🏞｜general-with-images

[dense tapir]

Someone I know did exactly that and it couldn't even work on 2.1

[bold cave]

2.1 is and forever will be poopy

[hearty karma]

im majorly hoping high res fix is going to work 🤞

[dense tapir]

I am shocked at zero programming knowledge and out pops new schedulers that not even Kohya was doing

Thanks to GPT

[hearty karma]

im already using 5.8gb of vram

[bold cave]

wtf?

u must be doing something wrong m8

cuz tht looks like pooh

[pure canyon]

I found some shared images in facebook group.Do you know any info about this model ?

[hearty karma]

its not done...

thats like 50 steps in

[bold cave]

dayum those are neato mosquito

[hearty karma]

out of 91

[dense tapir]

I know this I am elated that I asked gpt to write me a regex and it worked where all tools failed that were designed for the job. I have despised regex for over two decades as it just doesn't click with me.

[bold cave]

SD can do regex?

[dense tapir]

No, was for a side project

[hearty karma]

crap... of course it didnt work

[dense tapir]

would be damn cool though

[bold cave]

we have dynamic prompts but I havent fudged wif dat shih

[kind quartz]

@hearty karma if you have slow gpu 50 steps can be overkill.

[bold cave]

also make sure ur using euler a

[dense tapir]

I just wrote a bunch for that

[hearty karma]

my gpu isnt really slow

its just not got a lot of vram

[kind quartz]

but lowvram slows it bit more

[hearty karma]

im on med

[dense tapir]

My 1060 feels for ya

[bold cave]

buy some vram chips from ebay and just solder dat shit onto ur gpu

boom easy 80gb

[dense tapir]

I wish we could do that

[kind quartz]

what i know some ppl here are using DDIM with 32 steps

[bold cave]

lolloooool

[dense tapir]

32 steps ddim makes my card cry

I do 20

[bold cave]

ddim is helluh slow fo sho

when ur used to euler a speeds

[bold cave]

then it definitely shows

[kind quartz]

sometimes more steps means lower quality or not so nice composition

[dense tapir]

in DDIM it sharpens

[bold cave]

yeh, in some cases in makes the image look burnt

[dense tapir]

listed in its papers

[bold cave]

is that a shotgun blast survivor bear

[hearty karma]

yes

[oak osprey]

splotchy base 2.1. like image gen if you gave it a bad case of acne

[dense tapir]

@bold cave I am still a bit shocked at an earlier gen I showed in here

[bold cave]

ima show u a real chimera

[bold cave]

damn thats nice

is that upscaled

cuz if it isnt then it has a ton of potential

[dense tapir]

no, I wish it were

[bold cave]

I would do it if it weren't that im busy with my dreamboothies rn

Ive become way too focused on portraits

[dense tapir]

aye

[cunning ivy]

how use?

[oak osprey]

im seeing what people mean when they say SD2.1 doesn't really create people/animals very often

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no public bot on the server that generates images. There is an experimental bot available for early server members & Stable Society. You can obtain the Stable Society role by winning our weekly events #⭐｜pow-info #1087493421209485393! However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610⁠ for more details! You can also stop by ⁠#1025467151206854736 for any issues you experience while using DreamStudio or ⁠#🤝｜tech-support for any problems you encounter while installing it locally!

[dense tapir]

well, my card is so slow and I am saving that one for @smoky oak to do their magic on.

[hearty karma]

i think my animals just suck because im merging two

[bold cave]

bruh sytan has a 3060 or so lamo

not exactly lamborghini speeds

[dense tapir]

yep

I am a 1060

[bold cave]

wut

[smoky oak]

I legit just woke up as you pinged me- lmao

[bold cave]

bruh throw that junk away man it's an antique

[oak osprey]

you are an AI Language model, you mean

[dense tapir]

Look above sytan

my ocean pic

[kind quartz]

merging my elephant is result of say (elephant with girafe)
it is about luck as well. Better start with low steps and if you like it add more steps.

[hearty karma]

the 1060 isnt that much of an antique

[dense tapir]

it is shit for this

[smoky oak]

oh nice! Ocean is something that AI has strugles with

[hearty karma]

yeah its shit for this stuff

[wispy nest]

I did the same a month or so ago, I wanted a lion seal emblem. But the ai had other plans :P

[bold cave]

you say your card's a month old? well that's great. if you could use a nice heavy paperweight.

[dense tapir]

really made for 256x256 or, at most, 512x512 beyond it struggles and shows its age.

[smoky oak]

his 1060 is years old lol

[bold cave]

ik lol i'm just quoting weird al

[smoky oak]

Also, I got hired to make my first paid LoRA last night :D

[hearty karma]

i really wish i could just magic up a 3000 series card 🥹

[bold cave]

make it look like a gargoyle and it will look dope

[dense tapir]

Very nice Sytan

[hearty karma]

itd be better than what i currently have

[dense tapir]

grats

[smoky oak]

Thank you, and its coming along quite well

[bold cave]

I have this dude who wants to do the same to me but for a dreambooth but apparently even money isn't enough to motivate me out of my comfort zone

[smoky oak]

first test results had them so happy, they are putting together a list of other comissions for me

at $50 a pop, I am more than happy lol

[hearty karma]

i think im going to try and make a spider-man and venom hybrid with this

hold up

[dense tapir]

I was just saying that ocean scene with fish has potential to be upscaled. I mean the camera and that certain underwater look floored me.

[smoky oak]

oh yeah, I am lucky that the things they want are a niche I fall into

men

[hearty karma]

btw this is a thing that i just got

[smoky oak]

They are one of the only living beings alive that want LoRA's of men, and not big tiddy anime girls

[dense tapir]

Yep, that's a "thing" alright.

[oak osprey]

me, thinking: you know, maybe now that i've figured out the basic quality issues, SD 2.1 isn't so bad
SD 2.1:

[bold cave]

bru

is there a way to block someone's photos from appearing

im tired of seeing cloned monstrosities here

[oak osprey]

i'm tired of your rudeness

[bold cave]

or lets ask the mods to make a #1072236442463518882-aesthetics-only

[oak osprey]

maybe an #elitists-only channel

[hearty karma]

woah....

[bold cave]

black spiderman is already a venom spidey mix

[dense tapir]

Okay, fainting

[hearty karma]

ok not all black spiderman are venom spidey mix

[dense tapir]

I think that is my cue to make the feeder I had been putting off

[wispy ether]

Did anyone here train a model before

A main model

Not lora

[dense tapir]

DB or go home

[wispy nest]

when people fight, cat chef get angry >:I

[smoky oak]

I am in a weird middle ground, cause I am actually working on making a model from LoRA's

[oak osprey]

foundational model is a keyword that might help

[wispy ether]

Does images work if training images greater than 1:1

[dense tapir]

Well, I have to say for the real elites if it isn't a full on Finetune then it isn't a real model. :/

[oak osprey]

https://tripleback.net/public/discord/1683569962.03121953ff3ddaf1072bc0cbabb243fd87fd5db.png

[bold cave]

^

[dense tapir]

360k images FT

[oak osprey]

that mouth looks so dumb

[hearty karma]

eldritch horror spider-man

[bold cave]

training a model from a model is called a metamodel

a classic dreambooth strategy for low-picture subjects

[dense tapir]

I like it and how nightmarish it is.

