🏞｜general-with-images

[magic bolt]

If James Doakes' actor Erik King was a GTA loading screen character in 2010, 47 years old

[hasty nova]

This is a fun prompt, a bit messy but neat nonetheless

[magic bolt]

If James Doakes' actor Erik King was a GTA loading screen character in 2023, 60 years old

[oak osprey]

https://github.com/huggingface/diffusers/pull/3386

controlnet now merged into diffusers, no need to use that guy's garbage impl anymore

[kind quartz]

Batman 🙂

[mystic marlin]

Welp, accidentally ended up with goth girl twins to begin the day.

[oak osprey]

man i've done 100,000 steps of fine-tuning 2.1 and the loss is reducing from .5 to .3

the images are looking incredible

i'm starting to think they stopped too early before releasing this

my learning rate is so low though it's basically doing nothing on each iteration. it takes 100 steps to show a bit of movement in an image it creates. over 1000 steps, shadows are becoming more refined and skin texture is getting better

[dense tapir]

releasing what?

2.1? Yes, good luck staying below 0.29ish.

[mystic marlin]

Who let the dogs out?

[oak osprey]

hey you know, it's kinda annoying, this shitty pessimism you have about other people's efforts. can you stop that?

try being supportive and see how you feel.

[dense tapir]

What in the fuck are you on about?

I was agreeing with your fucking ass

[oak osprey]

oh yes that's totally the response you need to have to that question, asking you to stop being pessimistic and endlessly negative about efforts to improve things

aggression! that'll fix it.

[dense tapir]

no time for your nonsense. good day.

[oak osprey]

no time for what, you're here on discord. in any case, don't worry. it's all over now.

[golden sorrel]

wrong channel

and message

mb

actual server in general

[oak osprey]

[wispy nest]

Demon girls

[smoky oak]

@dense tapirWE HAVE INTERNET!

WOOOOOO

[crystal bloom]

testing character consistency using controlNet reference only.

[dense tapir]

SWEET!!

I am migrating my data now to the new Nvme right now.

nvme to nvme gen 3 to gen 2 is still fast af.

[smoky oak]

yea haha

I have 2 gen 2's and a gen 4

[oak osprey]

after 15,000 steps fine-tuning SD 2.1

❤️

[smoky oak]

dude that looks awesome!

You improved it so much!

[oak osprey]

thanks you, thanks you

your constant insistence it couldn't be done drove me to the brink of insanity, getting it done

used a learning rate so low i wasn't even sure it was doing anything

i know more now, and this'll be a baseline i can train further from since it's so good without getting burnt at all

[dense tapir]

970evo plus runs 20c hotter :

[smoky oak]

I never once said it couldn't be done, I said it is way harder and more inconsistent than 1.5, and will probably not work as good 😅

which is still true 😅

[oak osprey]

that seems to be because the model wasn't finished by stability to the extent that 1.5 was. gotta remember, they produced four checkpoints with the first SD release

i think SDXL is their follow-up but we don't really know yet lol

apparently you can dreambooth 2.1 in 6gb of vram with DeepSpeed

[smoky oak]

I am just patiently waiting for SD 3.0

[oak osprey]

well i am happy to do the training if the community is willing to put together a dataset to 'fix' 2.1 with

i'm adding vote tracking to my bot so that images it produces can be selected for fine-tuning

1800s painting of a puppy bouncing through a field lmao

[wispy nest]

cute :D
Why did you pick a dog in a field? :)

[dense tapir]

Yep, that is a bad spot in this case as it has no fan movement there until the vid card turns on. New case will take care of that.

As soon as I turned on the vid card to 100 the temps began to fall much faster.

The first slot for nvme gets the cpu spill over and the other fans too.

[smoky oak]

after undervolting my GPU, I have had no issues with power, and I have lost no performance too :>

[wispy nest]

I had major issues with over heating months ago, but then when I started using the ai again, it seems some update or something reduced that a lot

[smoky oak]

it shouldn't be hot at all when generating

well, if you are using a GTX card, then it would be a lot hotter, but RTX cards should be quite cool when generating

[wispy nest]

yeah, but it did, but also it did so about 4 months ago, now it just get hot from using too much upscaling power ;P

[smoky oak]

ohhh, that long ago may have been before proper xformers, which makes a huge perf and resources difference

[wispy nest]

yeah, I took a break for about that many months. I just started using the ai again about two weeks or so ago :D

Can we ai generate this person

[wispy nest]

are these shitstains on her pants

https://tenor.com/view/vsim-sniffing-rat-smells-great-gif-17444805

[solar pebble]

Found this AI generated image on Instagram, how does one go about creating something like this?

[wispy nest]

what makes these images feel unreal

idk i cant put my finger on it

it has a certain quality that looks unreal

maybe it is because i didnt specify a lens and its just taking a mix of perspectives

Can't answer that. I'm more for a painting style than photo, but at a quick glance, I'd say it's because the details look blurry and some details seems to be more of a copy paste

i mean it has some kind of computer graphics quality

might be something to do with the upscaler. I'm fighting mine still to get the details to look less "glow?"

