🏞｜general-with-images

[dense tapir]

It does now

at first no way

I am standing back and watching and really interested in the ones coming (Battlemage, and Celestial and eventually Druid).

I am surprised that the A750 beat the brand new 7600.

that gives me great hope for BM

[oak osprey]

bowel movements?

[sterile temple]

hehe BM

[oak osprey]

personally i think youre shooting yourself in the foot going to intel or amd

dunno why you hate nvidia so much that you would praise arc

[kind quartz]

to head with intel

[dense tapir]

personally I don't care what you think and besides I am not going Intel I am just watching on the sidelines with high hopes they pull it off.

[oak osprey]

intel will not

[dense tapir]

Yeah, right

You just hope they don't

[oak osprey]

you absolutely care what i think or you wouldnt be discussing any of this here

why would i want less competition

[dense tapir]

No, I don't because I didn't address you I addressed Sytan

[oak osprey]

this is a public channel lmao

[dense tapir]

anyway I am out.

piss off

[oak osprey]

if yiu want it private, use dm

[kind quartz]

would be nice if they are competitive. Unfortunaly they are bounded with certain CPUs to benefit more with intel gpu. But if they are great, they can move prices of nvidia down. So i wish them good luck

[dense tapir]

now go cry and throw another temper tantrum.

[dense tapir]

I agree

[oak osprey]

uhmmmm where is this hostility coming from

[dense tapir]

Just saying that when things don't go your way your other personality comes out and you throw a tantrum.

[oak osprey]

we are discussing gpus and you are;going to personal attacks

speaking of tantrums.

anyway i pinged the mods for you. hope they can straighten that out

[oak osprey]

v1 vs current checkpoint on v2 (in training)

[cyan snow]

also, 2.1 seems to understand cars better

[oak osprey]

yeah it totally does

if it starts to make a double car melded together, it tends to separate them

https://images-ext-1.discordapp.net/external/vBYZH1c559ZSXHdctN2Q8c0M74CrxHVsftmkAU4Vlh8/https/tripleback.net/public/discord/1684879063.6033475e9f754b7d6718ff3cf0069db0da6350.png?width=1267&height=844

[cyan snow]

also when does the SDXL model release?

[oak osprey]

clearly two cars

it's 50% trained, homie

probably next month

[cyan snow]

hype

[oak osprey]

they might restart from scratch again but at this point they seem to be past their experimental stage and are into fine-tuning

they have also implemented some new research papers for terminal SNR based noise gen instead of offset noise, to improve outputs monumentally

that was like, a week ago

[cyan snow]

just the base sdxl model seems to top most current finetunes

[oak osprey]

the base SDXL model is unlikely to be as easy to train due to that, though

like, the SDXL team is fine-tuning it to a crazy extent with user preference layers

[cyan snow]

yep, way more parameters

[oak osprey]

they are not releasing this one as a total garbage model but its flexibility will probably be limited for fine-tuning purposes

[cyan snow]

but it will be able to be finetuned, right?

[oak osprey]

like DeepFloyd is?

aka poorly and with little understanding of what to do to achieve X or Y

i doubt Joe Penna's team is going to tell us how to do that

they keep saying all of the fine-tuning they're doing is to make it easier for us to fine-tune later but eh

[cyan snow]

well, i think the sdxl model is going to be a huge step, worst case scenario we would just make more loras

[oak osprey]

i will do an experiment sometime and try to reproduce Pseudo Journey v1 on SD 2.0 instead of SD 2.1

and then we'll know

2.1 is a fine-tuned resumed checkpoint from 2.0

[cyan snow]

pseudo jorney?

[oak osprey]

my current model i have published

#🏞｜general-with-images message

a comparison against the next version which is training atm

[cyan snow]

i see, wont you wait for sdxl to start mass finetuning?

[oak osprey]

why would i? 2.1 is here now and i wanna have fun. i don't want to wait for SDXL to come out and find out i have to wait for tools to be developed for its fine-tuning and then see that they're not ready for use yet etc

if anything i'll do the same fine-tune on 1.5, 2.0, 2.1, deep floyd, and SDXL

and just maintain my sort of aesthetic on any one of those models.

[cyan snow]

ah, i get it

good luck m8!

[oak osprey]

oh i don't need luck at this point i know what i'm doing and how it'll turn out at which parameters. that's after burning 300 models or so this month

[wispy spindle]

keep it nice, @dense tapir. Uncalled for and needless behavior such as this doesn't exactly foster a great space. Timed out for a bit since this seems to be a running theme.
Have a nice day!

[oak osprey]

was 1.5 a continuation of the checkpoint from 1.4?

or like, a fine-tuning of it

https://images-ext-2.discordapp.net/external/KorqIYuKjEtYG6MvSaCeFPGfIiPTV31pghuGq8MKAKk/https/tripleback.net/public/discord/1684939640.8216538d849373ca87aa5971cca051a348c1eec.png?width=987&height=987

friend prompted something about swine and i think it made Congress?

[wispy ether]

Can someone generate the mark of cain

[kind quartz]

Has anybody tested if ex-post generated SBS image in depth script, is properly working? Could be tested in VR in some image browser probably.
Or is it just gimmick not working? Anybody knows?

[oak osprey]

have you ever heard of steroscopic images? 😄 It's quite cool

cross your eyes slightly so the ferames overlap

your dog image DEFINITELY works

that has so much depth to it. good job

[wispy ether]

Thats raytracing.

And fake vr

😎😎

[oak osprey]

the ball?

that is from Amiga Workbench

ray tracing wasnt really a thing

man, this next version of my model feels like it's gearing up for something amazing

lol current result for leopard gecko

are them's fuckin whiskers?

a group of men prompt is improving monumentally over what baseline can do, and also the v1 of my model, which is something i was hoping would happen with a much larger training corpus

v1

holy crap, step 3500 is something special. it has outputs that are radically different from any other

like this step seems more tilted toward realism

[kind quartz]

@oak osprey i know about stereoscopic images and movies. Just it is made different way. So curious if those made from single image are working properly as well. 🙂

[oak osprey]

yeah when i crossed my eyes it worked for me

[kind quartz]

that dog as well? In VR one dont need crossing anything. L picture for L eye R picture for R eye 🙂
Also there are two cross eye and no cross eye. It is order of Left and right images.
Thank you! It is in depth script if you like SBS images

[oak osprey]

yeah the dog

i don't have VR

[oak osprey]

https://images-ext-1.discordapp.net/external/Q8jsWNYeqhSTujZHgm3jMxbMgbZPZKQQYjmrU-_W2Gc/https/tripleback.net/public/discord/1684942715.151610635ada11f187ceae505ab1be8c07df378.png?width=987&height=987

baseline 2.1 with Karras sigmas is so clean

[wispy ether]

https://tenor.com/view/meme-gif-25383714

https://tenor.com/view/byuntear-incrives-meme-meme-react-gif-25029710

[oak osprey]

https://images-ext-1.discordapp.net/external/VjihrXqNdOcfwH0m_3uhztGAYHQU-o8X-Tu5r6UDQic/https/tripleback.net/public/discord/1684949652.00340441b9dbad1fe06276220fb025485c55b53.png?width=987&height=987

porcelain doll version of john belushi, didn't turn out how i wanted

[wispy nest]

ah, perfect :P

[oak osprey]

i wish my final model would make the child fly in the kite prompt like it does on step 6500

[neon crystal]

Ooh nice stereoscopic effect there

[kind quartz]

So it is working? I have to dust off my VR but it is quite hot for using it...

[neon crystal]

Wow that’s very cool. Haven’t tried your latest HUGGINGFACE model yet gonna do it soon.

[neon crystal]

Yep yep

[oak osprey]

7k steps now

[neon crystal]

The earthworms, once mere soil scribes, become neon streaks across a cybernetic sky, their squiggly trails leaving behind echoes of an alien melody.

[wispy nest]

I wanted one of those cool jackets with a fluffy mane around the neck area. What I got almost made me drop off my chair from laughter :P

[smoky oak]

Hey look, you made @glossy herald

[wispy nest]

If it's a good thing, then: Yay! 🥳
If not, then: I'm so, so sorry! 🫣

[smoky oak]

He loves cats lol

[wispy nest]

here a, I think anyway, normal one :P

[kind quartz]

expected something bit different. Yes i should use pilot instead of racing...

[woven dawn]

Guiz has been away for a while. P sure he got slapped on the wrist for having an alt account and perusing through the nsfw animeai server

[smoky oak]

If thats true, thats BS

[wispy spindle]

?

stop trolling lol

[neon crystal]

Ravana's penance in a food bowl.

