🏞｜general-with-images

[boreal falcon]

Gigabyte and ASUS are in some deep sh*t right now

i mean MSI and others have their controversies too ofc

[dense tapir]

ASUS burned me once before with their exploding electrolytic caps but I was just about to buy a TUF 4090 when tshtf on them again. I just can't

Decided enough years had passed so give them a new shot and kerpow

[boreal falcon]

well hopefully something changes at them

[dense tapir]

Everyone misses EVGA for their gpus in all surveys.

[boreal falcon]

that MSI Suprim 4090 Liquid is so epic

[dense tapir]

Sapphire for AMD and EVGA for Nvidia ALWAYS been that way for me

[boreal falcon]

beauty

[dense tapir]

I do not want an AIO as I would rather have it have a block I can use my open loop with since I keep cards for a very long time.

if that had the pump out side the card it would have been sold to me

[boreal falcon]

i dont have any experience with building, it will me my first build

later i would do some "sick" custom stuff

[dense tapir]

I am just not into AIO (closed loops) as they lose liquid over time and they break down far more than an open loop does.

open loop loses liquid too BUT you can see the reservoir and top it off when needed.

[fleet goblet]

Does anyone here use the current DirectML fork of Automatic1111's WebUI?

[dense tapir]

Gigabyte has a model that is just a block for people like me but no to gigabyte

[fleet goblet]

It recently was given Olive Support for optimized ONNX models

I'm trying to figure out how to set it up proper.

[dense tapir]

There is a video on YT setting it up as I stumbled over it.

[fleet goblet]

Is there?

[dense tapir]

I don't use it so skipped it

Yep, due to olive support now

[fleet goblet]

https://www.youtube.com/watch?v=r-ZYOlZTLDA

Just found it.

[dense tapir]

Yep, that is it

[fleet goblet]

There's a repo called Stable Diffusion XUI that supports converting the models to ONNX

This will be interesting for me.

[dense tapir]

Good luck and come back and let me know how it goes. Just ping me as I am interested.

[fleet goblet]

👍

Seems it has everything you want too.

[dense tapir]

Does it do Lycoris/Loha?

[fleet goblet]

That I am unsure.

[dense tapir]

More are done with Lycoris than plain Lora now

too bad my old card can't use any of the speed enhancers but it will eventually be replaced.

[fleet goblet]

Since I'm on Arc, I've been trying to find solutions for running repos through other than CUDA.

I've found KoboldCPP for OpenCL LLM inference, and now I've found this for DirectML.

[dense tapir]

I sure wish Intel would rock with Battlemage but I fear with them jumping in bed with Nvidia the dream of a viable 3rd gpu maker is dead.

[rain gazelle]

What's up? New repo?

[oak osprey]

i don't think so

just over-hyped onnx models

[languid pagoda]

alright one last one with the pixar one, if I mess about with ai again I'll be more productive and install add ons like controlnet

[smoky oak]

Just landed

Not really IMO, but it doesn't look terrible

But it is overdriven with contrast tho

[oak osprey]

i think the prompt asks for high contrast

just to make sure it's possible to

it wibble-wobbles toward looking like a real person lmao

the last pic is the newest there, and imo the contrast is better

it's a little too stark at first

this was an interesting one

cyborg concert

[fleet goblet]

Terribly, sadly.

Some of the outputs are black. Some of them aren't to do with the prompts placed.

Oh well.

🤷‍♂️

[dense tapir]

Black I can understand as that has to do with not being able to use xformers, or SDP, so you are forced into fp32 for everything.

Use far more ram but it is the same deal for 1650/1660 cards long ago AND AMD cards currently.

Since phanteks is being an ass with raising their chassis price by the exact amount of their mail in rebate is there any other case that can handle even next gen length cards, has a real hardware reset button, and has a drive activity light?

[oak osprey]

use google, rofl

just more hardware purchase roleplay

[dense tapir]

I love this

paying 40 usd more in under 18 months to 179 plus I need 5 fans to purchase it will be well over 200 USD just so I can hold a modern gpu with a little elbow room.

Apparently Phanteks is only sold in Amazon and NewEgg with the rest being system builders. Bit odd.

[fleet goblet]

https://github.com/lshqqytiger/stable-diffusion-webui-directml This repository mentions it has olive support

The problem is I don't know how it's implemented.

There's an Olive_Optimize folder in configs, as well.

[dense tapir]

Always head, first thing, to tickets and the tickets I am seeing is a bit much imo

Lots of bugs with one even being setup failed.

OUCH

https://github.com/lshqqytiger/stable-diffusion-webui-directml/issues/170

[oak osprey]

you're always so negative about everything, GA

[fleet goblet]

I'm probably just going to re set back up the IPEX WSL version of the WebUI

Since that was quite good.

[dense tapir]

WSL of anything takes a hit on speed. Why not dual boot?

[fleet goblet]

Because I don't want to do that.

Lol

That's why I even have WSL in the first place.

[dense tapir]

Well, your choice as dumb as that is since WSL is a stupid contraption Microsoft made.

A ton of people just left that thing and dual boot

[oak osprey]

he does what he wants, why belittle him

[fleet goblet]

I don't believe what he's saying is true though, since WSL2 is quite literally a native kernel.

The only thing that would bog it down would be the fact that I'm also running windows.

[oak osprey]

yep it's a hypervisor

[dense tapir]

yep

[fleet goblet]

The thing with that argument though

[dense tapir]

it does bog it down too

[fleet goblet]

is that my windows installation was de-bloated

If I don't run anything windows-wise it takes 0% CPU.

Lol

[oak osprey]

yep and due to the optimized libraries on linux some stuff runs faster on WSL

GA just talks out his ass and doesn't want to see anyone else have anything he can't have

[dense tapir]

Well, argue if you like but too many others who tried dual boot once would come back in here and just flat out say to hell with WSL and never again.

[fleet goblet]

If you can provide a proper reasoning as to why that's the case, go ahead.

Otherwise, no.

I will not be installing a secondary operating system specifically for running Stable Diffusion.

[dense tapir]

As I said it is your life, your choice just live with it.

[fleet goblet]

No, I am asking you specifically.

If you can give me a reason to swap to dualboot that isn't the difference between 1-2% CPU usage, be my guest.

You can't say it's just "my choice". That isn't an answer.

[dense tapir]

It is an answer as I am not your God, or parent, I am just saying you do you as what you do with your equipment is up to you. I am not here to persuade you to switch or not to switch I simply said most said fuck WSL/2 once they tried it.

[fleet goblet]

Gimme facts, or don't tell me anything.

🤷‍♂️

[dense tapir]

Done, nothing more from me will tell you a motherfucking thing. Bye

[fleet goblet]

Good.

[dense tapir]

You are now blocked so have a nice day if you are still there.

[fleet goblet]

Good for you. Please accentuate your ego further since you've failed to provide a true, factual answer.

Immaturity is something I heavily dislike.

Especially when it involves wasting time for no reason other than "it's your choice". Doesn't solve a thing.

[oak osprey]

and he calls other people thin-skinned

he just made up an argument for no reason

[fleet goblet]

He couldn't just say he didn't know.

[smoky oak]

Finally, back in America, my phone finally works again

[fleet goblet]

Nice.

[oak osprey]

i can't even understand any of the argument reading it back it's like what the heck was the issue even

[smoky oak]

I am gonna go after Verizon hard this time, cause this bullshit is ridiculous

[fleet goblet]

I asked him to give me a reason to use dualboot over WSL2.

[oak osprey]

you make a personal choice and then ... anger?

[fleet goblet]

He wouldn't give me a factual answer.