[wispy nest]

literally an ai generated human irl

[keen terrace]

i feel like im missing something obvious, trying to make a sketch of a image, doenst have to be spot of but close enough, and img2img just keeps getting no where near close, like its not using my reference

yeah was missing something obvious

[wispy nest]

i like to think that there are technologies within our hardware capabilities that have immense applications

all that stands in between us and that is knowing how to assemble them

[kind quartz]

lower denoise strenght? Around 0,3-0,5?

[wispy nest]

do you use controlnet ?

[keen terrace]

i am now

[wispy nest]

did this earlier

to try scribble

[smoky oak]

@dense tapir https://www.reddit.com/r/StableDiffusion/comments/13joe98/sds_noise_schedule_is_flawed_this_new_paper/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button

Have a look at this. Very very very interesting. Noise offset, but better, more integrated, and it has additional coherency benefits

Oh! The change was implemented into SDXL

This is apparently what it does in SDXL

[smoky oak]

That left side looks straight out of MJ

[dense tapir]

I will as soon as I can get back into windows as I cloned and formatted the old drive (which changed from C to F) now I can't get into it nor auto repair. I am in the linux partition right now.

Everywhere I looked just screwed it up so must be old info.

Yes, uefi is pointed at the right drive too

[smoky oak]

hmmm...

[dense tapir]

it's an old boot record thing that seems to really affect windows when cloning for a long time

I would think Ubuntu could fix this. hmmm

[smoky oak]

I remember, I had a situation at long time ago was a little similar

I just booted my PC one day, and it stopped working

windows managed to nuke its entire file system, and it took several days and tons of recovery software to get it all back

[dense tapir]

been there before too

see, this is what cloning software did to me

/dev/nvme0n1p1@/efi/Microsoft/Boot/bootmgfw.efi:Windows Boot Manager:Windows:efi
/dev/nvme1n1p2@/efi/Microsoft/Boot/bootmgfw.efi:Windows Boot Manager:Windows1:efi

[smoky oak]

luckily I got everything back, outside of videos, as they all got corrupted

[dense tapir]

yep

[smoky oak]

oh man, that is annoting

*annoying

[dense tapir]

Windows doesn't know which one and the old nvme is formatted so goes to nothing

[kind quartz]

🙂

[dense tapir]

I was in windows though

then formatted F and rebooted to update nvidia drivers and couldn't get back in

[smoky oak]

https://tenor.com/view/patrick-pat-back-pat-back-gif-5620720

[kind quartz]

I think NVME m.2 cant be easy used as uefi booting device... not sure how it is

[smoky oak]

it can be, but the issue here is that windows is not knowing to go to it cause the values seem to be off

[dense tapir]

both drives are nvme m.2

[smoky oak]

I need to get an external NVME dock for my PC so I can finally use my 2TB drive I have had for months

[kind quartz]

i told you later why it is will ask pal which is building PCs. Probably it need win11 or something.

i am happy i learned how to mix characters. This is dracula i think mixed with frankenstein and small fat body.

[dense tapir]

As I thought the bad entry is the second one which is now formatted

[kind quartz]

@rain gazelle you are Degas 🙂 That painter painting ballerinas

[smoky oak]

it looks like civit has started throttling speeds

honestly, better than it being down for several hours lol

[rain gazelle]

yeah was just tryin' something different 😛

[kind quartz]

i got fear to post ballerine i did to not be wrongly understand....

[rain gazelle]

Black Swan style might be cool.

now I'm doing Star Warsy stuff.. 😛

[storm radish]

Football anime

[kind quartz]

bottero like but photo

[dense tapir]

@smoky oak Fixed it thanks to having Ubuntu

[smoky oak]

nice man, glad to hear

[dense tapir]

All updated with my nvidia drivers so now on to your link. @smoky oak

btw, if I hadn't had ubuntu dual booted I would have been screwed or at least spend 12+ hours as I once had to.

First is new drive C other is old drive C that is now gen 2 x4. I was doing a defrag of my mechanical at the time of those diskmarks.

@smoky oak This part is PRECISELY what I was saying since the first moment I saw it used. "Offset noise does enable Stable Diffusion model to generate very bright and dark samples but it is incongruent with the theory of the diffusion process and may generate samples with brightness that does not fit the true data distribution, i.e. too bright or too dark. It is a trick that does not address the fundamental issue."

The bandage approach, as I called it.

this here is shit myself worthyhttps://imgur.com/a/hltcdEb

I am only seeing this in the trainers, so are we not able to get this on already trained models, loras/lycoris/etc...? @smoky oak

I suspect that once cooked we can't do it?

[smoky oak]

I have no idea

I do know its in SDXL now tho

[dense tapir]

Unless they improve upon the textures I have seen in the examples, and shrink the model so I can use it that becomes moot for me.