[smoky oak]

ok, that is at least assuring that its not true lol

[wispy spindle]

trexdel is kinda known at being the server's Resident Goofy

it is endearing in its right hahaha

[smoky oak]

Made me worried my favorite mod had gotten in trouble

[oak osprey]

7500 steps vs 0 steps

[woven dawn]

[smoky oak]

@dense tapir

Did you happen to see the new Nvidia driver announcement for SD?

[oak osprey]

https://images-ext-1.discordapp.net/external/ky7LRaaLsXWOpMldGw7QH6sfqwvaZYddG-oSulucd8w/https/tripleback.net/public/discord/1684961510.9688792f21cd3ea8bc175430bece90f6eae4123.png?width=987&height=987

man this is the worst image i've made so far from my model

i don't think it got a single thing right, but i used 'in the style of arnold palmer's artwork' and i don't even think that's a real person let alone an artist it would know

[smoky oak]

Does anybody have any info on SD models with Olive support? I really wanna try this out

@oak osprey wanna try something fun? I have a task for you :p

Apparently there is a way to apply olive optimizations to existing SD models

[oak osprey]

i don't want it olive toned, wtaf lol

is that some new hot shit

everything being green?

https://images-ext-1.discordapp.net/external/BqnvT3g6a5WkVN2NseXhNPQPG8MpYMfRwpR9GggfkME/https/tripleback.net/public/discord/1684961850.309922cf67fdf162260c436cafa27d5ad07d49.png?width=987&height=987

i tried to spruce sad keanu up

[smoky oak]

No, Olive optimization lmao

Nvidia just added it to drivers today, and apparently it's supposed to show a monumental performance increase on models that have been optimized for it

(the paper says some cards and models saw a 6x performance increase) and it was around 70% on average

[oak osprey]

idk waht that is

[smoky oak]

https://devblogs.microsoft.com/directx/optimize-directml-performance-with-olive/

[oak osprey]

god damn it. it needs a new library? fuck that

[smoky oak]

So let me get this right

You're turning down up to a 6x performance gain cause it needs a library? Wow lol

[oak osprey]

it's a Windows thing for DirectML

[smoky oak]

Oh right, you're not on Windows, nevermind

I'll see if I can somehow figure out how to run it

[oak osprey]

like, ONNX has been a thing for a while now

it hasn't taken off

idk why

but automatic1111 i see constantly still using pytorch 1.13 so, it's probably like that. new good thing, no one cares

[smoky oak]

It just got direct support in NVIDIA's game ready and studio drivers to be used for SD and other AI programs, so clearly it's getting some form of recognition

[oak osprey]

i don't think you're going to be able to get this in without radical changes for automatic1111's sd-webui

looking at their library code

[smoky oak]

A1111 uses torch 2 now, I just stick to 1.13 cause it runs better

[oak osprey]

    pipeline = OnnxStableDiffusionPipeline.from_pretrained(
        optimized_model_dir, provider="DmlExecutionProvider", sess_options=sess_options
    )

it uses transformers to load, like i've wanted A1111 to do forever

definitely can't make this work in there

ohhhhh this saves the compiled models

that's i think what General Awareness was talking about with SHARK

[smoky oak]

Yeah, shark also does that

It's supposed to be way faster

@oak osprey it comes with its own python script for running SD, and you run it in there to reap the speed benefits

[oak osprey]

YESSSSS

FLYYY

well its nice to see windows getting a 6x speedup since it has been 20% slower for some time now

[smoky oak]

I make and consume a lot of parallel view images

its not the exact same as what you just send, as those are crossed view images

[smoky oak]

https://tenor.com/view/salutations-salutations-my-good-chum-seal-the-creature-lovethisseal-gif-24195290

[feral pebble]

anyone know what happened to the controlnet preprocessor preview?

the preview area is above, but no preview button

[static tusk]

the little explosion icon

[feral pebble]

@static tusk thanks bud!

I'm making a birthday "card" for a friend of mine. made the image, then made some text (not shown here since it has his name in it) in inkscape and curved it.. and am putting it on a "banner" held up on sticks I made in blender.. with some scribbled in "grease pencil" grass..

then controlnetting it back in SD again to make it all uniform and perdy

[oak osprey]

ugh paperspace is having issues again

[oak osprey]

sdxl lol

[static tusk]

neckXL

[oak osprey]

it's clearly Slinky Donks XL

[wispy nest]

is XL out or is that from the official site?

[wispy nest]

cinematic portrait of a woman with ice-blue eyes in an Aztec outfit. Her gaze is intense, her features strong yet delicate. The outfit is richly detailed, the background subtly blurred. Lighting is soft, highlighting her features and outfit. The image is hyper-realistic, with a cool and warm color palette

[tired peak]

is there some easy way to get SD 2.1 on automatic 1111?

[kind quartz]

it is matter of model only, isnt? Download model trained on 2.1 on civitai for example @tired peak

[wispy nest]

if you want photo-realism don't use realistic word, that doesn't really work, your example is a good example for that.
It looks like game graphics or art

photo, raw photo, 35mm (or whatever mm), is better option

[tired basin]

interesting take - will have to investigate this...

[wispy nest]

well, at least I got better results this way on my models when I was testing it

"realism" - realistic looking graphics
"photo" - actual photos

[tired peak]

i dont know where to find the yaml 😦

[kind quartz]

have you downloaded it from civitai? It should be in text under button where you downloaded model.

[tired peak]

gives me this error. : RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper__index_select)

i put the yaml file in there too. both versions

[tired peak]

oops. nm this is the actual error and it goes on for 7x the length of this: RuntimeError: Error(s) in loading state_dict for LatentDiffusion: size mismatch for model.diffusion_model.input_blocks.1.1.proj_in.weight: copying a param with shape torch.Size([320, 320]) from checkpoint, the shape in current model is torch.Size([320, 320, 1, 1]). size mismatch for model.diffusion_model.input_blocks.1.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([320, 1024]) from checkpoint, the shape in current model is torch.Size([320, 768]). size mismatch for model.diffusion_model.input_blocks.1.1.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([320, 1024]) from checkpoint, the shape in current model is torch.Size([320, 768]). size mismatch for model.

i applied settings to 2.1 and that is what i get

[static tusk]

the file name of the checkpoint model and the yaml should be the same. I am guessing they arent

[cyan snow]

jar-o' space

im hungry

[wispy nest]

cinematic portrait of a woman with ice-blue eyes in an Aztec outfit. Her gaze is intense, her features strong yet delicate. The outfit is richly detailed, the background subtly blurred. the lighting is soft, highlighting her features and outfit. The image is highly detailed, with a cool and warm color palette. (raw photo, 35mm)

[wispy nest]

cinematic kinda goes against photorealism too 😄

but in this case seems like it's just not really realistic model, but still not bad.
What model is it?

[wispy nest]

going to run another process by removing cinematic

[wispy nest]

how do you control camera angles and body expressions? with the last two images they kind of look non dramatic looking straight into the cam

[wispy nest]

Model: DreamShaper v6
portrait of a woman with ice-blue eyes in an Aztec outfit. Her gaze is intense, her features strong yet delicate. The outfit is richly detailed, the background subtly blurred. the lighting is soft, highlighting her features and outfit. The image is highly detailed, with a cool and warm color palette. (raw photo, 35mm)

[wispy nest]

oh yea, dreamshaper is more of an art model, but still not bad

[wispy nest]

idk , the only angle I did on purpose was high angle or high angle PoV
Both works

[feral pebble]

I can't get a prompt to do what I want. :/ "A blue whale leaping out of the ocean with a cowboy riding it. Saddle. Beautiful Island in background."

posted in #prompting-help. Wrong place.

[stone cipher]

Dark souls

[wispy nest]

or wide angle?

wide angle doesn't work, just tested it

trying with high angle now

What is wide angle?

[wispy nest]

supposedly panoramic with wide view

oh like high FoV thing

idk, never tried it with sd

[wispy nest]

yeah its called wide angle in terms of photogrpahy

but didn't work with sd, so lets see what i get with high angle

[astral flax]

Camera angle term recognition can be dependent on the model you’re using from what I understand

[wispy nest]

#🏞｜general-with-images message

#🏞｜general-with-images message

#🏞｜general-with-images message

#🏞｜general-with-images message
That was high angle

[wispy nest]

very likley, i didnt see any camera specificaiton in the models page tho

[wispy nest]

ohh thats great, im trying it, almost done rendering

[wispy nest]

so i got this ...

high angle photography of a woman with ice-blue eyes in an Aztec outfit. ((captured in a dramatic moment, the wind blowing her hair)). Her gaze is intense, her features strong yet delicate. The outfit is richly detailed, the background subtly blurred. the lighting is soft, highlighting her features and outfit. The image is highly detailed, with a cool and warm color palette. (raw photo, 35mm)

With that amout of wods in the prompt you'll need higher weight, like (high angle:1.3)

But might be also just model doesn't know what that means

maybe model interprets that differently, let me try with Deliberate v2

yeah

[marble wasp]

First off, that image is superb. Secondly I too have had difficulty getting various angles but have had some luck using lens terms. Like a 35mm lens, 25mm lens, etc.

