a lot of steps and variables, and it's going to be laid out in a Google doc with lots of information and images for people to search through.

The guide has been delayed a bit as I discovered that new step (it's outside of SD, but something that takes legit 5 seconds to do once you have it set up, so I will explain along with the guide)

[runic zephyr]

So it's just tiled vue and tiled diffusion in effect?

[smoky oak]

Yes, but it requires various different variables, a little bit of post-processing out of stable diffusion (The part that allowed me to get significantly higher detail)

As well as lots of interchangeable values that you can utilize in order to get a result that closer matches the original input, or a much more creative result that is inspired by the original input.

It can also be utilized to fix generation artifacts such as an overtrained LoRA with shockingly good accuracy and detail

I'm approaching 400 hours of time invested in this process, and I found several different bugs and issues that I need to detail on how to fix, and I've also found several different niche use cases that I would like to include in the guide as well

[runic zephyr]

That's interesting.

[smoky oak]

@Oauabei#6406 please have a look above to see my explanation of what exactly it is, and how I'm going about releasing a guide

Oop

Did they leave?

[runic zephyr]

Or get banned 😄 One of two.

[smoky oak]

Oh yeah, they are gone. Welp

Nevermind

Haha

But yeah, I've recently discovered that certain upskill increments work better with certain upscalers, and then certain upscalers are able to improve the artifacts of some of the other upscalers.

It's really not a hard process, and I can get from a 512x512 image to a 3840 x3840 image in a few minutes really, but it's just understanding the variables and knowing how to play them to the strengths of what you need

So really, I could detail my entire process right here, but without explaining what needs to be changed, it won't be ideal for most other generations

[slender needle]

Thanks, looking forward to the guide, the results are really amazing.

[smoky oak]

Hey guys, anybody know the command to run SD on CPU? I have a friend who has a pretty slow PC, but hes damn determined

Ok, think I got it

seems to be CMD arg --use-cpu all

https://tenor.com/view/crazy-cat-attack-gotcha-cat-jump-gif-15732712

Oops

I didn't mean to send that here, but I feel like it fits lol

[smoky oak]

@dense tapir ohhhh my god lmaaoooo

My friend has no GPU and needed to run on CPU

hes at 11.2s/it

now 15.22s/it

17.27

22.3s/it

[dense tapir]

Same on my old 1600

I haven't tried on the newer 5600

I have to say the 4070 has landed with a pretty big thud.

[smoky oak]

Can't say thats surprising

[smoky oak]

@dense tapir how much VRAM does it have, and how big is the bus?

Sorry, I would Google it myself but my hands are occupied right now, I only have voice typing at the moment

[dense tapir]

12gb 192 bit bus

as we saw it is just a slightly faster (for SD) 12gb 3060ti

Man, I just can't find a case I like. Found one only they stupidly lock you in with the control panel on the left hand side, then the lian-li no longer offer HDD/Power on lights and a lot of the new boards are using the power on light to blink for errors. Another stupid.

[calm bay]

hi im new here

[crimson dawn]

sd ultimate upscaler is awesome. the author there deserves a lot of credit for making such a great tiled upscaler with diffusion tech, and releasing it entirely free

https://github.com/Coyote-A/ultimate-upscale-for-automatic1111 i'd at them for thanks but i don't see a coyote-a on discored

[rain gazelle]

Did anyone create a decent Terminator/T-800 model?

[smoky oak]

I could probably do it

[smoky oak]

right right, I was thinking it was the 4070ti we were talking about, my bad

@dense tapir@glossy herald Please please please wish me luck. I am buckling down and redoing my whole guide to explain and showcase the new workflow. First I need to remake my example images

[glossy herald]

You didn't release a first version before redoing everything ? Oo

Don't prepare your jump for too long though

[dense tapir]

That is a lot of work you are putting in there so best of luck.

[dense tapir]

btw, I am watching Moore's Law is Dead channel and now I see why the 4070 showed to be just a glorified 3060ti with more ram because the die size was cut down just about the same as the 3060ti was. Add that with a 192 bit bus and it explains the numbers I posted yesterday and it showing it isn't much better than 3060ti in SD at least.

[smoky oak]

When the results are as big of a change as that breakthrough I showed you in DMs, I definitely have to build off of that framework first

I mean, just seeing my old process next to my new process should show that we're not dealing with a minute improvement 😅

[dense tapir]

Oh, he just said it that the only thing it has going for it is its 180w power consumption, like the 3060ti and should have been called the 4060ti. BOOM, I said that last month that this is what the 4060ti should have been.