[keen terrace]

Trained a model! Looks very good, if I up the sampling it pretty much is a photo

[nova galleon]

failed diffusions more like failed my discord

[mystic marlin]

In response to the fat Superman, I raise a healthy Supergirl. 😄

[modest vine]

Finally got around to testing controlnet, here's the steps I took in creating my DnD character with my face from the hand-drawn token that I use during our sessions. Hand-drawn token > Rough outline > Controlnet generated character > Inpaint using dreambooth trained model, pretty happy with the result! 😄

[wispy ether]

uhm

[elfin plaza]

#🏞｜general-with-images lion

😭😭😭

Noo

😭😢

[elfin plaza]

how to use it please tell me

[wispy ether]

youtube

[elfin plaza]

No

[fresh hound]

You like? 😄

[wispy ether]

what

[wispy ether]

then leave it

[fresh hound]

nevermind. haha 🙂

[wispy ether]

[oak osprey]

Other optimizations
In addition, we have improved efficiency of GPU memory operations by eliminating some common pitfalls, e.g. creating a tensor on GPU directly rather than creating it on CPU and later moving to GPU. The places where such optimizations were necessary were determined by line-profiling and looking at CPU/GPU traces and Flame Graphs.

didn't realise, PyTorch 2 fixed the deterministic GPU tensor creation issue

[wispy ether]

ok image broke

[wispy nest]

I dislike the latent upscaler, and all similar ones as they blur too much :/

[wispy ether]

just sometimes

if that happens i use another

mmmmmmm

MMMMM

m

[wispy nest]

my temps reach 87 after less than 1 minute of run time :(

Ueu

[kind quartz]

My first open outpainting test

[fresh hound]

PER MY LAST EMAIL

[wispy ether]

john wick

[fresh hound]

RO OH! TOO MUCH WORK

[wispy ether]

wtf is your model

jesus

[oak osprey]

@smoky oak Windows has about 20% lower perf than Linux on the same hw

😮

[wispy ether]

https://cdn.discordapp.com/attachments/942861254157029457/1108781089645940766/00606-2546228440.png

[neat zinc]

what am i doing wrong

[smoky oak]

From my experience at least, any of the implementations that I've used running stable diffusion on Linux have been around the same performance or less. I'm not saying that it's not more efficient, but I am saying that people don't put in as much to try and get as much out of it

Also, I would just flat out rather have lower performance and have to deal with all the BS that comes with running Linux for stable diffusion

It took about 7 hours in order to get Volta ML working, only for it to be about 15% slower than a1111 on Windows

[oak osprey]

well PyTorch 2 itself, on Linux, has a demonstrated 20% perf gain over Windows due to the drivers being better

it has to be correctly configured, and on A1111 that's probably tricky to get all the bloody commandline switches correct. and i don't know whether A1111 is compiling the models.

my goal in telling you that wasn't to convince you to switch OS, but rather, to help you not worry if you're about 15-20% under the reported performance values that people achieve on the 3080

those tests are generally being run on Linux

[gritty cedar]

this is an ai image with a black border on left and right. drops to only 9.4%

[gritty cedar]

and this @boreal falcon is not an ai image

[wispy ether]

what

[gritty cedar]

its the open pose thing

[wispy ether]

nice

how do i get that

[gritty cedar]

this site is a detector of ai images

that is flawed on my first attempt

[gritty cedar]

go on yt there are better tutorials than I can give

[wispy ether]

i mean just the link

[gritty cedar]

also not an ai image that registers as ai

[gritty cedar]

no

[wispy ether]

what

[boreal falcon]

@gritty cedarmost of the detectors arent good right now

[gritty cedar]

but all i did was add a black border

and it went from 95 to 9%

there will always be another way around it

[boreal falcon]

for the humans the AI images are especially right now far easier to recognize them

especially those who are skilled and involved in creative area

[boreal falcon]

detectors not so much

still underdeveloped, at least from what i saw until now

[knotty oar]

/image. a woman of medium height, dark skin color, long curly hair, light blue eyes, very plump lips, well contoured body wearing a wedding dress with a bouquet in her hand

[gritty cedar]

real or fake

@boreal falcon

this one should be easy

[wispy ether]

damn

[boreal falcon]

the second one is fake

[wispy ether]

why so pixelated

[oak osprey]

ohhh snap i just figured out how to get pytorch 2's compile feature working

the A100 does inference like night and day faster without Eager mode

[wispy ether]

ur doing alot wrong

[fresh hound]

What do you mean?

[wispy ether]

whats your prompt

[fresh hound]

A movie scene of Yoda, in the style of anime, bad-ass, detailed, 4k, SUPER powered, full body, (((Hand drawn style))),

im just playing about on it

[wispy ether]

what seed

[fresh hound]

idk

[wispy ether]

LMAOAOAOA

[wispy ether]

the colors kinda suck

[wispy nest]

hey guys why all my stable diffusion 1111 generations looks like shit? example:

[wispy ether]

use VAE

[wispy nest]

I had to add this because it was giving error set COMMANDLINE_ARGS= --no-half --no-half-vae --disable-nan-check

is it because of that

[wispy ether]

no

just do what it says

go in folder

edit webui-user.bat

do

COMMANDLINE_ARGS= --no-half --no-half-vae --disable-nan-check

instead of just COMMANDLINE_ARGS=

[wispy nest]

yeah I already did that. I was thinking maybe that causes bad images

[wispy nest]

come on man look at this is this normal? prompt: Julia wright red dress, in the style of magali villeneuve, influenced by ancient chinese art, concept art, whirly, dau al set, spirited movement, andrew ferez