[oak osprey]

he follows the trends

"most people do this"

80794MiB / 81920MiB

sweating

that's close

[smoky oak]

Yeah it is hahs

[fleet goblet]

Big boy page/swap or actual RAM?

[oak osprey]

vram

[fleet goblet]

VRAM?

How many cards?

Or what card has 80GB of VRAM?

[oak osprey]

a single A100 80G, i tried two again earlier and i just haven't managed to make multi gpu training work

[fleet goblet]

Goddamn

Expensive-ass card.

Nice.

[oak osprey]

i'm training 2.1 on like,

2023-06-22 02:45:09,742 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:09,818 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:09,886 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:09,955 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,041 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,132 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,209 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,284 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,371 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,453 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,546 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,644 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,737 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,829 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.
2023-06-22 02:45:10,910 [DEBUG] Inspecting image of aspect 1.498 and size 1618x1080 to  1533x1024.```

batches like that

[fleet goblet]

I really wish I had a workstation GPU like that.

[smoky oak]

Still trying to get my friend to let me use his a100 lmao

[fleet goblet]

That would be nuts to have for SD stuff.

Or DeepFloyd.

[oak osprey]

i wish i had 8 of them

[smoky oak]

I don't think deep Floyd would benefit like at all

[oak osprey]

but i wouldn't know how to make it work

trust me i'd fuckin make it work

[fleet goblet]

The amount of images generated at once would, wouldn't it?

[oak osprey]

but it might not be right

[fleet goblet]

Or does DF not work that way?

[oak osprey]

i still think deep floyd's stage 3 is coming out soon

[smoky oak]

Maybe I guess, but I don't think that's like a huge VRAM cramp

Also, what's up with all this talk of SDXL this weekend?

[fleet goblet]

If there was a way

to run DF on Intel Arc

[oak osprey]

it's a myth

[fleet goblet]

I'd be trying it out

[oak osprey]

did you see the watermarks on all the SDXL outputs today? they still figurin things out

[smoky oak]

I have seen several things saying that SDXL is coming out this weekend, and I am not sure where they came from

[oak osprey]

it's people guessing the meaning of Emad's countdown

[smoky oak]

Ahhhh, right

[fleet goblet]

Well

If it does come out this weekend, thatt'l be great

[oak osprey]

but the text and counting his images portray are also DF-IF with Stage 3-esque

[fleet goblet]

Since I'll have the IPEX WebUI set up by then.

Intel Pytorch Extension*

or Intel Extension for Pytorch*

[oak osprey]

his images have some not-so-great textures in them but i'm not sure if deepfloyd can generate a quadcopter

oo

[smoky oak]

I am curious to see how much stage 3 can fix stage 2's fucked images

[oak osprey]

the deepfloyd demo is on an A100 now, not an A10

it's cominnnggg

well

maybe not. the quadcopter i got from SDXL looks just like the one Emad posted

the one from DeepFloyd is totally different

but knowing him it's just an announcement

still, deepfloyd does props correctly

@sonic arrow ^

[smoky oak]

DF is interesting, cause it can do some stuff very good, and other stuff like shit

[oak osprey]

if you stop at stage 2 it's easier to see the promise

when you plug it into controlnet for upscaling, DF is great, but it's very peculiar and needs to be done carefully

[smoky oak]

Deepfloyd can't do realistic people at all lmao

It can do really good some things, and abysmally bad others 😅

[oak osprey]

yeah i was quiet for a bit there cuz i was trying

it doesn't do it lmaoo

cartoony faces

[smoky oak]

Mutilated faces lmao

They look great at low res, but then they get mutilated in stage 2 lmao

[oak osprey]

the dark arts of plumbing

[smoky oak]

Yeah see, those look neat

[oak osprey]

those are 2.1, sorry

[smoky oak]

Now generate a portrait of a woman with red hair lol

[oak osprey]

i got bored with DF lmao

[smoky oak]

Ah, nice!

[oak osprey]

red hair

oh

[smoky oak]

Red hairn't

[oak osprey]

well if ya squint, she mint

i fucked around with the guidance levels

[smoky oak]

These look considerably less bad than when I tested lol

Still bad, but much less horrific lol

[oak osprey]

she wants to burn a house down i bet

[smoky oak]

See, they look great at low res lol

Wonder how well that would work with tile upscaler

[oak osprey]

i'm not even that good at prompting

you shoudl probably try it again

therrreee we gooooo

the eyes but oh well

maybe changing upscaler seed

[smoky oak]

Yeah, those look way less pathetically bad than when I tried it lmao

[oak osprey]

upscaler seed matters a LOT

wow, interesting results though

seems simpler prompts work better

[smoky oak]

These were from when I tried lmfaoooo

R A N C I D

[oak osprey]

9.2 cfg for stage 1
5.7 cfg for stage 2
5.6 cfg for stage 3

gorgeous red-hair girl, 1977, ethnographic photography, sweden, best quality, highest detail, balanced contrast
negative: (bonnet), (hat), (beanie), cap, (((wide shot))), (cropped head), bad framing, out of frame, deformed, cripple, old, fat, ugly, poor, missing arm, additional arms, additional legs, additional head, additional face, multiple people, group of people, dyed hair, black and white, grayscale

[smoky oak]

R A N C I D

[oak osprey]

i try to prompt a little more like a language model would

subject, location, attributes

let's change the year

[smoky oak]

That reminds me, I have a handsome tiger boy commission I have to get done soon lol

[oak osprey]

oooohhhhh deepfloyd loves the year 1992

[smoky oak]

Probably gonna start working on it tomorrow

[oak osprey]

first try

that woman would destroy my life, i know it

the eyes again tho

[smoky oak]

Definitely doesn't look great, but it's a lot better than the shit ass results I got from deepfloyd back then lol

[oak osprey]

it's not really a lot of work i put into it either

i didn't steal a prompt to make it work, which i'm proud of 😛

the upscaler seed can be like totally different results

so i'm pretty sure stage 3 is gonna be ❤️‍🔥

this one is trained on a super small dataset

[smoky oak]

Yeah, these results are giving me mucchhhhh more hope for DF

[oak osprey]

i think you have to remove stuff from your prompt

the one that gave you the freakshows up there

try the different cfg settings i gave ya too

high cfg for early stage, low cfg for later

[smoky oak]

It was a shorter prompt, don't know what it is anymore

[oak osprey]

ohh ok

sorry for assuming you prompt like a typical 1.5 goon 😄

trending on artstation

[smoky oak]

Also, I don't remember the demo having much control either

[oak osprey]

you have to click the button for it

[smoky oak]

My prompting styles are diverse, I just remember not prompting much for those images haha

[oak osprey]

its so well designed. it blends right the fuck in

[smoky oak]

My anthro gens are 100's of tokens pos and negative, but it pays off massively in the results

[oak osprey]

the trick is don't click Generate when you need to fix the upscaler

click Back to Selection

[smoky oak]

But that's cause that model has so many trained in concepts, and it's so damn good at mixing them

[oak osprey]

then click upscale again

hmm true

you could try a higher CFG as well

on your anthro model

if certain elements become too baked with a higher cfg but an overall improved image otherwise, you can just de-emphasize that term

[smoky oak]

For realism models, you can't go too hard into prompting or they start to fall apart, or they get messy

[smoky oak]

I could give it a try

[oak osprey]

yeah they're so heavily trained on coherent people in real situations that they really explode when you ask for stupid stuff

[smoky oak]

I use it at 7.5 and get phenomenal results, but it has a shit ton of headroom

You can use 1.6 token weights without getting baked images

[oak osprey]

interesting, i use 9.2 on my better models and i get great results, but, v_prediction loves higher CFGs