[oak osprey]

https://tripleback.net/public/discord/1683569966.53666118b03a293865798459923031155f5a2a6.png

TIL I have no idea what bears really look like

[smoky oak]

You can train images at an aspect ratio, but they can add some bas distortions.

I do have a friend that is working on an aspect ratio squisher LoRA tho

So you can squish aspect ratio images down into a model, it can generate squished images at ratios, and then you desquish them

[hearty karma]

im going to try and controlnet it

and get a more humanoid creature

[oak osprey]

i can feel controlnet sweating already

[hearty karma]

i can feel my gpu sweating...

[oak osprey]

i need to rip my garbage tiling code out of me controlnet flow so i can use it again

[hearty karma]

without controlnet im already using 5.8gb vram

[wispy ether]

You cant

[hearty karma]

💀

why

[dense tapir]

I couldn't DB 2.1 as colab didn't give me enough mem for it. I finally managed it with a lot of tweaks but the output was nothing like my work in 1.5

[wispy ether]

5.8 is nothing

I use 7

[hearty karma]

i have a 6gb vram gpu...

[kind quartz]

if you feel realy bad, remember this horse 😦

[wispy ether]

7 of 12

[hearty karma]

wdym 5.8 is nothing

[distant cloak]

she looks so real. its crazy

[hearty karma]

it is for me

[wispy ether]

Digital waifu

[bold cave]

upscale that joint with a realistic upscaler

[wispy ether]

can i use this one for training a freddy model

[bold cave]

it will give her wrinkles and veins and all this realistic shit

if u pass the prompt I can do it, I'm curious

[wispy ether]

What resolution does it have to be tho

[bold cave]

why do u need to train freddy. SD can already get it right

[hearty karma]

someone needs to train bonnie foxy and chica

:)

[wispy ether]

I wanna train a general fnaf model

[bold cave]

oic

[hearty karma]

ok... controlnet got the preprocessing done

[distant cloak]

||(photo realistic, RAW photo of woman), ((Bokeh Photography)), (masterpiece, high resolution, 8k, top quality, best quality, finely detailed)
Negative prompt: (bad_quality:0.8), (worst quality, bad quality, child, cropped, white edges, image borders:1.4), ((monochrome, grayscale)), (badhandv4:1.0), (nsfw, nude, naked, revealing clothes:1.2), (easynegative:0.8),(disformed limbs, broken limbs:1.1), (ugly teeth, broken teeth, imperfect teeth, bad smile, forced smile, dirty teeth), (fake skin, bad skin, broken skin, ugly skin:1.1), (text, logo, watermark:1.2), (ugly eyes, death stare, cross eyed, imperfect eyes, redeyes:1.12), (too many arms, too many legs:1.3), (overexposed, oversaturated)
Steps: 23, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1187429427, Size: 960x540, Model hash: e6415c4892, Model: realisticVisionV20_20, Denoising strength: 0.47, Clip skip: 2, Hires upscale: 2, Hires upscaler: 4x-UltraSharp||

[hearty karma]

now im just hoping i can actually gen the thing i want

[wispy nest]

as long as people remember that before bad ai art, there was bad REAL art ;P

[dense tapir]

Smile, hands down below the camera, down below. Hmmmm

[bold cave]

aight gimme a mo

[hearty karma]

aw man..

[distant cloak]

mo?

[hearty karma]

moment

[distant cloak]

oh lol ok

[bold cave]

did you render natively at this rez? 960x540

@distant cloak

[distant cloak]

960x540 + highres fix

[oak osprey]

BlueWillow just added your upscaler as their "v4 model" @bold cave 😄

[hearty karma]

hamburger cake

[upper inlet]

Soldier

[smoky oak]

yeah, will be fine

[wispy nest]

oh, well, time for le bed! but before that, this was one of the last 2.1 images I made. I think :P

[dense tapir]

Well, gpt just wrote a lora trainer for me. :/

[wispy ether]

Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
RuntimeError: CUDA error: invalid argument
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

[wispy ether]

[dense tapir]

In this example, we first define the LoRA model configuration and tokenizer. We then load a base Stable Diffusion model checkpoint (in this case, distilgpt2), and use it to initialize a LoRA model with the same configuration. We then define a training dataset (in this case, a small set of example texts), and use JAX and Optix to train the LoRA model for 1000 steps. Finally, we save the trained LoRA checkpoint.

Note that this is just a basic example, and you may need to modify this code to suit your specific use case. In particular, you may need to adjust the training data, hyperparameters, and optimizer settings to achieve good results.

Interesting but I don't think that is it

[bold cave]

alright its upscaling and its looking promising

might even have me a new wallpaper here

[dense tapir]

although this looks very familiar

rought but familiar

*rough

[hearty karma]

its going to be a dummy thicc spider-man

[wispy ether]

big ass booty man

[dense tapir]

Does that work with images?
ChatGPT
No, the example code I provided is for training a text-based Stable Diffusion model using LoRA. To train a Stable Diffusion model with LoRA on image data, you would need to modify the code to work with image data and a suitable Stable Diffusion model for image generation.

One option for image-based Stable Diffusion models is the "BigGAN" model, which is a large generative model for high-resolution image synthesis. You could use the "BigGAN" model as a base checkpoint, and apply LoRA fine-tuning to reduce the size of the model.

Here is an example of how to fine-tune a Stable Diffusion model based on BigGAN using LoRA in Python using the "Stable Baselines 3" library:

In this example, we first define the LoRA model configuration and load the base Stable Diffusion model checkpoint (in this case, the "BigGAN" model). We then define a training function that trains the LoRA model using the "Stable Baselines 3" reinforcement learning library, and saves the trained LoRA checkpoint. We use the PPO agent to train the LoRA model for 10000 time steps.

Once the LoRA model is trained, we can use it to generate images by calling the sample method with a number of samples and the number of timesteps to generate each sample.

I'll be damned. lol

[kind quartz]

there is phind.com, not sure but it is free should be chatgpt 1.4 and no need registration. Suggesting enable expert mode.

[wispy ether]

vram 4.7

[oak osprey]

interesting result from SD2.1

[distant cloak]

nice use of my prompt :)

[dense tapir]

Thankee. 🙂

[wispy ether]

i have 0 use of fp16

i get 7it/s at fp32 and 7it/s at fp16

crazy

[dense tapir]

Your prompt came out nice

[distant cloak]

thanks

[distant cloak]

ive been so focused on unrealistic photos. but then recently I started getting into photorealistic ones

[hearty karma]

whats an.. open clip?

[wispy ether]

feddy

[hearty karma]

ew :)

[wispy ether]

ikr

i cant train a model

[hearty karma]

its got the colour for the main body down

[wispy ether]

its cursed

[hearty karma]

and its got the hat down

but everything else is horrible

[bold cave]

who tf even uses openclip

control net has always been the master race tool

they should just call it magicnet at this point cuz only god knows how it works

[distant cloak]

@bold cave is it still upscaling?

[bold cave]

good for them! although I wouldnt call it my upscaler

[bold cave]

I used ddim and it turned out a little overburned

Im doing karras to see if its better

[distant cloak]

ah okok. what gpu do u have?

[bold cave]

3090

[distant cloak]

geeezzzzz thats long for a 3090

ig u r upscaling to 4k or smth?