Nvidia games to screw us all over is not appreciated. I already was disgusted at them during the Crypto mining days when rumours were saying Nvidia was pushing what was happening that have now been confirmed to now these shenanigans. If only I didn't need CUDA.

[crimson dawn]

maybe gpu conspiracy chat can go into the gpu chat channel instead of polluting gen? i see it in so many servers. fanboys drop gas bombs until they catch a bite and lite the whole room up about the tragedy that is [insert gpu company here]. Having a gpu room for you guys is a good compromise and is more polite than how i see y'all treated on other servers. banned.

gamers being mad about expensive gpus is nothing new even before crypto hehe

[glossy herald]

It's a recurring subject for sure, and people can go in the #1011228477954998273 if that becomes prevalent, but we try not to stifle dialog until it's necessary. Some GPU talk is ok in gen imo, I didn't feel it be a problem right now. But let's share more pics :p

A digital version of the Mad Hatter's hat, featuring a built-in AI assistant and holographic displays, illustrated in the sleek and minimalist style of Jonathan Ive.

[wispy nest]

i tossed 2 cutesy embeddings with a negative embedding into the positive prompt and apparently made a low effort pokemon

[glossy herald]

hehe it's so cute !

catpugmon

Trying to get a good robotic Queen of Hearts with Alice

[wispy nest]

robots can be hard. so many different types muddles the field

[glossy herald]

Yep. and the style itself, not sure where to go with it

a little star wars feeling would be cool

[wispy nest]

even among the various subdivisions there are just -so many-

[wispy nest]

that's a pretty sweet style imo

[glossy herald]

yep I'm late this week

hard to do something not done already

I may start up comfyUI soon and let the LLM try things

[wispy nest]

very aesthetically pleasing with the red contrast against the cold background

[glossy herald]

the prompt has good constrast there

I need to find the good seed I think, maybe adapt it a little but not much

the red and blue is cool

An Alice and a robotic version of the Queen of Hearts, featuring sleek lines and glowing LED lights, inspired by the futuristic designs of Syd Mead and Neil Blomkamp.

Steps: 50, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 3534123143, Face restoration: CodeFormer, Size: 1368x768, Model hash: fc9afc15a5, Model: perpetual-v1-sun, Clip skip: 2

[wispy nest]

out of curiosity, anyone know the decision behind discord removing metadata from images?

was a big "oof" to sharing sd images

[glossy herald]

they have gone on and off about this in the past

the reason has always been the same though

[wispy nest]

i assume a safety thing

[glossy herald]

protect the less tech savy users that were sharing pictures with metadata that could doxx them basically

[wispy nest]

figured

[glossy herald]

nice to share a pic of your dog, less nice if you give up your adress and all of that

(and yeah, I can attest that I removed quite some content on that basis in the past)

sometime you get a new user coming in, using their real photo and name as username/profile pic

not everybody is knowing what discord is

damn, another fire one, I'll upscale it

[wispy nest]

there was a big assembly at a school near me about 10 years ago that discussed the dangers of geotagging your photos even then.

kids would do it because they thought it was cool and were basically giving gps coordinates to find them -_-

[glossy herald]

yeah... tech has gone fast, and there has been lots of traps like that
If you didn't use all this when it came out, evolved with it, you don't know those. You get owned on so many sides before knowing. And even then, you can still get owned, social engineering and just sweetalking can be quite effective on humans

[wispy nest]

even more so now that you can fake voices pretty realistically. story came out a few days ago about an older couple that got a phone call from their son that said he needed bail money to get out of prison. was AI generated. scary stuff for those unaware

[glossy herald]

hell, I have no excuses for having been hacked 2 months ago. I have 15 years in professional IT. And still, a weakness, someone pushing on it, and I broke just long enough to loose a lot

[wispy nest]

i had a similar incident a long time ago when i used to play Ultima online. been doing computer stuff for a while at that point and still got tricked into putting a keylogger on my pc -_-

[glossy herald]

I ran a .exe .... they stole my auth cookies, transfered some of my authenticators to their phones, locked me out of my main email,... x) it was a fun time

[wispy nest]

honestly feel like the only reason it hasn't happened much more often to me is people realized i don't have much that they can take lol

[glossy herald]

I was lucky it was some discord kid scripty in the end, because they technically had a hold on my paypal, and didn't realize it

[glossy herald]

you don't need to have a lot to loose honestly... even just a way for them to spam is a reward enough