[smoky oak]

I have used 12 CFG as well, but I don't remember any major benefits

[oak osprey]

epsilon models behave differently with CFG

[smoky oak]

I guess I could do a CFG comparison on the same prompt

[oak osprey]

eg. 2.1-base is epsilon 512x512 and 2.1-v is v_prediction 768x768

[smoky oak]

All of my exceptional images from the anthro model are fully NSFW, otherwise I would share here lmao

[oak osprey]

well after doing some negative-free prompts on various models i realised maybe i should just re-train from the beginning on my current code/dataset using the base 2.1 model, since i can never know what damage i've done along the way and how it contributes. always good to get a clean attempt at something

[smoky oak]

That model, and the new gay focused fine-tune/prune of it I have been along with still leave all other models I have used so far in the dust it's kinda depressing to go back to them

[oak osprey]

had so many issues with my aspect bucketing because i didn't want to have to do too much pre-processing of a dataset anymore, i just want it to handle everything as it goes

there's a lot of issues you run into with images, one of which is that there's EXIF rotated images that aren't at their proper orientation/resolution/aspect ratio

so someone held a camera sideways and instead of just taking a portrait image it takes a landscape one and marks it as portrait

you have to know that exist, ahaha, and, rotate the image when you find those. but i'm thinking i'll make some small photoreal datasets by filtering out images with good EXIF data from cameras i like

laion's datasets are actually pretty great, i'm disappointed i ignored them for so long. apparently SDXL is trained on this too

DF isn't bad tbh, just needs tweaking the settings for photo stuff

[wispy ether]

#1101520089980882994

https://tenor.com/view/george-floyd-truly-lawst-big-floyd-gif-23297072

[oak osprey]

god damn you suck

[oak osprey]

the no negative challenge

wild how much better positive prompts work with zero negatives, kinda hard to get that kind of photo style from pseudo-journey-v2 otherwise

[sterile temple]

how does this look?

[shut sinew]

Looks like an expensive pc to me idk

Also isn't the 7800x3d the best gaming chip

And where is the motherboard

[sterile temple]

I have no idea about the CPU haha, I copied the shops top build PC and swapped out a few things to save a little bit of money, less RGB bling

[shut sinew]

Take some time to take everything into account, that's my advice. Took me 2 or 3 months to assemble my pc as it is now, swapping parts all the time until I ordered

[sterile temple]

the mobo is the ASUS TUF Gaming , item 4

[shut sinew]

Oh im blind

[sterile temple]

they have a sale on atm, so yeah don't wanna miss out

[shut sinew]

I mean it's your call
Just trying to be helpful

Also prices are steadily decreasing from what I've seen

I almost bought a 7900 xtx for 1.2k and it would be less than 1k now a few months later

[past pelican]

Speaking of motherboards, I need to order mine so I can finally install my CPU I bought.

[shut sinew]

What did you get

[twin flame]

How can you make 4kx4k I can barely hi res 2x

[jovial jungle]

By cheating, use upscaler in extra's tab to multiply the size by 3-4x, then inpaint it at a resolution your card can handle one section at a time. In essence I'm inpainting and cramming 10+ sections of 1000x1000 into a 4000x4000 image.

[twin flame]

Oooh

Nice. Can I know do you use any embeddings for the fingers

I'm using negative hands neg embeddings but fingers still comes out bad

[midnight mirage]

Anyone who is capable of creating different versions of animated/cartoonish faces of people/dogs/cats if given pictures of the subject. In the future I have consistent work load for creating these pictures. Searching for people who are motivated and consistent. DM me for more info

[jovial jungle]

I'm just using a few generic "make weird things go away" embeddings. They reduce weird shit by only about 30-40%, but that is enough for bad hands to be salvageable in photoshop using patch and clone stamp tools.

[faint skiff]

I keep reading conflicting information about how to blend prompts. Some things say to use ':' some '|', some '[]', some '()' and some say a combination of all of them. I did some testing this morning and now I think I'm more confused than ever about what does what — exactly

[sterile temple]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing

[oak osprey]

is the prompt maybe, lonely bachelor meals on instagram? 😄

using no negative prompt and asking for a manager looking at a file folder

[wispy nest]

so is Emad's counter going to 1 today and a reveal or is that tomorrow? 🤔

[oak osprey]

aw shit the papers containing the printed out bytes of the model were dropped on the sidewalk when being transported this morning

[pastel mango]

I made these like 6 months ago and photorealism has come a long way since @covert nymph so don't worry about SD when it comes to photorealism. Just be concerned about the hands. This is a problem MJ also struggles with, it's a seemingly impossible task for the AI 😛

[covert nymph]

Do you just add "photorealism" to the prompt? I usually try to add "hyperrealistic"

[wispy nest]

MJ 5.1 is pretty solid with hands. And 5.2 is bout to release any moment I guess hopefully not another time where it eats sd lunch again (XL)

[pastel mango]

I tend to weigh photorealistic, photorealism, hyperrealism, reality, and whatever word I can cram in pretty high. But that was then, models probably each have their own suggested prompts

I also remember adding photogenic

[covert nymph]

ok I probably need to start wheighing my prompts because I havent even tried yet

and adding more realism prompts

[oak osprey]

photorealism is an art style, don't add that to prompts lmao use like, ethnographic photography, add a year and a location, a type of camera, a film product name like kodak 300 or kodachrome

[pastel mango]

no hyperrealism is the art style

[oak osprey]

photorealism is an art style for 3D rendering and hand drawn art

that's sort of why it's called that and not just photography

[pastel mango]

so photorealism is the digital age hyperrealism?

[oak osprey]

they're two different things

there's a venn diagram overlap i suppose

[pastel mango]

hyperrealism goes for black and white mostly and portraits

[covert nymph]

I'll just smack my PC really hard and pray it gives me a more desired result

[pastel mango]

but I dunno what new models respond to, I know adding photogenic, studio photography, Sony A7R IV 30mm,

or Nikon Z9

if they have those trained

[oak osprey]

laion dataset has a bunch of exif data in it

don't know what they used as captions from it exactly

[oak osprey]

@smoky oak

SDXL 0.9 has one of the largest parameter counts of any open source imaging model, boasting a 3.5B parameter single model and a 5.8B parameter model ensemble pipeline (the final output is created by running on two models and aggregating the results). The second stage model of the pipeline is used to add finer details to the generated output of the first stage.

To compare, the beta version runs on 2.4B parameters and uses just a single model.

SDXL 0.9 is run on two CLIP models, including one of the largest CLIP models trained to date (CLIP ViT-g/14), which beefs up 0.9’s processing power and ability to create realistic imagery with greater depth and a high-resolution 1024x1024 resolution.

uses a larger OpenCLIP (but not THE largest one)

[smoky oak]

So I was right, what we have been seeing is just a fraction of what it can do, as evident of the results they have been giving from the fine-tuning tests lol

That is still insane that they are saying it runs at 2048x2048 on 8GB VRAm

I'd say it was a fake claim, if not directly from one of the developers

[oak osprey]

the 8gb vram use is just for the first stage

the two models together need like 16gb

[sterile kiln]

SDXL is available???

[oak osprey]

on clipdrop

[sterile kiln]

?