[bold cave]

yh this upscaler for some reason is slow as balls

yh I think its 4k

the girl will look different tho cuz im not using RV2.0

[distant cloak]

ah ok. @ me when its done pls

[dense tapir]

DDIM is faster, for me, than euler_a

[bold cave]

impossibru

[dense tapir]

well, what step count? A=A?

[bold cave]

Aitrepreneur might go on my shitlist soon cuz all he posts about now is GPT punk ass

[bold cave]

I used to revere this dude now Im annoyed by his feed

[hearty karma]

why?

[dense tapir]

Snoop

[hearty karma]

yes?

[distant cloak]

ya hes getting really annoying with all the nsfw chat vids

[hearty karma]

woah...

[dense tapir]

I am about to unsub him tbh

[hearty karma]

looks like venom attached to a robot

[dense tapir]

Yeah

I was not expecting that

[hearty karma]

btw heres a cooked mustang delorean mix

[wispy ether]

4k is 3840x2160 or 4096x4096

[bold cave]

daaaaaaaaaaaamn

thoughts

@distant cloak

[distant cloak]

is the one on the left the original?

[bold cave]

yeh lool

[crimson dawn]

i heard the term "MPCS" for the coming age of AI characters that can interact with you in a consistent natural way. "Machine Player Characters" instead of "non"

i dig it

[distant cloak]

the skin on the right looks more realistic and the eyes look better

[dense tapir]

I like that

[distant cloak]

imo

[dense tapir]

eventually we will all live in a pod never interacting with any other real person

[bold cave]

yeh unfortunately she has makeup, if u open up the img u can see that classic crackly makeup texture

[distant cloak]

ah

[bold cave]

I always use (heavy makeup) in my neggies for this reason

[crimson dawn]

the movie surrogates basically xD

[dense tapir]

Snow on roof top is sus

Yes

[distant cloak]

ya that makes sense. so how did u do the upscale? did u use an extension?

[hearty karma]

i just realized why my results look so baked

i keep bloody forgetting to set the resolution to 768x768

[bold cave]

yeh use the cnet tile upscaler. keep in mind I only used 0.3 denoise. With this tool u can do 0.75 and still get a coherent image

With a crapton of detail ofc

I would do a reupscaler with .75 but unfortunately I got some models I need to catch up to

[crimson dawn]

ui-config.json edits can fix that one. hits me all the time on fresh installs too

[dense tapir]

What I did a long time ago.

[bold cave]

wait what

where do u get that file @crimson dawn

[crimson dawn]

root folder of webui

[hearty karma]

in the sd web ui folder

[distant cloak]

ok thanks for doing that. the upscale turned out really nice. i didnt even know it was possible to get it that realistic looking lol

[crimson dawn]

all the upper limits and defaults i comb through and change whenever i have to. well like, i know what i like now so if i fresh install i just go through to what i want.

[hearty karma]

the doors are huge for this thing

[bold cave]

oh shit I totally missed that file lmfao

Here I was editing the python code file loooooooool

[distant cloak]

@bold cave quick question. what sd extension did u make? its cut off in your bio

[distant cloak]

oh i clicked and now i see

[bold cave]

It wasnt my extension, its ultimate SD upscale

With the 4x NKMD siax 200k upscaler

I would redo that prompt with (heavy makeup) on the neggies

for maximum realism effect

[hearty karma]

oh btw sen i found that nkmd siax upscaler

[distant cloak]

oh neat neat

[crimson dawn]

haha. i think the git pull is even configured to ignore it. i back mine up too but like, i'm never sure if the settings names have changed or not with this stuff, so i just pull fresh ones usually and edit them

[bold cave]

also use that tile resample tool

[dense tapir]

I wonder how large is the lexica prompts if scraped?

[vagrant zinc]

where is help # ? i got some problems to make it work

[dense tapir]

A lot of trash in there

[bold cave]

#🤝｜tech-support

[vagrant zinc]

thanks

[dense tapir]

Oh

[crimson dawn]

i found some "prompt helper" characters for oobabooga llm's. i might try one out tonight and get a collection of prompts to push through SD

[dense tapir]

Look at all that glorious SSS

[bold cave]

yo if u got them share them

Im looking for new prompts to spice up my gens

[dense tapir]

oogabooga does prompts?

[bold cave]

its oobabooga

[dense tapir]

nah, to me it is oogabooga

[bold cave]

but its a r3tarded name so I dont blame u

[dense tapir]

yes

I wonder what the name really means?

[bold cave]

idk. it pisses me off when devs give their projects idiotic names

like android naming their versions after candy

[dense tapir]

Yeah, oogabooga like some damn caveman

[bold cave]

AnDroId LolliPoP

[crimson dawn]

https://civitai.com/models/33910/oobabooga-text-chat-prompt-helper-me this thing

[dense tapir]

KitKat 4.2 I still say was the last one worth a damn for rooting and shitr

[crimson dawn]

i got vicuna+gpt4 loaded into oobabooga and want to see what i can do with it

[dense tapir]

Always ticks me off I buy a phone but it isn't really mine to do with as I wish

[crimson dawn]

its oobabooga lol

[dense tapir]

After Marshmallow it got so locked down

[oak osprey]

maaan some youtube'rs have really revealed a weird side with the AI stuff. like they're totally different people now. or just feel like it's cool to be disgusting, idk

[bold cave]

its bc it gets views

[dense tapir]

click bait

[crimson dawn]

i love watching all these clickbait youtuber producers stumble over all these project names . LOVE IT

[oak osprey]

yeah that's the cool part

[bold cave]

for some reason SD videos dont get as much views anymore ig?

[dense tapir]

tiktok like behaviour

[bold cave]

even though SD technology is still alive and evolving

[oak osprey]

Sabine Hoffsteader or whatever her name is, started railing about controversial subjects incl gender and AI lately and she's a gd physics professor

[dense tapir]

no longer new its 5 seconds of newness wore off

[bold cave]

she did?

ive seen her videos and theyre all non-clickbait af

[oak osprey]

yeah man a week ago she had a pretty garbage take

[bold cave]

damn

[wispy ether]

is fp8 a good choice?

[crimson dawn]

watching aitrenpreneur for how to set this stuff up, and he demo'd it with the most toxic bro character that he created. like, it wasn't the ai doing it. He crafted the toxicity himself lol. It's weird that he goes that way now.

[dense tapir]

smoke fp8 if you have it

[distant cloak]

ya instead of focusing on some really cool groundbreaking opensource AI stuff, they target the sellout clickbait stuff 🙄👎

[oak osprey]

it's been getting there for a month. she is testing the waters. she got 3.8M views on controversy and since then, it has become a part of her toolchest

[wispy ether]

[bold cave]

wait is fp8 better than bf16

[oak osprey]

is that accelerate config?

[wispy ether]

wtf is bf16

[bold cave]

bc im pretty sure bf16 is the best possible

[dense tapir]

Yes, it makes me sad and I unsubbed and ones I didn't I use my blocktube.

[oak osprey]

bfloat16

[bold cave]

a superior version of fp16

[oak osprey]

mixed-precision

[distant cloak]

blocktube???? I need that rn

[bold cave]

ive never heard of fp8 tho

[oak osprey]

it's very odd and might be considered garbage if you care about determinism (bfloat16 that is)

[bold cave]

no, bf16 > fp16

but it uses more vram

[crimson dawn]

youtube is probably adjusting the algorithm to drive more clickbait nonsense in the space. Alphabet are in a gigantic panic about losin control to the grass roots open source efforts

[dense tapir]

Oh, lord yes. I block foreign language vids all the time as I can't understand them so search and they appear does me no good.