In conjunction with "wide angle"

[wispy nest]

good idea about lens, will have to try that out too. 🙂

try to do few quick gens with low resolution, 512x512 and cheap \ quick sampler, like DDIM to see what which words works and which doesn't

DDIM is great for quick tests, since it's fast

[wispy nest]

ok

20-25 steps will do too

20 for a test good enough

ok im querying bard about

what is a good camera and lens for wide angle photography

going to add those detail in too with a specific cam and lens

so choosing this cam and lens sounds like a good idea

Nikon Z 6II with Nikkor Z 14-24mm f/2.8 S lens: This full-frame mirrorless camera and lens combination is perfect for landscape, architecture, and interior photography. The wide-angle lens allows you to capture more of your scene, and the fast aperture lets you shoot in low light.

but the image came out face focused

do i need to add more description to the background?

[astral flax]

yeah, lots of models are portrait face focused

[wispy nest]

here goes nothing ... 800x480

wide angle photography of a woman with ice-blue eyes in an Aztec outfit. ((captured in a dramatic moment, the wind blowing her hair)). Her gaze is intense, her features strong yet delicate. The outfit is richly detailed, (the background is a panoramic landscape). the lighting is soft, highlighting her features and outfit. The image is highly detailed, with a cool and warm color palette. ((shot in Nikon Z 6II with Nikkor Z 14-24mm f/2.8 S lens))

im not really specifying portrait, im using the word photography

but her front profile covers most of the image

[marble wasp]

i'd keep in mind, anyting below 17mm is generally considered fish eye because of the distortion.

[wispy nest]

ahh ok, will have to correc that

[marble wasp]

24mm to 35mm prime lenses are generally good wide angle lenses, but, that's photography, idk if the SD makes that distinction exactly

[wispy nest]

i changed it too 24mm as per cam model

the width and height in settings also seem to affect the render a lot w/o additional prompt

[astral flax]

I've also heard people put "x runnning" (x being woman/whateva) just so they can get the longer shots but yeah that's been my experience as well

its hard to get consistant wide angles/fisheye & everyone looks the same & no one is dirty ever 😛

[oak osprey]

CUDA out of memory. Tried to allocate 136.00 MiB (GPU 0; 79.35 GiB total capacity; 72.67 GiB already allocated; 90.12 MiB free;

😭

78685MiB / 81920MiB ah there we go

riding the razors' edge

[cyan snow]

WTH what is that gpu?? 79.35 gb vram?

why are you complaining?

[neon crystal]

Welcome back Peter

[wispy nest]

i think whenever i target facial features, like gaze, eyes, or mood, that will work counter to a wide angle shots

[stone cipher]

Thanks

[cyan snow]

A100?

[oak osprey]

yup, SXM4 version

[cyan snow]

Where do you even get that?

[oak osprey]

best buy

[cyan snow]

that's insane

[oak osprey]

and my complaint was that i can't train at batch size 24 with this without gradient accumulations

[wispy nest]

what i came up with when removing all facial features in the prompt

((wide angle photo of a woman standing before a mountain)) wearing Aztec outfit (full body shot). her features strong yet delicate. The outfit is richly detailed (the background is subtly blurred). the lighting is soft, highlighting her features and outfit. (highly detailed image) with a cool and warm color palette. ((shot in Nikon Z 6II with Nikkor Z 24mm f/2.8 S lens))

[cyan snow]

how much does that even cost? aren't those used to make SDXL?

[fresh raven]

Goku with sword

[frozen socket]

a little bit of calming Zen

[oak osprey]

https://images-ext-2.discordapp.net/external/5UkNhcuR_afcQFjyVGLCXXRzePexhKcu497Hst3HDlY/https/tripleback.net/public/discord/1685043806.006119784afa12499fb91d99f58ac06e3dcc719.png?width=844&height=844

what my pseudo-journey model does

[wispy nest]

ah ok, im using dreamshaper btw

bot is back

[oak osprey]

a gentleman doesn't ask and a lady doesn't tell

[wispy nest]

i changed the prompt a bit so she faces the viewer

[oak osprey]

it's more expensive than my A6000 and 4090 combined though.

[cyan snow]

How many it/s does that even do

[oak osprey]

i don't know, i've never bothered to measure, that's not why i got them

are you talking about during training?

[cyan snow]

then why did you buy it? to make finetunes?

yeah

[oak osprey]

that's not how training speed works

it's not directly comparable at all, if you're seeing iterations per second, you're probably creating a LoRA

this is seconds per iteration

[cyan snow]

Dude, that's awesome, this community needs people like you

[oak osprey]

thanks, that's why i am here 👍🏽

it is 9.62 seconds per iteration at full 32bit precision training, and a learning rate of 1e-8, and batch size 18 with 4 gradient accumulations, and zero gradient checkpointing. TF32 is enabled, and so on

the goal is to drive the VRAM use as high as you possibly can without blowing out an O-ring

this current run is as close as i've been able to get to 80GB without slowing it down beyond 10 seconds per it

[cyan snow]

are you finetuning 2.1?

[oak osprey]

not much point in fine-tuning 1.5, is there 😛

dead platform is dead

long live OpenCLIP

[cyan snow]

also, what dataset are you training on?

[oak osprey]

22,000 pics of my grandma

[cyan snow]

let me guess, the name will be- grandma diffusion

wait, why do you have 22000 grandma images

[oak osprey]

well now i kinda want to do that for real but i'm kidding, it's midjourney 5.1 data i've processed to remove a lot of the garbage out of about half a million total images i had

[cyan snow]

ah, i see what you're making

[oak osprey]

they're all captioned with their initial prompt which is interesting because MJ uses CLIP and SD2.1 uses OpenCLIP so i'm essentially teaching 2.1 how to prompt like MJ

[cyan snow]

dude, that model would probably be one of the best

[oak osprey]

i did 4,000 steps training OpenCLIP and the unet concurrently at a low training rate and then froze the TE and trained just the unet for another 10,000 steps, and then unfroze the TE, and now doing another maybe 3 epochs of training the TE

hence the huge vram use and huge batch size, trying to slow down the learning rate even more while providing the optimizer with more context in one shot. at least that's my theory

i need to ask GPT4 how to freeze specific layers

[oak osprey]

it's so good i've stopped using Artius v2.1 and i don't remember the last image i generated on a 1.5 model

81W / 400W | 78977MiB / 81920MiB | 100%

weird that the SXM4 A100-80G uses just 81 watts when it's pegged at 100% during certain parts of training

[cyan snow]

when will you release it?

[smoky oak]

It's almost all VRAM load, and very little die load for such a powerful GPU

[oak osprey]

the v1 is already out but it's not really useable for A1111 users because that tool sucks

[cyan snow]

what other tool is there?

[oak osprey]

i don't know, i wrote my own into a discord bot

[smoky oak]

It's really unfortunate cause his 2.1 fine-tune is super solid, but not usable by most people

[cyan snow]

wait, so it is impossible to run that model in sd?

[smoky oak]

In A1111 or any other common UI, yeah

[oak osprey]

InvokeAI works with it

-_-

they switched from ckpt files to Diffusers

[smoky oak]

Maybe I'll give that a shot for a bit and then delete it when done

[oak osprey]

i've never used it so it might just be a huge waste of time

[smoky oak]

Maybe

Invoke AI is childproof SD from what I have seen in the past

Not something I would continue to use, as it's not nearly as supported for feature rich

As much as I dislike A1111 and the dev, it still is probably the best option for GUI in terms of support and features. I wish I could have given that to Vlad, but he's considerably worse IMO

[cyan snow]

what did automatic1111 do?

[oak osprey]

he steals code and violates licenses left and right, calling it "borrowing"

[smoky oak]

He's just a terrible dev, bad at coding

Crappy person

[cyan snow]

oh, is he the person responsible for the NAI leak?