[oak osprey]

clipdrop.ai

you can go pay to gen images with SDXL for now

[sterile kiln]

Ah, i'm talking about a free safetensor

[oak osprey]

you can apply for researcher access but who knows how they go about that

i don't think it'll be like that even when it does become available 😛

it's two models

[cyan snow]

wdym

[sterile kiln]

☹️

[oak osprey]

it's two different models

one uses OpenCLIP and other uses CLIP

they need about 16GB of VRAM to work together too

you can run the earlier stage and get lower detail images on 8GB VRAM

[cyan snow]

huh, i thought it will just be a huge .safetensor file, just like recent releases

[oak osprey]

nope it sounds like there's 4 models in this thing, two text encoders and two unets

[cyan snow]

i'm confused. what will the community be finetuning after it releases?

will the A1111 webui even support it?

[oak osprey]

you know how i feel about that program lmao

you should always just fine-tune the unet and freeze the text encoder

you can finetune the models separately

[cyan snow]

that sounds like a diffusers format. this means it can be compressed to a single .ckpt/.safetensor file like the other models

[oak osprey]

it won't use the same pipeline to run

[cyan snow]

=[

[proud dagger]

once it's fully public, yes

[proud dagger]

Not quite - each model uses both

[oak osprey]

😮

[cyan snow]

which model will be the one this community uses

[proud dagger]

there's a main model and a refiner model, so presumably the main one

the refiner has cool magic tricks built in tho

[cyan snow]

sounds great

[proud dagger]

and the vae

[wispy nest]

so for us dummies when will this XL be available on auto?

im confused how its released atm, like just a bot or what

[proud dagger]

once it's public, which is Soon™️ but not now

[oak osprey]

i like it, chaz. straight to the point

mid-july the blog post speculates

[proud dagger]

it's currently released for Bot/webapps, and researcher access

[wispy nest]

oh mid july

so todays countdown was just for an announcement no release 😭

[oak osprey]

-_-

it has academic access, and clipdrop, and api access is coming

[cyan snow]

https://tenor.com/view/brain-gif-22265960

[oak osprey]

those have higher rate limits and so on

[wispy nest]

whats clipdrop?

[oak osprey]

clipdrop.ai their image gen webapp

idk how to brag about what i've done in my research access request so i just hope they look at my profile on HF and see my efforts 🥹

[cyan snow]

POG

wait, can we access 0.9 already?

[wispy spindle]

via clipdrop!

[cyan snow]

i meant locally..

[wispy spindle]

not yet ^^

[cyan snow]

also, how can you trade-mark the word- soon?

[kindred parrot]

Apply for it at an undesignated point in the future

[wispy spindle]

say soon so much you become the soon

[cyan snow]

this is like how monster energy trade marked monstertm

[wispy spindle]

soon-energytm

[cyan snow]

well, idk about trade-marking words, but check out this Splatoon gun i made =]

[proud dagger]

if you use the word soon without our approval, we'll sue you... eventually.. once we get around to it..

[oak osprey]

i keep meaning to release my thesis on how distraction isn't a thing, but i've not had much time with all these other things that keep popping up

[rugged spear]

yooo thats cool

looks kinda like the tranq gun from mgs5

[cyan snow]

Thank you, thank you.

[zealous socket]

A week ago I didn’t even know about SD, today I trained my own model on 30 photos of me, then turned myself into fortnite

[rugged spear]

What kinda magic did you have to pull to get those splashes with the splatoon gun? Makes me wanna make some mock 90s soda ads

[cyan snow]

lora:add_detail and some simple prompting.

But SDXL won't even need any fancy techniques to pull off that kind of stuff

[edgy otter]

Any battle cats fans here?

I made lora inspired by this character and I was skeptical while training it due to very low resources but it suprised me how well it turned out!

This is the char from the game:

And this is my fav one I made using the lora I trained of her (and some inpainting):

[kindred parrot]

Awesome work!

[rugged spear]

the color theory in this one turned out really good lol

[edgy otter]

I mean
I used value of 0.7 for her lora
Any bigger value would make her more true to the char but less sharp coz all her pics are pixelated and small

Like this for example

[zealous socket]

thank you 🙂

[edgy otter]

Oh yeah this IS awesome indeed

😁

[oak osprey]

interesting transition during widescreen training

[fleet goblet]

If anyone here uses vladmantic's automatic, could someone tell me where the launch commands actually go compared to A1111 repo?

[edgy otter]

Ok it didnt send for some reason so here

[oak osprey]

look ma, no dupes

so much room for activities

[smoky oak]

Damn, SDXL is looking extremely good from what people are sharing on Reddit, but my question is... Where are they getting such good results from? As is, the results they are posting blow all 1.5/2.1 models and MJ out of the water handily

[oak osprey]

probably the clipdrop version

[smoky oak]

Looks like SDXL's no dupes haha

People have been messing with that a lot

[oak osprey]

can't run it locally yet

[oak osprey]

thats what made me realise this is possible

18 hours to do 115 steps of training so far

[smoky oak]

Ahhh, I see haha

But God damn, the images on Reddit right now are so fucking good

[oak osprey]

cherry-picked 😛

it's a 1024x1024 model and apparently doesn't do smaller than that so well

[smoky oak]

Yeah, but no amount of cherry picking could get 1.5 or 2.1 to look half as good regardless lol

These look amazing

Oh ew, excuse the discord compression, they don't look that bad normally lmao

[oak osprey]

yeah they're really good if you don't want a jack hammer in the output

[smoky oak]

Again, I find your critiques so Interesting lmao

You give 2.1 a golden star when it occasionally doesn't look terrible, but any other model you have soooooooo much bad to say lmao

[oak osprey]

whut

i complain about 2.1 all the damn time

1.5 is monumentally worse

SDXL, i don't have access to, so i can't see how good it is and what it runs like

[smoky oak]

Yeah true, but you also praise a lot of it, and the second SDXL looks phenomenal, you can't give it even the slightest bit of cudos lmao

[oak osprey]

also not true, but you see what you wanna see

pretty sure i've done more SDXL gens than you have, did you even play with the bot 😛

[cyan snow]

if you calculate average fuck up rates of diffusion models, 2.1 fucks up way more then 1.5, and SDXL is way better than both.

[oak osprey]

i'm not talking about the overfitted 1.5 models that no longer listen to your prompt

i'm talking about base models

[cyan snow]

then SDXL is BY FAR the best.

I didn't see SDXL0.9 fuck up even one time.

[fleet goblet]

SDXL will be fun to use.

I have been patiently waiting for a while to try it.

[oak osprey]

i do but you have to go into the obscure prompts that enter training gaps

[cyan snow]

you probably used the bot, the bot uses a much earlier version of SDXL

[oak osprey]

no it doesn't

the bot, if anything, has a better version than clipdrop does. mcmonkey stated as such. it's being worked on

[cyan snow]

idk man, the images generated here: https://stability.ai/blog/sdxl-09-stable-diffusion look way better than the ones the bot makes

[oak osprey]

skill issue

[smoky oak]

I am messing with this as we speak, and the results looks amazing

[cyan snow]

my theory is that the bot uses the beta version, the images it makes look more similar to SDXL beta than SDXL 0.9

and it's not even done yet, SDXL 1.0 will release in less than a month

[oak osprey]

yep literally anyone can just make pretty photography of people, no problem

[oak osprey]

it's kind of wild

[smoky oak]

Just imagine what we will be able to do as we learn to get more and more out of it over time like 1.5

And then also the fine-tuning we should be able to do as well

[oak osprey]

i see they're using deep fried in their negatives now

[fleet goblet]

Isee.