[oak osprey]

i run a worker network and i care about determinism so if something isn't working across diffusers versions / python versions / hardware, i cannot really use it reliably

[crimson dawn]

paying youtubers with stupid amounts of adsense dollars probably too. so they're encouraged to lean into the nonsense content

[wispy ether]

sooooooo

which one

[crimson dawn]

Alphabet are in a huge strategic war right now in the AI space, and i have no doubt the new CEO of youtube is making moves under direction from the top

[dense tapir]

I am waiting for "let's build a bio weapon, but don't do this at home" shortly.

[crimson dawn]

that guy has been entirely silent since he began and that's not good for community

[dense tapir]

What guy?

[hearty karma]

how do you change the gp16 thing

[crimson dawn]

the new Youtube CEO

[oak osprey]

i had to eat some indeterminism to include xformers already but it has a sliding scale of determinism with for example, the 4090 24G and A6000 48G and A100 80G all having very similar output. the A100 and A6000 are identical. the 4090's fourth picture in a batch will be different, the first three will be the same. but a 3070 8GB has a different photo from the 2nd. all four have identical output for the first image in a batch.

[bold cave]

if u got good vram go bf16

[dense tapir]

Oh, yes. Silent is deadly

[crimson dawn]

Neal Mohan

[bold cave]

if u got puny vram fp16

[dense tapir]

I despised Susan, but this new one where is he or is he really AI?

[hearty karma]

this new one is likely just as bad

possibly worse

[dense tapir]

I expect worse

[hearty karma]

i mean he signed for the removal of dislikes

[oak osprey]

iirc, if i use bfloat16 and long prompt weighting, i can't even get the same image twice from the same seed.
also, on some hardware, bfloat16 is an incomplete implementation. it can crash, or produce odd outputs.

[crimson dawn]

lol google woudln't let their AI models run the company. god damn, that'd be bad

[dense tapir]

For a lot of content creators, and general users YT began to suck in 2013 then 2014 had gamer gate and it is just suck city since.

[hearty karma]

what is gamer gate???

[dense tapir]

be glad you don't know

[hearty karma]

ive heard about it but ive not heard about it

[crimson dawn]

i think susan was told to retire so that strategicians could enact plans. Google is losing the AI arms race right now and they're coming into a huge pivot soon. They've got a few products keeping them relevant and those are suddenly expiring soon giving the new advancements.

[oak osprey]

gamergate was about the time as elsagate

[hearty karma]

elsagate

isnt that those crappy kids yt channels that are not kid friendly

but pretend to be

[oak osprey]

🎯

[dense tapir]

The stupid beoytch who caused all the ads to be pulled because we were toxic males in the gamer realm is now gone into obscurity.

[dense tapir]

Yep

[oak osprey]

geez mate you don't have to violate the server rules to get your point across

[dense tapir]

What?

[crimson dawn]

gamergate started because a trans game designer was dating a writer at giant bomb or some clickbait gamer rag place. That writer wrote a favorable review about a walking simulator game that aims to do nothing but tell a story in an interactive way. Some gamers got mad about that and started acting like asses. THEN though, THEN 100 game journalism sites all published "Gamers are dead" articles and that pissed everybody off

[oak osprey]

why are you going to rehash all of that right now, right here?

[hearty karma]

dont blame him im the one who asked

[oak osprey]

uhmmm it's weird but you can use a custom pipeline within the Karras diffusers pipeline. but i'm not clear on whether this breaks all the cool Karras stuff.

[kind quartz]

nice b&w

[hearty karma]

im ngl

on certain steps the delorean and mustang mix looks a lot like the blista compact from gta

[oak osprey]

i love that car. apparently it's modelled after a Daihatsu JDM

[hearty karma]

really?

[oak osprey]

yup

[hearty karma]

its modelled after this?

[oak osprey]

one of the other body styles but yes

wonder which year that one is

[hearty karma]

oh

[wispy ether]

why

is it faster?

than fp16

[oak osprey]

it is slower and uses more vram iirc

[hearty karma]

the mustang deloeran mix looks better with high res fix

[oak osprey]

uses more vram because it's storing the weights in the GPU as fp32 but it inferences them using fp16, while upcasting for the parts that work better with fp32

at least, that's my limited probably-incorrect understanding

[crimson dawn]

Fantastic, 8K, Illustration, low angle, Prompt for a DMC Delorean Dragster, detailed, powerful, gleaming, racing, futuristic city, shimmering lights, dynamic lighting, studio light, alternate reality, ink, markers, Hot Wheels, Kustom Kulture, Retrowave, Cool Colors, Chromatic Aberration, Jean-Michel Basquiat, Ed Roth, Darick Maas.

first go

[oak osprey]

is this because i changed the VAE on 2.1?

[crimson dawn]

i dont' have sd loaded though so if someone else could try? xD

[hearty karma]

what is this?

[bold cave]

more accurate

in other words, better trained model

[crimson dawn]

a character for making prompts for SD, loaded into oobabooga with the gpt4-alpaca model that people are saying is really good

now there's a new model that some researchers have put out open source and on an entirely unique dataset. whole new arch that isn't tied to research licences. it has soemthing like 70,000 token memory to it which is stupid huge. gpt4 has like 8000 or something much much much much lower

stable lm is cool but it landed softly. just touching base to say they were here really. i'd say it's marginally better than gpt3 in my short experience with it

[oak osprey]

https://tripleback.net/public/discord/1683574143.5965683a231b8be8d17542dbfa00037c9055ae0.png

oh hot damn we're getting places now

[oak osprey]

tile upscaled

i see almost no reason not* to put controlnet tile on every single image output lol

here's the original. you can see it really cleaned it up

[bold cave]

exactly h

G*

Ive said it months ago

there is no reason NOT to use cnet on every img2img these days

[graceful island]

Any general guidelines for keeping images coherent when increasing resolution? (first pic is beautiful hi rez disasterpeice) second is what it should usually look like (low rez)

[hearty karma]

dont generate it at high res

[graceful island]

So I take it everyone just upscales?

[bold cave]

yes

[hearty karma]

yep

[graceful island]

Got it!

[hearty karma]

because thats the best choice

[bold cave]

upscale is the incorrect term tho

resampling is more accurate

[hearty karma]

what is the best 'resampler'

[bold cave]

CTU

[hearty karma]

?

[oak osprey]

when you use CTU are you putting out >1 img per prompt?

[bold cave]

nope just one. no one has been able to fix it to get more per session

If I had the source code I would fix it

but who knows where to find it

bruh. I want u guys to see this

I was able to restore a lost picture to insane detail

thats the power of controlnet

[oak osprey]

i have it working with >1

[bold cave]

with UU?

UU can only do 1 at a time

[oak osprey]

i don't have a way to get UU working yet

[bold cave]

hmm

[oak osprey]

but the original tiler i can

[hearty karma]

what is ctu?