(yeah, like I said, I have no excuses, .exe is .exe. I scanned it and took the risk I shouldn't have taken)

[wispy nest]

i kind of feel like that recently. There is a pretty cool extension for auto111 that lets you animate a face to match (pretty well) an audio clip. Requires you to run --disable-safe-unpickle though. I ran it anyways...I'll see in a year or so if it was worth the fun little clip i made -_-

[glossy herald]

yeah... I don't take those risks anymore. And the tools that require me to do, I run in an empty linux instead

too scared right now

[wispy nest]

i should've at least done that--ran a vbox of some sort

[glossy herald]

I should have least have cut my internet before trying anything

lol

[wispy nest]

hah

[glossy herald]

but so many things you think off once you got owned ^^

but no, I knew those

[wispy nest]

just thought "it can't happen to me", eh?

[glossy herald]

I just didn't do the effort after a point given how many we run tbh

and I was being pushed toward it, I fall for the old "no trust me"

like the first time I gave my full armor in T4C to someone to "upgrade it for me"

given I was 12

but still x)

felt pretty stupid

[wispy nest]

I never played that one. I played Realms of Despair, Aardwolf, and a ton of others though

heck, i still jumped into Aardwolf again just a couple years ago. i don't usually play them for long, but MUDs are pretty great for gameplay so long as i can still type

ah. thought T4C was a MUD based on it saying Realmud. Seems I was mistaken lol

[glossy herald]

it was the first MMO I played

back when we were paying the Internet by the hour

on the 45kbits :p

it was my first online game, and mostly first real game I think, on computer ^^

[wispy nest]

I still recall getting AOL cds in the mail regularly lol. I actually had a service called "Juno" for my first internet. Was basically offbrand AOL. Dialup connection and allowed to stay online for a certain amount of time before it killed the connection.

I held off on AOL for quite a while, though eventually switched to them after giving Juno a chance. I'll be honest, i didn't expect much from the internet in the early 90s. Was a cool thing that interested me, but couldn't justify the expenditure when i was mostly using it for BBS and the occasional game

[glossy herald]

it still was such a magic world even back then. no real big uses at first, except knowledge, but such a new thing. I remember getting into problems with my parents for using the whole 30 hours we had for the month in 2 weeks on my own

[wispy nest]

this is exactly how i feel about ai generative technologies at the moment. reminds me so much of the early days of internet

difference is i expect really big things from this

[glossy herald]

yeah, and quite fast, a lot faster than the 10 years internet needed to mature at first imo

I'd be surprised if our lifes weren't transformed by this in the 3 years to come already

[wispy nest]

the rate of development, ever increasing use cases, and large adoption rate all seem to mirror it quite well

[glossy herald]

did you start playing with LLM too ?

[wispy nest]

it's been a lot more explosive recently though

i have

[glossy herald]

I tried a little chatGPT online and was like "yeah this has potential"

but installing it, and being able to fine tune a character already so easily, it's insane what it said to me with the few knowledge I have of this

I'm also plugging it directly into SD

in this workflow (comfyUI), you have my input at the bottom "a cat", then top left the whole LLM thing, making a prompt
and top right, SD making the pic

[wispy nest]

i run music generation with Riffusion, LLMs using Vicuna, Image generation with Stable diffusion, voice generation with Tortoise, and frame interpolation with FILM. all of which were made possible with AI in basically the last year. crazy how it's leapt forward when development was going on pretty heavily for like 5 times that long

that node setup seems interesting. something else for me to eventually end up learning, probably lol

i know you mentioned comfyUI before, but it seemed like just a reskin to me at the time. looks far more capable than i originally thought

[glossy herald]

it's a series of nodes

basically it's just SD on steroids

but it has extensions, custom nodes, just like automatics

since I started adding some, like this LLM, I get more and more things I can plug together

being able to make the workflow really yourself is quite strong in the end, but it sure requires more time

and coding a custom node is really not hard

[wispy nest]

I honestly expect most of what we've learned in the last year is going to be irrelevant with the rate LLMs are advancing. ChatGPT-4 with auto gpt is quite frankly mindblowing to me. A few decades ago we had whole teams working on the workflow that gpt4 can do autonomously now

and it does it crazy fast compared to people. the rate it can receive and act on information is just completely insane to me. I expect there will be other competitors in the very near future, and look forward to the potentialities

[glossy herald]

yep. the "baby agi" or auto gpt are doing things incredible, and it will go even faster because of it now. it's mostly computing power that will become limiting

let's build a character for the LLM to chat with :p

[wispy nest]

i frequently use notebook mode with oobabooga webui, though i admit the character chat works ridiculously well