[smoky oak]

I am trying to use clip drop at the moment, but nothing is loading

Weird

Yeah ok, so the site is being like bombed now lol

[cyan snow]

yep, SDXL 0.9 makes MJ and 1.5 finetunes look like wish.com. and it's not even finished, or finetuned

[oak osprey]

the queue is deep, honey lol

[smoky oak]

It was 8 images ahead, now it's 1000 images ahead

[oak osprey]

just use the bot channels imo

[smoky oak]

It really does haha

[smoky oak]

Significantly worse results

I'm willing to wait to see the real deal

[oak osprey]

you have to re-roll sometimes because the bot has 4 different SDXL models

[kindred parrot]

The bot uses a few different models to collect preference data so while it can be close to, or on-par with 0.9, it's hard to compare any output it gives since it may not be a comparable model 🙂

Basically what Pseudo is already talking about

[oak osprey]

it's a funny concept of whether a model is finished or not because we were observing Pope Francis being epicly burnt-in and other celebs are under-trained. so if your goal is to generate nothing but Pope images, you're good to go, maybe a little hard to make him do exactly what you want as he was overfit so much though

[fleet goblet]

So wait.

[oak osprey]

there's always something it won't know even when it knows all of the planned test surface area concepts already

[fleet goblet]

What methods are there currently to access the 0.9 beta?

[cyan snow]

it's not released yet, they said that SDXL 1.0 will release in less then a month

[fleet goblet]

Open-source or Dream Studio only?

[kindred parrot]

For specifically 0.9, clipdrop & some specific research groups that have been given access.

The bot can be used as well on the server of course, but you won't get purely 0.9 outputs from it

Full release (coming ) will be Open-Source!

[fleet goblet]

Awesome.

[smoky oak]

Everybody and their mom keeps saying it's dropping tomorrow, and I'm pretty sure they are just smoking copium lmao

[fleet goblet]

Now the only problem I will have

Is that I'm on Intel Arc.

[oak osprey]

Dan, i asked about access for research and was chuffed-off as "some random discord user that likes fine-tuning"

[smoky oak]

Oh God, good luck

[fleet goblet]

I joined almost a year before you.

As an early access member

[smoky oak]

Gonna need minimum 16GB VRAM and at least 3 months for somebody to get one sampler working 20% of the time

[cyan snow]

bro

[fleet goblet]

I was already messing with Dall-E 2 and Dall-E flow in Collab.

[cyan snow]

they literally said it

[oak osprey]

i write proposals and try and do big research, and am working on a thesis next quarter

[smoky oak]

Uhhh, tomorrow is not July lmfao

[fleet goblet]

And yet they treat you that way.

Interesting.

Maybe it's because of your discord image? Lol

[smoky oak]

It's June soooo

[oak osprey]

shrekscrimination is a real problem

[cyan snow]

yeah, i didn't say tomorrow LOL

[oak osprey]

tomorrow is july though. for extremely large values of 'tomorrow'

[smoky oak]

Yeah, everybody is saying tomorrow lmao

[fleet goblet]

Tommorow***

*1. Next month

[smoky oak]

This looks insane

[oak osprey]

that dude is smuggling an eggplant

[fleet goblet]

How would you say it compares to any other T2I ncluding DeepFloyd?

[oak osprey]

oh my god i don't even think i can show my clipdrop results here

is there no nsfw filter?

[fleet goblet]

A watermelon made of icecream.

[cyan snow]

not on the open source that will release, as far as i know

[oak osprey]

no, no, on clipdrop you can prompt for nsfw and it works. but the model is incapable of it, like 2.1 was

it shows horrendous stuff instead of actually doing it

so i'm assuming the model is filtered

try it yourself and see 😛

[fleet goblet]

Hm.

Compared to Bing Dall-E

It's pretty damn close.

[cyan snow]

i bet after it goes open source the community will already make hentai models with SDXL 1.0 as base , then upload them on civitAI

[oak osprey]

trying to train NSFW into 2.1 breaks it pretty good so i wish them luck, there's TWO text encoders to contend with now

[oak osprey]

i guess it's possible but it won't be very generally capable after you're "done training"

[cyan snow]

i'm not sure if the open source release will even be filtered, if i recall correctly they said it will be as capable as 1.5 when it comes to nsfw

[oak osprey]

damn, Bing can do pixel art pretty well compared to SD (1.5, 2.x, SDXL included)

[wispy nest]

Bing seems better than dalle2 which is what I thought it used

[oak osprey]

Bing is DALLE2 indeed

[cyan snow]

AGAIN?!?!

[fleet goblet]

It's improved.

[oak osprey]

sorry

[fleet goblet]

An improved Dall-E 2.

[oak osprey]

the SDXL devs acknowledged this though

pixel art requires very specific training

[wispy nest]

interesting. I thought the results on bing were much better guess it wasnt all in my head

not that i use dalle 2 much these days but I played with it last month and it seemed kinda ass

[fleet goblet]

It's Very good.

A boss from the game "Dark Souls", Computer-Generated realistic art.

[wispy nest]

noice

[cyan snow]

there is no way it beats SDXL, i can't stand that

[oak osprey]

i was messing with DeepFloyd yesterday and tweaked guidance values and got much better photorealism out of it so i could see that being the case for Bing's DALLE2

[cyan snow]

it only beat 1.5 with making babies, so, yeah

[oak osprey]

9.2 cfg for stage 1
5.7 cfg for stage 2
5.6 cfg for stage 3

gorgeous red-hair girl, 1977, ethnographic photography, sweden, best quality, highest detail, balanced contrast
negative: (bonnet), (hat), (beanie), cap, (((wide shot))), (cropped head), bad framing, out of frame, deformed, cripple, old, fat, ugly, poor, missing arm, additional arms, additional legs, additional head, additional face, multiple people, group of people, dyed hair, black and white, grayscale

you might have to randomize the upscaler seed a few times but you'll get a good result from DeepFloyd with this, even with the x4 upscaler

[naive wren]

Does anyone knows how to replicatethis? or what plugin setup was used?

[oak osprey]

just use that one lora everyone uses that overfits on that exact face i see everywhere so annoyingly

[cyan snow]

this is what a 1.5 model does

[fleet goblet]

eh

[wispy nest]

i been hung up on 1 image waiting to be processesd for like 5 minutes 😭

[fleet goblet]

Yeah, SDXL is actually quite close in quality.

Damn, what a cool concept.

[cyan snow]

SDXL should be way better then bing, it was trained way more

[fleet goblet]

I don't know where you're getting that from.

[oak osprey]

openai doesn't tell us that stuff anymore

[fleet goblet]

SDXL is what. 2.3B parameters?

[oak osprey]

dalle2 could be huge at the moment

[fleet goblet]

DAll-E 2 is above 3.5B parameters.

[oak osprey]

it's like 3.4 plus 5.3. there's two models

it's on the SAI blog post.

[wispy nest]

as long as people will fine-tune it somehow, then it will probably be popular. If not, then it might become 2.2, I'm hoping and wishing for the former :D

[oak osprey]

the refiner model is purely about adding fine details to images

[fleet goblet]

Both models nail the dark souls style.

[cyan snow]

so, you gave up that easily? next thing you say is you are buying an MJ subscription

[fleet goblet]

What in the hell are you talking about?

"Give up"

Like I have to choose between models.

[oak osprey]

tdg8uu has a strange sense of humour

take it not personally 😛

[wispy nest]

nobody else hung up on spinning wheel and 1 images i take it huh 😭

[oak osprey]

mine's behaving poorly too

it gave me three NSFW images and one dude with a pipe for "a man using a jackhammer on a construction site"

[oak osprey]

tryina get the dude with the pipe in HD at least and it's stuck

[fleet goblet]

I'm just generating both Dall-E Bing Images and Clipdrop SDXL 0.9 images to compare

Both are good. Let's be straight with that.

SDXL will be straight up better though with the open-sourcing of its model.

More customizability is always better.

Neither of which however can do text.