[oak osprey]

controlnet tile upscaling

[hearty karma]

ah

[oak osprey]

original, 0.3, 1.0 strengths

that's a national geographic photo

i haven't really been changing my tiler prompts but i imagine i could improve it substantially if i do

@bold cave have you had any luck feeding an image back into CTU? it seems to poop all over it

[bold cave]

wym

wtf is that atrocity

ur supposed to use CTU with UU otherwise it doesnae work

[oak osprey]

it's three times fed back through CTU with strength at 1.0

oh it definitely works but you just can't feed it back to itself without changing the strength or using minimal numbers lmao

[bold cave]

image restoration done like a pro

and I'm thinking of even more accurate ways to restore an image

perhaps I could start charging for it o:

[hearty karma]

should i use seam fix?

and if yes which one?

[oak osprey]

cant wait til i can blame AI for the bottle of grey goose vodka in the shot behind me on my work meeting videos

snoop, just check the seam fix box

[bold cave]

it will take longer if u use seam fix

but it can do some good stuff every now and then

I do 50-50

[oak osprey]

is that for latent upscaler only? or CTU as well?

[bold cave]

I dont use latent upscaler

[hearty karma]

maybe try 'slightly tan complexion' idk

¯_(ツ)_/¯

[oak osprey]

CTU really enhanced the tile artifacts from the VAE -_-

[bold cave]

ur doing it wrong m8

[oak osprey]

how so

[bold cave]

watch

no artifacts no seams, just upscaled goodness

If the senman can do it, so can you I'm the scatman

papapapapeeebababababop

[crimson dawn]

lol this is like goldilocks the bear

[oak osprey]

is MSE bad?

[oak osprey]

Bearpunzel

https://huggingface.co/stabilityai/sd-vae-ft-mse

[bold cave]

I use the vae that came with the Ultra model

perhaps changing to that one may help

[oak osprey]

which is ?

[hearty karma]

theres a vae that comes with it??

[bold cave]

vae-ft-mse-840000-ema-pruned

idk if its the same one

[oak osprey]

puddin' pops. i don't know where that is

problem is i don't really have a good way of pulling that VAE down

oh wait, that's a part of the one i linked to

[bold cave]

yh jus download the 84000 one and pop that mf up in the same folder as ur other models

[oak osprey]

my stuff all pulls using the huggingface_hub module

[hearty karma]

wait.. how are you going to use the sd ultimate upscale?

your on a discord bot arent you?

[bold cave]

?

u asking shrek or me

[oak osprey]

yes it hosts its own images

[hearty karma]

shrek

[oak osprey]

so the mse link you use is the checkpoint file and i'm using the safetensors version via diffusers

so i guess that's not "the" difference

[bold cave]

true

[oak osprey]

wonder if you have VAE tiling disabled? do you know?

takes a lot more vmem that way

[bold cave]

Im not sure

Here she is with her eyes fixed. God this is probably the best job I have done in terms of skin detail

It's going to the hall of fame for sure. I love that fingerprint pattern

[oak osprey]

jeez the light outline on that hat seems intense

[bold cave]

it's an image restoration

[hearty karma]

i am so damn confused on how you are getting good upscales

mine looks like patchwork

[bold cave]

Im uploading a video TuT

But I will need time bc I need to rebuild all my dreambooth models

[oak osprey]

its the ultimate upscaler. are you using that script Snoop?

[bold cave]

I have this dude waiting on a TuT from scratch for dreamboothing

[hearty karma]

yes

i know its the ultimate upscaler extension

i just dont know how to properly use the damn thing

[bold cave]

its finicky

Also some tips

well its hard to explain via text

I used a photoshop technique

to "prime" my image

then used 5 on cfg which is low for this model and 0.25 denoise

used the special NKMD upscale model and kaboom. realistic skin

[oak osprey]

https://giphy.com/gifs/filmeditor--the-hangover-movie-3owzW5c1tPq63MPmWk

[bold cave]

this is the original picture I restored

I essentially used a metamodel, a process I will explain in my TuT

[hearty karma]

well i cant wait for that tutorial :-)

[bold cave]

Sub to my channel

let me fish it out

https://www.youtube.com/@AI-wf9he

Im not a youtuber, I will just use that channel to paste video tuts cuz theres high demand for them

and also cuz this dude wants to pay me to do a db tut

[oak osprey]

man i think i fixed my image gen quality?

[hearty karma]

those bears got some ugly outlines

[oak osprey]

😦

[hearty karma]

on the lighting i mean

heavy rim

[oak osprey]

well this is pre-CTU

so for a base default output i mean, it's pretty good, not even cherry picked

the VAE tiling seems to really really screw with the output

[bold cave]

another reason why more vram is always a good investment. worth its weight in gold

stuff that helps reduce vram usage can backfire

[hearty karma]

whats a good 2000 series gpu?

[oak osprey]

i mean i have plenty i just also couldn't get xformers working across the board before when i was at first messing with the tiling option

[bold cave]

whichever has the biggest vram

[oak osprey]

^

and then get 4 of them

[bold cave]

r0fl

and pool them?

[oak osprey]

hell yeah borther

[bold cave]

thats not viable unless u have a server

[oak osprey]

we have the technology

[bold cave]

2 cant fit on a standard pc case

yeh but who wants this ugly contraption in their living room

jus get a 3090 trooolol

[hearty karma]

yo wait.. 2060 has 12gb

[hearty karma]

or well some do

[oak osprey]

[bold cave]

SO u just have that right in the open

for all the dust and debris to settle on it

[hearty karma]

thats just horrifying

[bold cave]

https://tenor.com/view/side-eye-dog-suspicious-look-suspicious-doubt-dog-doubt-gif-23680990

[oak osprey]

so i can blow my cigarette smoke directly onto it

[hearty karma]

im not even sure if anywhere sells 2060s anymore...

[bold cave]

those are mining rigs

theyre not even meant for 2 gpus, theyre meant for series of them

[oak osprey]

you won't get the speed, but you'll get all the memmies

[bold cave]

can you though? because vram works in segments, and if it needs a big ol' segment there may be issues with having to split it across GPUs

thats just a guess though. I just dont know anyone who has it that way

[hearty karma]

319 for a 2060 🤔

is this a worthy investment?

[bold cave]

is it new or used

either way its a good buy

at least in the meantime while u save up for a 3090

[hearty karma]

huh surprisingly new

yeah im definitely not ever going to get a 3090

those shits are over 2000 pounds

[bold cave]

no theyre not lol

I got mine for 900usd

[bold cave]

or better yet like 150usd cuz I sold my 3080ti for it

lmfaooo this dude buys gpus from amazon

rookie stuff.

[hearty karma]

hey i dont know where to bloody buy gpus

[bold cave]

offerup, fb marketplace, or as a last resort newegg

[hearty karma]

fb marketplace???