[frozen socket]

there are some people who would say that that sentiment applies to the whole accquisition of images used for training the datasets initially ;o)

[oak osprey]

i would agree with that, Winston

[wispy nest]

using images wouldn't really be stealing in my eyes as stealing implies it's gone from the owner and now in the hands of someone else. But yeah, there's some issues with them still. :/

[oak osprey]

my code may be garbage but it's all mine

[frozen socket]

except for the bits you googled or chatgpt helped out with ;o)

meanwhile, lets see if I can get some more like this

[wispy nest]

these prompts are pretty specific to cam angles and im seeing them in effect

low angle shot, high angle shot, dutch angle shot, tilted shot, bird's eye view, worm's eye view

something like

dutch angle shot of a beautiful young woman
looks to be working good on all

[smoky oak]

Some models have even more camera control

I've seen some models with camera angle, focal length, and aperture control. Cool and useful stuff

I even have one model that has 104 camera models across 12 brands, and specific photographer styles

[wispy nest]

from my experience those things don't work, or rather, they work just as well as other prompts I've seen people use. Though, my experience is also months old from all sort of updates :P

[wispy nest]

sounds cool, im trying to stay with one model for learning purpose, but eventually will try out other models

chose to download dreamshaper since its talked about a lot in the community

[smoky oak]

I'd recommend spreading your learning across a few popular ones, as they are rarely well made all around, and you may learn bad habits to combat specific issues in them

Dreamshaper is one I have never touched, on the account of their messy licensing

[wispy nest]

i downloaded a few specific ones based on their ratings and popularity

tryng not to get too distracted with too many options since i started playing with this tech a less than a week lol

[smoky oak]

The biggest one I would stay away from is chill out mix. Horribly unethical

[wispy nest]

ahh these are what i currently have ...

deliberate
realistic Vision
protogen
dream shaper
majic mix realistic
those pretty much cover my current level of interest

[oak osprey]

i just used the documentation for each of the libraries. literally in VSCode you can eg. ctrl+click into the asyncio library and read how it works. i've done development for 2 decades now

i get that you're just trying to be edgy, but it's kind of tedious

just because "other people steal too" doesn't give A1111 a free pass

that mindset is reminiscent of someone who contributes nothing to society. because you create nothing, you own nothing, and you expect others to own nothing that they create. but your wishes don't override the rights of others.

[frozen socket]

#confusedoldmanhere: who was that directed at ?

[oak osprey]

you, the one who writes whataboutism edgelord comments with that ";o)" emoji at the end, which indicates that you know you're being an edgelord

[frozen socket]

sigh how things have chnaged since 1995. Indulge in a little light hearted teasing and it all gets taken so seriously ;o)

it was a tongue in cheek comment

[wispy nest]

I read the emote as a way to identify the text as a joke :)

[oak osprey]

idk how you expect people to respond to lighthearted jabs at their intellect

[frozen socket]

TY, there was a world before animated gifs and snowflakes :p

[oak osprey]

nice dogwhistle, bro

[wispy nest]

I always read something a person say on the internet as a joke. It's easier that way ;P

[oak osprey]

man, you're making this an uncomfortable place to be right now

please stop

[wispy nest]

I actually thought the :P was aperson's eye below a baseball cap before someone told me it was a tongue and eyes. :P

[frozen socket]

Danger danger Will Robinson, sense of humour (spelt correctly) failure ahead 🙂

drat that one translated into a animated one

[oak osprey]

it's something called consent and good faith, and you're just a bully

who cares about neither

you know i am unamused by your antics, and you continue them. that's what a bully does

[frozen socket]

geeez will you just just take a deep breath and chill a little

[wispy nest]

yeah, I'd also agree not to really continue if someone isn't feeling comfortable. I just want everyone to have fun. It's more fun that way.

[oak osprey]

oh it's probably more fun for him when i'm not into it

[frozen socket]

It's interesting how much the internet has evolved over the years (and this is a general observation). IRC and Usenet was a Wild West and you needed thick skin. Also Real Life has chnaged , back in the day if a young kid was in a shop crying you would have no hesitiation in going up to help them, these days you (especially as a man of a certain age) you wouldn't dare help for fear of being reported.

Worlds gone mad I tell you

[oak osprey]

this is a learnable moment for you and you're missing the mark

[frozen socket]

who's Mark?

[wispy nest]

And I just want everyone to smile and have fun. Don't need to be longer than that :D

[oak osprey]

so when someone's not liking your jokes, maybe they're not in the mood, maybe their dad's just been diagnosed with cancer and they don't want to put up with you the answer isn't to call them a snowflake and refer to a tampon being shoved inside them, you know, giving them a graphic visual of being violated by you. it's maybe to just be quiet or move along, be supportive and kind. is that hard?

@limber remnant can you weigh in on that

[frozen socket]

the two before couldnt be posted as has exposed bits

[frozen socket]

Ive been in a car accident and suffered 30% burns, my first wife died in my arms, my current GF is recovering from Breast cancer

[oak osprey]

oh, @tropic shell is around now too.

[smoky oak]

I passed my permit test!

[frozen socket]

I'm sure there will be plenty of people around 🙂

[oak osprey]

oh god we're all that much more unsafe

[frozen socket]

Permit for what ?

[oak osprey]

congrats

[smoky oak]

Yeah, those models are fine, if a little outdated by now. I just don't like chill out mix cause it's made for literally making loli/CP kind of stuff, and it's disgusting

[smoky oak]

Driving, which I am not too excited for lol

[oak osprey]

Deliberate really messed up the text encoder

[smoky oak]

Deliberately messed up

[frozen socket]

driving is fun 🙂

[smoky oak]

I have 0 interest in driving, and I know it's just gonna be leveraged for other people to get things out of me lmao

[tropic shell]

what about me?

[smoky oak]

Whut

Also, did you guys see that new SD to vector addon that does what adobe and firefly were flexing about? Lol

I wanna mess with it

[hallow cape]

Is it good for all models or smth?

[smoky oak]

Should be

[oak osprey]

i was told to shove a tampon inside by @frozen socket because i asked them to stop bullying

[smoky oak]

I am not sure if it can do color, but still

[frozen socket]

oh its nothing, I've just been accused (falsey) of being a bully

[tropic shell]

I'm around....a round shape. Now, let's all go for a nice round of being nice to each other.

[smoky oak]

This addon could be used for 3D printing though, which is cool

[wispy nest]

aren't those normal ai images? Or did I miss something? :O

[smoky oak]

They are PNG's turned into SVG's, or vector images

[oak osprey]

vectorised

[wispy nest]

ahhh, if they are turned into SVG's then it's a whole othe...kettle...of....fish?

[oak osprey]

kind of yes kind of no

[smoky oak]

What lmao

[wispy nest]

I thought they were normal PNG :P

[oak osprey]

doing SVG conversion efficiently is hard

svg is scalable vector grahics, it is not a bitmap

it is a bunch of definitions of shapes (which are just lines) and lines

[neon crystal]

Leaving this ridiculous Bollywood thing here to lighten up the mood

[wispy nest]

yeah, vector stuff is awesome! :D

[smoky oak]

Bruh lmao

[oak osprey]

is that Amitabh?

[smoky oak]

I could utilize controlnet to make a new iteration of my logo, and then vectorize it, potentially

[frozen socket]

I like a little bit of Zen

[neon crystal]

Nah I was just fooling around with some South Indian names there

[frozen socket]

helps me chill out

[oak osprey]

एलजीटीएम 🙂

[wispy nest]

I thought vector art was just normal ai images made to "look like icons/vector thingies" :P

[smoky oak]

It's a style of art, but this makes them actual vectors

[wispy nest]

yeah, and if they do it in a good way, then it's going to be a game changer I believe :D

[smoky oak]

My little goober

[oak osprey]

it can result in an art style but i don't think it is one on its own

[smoky oak]

Funky little man lol

[oak osprey]

maybe it's a johnre

[smoky oak]

It's common for artists to go for a "vector compatible" art style, which is what I am going off of doeectly

*directly

[oak osprey]

i know those are english words. i even know some of them. but i'm not used to them in that order

[wispy nest]

I can only make normal cat ai art instead of taking a photo of a real cat >:I

[smoky oak]

Lol

[frozen socket]

I rmeber when Asteroids and Tanks and Tail Gunner first came out and used Vector Graphics on specia;ist monitors

[smoky oak]

Yeah, I think vector graphics and displays are awesome

[frozen socket]

a man of infinite taste 🙂

[smoky oak]

I really wanna get into electronics at some point so I can do a double LCD OLED like mini clock on my PC