[wispy nest]

I'd rather have everything else, such as items in hands, and fingers in the first place, correct, before getting some writing correct :P

[fleet goblet]

Indeed.

[crimson goblet]

not what I wanted but pretty

[fleet goblet]

All of these models are great, though.

[cyan snow]

1.5 finetune

bing

[fleet goblet]

Give me your prompt

[wispy nest]

and if it's still a limit on the aspect ratio and duplication, then the use of a good upscaler will still be a must. Really want to ditch using one one day though :P

[fleet goblet]

Alright, nah. I just got a really well-detailed image from SDXL 0.9.

What a beauty.

Havel if he was reincarnated as a mix between his armor and the tree guardians.

Nope. No model comes close to this intricate detail.

Damn.

Especially this one.

How much VRAM would SDXL 1.0 require anyways?

I understand the plot perfectly in this story.

[wispy nest]

funny that I was testing some armored prompts yesterday :D

[fleet goblet]

Very well-done.

Model?

[cyan snow]

a 1.5 model i finetuned

[fleet goblet]

I wonder what a SDXL finetuned model would output in comparison.

The fact a base model can output that level of detail though.

[oak osprey]

lol fuck

[cyan snow]

we will find out in about a month

[cyan snow]

16GB, i think

[fleet goblet]

Well I guess if I figure out a way to run it on my A770, I'm covered.

[cyan snow]

but i think that's for running both variations, some people say 8, some people say 10, there is only 1 way to know

[wispy nest]

have they stated plainly, and clearly that the XL version will be open to use like their previous versions? I.e., free to modify, or is it just an assumption that it will be the same?

[fleet goblet]

It is open-source on release, supposedly.

[oak osprey]

they say "full open source" but i doubt that includes their dataset

[cyan snow]

they said that this time, they would also make fine-tuning tools for it

[oak osprey]

https://github.com/Stability-AI/generative-models

[fleet goblet]

A caveman wearing headphones and glasses made of stone. Gamer Caveman. Analog.

[wispy nest]

so, there will be a similar "file" to download such as the .cktp? (or how it was spelled, can't remember from the top of my head :P )

[oak osprey]

if they do that it's going to be a 20GB file or so

[oak osprey]

SDXL is all about modularity. aka diffusers style checkpoints

[cyan snow]

there will be 2, one is the main version and the other has some tricks

[wispy nest]

so, is that a yes, or? :P

[oak osprey]

it's a "no one knows" but "likely no" but still "someone will do it anyway even if it doesn't make sense"

so don't worry

[smoky oak]

Likely there with be a base part of SDXL, and then you download smaller top layers, like huge LoRA's

[oak osprey]

whatever the case, A1111 will be forced into working with it

[smoky oak]

That's my guess, to save disk space and make things way more efficient to share

[fleet goblet]

I'll be using vladmantic's automatic webUI

Since it's just a version that is usable by all GPUs, basically.

[smoky oak]

Blegh, Vlad

[wispy nest]

all I want is to not have to log in somewhere each and everytime. I.e, subscription/cloud/internet of things/etc :D

[smoky oak]

Unfortunate that a POS like Vlad has a fork of a webUI from a POS like Auto lmao

[cyan snow]

a dev said today in this chat that there will be 2 models, both in a .ckpt format or something, scroll up, i don't remember

[fleet goblet]

Nice output from Bing Dall-E and SDXL.

[oak osprey]

they'll be huge files, and a1111 users will suffer if that's how it's implemented. but A1111 users suffer anyway, so business as usual

[weary oasis]

Please is the bot here using sdxl 0.9?

[oak osprey]

kinda yes, kinda no. that's one of the four models it uses

[cyan snow]

that's most of the people in this server

[fleet goblet]

Alright, well this one just straight up looks like a 3D render.

[oak osprey]

you have to re-roll a few times to experience all 4 versions

[weary oasis]

I see ,thanks for the feedback

[oak osprey]

most of them are just here for generating waifus and other sick shit lmao i have no concerns over their abilities to generate images. they're degenerate enough to find a way

[cyan snow]

well, i'm willing to bet that people will like SDXL wifus then 1.5 wifus

[oak osprey]

i'm curious what "research" i need to be doing in order to get access to the model weights

[fleet goblet]

This above image

is like a combination of two things from Demon Souls.

The default knight set, and the tower knight boss.

[oak osprey]

@cyan snow SDXL makes overly skinny women though the few accidental NSFW's ive seen have really curvy broads with zero areolas again

[oak osprey]

huge boobies with nothing on them

[wispy nest]

lol i love gamer caveman

[cyan snow]

isn't the open-sourced release going to not be filtered, like 1.5?

[oak osprey]

the haloing though

[fleet goblet]

I like it heavily.

It's like a profile picture.

[oak osprey]

when they say that they mean "do whatever you want with your own fine-tunes", not that the model magically generates the ability to do NSFW

[oak osprey]

remember they keep talking about responsible AI, and filtering boobs is one way they feel they accomplish that

[fleet goblet]

Sitting patiently.

[smoky oak]

I was reading some discussion about NSFW in SDXL, and it appears as though the model has no major censorship, from what I was reading

Tigur

[oak osprey]

you can simply test it for yourself

[oak osprey]

no need to guess

[smoky oak]

They blur any NSFW results

[oak osprey]

not all of them apparently

[fleet goblet]

That's an interesting outline.

It's like he's passing through a fog barrier.

[cyan snow]

they will release it in a month, we can't test it

[oak osprey]

okay, you keep hoping, lol

[smoky oak]

They did say that their goal is middle July, so it's at least a reasonable assumption

[oak osprey]

i don't think they're going out of their way to add cock monsters and boob goblins to the model but i could be wrong

[fleet goblet]

Heheheh.

[oak osprey]

this is the v0.9 model on clipdrop, you can see the kinds of issues they're wanting to resolve

that

[fleet goblet]

PyroSniper.

[oak osprey]

stuff like this i'm disappointed they haven't been able to resolve yet, and i'm not ready to give up on it being fixed but it does look like some kind of fundamental issue that their current architecture hasn't yet resolved

[oak osprey]

and anything it has issues with can just be passed off as "something to fine-tune", but you can't fine-tune small faces into SD 1.x or 2.x

[fleet goblet]

LOL

[oak osprey]

so it's going to suck if "dust particles" are another "currently unsolveable issue"

[cyan snow]

they said they will make tools that are made specifically for finetuning SDXL

[smoky oak]

Their upscaler or post processing really messed up images

[oak osprey]

it's just not actually possible to train it into there, vs, the tools not being available. two different issues

the upscaler is the stage 2 for SDXL

[fleet goblet]

That's not a knife. THIS is a knife.

[cyan snow]

PEW

[fleet goblet]

SDXL nailed SCP-049 pretty well.

God damn SDXL really got it.

SCP-173.

He's like a weird, rebar/congrete made humanoid.

[cyan snow]

huh

[fleet goblet]

Animated CS:2 skins.

[wispy nest]

never seen a gun tilt its "head" before, or rather, never thought about it before. For obvious reasons :P

[fleet goblet]

Hell ye.

[cyan snow]

are the SDXL styles LoRA's?

i mean, what else can they be?

[oak osprey]

no

they're literally just prompt pieces that get tacked on

if you start a prompt with "A man, holding a sign that says:"

it'll print pieces of it

[cyan snow]

DAMN SDXL is good

it understood the prompt perfectly

[oak osprey]

i heavily weighted jackhammer and it gave me more naked women

[wispy nest]

still getting some funky hands

[cyan snow]

with negative prompting?