[bold cave]

yh. u wanna buy used m8

[hearty karma]

ah

[bold cave]

this puppy was purchased used

I didnt even have to see it running in the dude's computer. I just checked out all the connectors, inspected for dust, and it looked good

[oak osprey]

yes absolutely you can put one layer of the model on each GPU

[bold cave]

plugged it in and has been working flawlessly. no need to bend over for Amazon's ridiculous prices for the same item

[oak osprey]

you can pin each part to a single gpu so that the work stays in one spot. i did this to try out LLaMA 65B on an 8x A100 system. what a waste of money that was

[bold cave]

bruh I want an a100 so bad

I will paint it gold for maximum flex

[crimson dawn]

if you're buying it in person, its a lot more trustworthy

[bold cave]

the dude that I bought it from said he bought it back in 2020 for like $2500 lmaoo

[crimson dawn]

thats how i got my vive headset a few years ago. went to his place and he did have it all set up to demo beat ssaber but i didn't really care. i coudl tell it was good shape

[bold cave]

thats literally what I'm selling my car for

thats a wise buy

[crimson dawn]

cars have shit resale because thers tons of shit cars out there

[bold cave]

yeh. everyone super sus even though it runs perfectly fine

I might have to drop it down to $2000 yet its a good vehicle

[crimson dawn]

i've owned my used crv for a year and i'm still expecting somethin to go

[bold cave]

you cant get cheap cars from dealers anymore tho

[crimson dawn]

it'll make a noise an i'm thinking "oh shit herrre we go I KNEW IT"

[bold cave]

its all overpriced bs

[oak osprey]

man i love CTU. do you see any tiles in the output now?

[bold cave]

loool mine makes noises but has withstood the test of time

[bold cave]

niiiice

that's actually wallpaper worthy

[hearty karma]

for some reason i cant use marketplace

[oak osprey]

big praise from you, thanks Sen

[crimson dawn]

yeah it's doing pretty good lol

[bold cave]

it doesnt look like those plastic clone pics u usually post

that one looks legit good

[hearty karma]

how... how u do dis

[oak osprey]

i haven't tested whether my prompt embeddings still do anything or not

[hearty karma]

https://tenor.com/view/math-calculate-confusing-figure-out-gif-6237717

[crimson dawn]

walker: texas bear

[bold cave]

they will

[oak osprey]

@hearty karma i can invite you to the server and you can use it there

[bold cave]

even using prompt editing works on upscalers

its just magic

[hearty karma]

oh right i forgor your using bot

🤦‍♂️

i keep forgetting that

nah ill just wait for sens tut 😄

[oak osprey]

yeah my lessons don't directly translate other than use karras 2M sampler and controlnet tile

[bold cave]

its gonna take a while m8. I need to do db tut first

[hearty karma]

yeah thats fine

[bold cave]

and its a looooooooong tut bc I have to do it from scratch

[hearty karma]

ill wait :)

[boreal falcon]

@bold caveyou do train large scale models and do in general AI things on a larger scale?

or why do you want to buy A100 xD

dont just tell me its only for Stable Diffusion

[bold cave]

bc it's super fast, fancy, has unlimited vram

hey I said I wanted one, not that I actually intend to buy one lolololol

[boreal falcon]

good xD

also RTX 6000 is coming out

[bold cave]

also, I can dedicate my 3090 for gaming while my A100 does training in the background

literal paradise

[hearty karma]

i have to find the gpu i want on newegg :-)

[oak osprey]

things i did to fix my output today: 1. switch to karras pipeline with karras sigmas, 2. disable VAE tiling, 3. pick the images i want to keep/share and put them through CTU first

even SD2.1 produces bomb output with this

[hearty karma]

oh holy crap

[bold cave]

what denoise do u use for ctu

[hearty karma]

the prices on newegg are great

[oak osprey]

the downside is, my 8gb laptop can no longer play ball

[bold cave]

I find better results on low denoise

[bold cave]

told u

[boreal falcon]

i guess i will buy the RTX 4090, 128 GB RAM and i9 13900K 24-core processor for my next or 2 years build of a PC

[bold cave]

and newegg sells new

[oak osprey]

what you mean. the number?

[bold cave]

yeh

[oak osprey]

1.0

[bold cave]

hmm

[crimson dawn]

https://www.techpowerup.com/gpu-specs/rtx-titan-ada.c3985 if i ever see one of these for sale, i am saving money now so that i'm ready to pounce

[oak osprey]

😄

[boreal falcon]

RTX 6000 would be a bit too much maybe but im not sure

[oak osprey]

it made the bear into the blonde long curly hair

[hearty karma]

damn 3060 for like 350

[oak osprey]

what CTU prompts you using, sen?

[bold cave]

ur not supposed to change the prompt

keep it the same as the original gen's

[oak osprey]

            "tile_negative": "blur, lowres, bad anatomy, bad hands, cropped, worst quality",
            "tile_positive": "beautiful 8bit DOS video game, willy beamish",

well this is the problem for me then maaaaybe ahahahaha

[crimson dawn]

careful with going for 4x32gb kits to get there. a lot of them have shit timings

[oak osprey]

i forgot i had that in my tile prompts

btw Sen the example for CTU has "best quality" as the prompt for a pic of a puppy

[bold cave]

yeh thats a good one. ((ultra detailed)) for me works fine

depends on what model ur using too

[oak osprey]

hmmm i don't even think i can use () weightings within controlnet

[bold cave]

u can

it just takes ur original prompt

[oak osprey]

that stuff is a huge pain to get working

like, under the hood

[bold cave]

no its not lamo

[crimson dawn]

depending on the vibe of the image you want too. ultra seems to bring out future qualities. bit of scifi influence maybe.

[bold cave]

it used to be worse when we used barebones UU

[oak osprey]

im talking about from a python perspective

[bold cave]

run it locally g

[crimson dawn]

intricate detail is like, royal classical stuff

[hearty karma]

i think my best option is a 3060

[crimson dawn]

extreme detail always makes people look lik ethey're sweating in my experience

[hearty karma]

wait why does a ti 3060 have less vram than a non ti?

[bold cave]

bc invidia is stoopid

nvidia*

their models make no sense a lot of time

just go for the best vram for the price

[hearty karma]

12gb on the non ti and 8 on the ti 💀

[crimson dawn]

they know their audience, gamers 😉

[boreal falcon]

really? O_O

[bold cave]

gamers need vram too

resident evil 2 remake barely ran at full settings on 12gb vram

[boreal falcon]

i dont build my new PC just for gaming btw

[oak osprey]

oh my god man, the quality 😭

haven't even upscaled that

[hearty karma]

wait i just realized i dont think id be able to use a 3060

[bold cave]

Ive been jonesing to play microsoft flight simulator after getting my 3090

[hearty karma]

i have a b450 mb 🤔

[crimson dawn]

yeah like, i got a build with the 12700 i7 right now, but in a ddr4 lga1700 bored. just cause fast timings are way cheaper on the ddr4 side. slower mhz but faster timeings == better latcency in this case. i'll swap up to ddr5 probably next socket gen

[bold cave]

boy im getting flashbacks from my last pc build

i broke the motherboard lmfao

from using the wrong screw. thank god newegg replaced it

[oak osprey]

i got ddr4 128gb 3600MHz CL16 for $600 or so last month

[crimson dawn]

yeah they're great prices right now on the ddr4 side

[oak osprey]

i am sure CL16 is like old geezer timings to someone else but to me that's magical, haha. i am used to ECC memory 2666MHz CL19

[bold cave]

bruuu why u need so much ram wtf

its a total waste

[oak osprey]

gave no shits for ECC this time since it's purely a ML rig.