[frozen socket]

could always buy a cheap wristwatch and blu tack it to the case ;o)

[oak osprey]

@smoky oak my meme

no stealing

[smoky oak]

I do have a Samsung OLED smartwatch, and I have it with an always on clock, but I am worried I will burn the OLED, so I think double layer LCD is the go, and then use some form of directly powered USB light behind it

[smoky oak]

Lmao

Good meme

[oak osprey]

its funy becuz its troo

😂 😂 😂 😂 😂 😂 😂

[frozen socket]

or just look at the RH corner of your monitor . TBF although I wear a watch thats what Im more likely to do(or use my phone)

[smoky oak]

I custom designed and 3D printed my own stand for it

[frozen socket]

sweet

[smoky oak]

My monitors are always in Fullscreen, and I am a notorious loser of time lol

[wispy nest]

and I just duct tape'd my headphones back together :P

[smoky oak]

I just don't want my little OLED display to burn in too much, so I use it at low brightness

[frozen socket]

It was Vista wasnt it had the always on top clock you could add as awidget

[smoky oak]

I also used my 3D printer to fix my studio headphones haha

[oak osprey]

that doesn't change the brightness for real, it just uses pulse width modulation to decrease the frequency of the power

i know this because OLED monitors suck the same power down at full or dim, which makes them not an option for my solar use

[wispy nest]

I would probably do the same, if I 1. knew how to 3d print, and 2. had a 3d printer in the first place :P

[oak osprey]

i should say traditional LG patented OLED stuff. i don't know about mOLED or any of the new stuff

[neon crystal]

Signing off now. Have fun.

[oak osprey]

mLED was disappointing, seems like LCDs are the pinnacle of power efficiency

[smoky oak]

I remember when neps dumb ass was saying that mini LED was inferior to OLED lmao

[oak osprey]

it depends on your use case

[frozen socket]

LED watchs with Red Digits that you have to press a button to see the time were pretty power efficient ;o)

[oak osprey]

there's things to say about scale eg. a micro display on some arduino gadget, a wristwatch, a cellphone, a laptop, a TV, a digital signage display that sits in a window of a shop, a roadside bulletin board, etc

i love my OLED Switch and the Steam Deck, but i wouldn't do OLED for my laptop because it just increases power use 24hrs and i didn't see much benefit when i tried an OLED laptop

[wispy nest]

three cam angles that work fairly well in the prompt below

tilted, high-angle, bird's eye view of a lion cub in a jungle

[oak osprey]

@wispy nest a good one to try is uhd

[wispy nest]

and isn't it technically above the jungle? 🤔 🙃

[wispy nest]

UHD to replace those 3 angles?

[oak osprey]

for more clarity

tilted, high-angle, bird's eye view of a lion cub in a jungle, uhd

[dense tapir]

It would appear that in Docker (iow non ROCm until 5.6.0) the 7900XTX performs like a 3080 in SD in Automatic1111 at 15-16it/s.

non optimized of course

[wispy nest]

trying it out 🙂

[oak osprey]

praying to RNGesus for you

[dense tapir]

@smoky oak Isn't yours a 3080ti?

[smoky oak]

3080 non ti

[oak osprey]

aka 3080

[smoky oak]

When I was OCd I got 18.3it/s

[dense tapir]

non oc?

[smoky oak]

15

[dense tapir]

okay

[smoky oak]

Closer to 16

[oak osprey]

every time i hear people worrying about it/sec i'm like that train guy in Back to the Future part 3. "Good lord! 90 miles per hour? Why in tarnation'd ya ever need to get anywhere so fast?"

[dense tapir]

Then, 7900XTX isn't bad at all then

I can imagine once it gets full on ROCm

[oak osprey]

it/sec can only be compared on the exact same models/pipeline etc

[wispy nest]

@oak osprey think i will use that technique with realistic images .. uhd would be good for image quality, what im doing now is experimenting with angles for better control on the subject and view

[oak osprey]

@wispy nest try looming overhead

towering over

generally worked for me to obtain a more downward look at a lower camera

[wispy nest]

are those terms used for photography or expressions?

[oak osprey]

it's more like a literary thing, i'd guess

[wispy nest]

ah ok

im kinda going by terms

[oak osprey]

you'd see it in a passage, eg. "they sat, watching; the eagle, looming overhead, aware of their every move."

[wispy nest]

yes

[oak osprey]

my point is this one is pretty prevalent in the training data tags, it works well

😛

[smoky oak]

Wabbit

Lol

[oak osprey]

wedonkuwus

[wispy nest]

or you'll get a literal loom but then you can always try again :P

those are for more advanced prompt descriptions, im learning the technical application for photography that i can use in SD

[oak osprey]

when "award winning photo" comes up with a kid holding a trophy i die inside

[smoky oak]

Just know that it's not something that can be transferred reliably between models

[proper hill]

#1072220168534642768

[smoky oak]

It all depends on the training data

[oak osprey]

unless you manually swap out the text encoder

[wispy nest]

of course

[oak osprey]

oh, sytan. i noticed. you can put a vae next to a model and name it like, model.vae.safetensors?

[smoky oak]

Yeah, you name it the same as the gen model and set it to auto encode, only issue is, 2.1 does support external VAE's outside the baked in one

[oak osprey]

maybe that's specific to a1111 again but i can definitely swap them out

[smoky oak]

At least not how it implemented in comfy/auto/vlad

[oak osprey]

yeah

but the text encoder can be

too?

because if so i think i can help you get my model running

[smoky oak]

No idea

[oak osprey]

man it's just wild to me that it does everything wth a single file by default, and loads that all into memory when it's not necessary to as a result

it's crazy inefficient and you just see people arguing against diffusers layout due to political reasons, their personal beef they have with the huggingface crew are the reason it's not implemented

[smoky oak]

Damn, I saw this dude with all sorts of Satanic iconography on him, and I wanted to chat, but he was on the phone and he had to leave before I could say anything ._.

Depression lol

[oak osprey]

that's when you walk up and mouth the words, "do you need me to get you out of that call?"

[smoky oak]

I have no idea anything about that. I just use what I easily have access to, and have had no issues with any other models

[smoky oak]

Lmao

[wispy nest]

or you could say that, "satan told me to tell you that you look very cool!"

[oak osprey]

if he says yes, you just start screaming help and stuff and he can be like "oh god i have to go!"

[smoky oak]

He was talking about work he has to do on his car lol

[oak osprey]

he sounds terrible, you're better off without him

[smoky oak]

WTF lol

[oak osprey]

i never know what to say in these situations, i'm just winging it, i'm sorry

[smoky oak]

I just wanted to compliment his pentagram tattoos

[oak osprey]

you should have then

[wispy nest]

if you did, then maybe he'd be happy :D

[oak osprey]

you're probably going to regret that for YEARS

[smoky oak]

Eh, it's not the first time I haven't been able to chat with somebody in our small community lol

Dude had the whole look

It was pretty cool lmao

[oak osprey]

i have the opposite experience that made me to stop communicating with new people. i live in a tiny surf town of 1300 people and it swells to 20,000+ in the summertime and no one i meet ever sticks around, and no one goes outside in winter, so i never meet any locals

[oak osprey]

i know these are the official numbers because they talk about it all the damn time. it's hard to miss when 20,000 new folks arrive in the same month

all summer i'm like, "just make it stop " and then in winter i'm like "i miss all the sounds and colours and people "

[smoky oak]

It's rare I get to see people who are so openly satanic in public, so I always try to go out of my way to compliment them and be the one good against a lot of backlash. Somebody did that me a long time and it stuck with me

[oak osprey]

you should join the Church

[smoky oak]

I am an ambivert. So I do really good with nobody or everybody. I go weeks with no contact outside discord lol

[oak osprey]

i mean the satan one, that help kids and teach them to be self reliant

[smoky oak]

Only reason I tend to stay away from people is cause people tend to want things or distract from what you are doing when trying to be productive

[smoky oak]

I am a member, even if I can't attend

[oak osprey]

well that's where hobby clubs come in

you can invite like-minded folks to a, say, mock sacrifice event

[smoky oak]

We don't have any churches of Satan anywhere close that I could go to lol

[smoky oak]

Oh my God no lmao

[oak osprey]

wat

why not

[smoky oak]

Last thing we need is people still thinking we sacrifice people lmao

[oak osprey]

it's a teambuilding exercise ಠ_ಠ

[smoky oak]

I already get that BS enough just cause I have pentagram earrings lmao

[smoky oak]

You are terrible lmao

I wish I was more comfortable wearing my stuff out in public

I am just not down for negative interactions like that

[wispy nest]

satan isn't that good anyway...when asking the art ai about it anyway :P

[smoky oak]

Ask for Satan in MatureMaleMix and you get hot demon daddies lmao

Now that is my antichrist lol

[oak osprey]

i thought Hot Topic fixed that

[smoky oak]

Hot topic has very little satanis stuff unfortunately

All of their pentagrams are pegan

[oak osprey]

i was in texas and Satanic wear was chic as hell where i lived.