[oak osprey]

i had to blur it myself

[wispy nest]

no just using that clipdrop

[oak osprey]

why is jackhammer so closely tied to obese NSFW

[wispy nest]

ben affleck dressed as vegas elvis holding a sign that reads "austin butler sucks"

probably it's because of the rest of the prompt. I'd test writing it as "construction worker with a jackhammer" or something :P

[cyan snow]

then why are you surprised, just the fact it rarely messes up hands without negative prompt is insane

[wispy nest]

idk mj 5.1 does hands pretty well without any other tricks i thought this would too 🤷

but point taken, i'll lower my expectations

[oak osprey]

i tried that at first and got the toasty image i shared earlier

[cyan snow]

MJ5.1 is heavily modified, there is 100% chance they have something that modifies the prompt

[oak osprey]

MJ 5.1 "does hands well" when you cherry-pick

construction worker with jackhammer has no jackhammer, so, i heavily weight the term and get NSFW

[cyan snow]

you can get results at least in MJ's level when running even a 1.5 finetune properly.

and heck, imagine what SDXL1.0 will do

[wispy nest]

i'll take your word for it

[oak osprey]

like what the fuck

[wispy nest]

ah, then it's probably because the jack word. Try maybe something like "pneumatic drill" or words not assosiated as much to nsfw stuff...hmm...are there ANY words in the english language that can not be seen as some dirty insinuation? 🤔 🤣

[oak osprey]

he looks ready to take on those other 3 images

who ya gonna call? jackhammer busters

🤫

well i hope they update clipdrop soon or stop the bot from using garbage test versions of the model, but ideally, both. until then, back to screwing with my degenerate 2.1 shit

[cyan snow]

idk man, looks fine to me

[wispy nest]

it seems it think jack is a person, and also "construction worker" is another person. That's probably why you get two people in the image. So maybe not use "with" as a joiner for it. Heh, well, it will be interesting to learn a new model if it has different thoughs :P

[oak osprey]

i don't want it to have thoughts

[wispy nest]

or was the consturction worker a prompt after an earlier one with nsfw stuffs 🧐

seems you can do one prompt solo over the other 4

[oak osprey]

i started out prompting "blue-footed boobies" and then no matter if i refresh or what, one of the image squares ends up with a new image following new prompt, and other 3 images become new following NSFW prompt

[cyan snow]

this is what my model makes

[oak osprey]

that looks better than whatever they're currently calling SDXL on Clipdrop

[wispy nest]

the third image look just like me when at work, a cross between wanting to hit myself with the hammer, and if anyone would notice if I left :P

[cyan snow]

maybe because i run it locally, maybe because i finetuned my model, maybe even my settings? i PROMISE if you make an SDXL fine tune the same way i made my 1.5 model, it will be infinitely better than any finetune

[wispy nest]

I'll trust it to be better in time. Gotta give it a chance to mature, in more ways than one if I know the community :P

[oak osprey]

but your model has late stage elder magic syndrome

@wispy nest 2.1 is still not really there because training it is "so hard"

people have a single text encoder and a single unet to train and they're smaller, so, easier. but these are huge models and people aren't going to bother

you'll see a lot of textual inversions because they're more powerful than LoRAs were in 1.5.

[wispy nest]

it probably wasn't hard enough if you get what I was saying before ;P

[cyan snow]

is that a pro or a con?

[oak osprey]

it was easy to get good aesthetics but dude all the faces end up looking the same on the fine-tunes i've had and played with

make a group of people and they're all clones

SDXL has the same issue of making three people and the middle one is a blend of the outer two

[wispy nest]

there's a lot of issues with ai art in general, but those things will become better in time. I'm sure of it :)

[oak osprey]

that's just a core issue to the way SD works it seems

[wispy nest]

maybe meta's new method will crack the code

[cyan snow]

he is so dumb

[oak osprey]

bing thinks refreshing the page will fix it but that just deleted our whole chat

geriatric bastard

again bing can do a jackhammer with proper dust particles, first try

[cyan snow]

can i even show this

[oak osprey]

he's just helping step-bro out

[sterile temple]

Teamwork makes the dream work

[wispy nest]

maybe you wrote the prompt too quick and "accidentally" switched jack and hammer? ;P

[cyan snow]

i deleted it, i don't want to be tracked down by the CIA

[wispy nest]

why are you guys so into construction worker motifs if i might ask

[oak osprey]

korean danny devito will find you

[oak osprey]

sdxl cannot do a jackhammer at all no matter how hard you try and when you try even harder it gives you NSFW tiddies

[cyan snow]

@oak osprey said that the model i use is better than SDXL at making construction workers

[oak osprey]

its sharper and has better contrast

reference that i'm looking at here that is all smeary

SDXL uses a LOT of bokeh/depth of field to hide the flaws in the background

[cyan snow]

i REFUSE to believe that a 1.5 finetune is better than SDXL, even SDXL beta at anything

[oak osprey]

it's easier to never see them if they never appear

@cyan snow trust me it makes me want to just give up on training if even SDXL after all this effort looks like this still, fixing the problems are well beyond my abilities

[cyan snow]

nah, it's better

[oak osprey]

i will try and get 'narrow aperture' to work

[oak osprey]

that looks like a painting to me

[wispy nest]

you're probably a lot better at writing a prompt for your model than this new one. It probably have different weights all over. Or at least that's what I've noticed happens to me when trying out new checkpoints :)

[oak osprey]

that is not good

what the hell is that

a vacuum operating in reverse?

blowing burnt AI fragments

[cyan snow]

and this is better?

i don't think so

[oak osprey]

i see no residual noise in your image unlike theirs

clipdrop makes SDXL feel like DALL-E before they fine-tuned that

there's a limit?

[sterile temple]

Can it do bicycles, motorbikes, car interiors? The stuff that wasn't possible before but fairly common in real life? Jackhammers is a bit niche

[oak osprey]

it wasted most of my gens on NSFW i didn't ask for

for fuck's sakes

[cyan snow]

still, my model is finetuned. i bet if you give me base SDXL1.0 i can make an exponentially better model

[oak osprey]

2.1 can do bicycles, especially it knows the difference between road bikes, trail bikes, enduro, and downhill. it is in my validation prompts list

[sterile temple]

Perfectly?

[oak osprey]

once you train it on wide screen images, it can do cars better too

mm, SDXL doesn't do bikes perfectly either

[wispy nest]

are you hating so much because they didnt give you early access or whatever? lol

[oak osprey]

that's also supposed to be John Cusack

[wispy nest]

I've not seen any perfect art from any art ai yet. There's always something wrong with them. Or at least something I dislike about the image :P

[split rover]

chatted with them, it should likely update soon

[oak osprey]

i'm literally just looking at the results i get

[wispy nest]

update to what?

[split rover]

better inference params

[wispy nest]

isnt it .9 now?

[oak osprey]

they should identify the issue that keeps old prompts around, maybe has to do with clicking the Generate button before "Back to grid"

[cyan snow]

idk man, i kinda like the sdxl model on clipdrop better than the model i finetuned

[oak osprey]

so it's like having a good vs bad pipeline for 2.1? that should be an exciting number of people giving up early on SDXL because of using their chosen ancient sampler ( looking at Euler A )

[wispy nest]

looks like it can be a lot of fun to use. I'm gonna wait until I can run it locally before trying it, but I really like the results I've seen so far :D

is XL like a whole new sampler?

[oak osprey]

it decoupled some stuff so that the samplers and CFG are separate

[wispy nest]

like say when/if its on auto would there be some new sampler(s) just for XL?