[glossy herald]

I'm loading the model right now, one with not that many restrictions, and fine tuned for role playing characters

[wispy nest]

i use it more for storytelling

[glossy herald]

I'm using the TavernAI UI on top right now

here is the character creation screen, I named her Essde (SD)

her/him/it

[smoky oak]

new demo for the new process

[crimson dawn]

all you need ot send them is boobs.jpg.exe and they'll open it 70% of the time, everytime

[glossy herald]

I tried to make it a funny AI trapped inside a computer, I got something else, interesting convo :p I'm stopping it there, it's becoming quite inconsistent, but fun read

[trail cradle]

I want Rick's butter butler

[wispy nest]

Really enjoyed the feel good bot for a bit there. It's too bad it seemed to go dark at the end 😦

[sinful mirage]

love inpainting with different models and prompts

[sand copper]

If I copy someones exact parameters including seed should I end up with the same picture?

[glossy herald]

mostly yes, if you have the same model too

some options like xformers can generate small differences

[sand copper]

hmm my guess might just be different versions from the same checkpoint then, thanks

well I guess different checkpoints from the same model

[sand copper]

nevermind I've tried it with many different checkpoints so idk

weird

or maybe I'm just missing some parts

[dry crow]

Mostly if you only copy the settings by Hand instead of Pressing the copy image data button on civitai

[sand copper]

copying data wasn't working so I was doing pnginfo

the loading boxes would show up but nothing shows up in the prompt boxes

[dry crow]

Hmm can you show an example?

[sand copper]

yeah 1 sec

It took me a second to find something it'll be done in a second

[smoky oak]

Anybody have any ideas for my demo images for my guide? I need 20, currently at... 3

[sand copper]

https://civitai.com/images/175092?modelVersionId=17233&prioritizedUserIds=104608&period=AllTime&sort=Most+Reactions&limit=20

and looking at it different model

this website is a bit confusing

or the images aren't tagged correctly

but speaking of any reason why my clipboard might not be working? If it makes a different I'm using brave

[dry crow]

Getting these settings if i press copy image data:

masterpiece, best quality, ultra-detailed, illustration, 1girl, solo, fantasy, flying, broom, night sky, outdoors, magic, spells, moon, stars, clouds, wind, hair, cape, hat, boots, broomstick, glowing, mysterious, enchanting, whimsical, playful, adventurous, freedom, wonder, imagination, determination, skill, speed, movement, energy, realism, naturalistic, figurative, representational, beauty, fantasy culture, mythology, fairy tales, folklore, legends, witches, wizards, magical creatures, fantasy worlds, composition, scale, foreground, middle ground, background, perspective, light, color, texture, detail, beauty, wonder.
Negative prompt: nsfw,(worst quality, low quality:1.4), (lip, nose, tooth, rouge, lipstick, eyeshadow:1.4), ( jpeg artifacts:1.4), (depth of field, bokeh, blurry, film grain, chromatic aberration, lens flare:1.0), (1boy, abs, muscular, rib:1.0), greyscale, monochrome, dusty sunbeams, trembling, motion lines, motion blur, emphasis lines, text, title, logo, signature
ENSD: 31337, Size: 512x832, Seed: 3080004317, Model: AOM3A3, Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Clip skip: 2, Model hash: eb4099ba9c, Hires steps: 20, Hires upscale: 2, Hires upscaler: Latent (nearest-exact), Denoising strength: 0.5

[sand copper]

the image is tagged under AOM3A1B but model says AOM3A3

[smoky oak]

@sand copperWhat GPU are you using to do these gens? that can massively affect your generation results

[sand copper]

3070ti

[smoky oak]

ah, nevermind

[dry crow]

Then you need the b model

[smoky oak]

non tensor cards have no consistency cause they run on emulated precision

General Awareness and I tested it. very weird

on my specific GPU,I have 0 Xformers noise

I have even done pixel value offset comparisons. I get the exact same image every time

[dry crow]

@sand copper you can find the model here:
https://huggingface.co/WarriorMama777/OrangeMixs/tree/main/Models/AbyssOrangeMix3

[sand copper]

yeah it goes into my clipboard but stable diffusion doesn't take it, so PNGInfo is the best I could do

though I'm sure there's many other factors I can't see so I don't know what it is

just going to assume it just is what it is

[dry crow]

Like how ? You need to paste the whole thing into the prompt and then click the blue arrow button