[smoky oak]

My people lmao

[oak osprey]

i didn't realise it was so fuckin specific jezues christe

[smoky oak]

It's very specific

Pegan pentagram is point up, satanic is point down

They have very different meanings

[oak osprey]

i didn't live in a city per se, it's actually quite hard to nail down in traditional terms, what the area i lived in is like. it was over-developed, and sprawling out. but a lot of it was suburbs and highways and shopping centres. it was a weird place. it is humongous by Canadian standards

[smoky oak]

We live in a vast sprawling slum of sorts

Lotssss of people, not much worth while

[oak osprey]

like Houston is absolutely a city, and if you were to take a random Canadian where I grew up in TX, they'd say, "this is a city", but it's not one

[smoky oak]

We are a several city conglomerate, in a sense that we don't have any separations

[oak osprey]

is that dallas fort-worth area?

friggen waco texas

[smoky oak]

Total pop is probably close to 400k I'd say

[smoky oak]

No, we are in cali

[oak osprey]

oh

[smoky oak]

Bro

I

[oak osprey]

i don't fuckin keep track of all your self-doxing

[smoky oak]

I can't believe how annoying some people can be lmao

This dude next to me

Has headphones, he just hit a puff on his vape and is now smoking a cigarettes, and he turned his headphones off to start blasting music on his JBL

Like dude

Can you be any more intrusive

And of all music, he's playing Drake

I guess it comes with the stereotype lol

[dense tapir]

Oh, lordie

[smoky oak]

Well shit, he just got up and helped a woman in a wheel chair get through the door, so maybe he's not as bad as I thought lmao

I didn't see her or I would have helped, but he beat me to is so I mean

[dense tapir]

Never judge a book by its cover.

[smoky oak]

I didn't think he's a POS or anything, just annoying lol

You can be a good person and annoying

[dense tapir]

old cliche' proven right so many times

oh, 100

[smoky oak]

The amount of times people have said that when they found out that I am either gay, or a satanist, or both is insane lmao

It's just that "oh... Well then"

I just look like a generic straight passing white guy, so I get why people think I would be lol

But the second I mention having the devel AND a man in my bed, suddenly there is a problem lmao

[oak osprey]

why is the developer getting involved

[smoky oak]

Who huh what

[oak osprey]

the devel and a man

[smoky oak]

Shit lol

[oak osprey]

OH

DEVIL

[smoky oak]

I messed up lol

[oak osprey]

i was so cornfused, man

[smoky oak]

Brain was going too fast l

[oak osprey]

https://images-ext-1.discordapp.net/external/nyFB01VtthmiKEi_JGWbARWs1sY200NS2dQQopvfJJc/https/tripleback.net/public/discord/1684908717.3577204ebead83a65dc0644588ed5c1f848e5d4.png?width=987&height=987

taste the cornfusion

[smoky oak]

Cause people are always "you're sleeping with the devil"

Precisely

But for now

[oak osprey]

https://images-ext-1.discordapp.net/external/3EVHexvXV72ZTaj4G38Fh9nxN0kN1vs_CdDGAfW5N9g/https/tripleback.net/public/discord/1684908717.38452723fa981e493380230e263892b3ff36d74.png?width=987&height=987

[smoky oak]

Have a shrimp pic

She's looking quite nice today, I'd say

[oak osprey]

cornhole comes up with interesting results in my model

[smoky oak]

Ummmm

Lol

[oak osprey]

that's that pic ^^

i guess it's like a hobbit hole but for corn

[oak osprey]

https://images-ext-2.discordapp.net/external/FDbKB5jHY6HI8FBPfB7VQHDYFuZZx0f4LJu4IEha7kg/https/tripleback.net/public/discord/1684908509.1287475e59c31b13c6d24aa3fbb5cae12e20e97.png?width=987&height=987

[smoky oak]

My gato is suave

I'm so bored lmao

[smoky oak]

I have to wait for my mom to pick me up lol

[smoky oak]

HOW ABOUT NO

[dense tapir]

Hehehehe

[oak osprey]

https://images-ext-2.discordapp.net/external/hHcbDZ66nDuj5IvOZWLdgxPrsFXsfDGB9rGmCkFgUj4/https/tripleback.net/public/discord/1684798082.8238342b96028a5eff0f022722b98df879e89dd.png?width=844&height=844

[smoky oak]

Oh!

I have internet again, I can use SD remotely now, I forgot

[dense tapir]

btw, on the 3080 non O/C what is the training it/s like?

[smoky oak]

I did it at various BS

[smoky oak]

At BS 4 with no gradient, it was about 1.7it/s

At BS 22 with gradient it was 3.1s/it

[dense tapir]

512x512 though, right?

[smoky oak]

Yeah

Haven't done higher

[oak osprey]

i'm training at BS=18 and gradient=4 and it's 8s/it

[dense tapir]

BS1 no grad has the same speed at 768x768 on the T4

[smoky oak]

You're doing dream booth tho

[dense tapir]

1.6it/s

[oak osprey]

[smoky oak]

Oh- lmao

[dense tapir]

probably because the T4 has 15gb of vram on colab

irl it has 16

[oak osprey]

haven't you tried the M5000 GPU on the free tier Paperspace?

[wispy nest]

a tip: don't ever write fish donut too close together! 😱

[smoky oak]

I still really want somebody to make a model or LoRA that can do super cool cyborgs and other synthetic life forms

I would, if I knew where to get a dataset

[smoky oak]

Dude is listening to a house/trance remix of Gangnam style

What is happening lmfao

I wanna train a LoRA on the NFL sports robot, cause he is too attractive for literally no reason lmao

[wispy nest]

sadly, my android images are made so long ago that they aren't that good anymore. This was in the middle of december

[smoky oak]

i am convinced the designers of this mascot knew what they were doing, cause they made him needlessly hot lmao

[dense tapir]

@smoky oak

[smoky oak]

What's that?

[dense tapir]

Someone showing their card non rocm

7900xtx

[smoky oak]

15.7it/s

Not bad

Oh well, not great either for $1000 lol

But it's not bad

[dense tapir]

will be almost double with rocm when they are done they said

[smoky oak]

Awesome, looking forward to it

[dense tapir]

5.6.0 when they finish the damn thing

I suspect 25it/s

7950xtx (if they do it) around 30

[smoky oak]

I am still hoping for broader adoption of tools like Olive and SHARK for better performance

[dense tapir]

I hate shark no matter what

[smoky oak]

You hate everything lol

[dense tapir]

I don't need TBs of models it compiles

no, I hate its implementation as they have said they don't care about size only speed for now.

if I change one dimension it has to recompile and save it

[smoky oak]

I mean, make it work, make it fast, make it efficient. Is that not the dev pipeline for most ML?

[dense tapir]

apparently not :/

[smoky oak]

That's the pipeline they are following it seems

[dense tapir]

no, per the devs it is not space efficient

[smoky oak]

Yeah, they are on fast, which is before efficient lol

They are focusing on fast

[dense tapir]

yes

[smoky oak]

Work. Fast. Efficient.

[dense tapir]

it is alright if all you do is at one dimensional set using the same model, and loras

change any of that and it has to recompile until you finally run out of drive space

[smoky oak]

Dude is smoking again

🤢

[dense tapir]

I don't think that can ever be done differently

[smoky oak]

It's blowing right into me bruh

[dense tapir]

I don't like the smell of vape smoke and it chokes me up

[smoky oak]

No, he's actually smoking

[dense tapir]

hit with this big ass cloud

[smoky oak]

His vape doesn't have a smell at least, but his cigarettes do

[dense tapir]

oh, nasty. not as dense but ugh

[smoky oak]

I'll take vape over cigarettes smoke any day. Cigarette smoke makes you stink

Oh, he's leaving!

Whew

[smoky oak]

He's gone

[dense tapir]

Hippy vibes

someone left the lights on

[smoky oak]

My data connection is so slow here lmao. It's taking like 2 minutes per gen to load the images

[dense tapir]

damn

I really like the 1970s hippy vibes one

[smoky oak]

Oh God, the crossed eyes tag in this model does some cursed stuff lmao

One eye looking up, the other looking down lol

[smoky oak]

1.2 seconds to gen an image, 15 seconds to see it

I have found a new bottleneck haha

Luckily I shouldn't have to be here for too much longer

[dense tapir]

I wish you would gen an image 768x768 and do DDIM step 30 and Euler_a. That would be a good set of numbers.