[gritty cedar]

what resolution

and model

[wispy nest]

magicmixralistic 512x512

[gritty cedar]

do 704x704 when exploring prompts, and 768 x 768 when you like it. or 800x800

try to zoom way in on their face, you will see that at 512 x 512 it has extremely few pixels to work with

its not going to turn out nice at 512

even with a good model

[wispy nest]

ok thanks

[wispy nest]

you are right I made 704x704 and it gives better results but I had to add --medvram because I use 1660ti lol

[gritty cedar]

i do to

i can make up to 832 by 832 tho

[wispy ether]

making a wallpaper rn lol

dawg

3440x1440

nvm

[wispy nest]

It looked bad aagain with 832

[wispy ether]

for some reason lower

[gritty cedar]

it depends on how big the person is in your image. If they are the size of an ant, then no resolution would work

make sure you control framing and distance with your prompt and negative prompt

also, you have restore faces on right?

[wispy nest]

no

[wispy ether]

mine also look bad rn

WHYY

the hand

[stable bobcat]

this was made here?

[oak osprey]

i've achieved the impossible

[wispy ether]

What

[stable bobcat]

made here, in stable

im new, sorry

[wispy ether]

@oak osprey

[oak osprey]

fine-tuning SD 2.1 to work with no negative prompts

[wispy ether]

Compression big

https://media.discordapp.net/attachments/942861254157029457/1108852443132985374/Herunterladen_19.mp4

There

[oak osprey]

@wispy ether i taught 2.1 how to do perfect boobies ahahaha

didn't even come close to overfitting the model

improved its overall rendering capabilities to have a better idea of what the human form is supposed to look like, who knew

[wispy ether]

You are disgusting

[oak osprey]

LOL

good one

i also trained it on the hobbit movie

[oak osprey]

https://images-ext-2.discordapp.net/external/NY67nE7J4MpgI39bw3tzROfYfRt26dqPZXrmu47q1yA/https/tripleback.net/public/discord/1684451494.8958151c9f4fb27e4f095be7e6ad46fdb6a56a6.png

Slob Ross

[dense tapir]

@smoky oak Official now, 4060 $399 8GB, 4060TI 16GB $499.

[daring kernel]

8gb vram in 2023 is laughable

[smoky oak]

oh, so the 4060ti does have 16 GB?

but its gonna be slower than the 4070 and 4070ti's VRAM... so its gonna be an even worse card

fun

[dense tapir]

precisely

[daring kernel]

I think the only good option is the 4080 or the 4080 Ti

[dense tapir]

128 bit bus. WE HAVE 30 more megs of cache. Proven sucks on 4070 so with even tighter bit bus forget it

[smoky oak]

both are still overpriced, but less bad I guess

[daring kernel]

Does the 4090 has heating issues ?

[dense tapir]

no

[smoky oak]

yup, 4060ti will probably be slower than 3060ti with AI

[daring kernel]

I heard that it burns or something

[smoky oak]

I haven't heard anything bad about 4090 therms

[dense tapir]

the connector if not properly inserted but if you hear it click and wiggle it no issues at all

[smoky oak]

actually, I usually hear about good 4090 thermals

like on the FE and stuff

[dense tapir]

thermals are FANTASTIC on the 4090

[daring kernel]

Good to know

[dense tapir]

I remember when they came out and had someone with one training and it never got above 60c with 37% fan

[dry crow]

Finally a new 16gb model

[smoky oak]

yeah, too bad its gonna be uselessly slow

[dry crow]

How slow

[dense tapir]

slower than a 3070 so far

[smoky oak]

probably between 3060 and 3060ti speed

[static tusk]

also the AMD 7600 with 8gb

[dense tapir]

even with 32mb vs 2 mb cache

[smoky oak]

it honestly might even be as slow as the 3060, considering it has an even smaller bus

We already know that the 4070 is about the same speed as the 3060ti, cause it has a smaller bus

so the 4060/ti having an even smaller bus does not bode well for AI at all

[static tusk]

tensor cores on the 4060 can do twice the TFLOPs. it will be faster

[smoky oak]

thats just wrong

the faster part, not the flops part

[static tusk]

no chance its slower

[smoky oak]

the tensor cores on the 4070 are faster than the 3070, yet its slower at AI cause the VRAM bus is so constricted that it can't get the data to the tensor cores to benefit from that speed

[smoky oak]

just going off the 4070 being slower than the 3070

[static tusk]

wheres the data showing that comparison?

[smoky oak]

the much smaller VRAM busses have had huge slow downs on lovelace GPU's

[smoky oak]

all over the community. @dense tapir and I did digging on Reddit and found that the people with 4070's were reporting about 12it/s, which is just a smidge faster than a 3060ti

and slower than the 13 or so on 3070

we talked about it a good while ago.

The new slower VRAM bus is the biggest limit on the cards

that cache does its job for gaming, but not for AI where the values are changing rapidly

[static tusk]

and did they use torch2?

with the new cudnn drivers?