[bold cave]

just get ddr5 32gb instead

[oak osprey]

uhm because i load LLaMA on there 😄

[bold cave]

ram or vram

[oak osprey]

ram

[bold cave]

I had no idea that shit ran on ram

[oak osprey]

Llama.cpp runs it in CPU, hence the 5800X3D

[bold cave]

oh

[crimson dawn]

cl16 at 3600mhz is nothing to scoff at. that's better than probably 90% of gamer builds at this point

[hearty karma]

ok im lucky

i should be able to get a 3060 onto my board 😅

[oak osprey]

i can hand you a hacksaw and then you can fit whatever you want on there once the back side of that pesky PCIe connector is gone

[hearty karma]

[crimson dawn]

i got 3600mhz at cl18 right now. only 64gb and i'll probably buy a 128gb kit that you were talking about if they still look nice. i want a new fast m2 nvme first though. my 1tb is getting tight

[oak osprey]

used a dremel to get rid of that artificial limit on a Xeon E5500 system

just notch the plastic where the card can't sit down and you can put a 16 lane card on a 4 lane port

it's going to melt stuff if you draw full power but you can do it

[bold cave]

or...just but a new gen motherboard

[oak osprey]

that costs money, son

[bold cave]

and a decent case which isnt even that expensive

yeh but it has more benefits than just space

[oak osprey]

bout time i learn you the value of a dollar 😄

[hearty karma]

is the b450m considered a 'new gen' board?

[crimson dawn]

i used the quad channel memory on threadrippers last build i had, and it actually sucked for day to day performance. only benefited me in a few circumstances so i just kept it on dual channel mode usually

[bold cave]

i would advise against it

I think I saw a speed comparison and the b series is way behind the z series

[hearty karma]

huh?

[bold cave]

a cheap Z series board isnt that expensive

[hearty karma]

yeah im not replacing the board that i only just got

[crimson dawn]

i always buy z because 1) B-team doesn't mean best team and 2) you can say Z like "Dragonball, Z!!!!"

[bold cave]

also because B only has 1 m.2 slot

tf am I supposed to do with 1 slot, and also no plate cover, so the ugly ass stick is just there for everyone to see

I like my pc to look fancy too

[hearty karma]

uhh as i said

im not replacing the board i only just got 😄

[oak osprey]

put a pcie riser on the slot and an expansion board and you can have a frankenrig that has 4 M.2 slots off of one

[bold cave]

but its gonna look ugly as balls loool

[oak osprey]

combine it with my mining card riser idea and wow man we're cooking with gas

do any of your computers have a CD-ROM drive?

[bold cave]

nope

[crimson dawn]

risers are official now. i think going forward we'll see a lot of vertical mounting

[hearty karma]

i have one

[bold cave]

just use an external one

[hearty karma]

not in my computer

but i have one

[oak osprey]

then my computer is better than all of yours i have a Zip drive, too

[bold cave]

but does it have a floppy drive m8

[crimson dawn]

"commmme on up for the rising" - bruce springstein. He called it in like 2009. he's the boss

[oak osprey]

i had to remove that for better cooling

[hearty karma]

ok what is a good card for a b450 then cuz idk

[crimson dawn]

3090ti

[bold cave]

all nvidia cards will fit old boards

pcie is backwards compatible

as long as it physically fits ur gucci

[oak osprey]

[hearty karma]

oh so i can run a 3060 on the board then

sweet

[bold cave]

yup. as long as it fits

[hearty karma]

guess i will get one of them in the future

[oak osprey]

you should get a W7900, Snoop

[crimson dawn]

physically fitting and getting proper airflow is why i think we'll see a lot of premium cases supporting the vertical gpus now

[hearty karma]

yeah im not gonna get a huge brick sized one

[oak osprey]

its AMD

64G VRAM

[bold cave]

price?

[hearty karma]

yeah but its got less power tho hasnt it...

[oak osprey]

$4k USD only. pretty good deal

[bold cave]

nty

[oak osprey]

well i paid $2800 for my 4090. i'd have bought the W7900 instead if it were out

[bold cave]

nvidia has all the good drivers

[crimson dawn]

amd on linux is a huge win

[oak osprey]

naw i have ROCm working fine on my rig

[hearty karma]

nvidia cards are more powerful compared to the amd cards its just that nvidia FUCKING SUCKS at giving people vram

[bold cave]

also g-sync compatibility with smart TV's

not to mention no compat issues with SD

the list goes on for team green

[oak osprey]

there's no compat issues for AMD. it was harder to get NVIDIA running reliably with the correct CUDA versions - especially on Ubuntu 23.04

i hate nvidia

[crimson dawn]

freesync works on anything that supports the standard. tons of tv's do

it's the standard

[bold cave]

Im sure the LG CX series has no freesync, only G-sync

that tv is next on my wishlist

[oak osprey]

no need to fear AMD, guys lmao. competition is good

[bold cave]

also nvidia dropped a dope auto-upscaler for video

which works pretty damn well

[hearty karma]

most gpus are all like the same width wise right?

[oak osprey]

is it using Diffusers library?

LOL SNOOP

[bold cave]

idk, it's built into the driver

[hearty karma]

i dont mean length

[oak osprey]

my 4090 takes up 3 slots

[crimson dawn]

no. many take up 2 bays. some future ones may take 4!!

[bold cave]

the answer is no to all dimensions

[oak osprey]

girth

[hearty karma]

i mean width going towards the side of the case

[oak osprey]

the 4090 is a girthy boy

[hearty karma]

well 4090 is just ugly

[bold cave]

and a power hog

[oak osprey]

[bold cave]

holy shit

is that a 4090

[oak osprey]

4090 has a higher range for power limiting options, homie

[crimson dawn]

leaked titan ada

[hearty karma]

isnt 4090 the gpu that has been like blowing up on its users or something?

[bold cave]

yup melting the cables

[hearty karma]

💀

[bold cave]

im good with my 3090 until 50 series comes out

[hearty karma]

you gotta make your computer a refrigerator

[crimson dawn]

no the blowups are the x3d cpus from amd. 4090s are melty boys

[oak osprey]

i'm using the new cable type on my 4090

not the weird one they filmed melting

[crimson dawn]

as always though, these situations are pretty edge cases

[hearty karma]

ok the 3060 i found will be able to fit in my case

length wise

[oak osprey]

yeah it took a concerted effort from GamersNexus to melt it

it almost reminded me of mythbusters the way they went out of their way to cause it 😄

"well, we're all here to see something catch on fire or explode."

[crimson dawn]

electricity and energy are dangerous things. look up lithium battery failures. spicy pillows aren't needed for one to suddenly erupt

[oak osprey]

don't say that, i live off-grid with 10kWh in a small space

[oak osprey]

looks around nervously

[hearty karma]

im good

i found one that should fit perfectly :D

[oak osprey]

@hearty karma did you buy it yet? did you buy it yet? did you buy it yet? did you buy it yet?

[hearty karma]

no

[oak osprey]

[hearty karma]

i shall bookmark that page for the future

[oak osprey]

omg

[dense tapir]

Snoop, you are buying an A100?

[hearty karma]

no...

[oak osprey]

this has been a major disappointment

[hearty karma]

a 3060

why would i buy an a100

[crimson dawn]

it "should" be safe

[hearty karma]

and im not kidding i had to take the side panel off to measure whether it would fit width wise :-)

[dense tapir]

My issue isn't width it is length

[hearty karma]

?