[smoky oak]

@dense tapir he's back again, smoking, and having a phone conversation on speakerphone

[dense tapir]

he is VERY annoying and self centered

[smoky oak]

Ok, he got picked up

My mom should be here any sec

[wispy nest]

dope

[dense tapir]

Thanks

[smoky oak]

Ok, I am finally free from sitting on that cement bench lmao

My ass will never recover

[wispy nest]

I think I've seen model with that purpose somewhere on civi some time ago

[oak osprey]

https://images-ext-2.discordapp.net/external/WDFUxON2cG7E1rGx28j6xsHizPpmptSu9rFG46DXdnI/https/tripleback.net/public/discord/1685055933.62588a07e4c75fc75a363c0e67c358ce7f6d7.png?width=844&height=844

lmfao

i did it

[smoky oak]

Lnaooo

[smoky oak]

👁️ 👄 👁️

[oak osprey]

perfect text

https://images-ext-2.discordapp.net/external/yghH62rx-yoCBxg3-KJIHffQCCmLFg6LzS7GyzIPZJs/https/tripleback.net/public/discord/1685055933.3613257d17d890a2ba80ac5dd35cee4ff5b8bac.png?width=844&height=844

every time

https://images-ext-2.discordapp.net/external/Mgvh3Hhb6QRhgY6GphSRSwnjkFF_IkELK1rW0HteYbA/https/tripleback.net/public/discord/1685056059.1519487cb52d7da20ef920f8868664946b4cab4.png?width=844&height=844

SD proves i was right to hate clowns as a kid

[smoky oak]

MIDDŌJORAEEY

[oak osprey]

only CD$VVRENY a month

https://images-ext-1.discordapp.net/external/AXrKiflX5TMmiuyZOzyd7n87dSbEDBPZZBsm7UGlbw4/https/tripleback.net/public/discord/1685056156.040829737e7e2cdf26aa59f4fbd84b9b35c5576.png?width=844&height=844

lol why does this happen when i prompt 8675309++

they be confusie

https://images-ext-1.discordapp.net/external/xo4S8sfJXgCBMgCujzXVJnrZYdidxwZJPF-YLvTQgSs/https/tripleback.net/public/discord/1685056219.8113344e6eac22b858477e9055e5cdc4615b3f7.png?width=844&height=844

it's not a seed thing

https://images-ext-2.discordapp.net/external/4Ud9Fjz0NO87SJE-b27R1lsr3_pc3K0xC9__v0bmScw/https/tripleback.net/public/discord/1685056277.8783295d5c6d0c1f01deced765965a966c92407.png?width=844&height=844

jenny 8675309 no wonder he wrote that song about her

https://images-ext-1.discordapp.net/external/9GZfb9KLUU7FzhY2h2CTyHJlufctxiuFj9dNExm7lT8/https/tripleback.net/public/discord/1685056277.632311324739bf6cfabe45a310153761019fca2.png?width=844&height=844

https://images-ext-1.discordapp.net/external/fyNmrvN8OJogUinAPKZCbTgnMVJ1YQgFW1zFe9TFFeo/https/tripleback.net/public/discord/1685056758.79890565c9f20e01221c5e40f73bf40018592c.png?width=1267&height=844

when i try to make hindustani joggers it always looks like the hunger games

[cyan snow]

well, when you will release the final version, will it support A1111?, i tried it on invokeAI and it did not work.

also sorry for delay

[oak osprey]

i have to use a new version of diffusers for it to work on my 4090

idk if that helps or not.

i can't do much with "did not work", i need an error message of some kind

i doubt my models will ever work with A1111, i can not get it running locally and no one there cares to fix the issue because none of them have the issue on their hardware

[cyan snow]

\invokeai\models\hub\models--ptx0--pseudo-journey\snapshots\a1f9f4835c378c1549220110c90f08ac201b213c\vae because the following keys are missing:
decoder.mid_block.attentions.0.query.bias, decoder.mid_block.attentions.0.key.bias, decoder.mid_block.attentions.0.value.bias, decoder.mid_block.attentions.0.value.weight, encoder.mid_block.attentions.0.key.bias, decoder.mid_block.attentions.0.query.weight, encoder.mid_block.attentions.0.proj_attn.weight, encoder.mid_block.attentions.0.proj_attn.bias, decoder.mid_block.attentions.0.proj_attn.weight, encoder.mid_block.attentions.0.query.weight, encoder.mid_block.attentions.0.value.weight, encoder.mid_block.attentions.0.key.weight, encoder.mid_block.attentions.0.value.bias, encoder.mid_block.attentions.0.query.bias, decoder.mid_block.attentions.0.key.weight, decoder.mid_block.attentions.0.proj_attn.bias.
Please make sure to pass low_cpu_mem_usage=False and device_map=None if you want to randomly initialize those weights or else make sure your checkpoint file is correct.

this is gibberish for me

i bet that model can be extremely good, but i couldn't get it to work unfortunately.

[oak osprey]

that's because you're on a broken version of Diffusers 😦

pip install git+https://github.com/huggingface/diffusers

this will fix that

[static tusk]

its looking for the VAE. see the first couple lines

[oak osprey]

trust me, this is a Diffusers problem

i just fixed that on my 4090 box because i hadn't started it up in a while

[static tusk]

so you got A1111 to work on your machine?

[oak osprey]

no lol

[cyan snow]

this worked =], i will now give it a try

[oak osprey]

@smoky oak ^

use InvokeAI with new Diffusers from git 🙂

[smoky oak]

I might

[oak osprey]

it lacks a bunch of features from A1111 but you won't use those anyway, i think, eg. dreambooth

[cyan snow]

ok, it did all sampling steps, but when finalizing the image, it says: TypeError: XFormersAttnProcessor.call() got an unexpected keyword argument 'temb'

[oak osprey]

weird, i uninstalled xformers recently because i'm on pytorch 2 now

[cyan snow]

wait, you can switch to pytorch 2? am i dumb

[oak osprey]

you on 1.13?

pytorch2 for some GPUs uses less memory and is a lot faster but for older GPUs might not be

[cyan snow]

i have a 4070ti, so i guess it should be better for me

also i have obsoletely no idea what pytorch version im on

[oak osprey]

it should be a lot better

how do you like that GPU so far?

is it a furnace?

[cyan snow]

nah, temp is good, it's an AUROS ELITE. decent cooling

[cyan snow]

fixed it, generations seem great, also cfg scale need change since it seems to behave differently than usual models.

[oak osprey]

lower or higher?

i keep mine around 9.2

[cyan snow]

1.5 finetune

pseudo-jorney

same seed and settings

[oak osprey]

it's pretty good at that kind of stuff

less good at photoreal people

good at paintings though

i typically add UHD+ 8K+ Sharp+ to the end of all my prompts with it

[cyan snow]

im not sure which i like better, but this is a good model over all

[oak osprey]

if you add the word midjourney to the prompt it should radically alter the output too

when an image in the dataset had an incoherent prompt i just replaced it with that keyword

so it's like a wildcard what the hell it'll do

[oak osprey]

thanks man, good to hear that. it's been (don't say a journey, don't say a journey, don't say---) a journey

the Karras sigmas from k-diffusion are amazing and make it much better too, if InvokeAI has that

[cyan snow]

already using it

this is with midjorney in the prompt

[oak osprey]

i did not expect that

i kinda like it more than the other one

it looks like one of those sand art things

[cyan snow]

also, InvokeAI seems to have different architecture that A1111, or is it just me?

i think it does upscaling to make generation time quicker

[oak osprey]

it's using diffusers, which is more efficient in general

can just run torch.compile(model)

i don't think A1111 has quite got that figured out yet

[cyan snow]

yeah, if invokeAI had as much features as A1111 it would defiantly be way better

[oak osprey]

those can be added, lol just have GPT4 help you write the PR

[cyan snow]

if it was that simple it would have been done ages ago

by ages i mean a few months ago

[oak osprey]

idk i only heard of it a few days ago

[cyan snow]

about gpt 4?