[dense tapir]

Thought so as it is that damn bus.

@smoky oak That cache, actually, is NOT even helping the 4070 in gaming. It is really bad and why they are not selling with a ton of returns.

[cyan snow]

i actually got my 4070ti about 5 hours ago, already installed cuda and it does about 15it/s

[smoky oak]

wait

15it/s on a 4070 TI!?

[cyan snow]

yep

[smoky oak]

bro what

thats

sad

[static tusk]

4070ti should be closer to 20it

[dense tapir]

4070 is ~12 and ti ~14

[smoky oak]

that means the 4070ti is slower than the 3080 as well

[cyan snow]

I don't think i downloaded the right cuda version

[smoky oak]

those are in line with the numbers I saw on reddit and in youtube videos

[dense tapir]

What I have been seeing real people showing

[static tusk]

most people dont know how to setup SD with the 40 series, hence the slower speeds

[dense tapir]

please, no, don't do it. No Jensen toe licking

[cyan snow]

idk, i saw a benchmark of 4070ti and it was a little more than 20it/s

[smoky oak]

I have yet to see a single person real world get that number

[dense tapir]

Yes, 512x512

[cyan snow]

so my guess i didn't set up my drivers correctly

[smoky oak]

unless you look at toms hardware guide, which is trash lol

[dense tapir]

Tom's is a joke

[smoky oak]

where it says the 3060ti is only 9it/s

and the 3090ti is only 19

[dense tapir]

512x512 is about 20it/s

[smoky oak]

on what?

[cyan snow]

with 4070ti?

[smoky oak]

oh yeah

[dense tapir]

4070ti

[smoky oak]

thats what tom says

[cyan snow]

for me it doesn't do that

[dense tapir]

768x768 the bit bus hits hard

[smoky oak]

but the rest of the graph is very wrong, so I wouldn't count on it lol

[dense tapir]

1024x1024 I saw one person say they were getting about 4it/s on a 4070ti

[cyan snow]

wait was i supposed to get the latest cuda toolkit? or 11.8?

[cyan snow]

that's true

[static tusk]

the toms graph has my 3080 at 14it, but I can see closer to 17-18it with now with torch2

[smoky oak]

yeah, I can get 18 on mine when its OC'd as well

and he also says the 3060ti is only 9, when I gto 11.5 on it

[cyan snow]

what cuda version do you use?

[static tusk]

thats with OC right?

[smoky oak]

but he also says that the 7900XTX is 19.3, when its really closer to 14

So I don't believe him at all

plus he doesn't say what sampler, and what optimizations

its an unreliable graph

[cyan snow]

amd doesn't support sd anymore. i know this from experience.

[smoky oak]

since when? The last couple days?

[static tusk]

its mainly outdated with new cudnn and torch2

[smoky oak]

yeah, then he needs to get off his lazy ass and fix it rather than just adding new lines to the graph and calling it "newly updated"

[cyan snow]

last couple weeks, everyone that got the latest ROCm got boned

[smoky oak]

interesting, I have heard nothing about that

[dense tapir]

Tom's does say in the Jan post but not in the followup

I am pretty pissed at AMD with how they are really screwing around with ROCm

[cyan snow]

also ROCm is trash, very difficult to set up and not even remotely close to CUDA

[smoky oak]

he also doesn't say what the 2 different values on RTX cards are

[dense tapir]

ROCm on linux difficult to set up? No, pretty easy actually.

Now if you are trying to docker it or WSL then yeah, good luck with that nonsense.

[civic gorge]

not difficult but cant use WSL2 to use rocm on linux so its not worth getting a new computer just for linux

[dense tapir]

Linux can run on a potato. wth?

[smoky oak]

you... don't need a new PC just for linux whut

[cyan snow]

idk man, when i used amd they screwed me over. had to format about 3 times because of them

[smoky oak]

I don't even like linux, but you don't need another machine for it lmao

just dual boot

[cyan snow]

very risky

[smoky oak]

bruh

[civic gorge]

so take my 6900 xt out of my pc, put onto potato pc, and buy new gpu? or just sell the amd gpu card and buy nvidia

[dense tapir]

Well, I know this I am done with Nvidia as I am no one's B. Jensen, Su's or anyones. I look at it like this. Is it faster than I have now and has 24gigs for a decent price? Sold.

[smoky oak]

DUAL BOOT

[cyan snow]

that's what i did, and im somewhat satisfied.

[dense tapir]

That is what I do DUAL BOOT or don't do it

[static tusk]

as long as you can do everything equally

[smoky oak]

that as well, no giving up certain functions

go intel arc, but only have support for 2 samplers half the time when the moon is in proper phase lol

[dense tapir]

Yep, if it does it slower but does it for half the cost I am there. I just no longer wish to go in the direction Nvidia is headed.

[cyan snow]

wait, so how do i increase my it/s? apparently it should be way higher than it is

[dense tapir]

I still have my hold wish that Druid from Intel does this stuff, which is about when it will be update time, or the one after actually. Tired of Lisa Su and Jensen to be 100.