[dense tapir]

12in and if I do a trick MAX 304mm

[hearty karma]

the thing i found is 235mm

[dense tapir]

damn cards are now too long for my massive case. 😦 No one figured cards would be so long

[hearty karma]

so ill be fine

[dense tapir]

yeah, I am getting a 4090 so I have to buy a new case.

[boreal falcon]

will have to look that up but it depends i guess?

[dense tapir]

The actual cards are still small it is the heatsink that is so long

[boreal falcon]

until next year or in 2 years they will probably bring the next generation already lol

[dense tapir]

Good luck affording next gen

[boreal falcon]

depends on how much

dont want to spend more than 7.000€ probably atm

[dense tapir]

a 5090 about 2500 to 3k

USD

[crimson dawn]

your build. just something to keep in mind. we're in this weird generational overlap of great ddr4 and sub par ddr5 being on the market

[boreal falcon]

4090 is already 1.700-2.200€

[dense tapir]

Yeah, do the math DDR5 at cl30 is worse than DDR4 at CL18

[dense tapir]

yeap, and 1.5 to 2x that for a 5090.

[boreal falcon]

i want the beast of PC for a) Gaming b) AI stuff c) creative stuff (2D and 3D art, animations, rendering, simulations, game dev, etc.)

[dense tapir]

the 6090, if they dare to go to 2nm, will be almost 8k USD. That is beyond stupid

[boreal falcon]

if 5090 isnt too expensive i might straight get it

[crimson dawn]

modern gaming right now is all optimized to the tits for unreal engine 4 generation

[dense tapir]

I am getting a 4090 then goodbye Nvidia for 12 years.

[boreal falcon]

if not then i get 4090 and thats it

[crimson dawn]

pc hardware is leagues ahead of what software is capable of right now. ue5 unlocks it somewhat but games have hardly begun to flex that yet

[boreal falcon]

for games maybe

depends tho

[hearty karma]

next time i pull my side panel off im going to bloody clean it

get some gloves

and put it back on

[crimson dawn]

i talk about unreal engine cause its' basically the driving force of engine development at this point. most games run on it and every other engine (source excluded) are trying to just keep up

[boreal falcon]

as soon as i start rendering and simulating and animating a massive environment in epic quality i can beg that 4090, 128 GB RAM and i9 13900K processor will handle it well

[hearty karma]

128gb is kinda unnecessary though

unless your using it for some large shit

[boreal falcon]

yeah want to be safe

im not on that level yet

but what if i build a massive environment

[hearty karma]

anyways this is the thing i found https://www.newegg.com/global/uk-en/msi-geforce-rtx-3060-rtx-3060-ventus-2x-12g-oc/p/N82E16814137632?quicklink=true

[dense tapir]

You will like the ram you will hate it for games

[hearty karma]

? why

[dense tapir]

I saw the benchmarks in vids it was sad

I almost bought that too

Nvidia doing stupid shit by not giving it enough bandwidfth

[hearty karma]

🥹

[dense tapir]

@smoky oak Can tell you precisely why the 3060 is a bad deal as they did me.

[smoky oak]

the 3060 is around half the speed of a 3060ti for AI and other VRAM intense programs, mainly cause it uses weaker VRAM, and has a smaller VRAM bus, so it is a lot less performant

[dense tapir]

I saw it in benchmarks for games it was rough

[hearty karma]

so the 8gb model 3060s are better then?

[smoky oak]

for example, a 3060 typically gets around 6.5 to 7it/s, where as the 3060ti I have gets 11.5it/s

[hearty karma]

wait huh

[hearty karma]

i am confused

[smoky oak]

The 3060 is much weaker with 12GB VRAM, the 3060ti is much stronger with 8GB VRAM

[hearty karma]

so if im going for a 3060 i should go for the ti then

[dense tapir]

I read 3060 is hotter too

[smoky oak]

if you wanna do AI, then I would say yes, its much more powerful and power efficient

[dense tapir]

I am sick of the vram games from Jensen.

[smoky oak]

Yeah same

[hearty karma]

ok with that knowledge in mind i might go for this thing instead of that https://www.newegg.com/global/uk-en/msi-geforce-rtx-3060-ti-rtx-3060-ti-ventus-2x-8gd6x-oc/p/N82E16814137792?Item=N82E16814137792&quicklink=true

[dense tapir]

I saw that shit when the 3k dropped during crypto

[smoky oak]

I would get an FE< if you can get your hands on it

[hearty karma]

FE<?

[smoky oak]

I have a 3060ti FE, and its the best card I have ever had

quiet, great temps, good overclocking, and about as affordable as you can get

[dense tapir]

FE, like trying to find a mythical Unicorn.

[smoky oak]

oh yeah, I see that nowOMG

$700 new, screw that

[dense tapir]

yep

and finding a FE 4090 is forget it

[hearty karma]

yeah im much safer getting the thing i found

[oak osprey]

lol CTU with 8bit prompt

[hearty karma]

since i dont have money

[dense tapir]

4090 FE would JUST fit my case

[oak osprey]

thank you for playing Oregon Trail. You have died.

[dense tapir]

LOL

[hearty karma]

yo 💀

[smoky oak]

yeah, for that price it makes way more sense

just lord, stay away from the 4060/ti and 4070/ti

[dense tapir]

If the 3060ti had 12gb I would have already had it

[smoky oak]

They are NOT the move for AI

[smoky oak]

right

god, I wish

[oak osprey]

if you really want to do AI good you should save up til you have like $2.5 million and then get a Cerebras wafer-scale rig

[smoky oak]

lmao

good point

Cerebras 800k core chip

[oak osprey]

lets all do this and really send nvidia a message

[dense tapir]

Well, for that kinda cash do I at least get kissed in the morning?

[boreal falcon]

for AI you better go for a 24 GB VRAM lol

and above

[smoky oak]

nah, not really, 8GB works for most things with image gen

[dense tapir]

Oh, kinky

[smoky oak]

you can do 4k and 8k gens with 8GB no problem

[hearty karma]

yes well 24gb vram = fuggin lots more money in the thousands

and i dont have that much...

[boreal falcon]

ofc but thats for AI

if you can afford 48 GB and above

[oak osprey]

and then, buy two of them

[smoky oak]

You don't need that much VRAM for stable diffusion

[hearty karma]

yes but what im doing it for is not just for ai it for gamering too

48gb is overkill

[oak osprey]

for deepfloyd 🧌

[hearty karma]

yes deepfloyd computer

[boreal falcon]

for some random SD model you ofc dont need that much

for other stuff you do

[smoky oak]

Deepfloyd is gonna be interesting, cause its gonna need to be optimized

[hearty karma]

anyways imma bookmark that little 3060 ti

[dense tapir]

we are at a horrible point as I can get Pascal server cards with 24gb for about 200 but they lack tensor cores and the turing server cards are hideously priced at about a 4090 level.

[smoky oak]

go for it, its about the best deal for AI in terms of speed and capability

[oak osprey]

deepfloyd is already hyper optimized

[dense tapir]

iow DF can go pound sand for most of us

[hearty karma]

DF would nuke my computer :-)

[oak osprey]

even with my 48G VRAM i don't see it very useful

[smoky oak]

Stable diffusion 3.0 is optimized deepfloyd

[boreal falcon]

i was thinking about a workstation or even AI special card but the both have the problem to sacrifice the gaming experience