[oak osprey]

most likely XL should have its own list of samplers and disable any that don't work. 2.1 should do that too, but it doesn't. so

[split rover]

Same samplers as before on auto/comfy/etc, nothing new on that front for XL

[cyan snow]

yeah, when it will be public we will make finetunes that can possibly be the best diffusion models to date

[oak osprey]

auto's samplers don't use the trailing timestep or rescaled pretrained betas

[wispy nest]

Batman standing next to Superman in some movie title I came up with, was few prompts ago. Not bad but the only one of the 3 tht didnt have them merged in some way

[cyan snow]

like, even the unfinished beta version is pretty much better than most finetunes currently and even MJ

[wispy nest]

idk Im impressed by XL from what Ive seen so far

for me I'd say that the most important part is if the results are what one would want. Some might want simple, strick, and similar results from the same prompt while others want more chaos! :P

[cyan snow]

well, if you want chaos just use base SD2.1 or 2.0

heck, even dall-e mini or what ever it's called now

that makes even better monstrosities

[wispy nest]

the good kind of monstrosities, or the bad ones, as in "what in the world is that thing?!" :P

[cyan snow]

LOAB

2.1 makes LOAB like there is no tomorrow

I'm almost 100% sure that's how LOAB was first created

[wispy nest]

what does it mean? I have never heard the word(?) before :P

[oak osprey]

bing now requires an Image Creator Account, which is currently free(? i think?) but who knows

[cyan snow]

bing makes less abominations than 2.1 and 2.0

[oak osprey]

true

you have to kinda ask it to do them

[cyan snow]

this is LOAB

[wispy nest]

ive dated worse

[cyan snow]

she is a result of AI going crazy

[oak osprey]

that's why they call you Chaz?

[wispy nest]

lol

[oak osprey]

SDXL can't do dinosaurs worth a fuck

but that's understandable

there's not many dinosaur selfies out there

[cyan snow]

I am almost completely sure she was created with sd2.0 or 2.1

[oak osprey]

we should bring dinosaurs back to life somehow

then we can photograph them for AI training data, and then, put them all down

[cyan snow]

but yeah, that is true

[wispy nest]

I…I regret ever asking what it was! :P

[cyan snow]

SDXL is probably the best diffusion model yet, and it's not even finished, or finetuned, so imagine what will happen after it goes open sourced

MJ is going to go bankrupt, unless they have something up their sleeve

[oak osprey]

there's finetunes that kick MJ's ass and it's so esoteric that the uninitiated find it overwhelming to even get started on.

[wispy nest]

I have about 10,000 words I'd like to see if it understands "correctly" that the other models don't seem to be able to. But I can wait, I enjoy what I already got so I'm not in a hurry. yet :P

[oak osprey]

using german is cheating

i found jackhammer, that one is mine >:|

[cyan snow]

but yeah, still. the moment someone like us will get their hand on the finished base SDXL model, it's completely game over for those guys

[oak osprey]

no it isn't, because people have been claiming The Year of the Linux Desktop for decades now and it has everything it needs to make it work, and guess what? mindshare and ingratiation are big factors in cementing people in place

midjourney has a community of chuds over there all chatting with each other about image gens just like we do here, and they kinda like each other in some cases, and wouldn't want to change their seat, as it has already been warmed up.

i'm curious what SDXL looks like without the refiner

[cyan snow]

but the images that we will make using SDXL finetunes will be exponentially better then what MJ can even hope to achieve, do you really think they don't care?

[oak osprey]

yes

you underestimate people

or overestimate, i'm not sure

you overestimate their ability to care about technical details when they can just open up a Bing chat window and say "make me an image of a welding baby ignoring OSHA protocols"

stable diffusion was around for a while and my friends and i were happy to use Craiyon because it met our needs for absurd, funny images

the first SD image i made looked so trashy i just couldn't even tell how i was using the same model that the demo images came from

[cyan snow]

that is because you had no idea what you were doing

[oak osprey]

well neither will they, which is my point

[cyan snow]

i see

[oak osprey]

bing chat and MJ are just discord bots. there's Blue Willow, which is like MJ, but it's SD 2.1

well bing chat isn't, i misspoke

but you get what i mean 😛

clipdrop is their best chance at grabbing that kind of mindshare

[cyan snow]

yeah, i see

[oak osprey]

even dreamstudio is too much of a barrier to entry

i am like you, you know? i'm a technical person, i like the fine details of how things work and why they're good. that is why i use linux. but i've been dealing with users for a while, i used to try and convince them to for example, switch from windows to linux. but i don't want to support those users. i don't want them clogging up our bug report forums. soooo i stopped trying. i accept that there's other tools for other folks that have very different needs than i do. it's always interesting to encounter a very different workflow that literally can't use Linux. i find it fascinating.

like this, i don't expect stable diffusion to become the best model for end users. i don't see any point in it, it'll end up clogging their support dept up with stuff they don't make a whole lot of money on, honestly. they're better off pursuing large customers like ILM (Industrial Light + Magic) or other heavy-hitters that would need their own models trained

i'm just happy i can mess with it however i want, because that's what i like to do. that's not what i expect anyone else to want though.. but like, making a bot that could replace Midjourney for someone is my goal

my model is currently in use on the "Fulljourney" service as their SFW model, which i'm quite proud of

[wispy nest]

I don't think any ai is good enough for real world "industrial" use. Probably not even close to it. :P

[oak osprey]

oh i don't know @wispy nest didn't you hear that NVIDIA indoctrination song??

[cyan snow]

or, there can just be a webui different from a1111 that is already set up, and ready to go. it will be no different from the bot they will use, except image quality will be way better

[oak osprey]

@cyan snow most of these people use like, compaq laptops from 2012

[cyan snow]

thats a fair point

[wispy nest]

it's not about I heard it, but if I listened ;P
Wait... hey, what's the different :O

[cyan snow]

but yeah, after SDXL fully releases the images we would make using SDXL finetunes will be insanely better than anything they used

i guess only time will tell

[wispy nest]

I hope so as well, but I'll wait till we get to test it locally :D

[oak osprey]

ah yay they were able to reproduce the issue with dogs and ducks. very safe for work 😄

it's such a weird bug because it seems like it made two images for me

i think it's just supposed to make one

probably wasting some GPU time on that

[wispy nest]

that's why I don't write it as firearm :P

[cyan snow]

oh, no. that was the intension behind my prompt, it did it perfectly

dude

[oak osprey]

@proud elk

the nsfw lady that appeared out of nowhere

was because the image got 'stuck' with the spinner

ahahaha

that time it got kinda jammed up for all of us, apparently my request went through and then showed up randomly when i'd prompted for a construction worker

soo that's not the jackhammer keyword being naughty after all

damn

a group of handsome man, standing proudly outside of a pub in Ireland at night, illuminated by neon lights

where's the handsome

[cyan snow]

also, will loras like <add_detail> work on SDXL models?

[oak osprey]

my WIP model's result

it has ALL THAT ROOM TO PUT THEM

instead, tiny derp faces

the prompt must suck, i know

photograph of a group of handsome men, standing proudly outside of a pub in Ireland during nighttime, illuminated by parking lot lighting

i also do this one for validation. maybe standing proudly makes no sense to the models

#5 from the right looks fucked but the rest look fine (when zoomed out) but they all fall apart the moment you start looking closely

[weak sage]

error

[oak osprey]

stable derpfusion

[oak osprey]

beep boop

seems like it broke

[sterile temple]

have you used controlnet before? just want to check you have fill out the settings before trying to use it

[weak sage]

yeah ibeen using it a long time

[smoky oak]

Complete side note, just found out I am considered actually triracial, which makes me a technical minority group lmao

[oak osprey]

that's a logic error

not a user error