"effective". No

4060ti vs 3060ti

[cyan snow]

if what you want is effectiveness, than google has something called a TPU, its the fastest when it comes to ai

[civic gorge]

aka u cant use both at same time unless u restart computer to switch

[dense tapir]

In all honesty the only thing holding off the 4060ti is the bit bus

I am so tired of Nvidia gimmicks because I want real world bandwidth not effective, or this new neural compression so we don't need vram.

[static tusk]

delete venv, delete repositories folders
change your launch ARGs to --opt-sdp-attention --opt-channelslast
git pull the latest A1111, dont install xformers
in that order

[wispy nest]

Delete system32

[dense tapir]

LOL, NO

[marsh hill]

mothman, dude stable diffusion is amazing

[wispy nest]

heh, I finally got some lora's, but I think I need to learn more about them before using them as the details aren't as defined as I'd hope. :P

[marsh hill]

LOOKS AMAZING

GG

[wispy nest]

thanks :D

[marsh hill]

I have been experimenting with stable diffusion for a while, and not going to lie, it's much better than it's reputation,
i mean, i wanted to create a mothman, and i am so satisfied with these results, and the fact i am just a beginner!! says a lot about the potential of this a.i

[wispy nest]

yeah, it has a bad rep from people. Lots of drama around it all. I myself have been taking a 4-5 month break, but I have been using the ai to make art since before/around version 1.4

this is the details I get without the use of loras. I hope it's just a strength issue, and not some "place it before X if Y is Z" :P

[marsh hill]

bro!!! Love the details

[wispy nest]

you'll learn in time, I've made over 600,000 images since last year. Also, having a "local version" such as auto1111's webui is a lot better than any online version. But you'd need a powerful computer to really use it sadly :(

[marsh hill]

I see ❤️ Luckily, I have a decent rig.
Ryzen 3600, and 3070ti 😃

[wispy nest]

then you shouldn't have any issues, other than getting the local version to run ;P

[rain gazelle]

evenin'

[dense tapir]

heyas

[hasty nova]

Well, that took longer than necessary but I just added 56 new images to the dataset for v3, mostly photography with some digital art. Just thought I would start working on it. I am planning on retraining from the ground up instead of building upon v2, so hopefully I can enact some new techniques to improve the model. I've been noticing a lot of things wrong with it and it might be because my learning rate was too high in the first two?

Now the total is 350 images btw

[static tusk]

good luck with v3

[hasty nova]

hopefully it works!

I want to get the dataset to 500 at least tbh

which will take a while

but I have some ideas

so

[static tusk]

time to make the GPU cry with intense training

[hasty nova]

yep :/
I only got a 3070 w/8gb so it takes a while... luckily I don't do much while i'm asleep

speaking of sleep, I need to go do that

got exams in the morning

Gn!

[static tusk]

nite

[crimson dawn]

Heres where the fun is. by having less defined details, but with strong context, you have this stuff. i pulled these off with simple prompts. "squirrels flying at each other" and "marvel cinematic fight". @boreal crown

[boreal crown]

This is an honestly great example of switching it about, definitely great for environments

[crimson dawn]

obviously better prompting can help, but i'm just showign you, it looks like a cheesed model with its big blotches, but those are powerful blotches

[boreal crown]

just wish the classification could do hair/face/tshirt etc. i know it has been done in other instances

and theres a tonne of training data for it

[crimson dawn]

segment anything could produce a mask like that, that you could feed into controlnet

[boreal crown]

yeah thats what ive been doing, chaining the pair together

[crimson dawn]

just as a simple mask even, its great for inpainting that sort of stuff

i mean to play more with segment anything

[boreal crown]

i really like the simple click point to mask thing.

[crimson dawn]

it's really the one feature that's keeping me paying for photoshop. i may end that subscription once this period is done

[boreal crown]

i just cant get grounding dino to work, i have cuda 12 and ive a lot of other stuff on it. im not downgrading and playing with 20 different venv versions of cuda

[oak osprey]

lol works great on my Vega 56

sounds like a skill issue

[crimson dawn]

yeah rocm worked on my linux machine beautifully. guess you just need skills.

[dense tapir]

Not sure why it can't do a Mermaid it gives them legs.

[marsh hill]

try saying like, "only 1 leg"

[smoky oak]

what in the hell is this combination of recommendations lmao

[sinful mirage]

i got the google text2music thingy if anyone wants to send me a prompt ill send the audio file here, think im limited to about 30 per day, its pretty impressive

[dense tapir]

@smoky oak Well, I did not know this "Everyone is focusing on VRAM. All you ignored that the 4060Ti and lower have PCIE4 8 lanes. 4070 has 16 lanes. Another potential bottleneck for the memory."

[smoky oak]

peeng

[smoky oak]

yeah, i saw that as well

tho its not like the memory is even close to fast enough to use that much bandwidth lol

[dense tapir]

128 bit bus, nothing else different from the 4060 but ram and 100 USD more and both use 8 lanes instead of 16 and we know how bad the 4070 is. smh

[smoky oak]

wait, they are the same perf?