[oak osprey]

no

InvokeAI

[cyan snow]

oh damn, it exist as long as A1111

[oak osprey]

i don't tend to go looking for these things, i just write my own code to do this stuff instead

[cyan snow]

Also it hasn't changed at all since released

[oak osprey]

i have a small standalone python program that does stable diffusion gen as a discord bot

[dense tapir]

@smoky oak

[oak osprey]

it's just so difficult to extend its codebase and only runs on a single system, i rewrote it to have a client and server, and connect multiple pytorch clients to the central server. that version is more complicated, and uses a SQL db to store some configs. the original discord bot code doesn't use DB or much else in the way of external reqs

[cyan snow]

this is not true, AMD doesn't support ROCm anymore. it was discontinued about a month ago

[oak osprey]

whut lol

you mean for consumer cards

[cyan snow]

I had an AMD GPU, never again

[oak osprey]

and it still works on my Vega 56

the latest pytorch-rocm stuff they dropped, made the Vega work with SD perfectly

[dense tapir]

WRONG

They Abandoned OpenCL and brought us ROCm

[cyan snow]

This is because you didn't update ROCm

[oak osprey]

i'm on gentoo, i always update everything to latest

i built it from Git

i get like 2.83 it/sec on the vega 56

oh i never tried pytorch2. i upgraded that system to a 4090 and now the vega just sits there on the shelf

[cyan snow]

Well, then i guess my old GPU was f**ked. ever since i tried to update ROCm it wasn't able to run pytorch, even after formatting my machine a few times

[wispy nest]

what is the constant bleeping sound i hear in stable channel, and only when i switch to nothing in notification it stops, but i can't figure out which is triggering the bleeping alert noise?

[dense tapir]

6k AMD cards are loving ROCm 5.5.0 and when 5.6.0 comes out the 7k cards get support. The only issue is how damn long it takes AMD to update ROCm.

[oak osprey]

dude it wasn't quite easy to get ROCm working so i was stoked for the 4090 because i was like, yay its gonna be so ez and then i had the wrong CUDA version and 12 is incompatible and i had to figure out how to get cuda 11.8 and so on, it's a mess on both sides

[cyan snow]

ROCm is a hot mess, the main downside is it's under developed, and very incompatible, and also inefficient. the list goes on

[dense tapir]

Well, are your eyes that of a Windows user or a Linux user?

[cyan snow]

both, i have been dual booting since 2016

[dense tapir]

Nice. I just know this much the AMD people I speak to love ROCm, have no issues, 5m to set up and running SD with.

The very positive thing is that they aren't licking Jensen's you know what.

[cyan snow]

i don't know than, for me ROCm was a hot mess. nothing compared to CUDA

[oak osprey]

v2 pseudo-journey vs v1 at the moment

[dense tapir]

I don't know on windows but on Linux it is so easy to setup.

[cyan snow]

it worked only on very specific versions of linux and its kernals.

[dense tapir]

ROCm demands Linux I think

[dense tapir]

not so now with 5.5.0 they seemed to have stopped that

made it more universal

[cyan snow]

ONLY linux, and ONLY with specific versions of linux

[dense tapir]

Oh, and Pytorch2 directly works with AMD now

[cyan snow]

huh, but still, they defiantly lost the software stack battle.

Nvidia is destroying them

[oak osprey]

this one improved a lot

[dense tapir]

I look at it like this that the world's most powerful supercomputer uses AMD for ML. I am not giving in to Jensen's extortion. Sure, I might not be as fast but I pay half the price and keep Nvidia away. My pocketbook says thank you master.

[oak osprey]

i like this one more too

as usual these are test images produced through the default SD scheduler, VAE, and pipeline, at 50 steps, and 7.5 cfg

so it's kind of amazing (to me) how clear the v2 model's output is. it gets much better when it is loaded with a newer scheduler

[tired peak]

there are no instructions for me. i had to search far and wide to find any yaml file in the first place. i dont know where to fidn this stuff

[static tusk]

OK. What 2.1 model do you have?

[tired peak]

let me check

v2-1_768ema

the yaml i dled is called v2-1_interference so i guess that is the wrong one

v2-1_768ema-pruned.ckpt

is the 2.1 model that i put in the models folder

[static tusk]

I wouldnt really recommend the base 2.1 model as its clunky and not that great quality. You should use a finetuned 2.1

download this model and grab the config file, the link is under the download button https://civitai.com/models/17277/realism-engine

basically the way it works it the yaml file need to be name the same as the model
In the case of your current v2-1_768ema-pruned.ckpt, the yaml should be
v2-1_768ema-pruned.yaml

[oak osprey]

it got super ambitious on this bicycle

[static tusk]

The above civitai linked model will give you much better results

[tired peak]

ty ill save the link

[tired peak]

so i just rename the yaml file? is that what you mean

[static tusk]

yes. once you download the files you will have both like this

but with civitai, they already have the correct names, if I remember

[tired peak]

is the realism engine a 2.1 model?

[static tusk]

check the details table next to the images

[tired peak]

are there any artsy fine tuned 2.1 models that you would recommend? most of my images are usually oil painting style

[oak osprey]

this is from my WIP 2.1 model update. but it's likely going to be difficult to make it run there in Automatic. no one has had success yet

[static tusk]

you can achieve oil paint styles with TI embeddings or Loras without a specific model that does that

[tired peak]

what is a TI embedding? sorry, im still learning

what i've been doing recently with Sd 1.5 and dreamlike photoreal 2.0

[static tusk]

Its smaller files that are trained on a particular style or subject without having to release a model. They can be added to your prompt as keywords to get a particular style or subject you want.

[oak osprey]

https://images-ext-1.discordapp.net/external/dzrchHrxP-oznYQmGwp2acbC22Ys9CENKWBUAbkaMng/https/tripleback.net/public/discord/1685063730.7975426db9d0df1cbb4a5ce96f6804bc8a69f10.png?width=1481&height=987

Bumblebee

https://images-ext-2.discordapp.net/external/nc-XNgBzjIyrPBxgpbkm4Pw0HxDZkme1np3a1SisNo8/https/tripleback.net/public/discord/1685063730.23336246f8970960ad0ecf09a5d3742b23ca861.png?width=1481&height=987

lmaooo

[oak osprey]

https://images-ext-2.discordapp.net/external/HYfJDbT6tBG2bfcR4CC8T-6a0oQh3SnP18Ay3uce1rg/https/tripleback.net/public/discord/1685064315.55285387cd12e090667a2ef5bc79f909425fe05.png?width=1481&height=987

[oak osprey]

elon's musk cologne ad

[smoky oak]

I had no idea Automatic is such a flaming racist, my god @dense tapir

[smoky oak]

I love how it says "Fellon"

Remove one L and its predicting the future lol

[oak osprey]

i knew you'd like that

[static tusk]

what did he do now? I know there was a big blowup around Dec/Jan

[smoky oak]

Nothing new, just saw his extremely racist mods on github

[dense tapir]

Yeah. btw, I called it last night when I saw the income quarterly report. We are now back, on purpose, to the same shit during the crypto mining days. https://www.youtube.com/watch?v=9e8XBqvM_C0

Only now it is due to AI not Mining but same shit different asshole syndrome.

[smoky oak]

yeah, now THAT is bullshit lmaop

[dense tapir]

Nvidia literally stopped making the cards

[smoky oak]

there are dozens of GPU's for sale around me in like a 10 mile radius

[dense tapir]

I agree

only shortage is of the FE

[smoky oak]

yup

cause they are the least bad

[dense tapir]

but they are going to run with it apparently

no, cause they stopped making them last month

My hatred of Jensen grows.

If this were all true, alright I can accept it as I did in the mining days but this is all BS

[smoky oak]

It seems like it could be a lie to make investors happy, honestly

[dense tapir]

I love this one comment

[static tusk]

I doubt there will be shortages as the server chips get the best silicon, where gaming cards get the lesser silicon, which will always be available with how wafers are made etc. Plus there is a glut of cards with not many people buying gaming cards

[dense tapir]

pure, thing is this is all artificial to keep the profits rolling in not a natural occurring phenom.

Are we sure Jensen doesn't work for the CIA cause he is a master at the Psyop game? LOL

[static tusk]

All manufacturers have admitted to reducing production artificially, because of the financial situation the world is in. Its not really just Nvidia

[dense tapir]

Thing is it has been shown that the cards are way over priced and reduced by 20-30% the profits would still be huge. AMD is guilty of this too

They simply refuse to let go of the crypto mining days

[static tusk]

same with motherboards, which are 2x overpriced

[dense tapir]

AMD5 shit is so overpriced I died looking at them last month. 200 is the new 85 dollar tier it seems

My 200 tier is OMG, 500 now. I see motherboard 750-1k

now

[static tusk]

and they dont even come with digital error code displays, unless you buy top tier