[dense tapir]

The creators saying it is great a nice card with vram and doesn't break the bank are the content creators that are highly sus. The first wave. Now the second wave are saying to stay the hell away from these two turds. Less audience so more able to speak freely.

[dense tapir]

Same clock speeds, same cuda cores, same everything except ram

[smoky oak]

wow lmfao

which means...

LMFAO

the non ti can theoretically be faster

[dense tapir]

Here is one for you but the 3060, not the 3060ti, actually had more cuda cores, and more bit bus, than the 4060ti

[smoky oak]

because the 8GB VRAM has access to the whole bus always, while the 4060ti has to split its bus usage across VRAM that nevert gets touched

this is insane lmao

[dense tapir]

for the same price

500 for each at launch

[smoky oak]

$500 for what?

[dense tapir]

at launch I read the old 3060 non ti was 500

same price as the 4060ti

[smoky oak]

no, it was $320

3060 was $320, 3060ti was $400

USD

[dense tapir]

well, ffs

even worse

[smoky oak]

yeah lmao

[dense tapir]

I lose about 1k cuda cores

[smoky oak]

the 3070 was $500, and the 4080 was $800

[dense tapir]

okay then it was the 3070

[smoky oak]

the crazy thing is the only real advantage lovelace has is its massively higher clockspeeds... which means that the hardware needs even more VRAM speed to get new information in and out as fast as possible lmao

[smoky oak]

That is actually insane

[dense tapir]

very much

the 4060ti is really a 4050

wow

the dude was right

10% more than a 3070 tflops

so it about 10% faster max if it had the same bus

in the end it should be the same, maybe a tad less we shall see, for the same price.

[smoky oak]

So its basically just like the 4070

same perf or lower than the next card up, for the same price, with much lower power draw, and a huge kick in the balls for AI

wow

[dense tapir]

I just don't get what Jensen is doing. I sort of think he is trying to drive the gamers away so he can close that division down to concentrate on AI/ML now.

[smoky oak]

I personally think its also a game plan to try and sell their massive overstock of ampere.

Make them equal on the same playing field and sell both

but they overcorrected too much, and now lovelace is ass lmao

[dense tapir]

4060 has 4 ram chips I think

btw, as one creator mentioned that had me doing a wtf as well is that Nvidia knows the 4060/ti is so bad they are actually comparing it to a 2060

2 gens earlier. that is seriously a wtf?

[smoky oak]

bro, a 60 card being tied to just 1080p at this gen is sad lmao

hell, even last gens 3060ti was a 1080p beast

[dense tapir]

4050 will be a 720p card

[smoky oak]

I could get 75FPS locked on that bad boy in CP2077 (before perf updates) at 2560x1080 with psycho RTX and only balanced DLSS

granted, thats not "high FPS", but thhats also still an extremely demanding game

[dense tapir]

4090 great card. 4080 great card but no more than 800-900 USD. The rest are just pure ass.

Trying to find a 4090FE for MSRP is impossible because Nvidia doesn't make it. If they do it is 100 at a time sort just so they can't be sued.

[smoky oak]

Yup, and it sucks cause the 4090/4090 launched so strong

Only problem with them was price on the 4080. Performance was stellar, size wasn't too bad (For FE), and they offered huge performance lifts from last gen, for not much more (specifically the 4090)

But then the 4070ti/4070launched and created a several tier GPU performance gap in the skew

[dense tapir]

yep, it is when the 4080 12gb was created that it all fell apart.

[smoky oak]

I mean come on, the performance gap between the 4070 and 4070ti is higher than the perf gap between the 3070 and 3090 lmfao

which makes the 4070ti a much better buy over the 4070, but only strictly cause its not as bad as the 4070

[dense tapir]

The lesser of two evils is still evil.

[smoky oak]

the 4090 at $1600 honestly isn't that severe to me. Its the best of the best, no corners cut, you have a REASON to get it, and you pay that price for it

But the 4080 at $1200 is absurd. $1000 MAX

[dense tapir]

1k for it is still too much. $1599 is doable but you just can't find a $1599 that isn't from a junk brand that removes chips on it to save money.

[smoky oak]

yeah, I agree

[dense tapir]

vrm is a big deal

Gigabyte is notorious for it and sure enough they did it on their 4090 too

I think their 4080 has 6 slots empty.

Zontac had 6-8

solder pads where it should go but empty

Well, shit

We have to wait for testing of the 4060 and 4060 Ti but because these cards only have 8 PCIE lanes, you most certainly need a motherboard with PCIE4 or you will take a performance hit. How big of a hit will be determined by testing. So if you have an older, cheaper mobo with PCIE3 this is not the GPU for you. You either have to upgrade your mobo or buy the next card up the 4070 with 16 lanes.

Most of us who would buy a 60 would be PCI3

Gen 4 maybe, but probably not.

@smoky oak This is what I was talking about. Pity.

[smoky oak]

BRO 1.15X 3060ti WITHOUT FRAME GEN LMAO

[dense tapir]

yep

[smoky oak]

That is insane lmfao

[dense tapir]

15% faster and only 60% faster than 2 gens before.

[smoky oak]

I just bought a GPU that is at least 30% better than my 3060ti for $320

[dense tapir]

As I said they are screwing everyone this gen.

even at decent prices the performance lift is just not there. Screw being forced into DLSS and that frame thing

[kind quartz]

Kitsch

[dense tapir]

you buy a fancy high refresh monitor then resort to DLSS and frame gen that causes latency so what did the high refresh montor do for you?

[smoky oak]

I have a 165hz monitor, and this 3080 can probably run any game I play at 165 max settings no problem lmao

[dense tapir]

not with <4080

[smoky oak]

Imean, it ran subnautica 2 at maxed out ultrawide at 160fps only using 60% of its grunt

[dense tapir]

gotta have DLSS and maybe frame gen

that is why people are not having any of this but Nvidia is standing their ground.

See that 4060 and the 107 die? 107 has always been the 50 class cards so it is a 50 class card with a 4060 sticker on it.

[kind quartz]

but dont forgot it is cutoff and nobody knows how it will work on pcie with 16 lines, this is only 8 lines, so on PCIe4 it should work well, but PCIe3 nobody knows

[dense tapir]

@smoky oak Look at this

[smoky oak]

what am I looking at?

[dense tapir]

Well, are you available to test the above on your card right now?

[smoky oak]

my card will whoop that lmao

[kind quartz]

no, just read about it, note x8 is key

[smoky oak]

what card is that supposed to be?

[dense tapir]

35 steps of DPM++ 2M Karras?

[smoky oak]

I am not gonna use that exact prompt, but let me try it

[dense tapir]

yeah

no embeddings or anything though

[smoky oak]

7 seconds LMFAOOO

[hearty karma]

would it be better for me to save up for a 3090?

[smoky oak]

768x768 35 steps DPM++ 2M Karras

[dense tapir]

good, much better than I thought

[smoky oak]

what card is that supposed to be in the screenshot?

[dense tapir]

that was ran in WSL which runs it like crap unlike a true Ubuntu

A770

[smoky oak]

ah, ok

[dense tapir]

not bad tbh

Especially since it is being ran in shit WSL

you do take a good hit with WSL

native ubuntu is better

[smoky oak]

do keep in mind, I am on a stock 3080, no OC, severe undervolt

[dense tapir]

doesn't matter I have seen what I needed to. Thank you.

[smoky oak]

actual time result

[dense tapir]

This gives me some hope for battlemage, then Celestial and finally Druid. Intel is far from stupid

[smoky oak]

its only .5 seconds faster to do DDIM, interestingly

[sterile temple]

😒

[dense tapir]

Well, for a first gen non red/green card I am actually impressed especially since it is being ran in WSL too

[smoky oak]

my main gripe with intel is that they are fine, but they like to lie out of their ass 10x more than all of the other tech companies put together

I will never forget the whole 10th gen mobile review fiasco lmao

[dense tapir]

on W11

I haven't been on an Intel CPU since Ryzen 1st gen dropped

[smoky oak]

Did you see their whole review fiasco? It was insane lmao

[dense tapir]

No, but Intel is scum. They are all scum.

Did you see what AMD did?

AMD is just as bad. Lisa Su takes cues from Jensen

[smoky oak]

if you mean as bad as Intel in terms of lying, no chance lmao

They reviewed their 10th gen CPU's using the same model laptop for AMD and Intel

and they showed an impressive 20% lead

except they took every single liberty to not disclose that they gave the AMD laptop a lower power GPU

2060 max p vs max q

All of those numbers were with Intel having the higher power variant of the GPU's

60 watt 2060 mobile vs 105 watt 2060 mobile

I believe they did it with the 2080 mobile as well

Like 80 watts vs 140 I think

I also remember when they strategically did that new tech demo 30 minutes before AMD so they could put "Vs the top of the line from our competitor" and make it look like they meant next gen

[dense tapir]

Just saying they all suck but which sucks the least as I don't like any of them but I need something to do a job without forcing me to mortgage my house.

[smoky oak]

Yeah true

[dense tapir]

The moment a good card hits from Intel is when their evil will come out. Watch.

[wispy ether]

@dense tapirdo you think fp8 will be a thing soon?

[dense tapir]

Would be nice

[dense tapir]

Another bad thing with Intel is they never made a 790

[wispy ether]

the hell

[dense tapir]

Tom's Hardware chart for speeds

[sterile temple]

what resolution?

18.12s on my laptop's 3060 6GB

for 512x512

[dense tapir]

@smoky oak I am waiting for someone to implement this https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#tensorrt-text2image-stable-diffusion-pipeline

[hearty karma]

Dynamic Threshold is funky

[wispy ether]

can u send 3060 vs 4070

[honest pendant]

Girl cowboy

[honest hatch]

My first photo I like.

[kind quartz]

Kitsch again but love outpainting! open one

[wispy ether]

https://cdn.discordapp.com/attachments/942861254157029457/1109098900180181092/00166-2506400077.png

[wispy ether]

[cyan snow]

KABOOM

[wispy ether]

what gpu

[cyan snow]

my rtx 4070ti arrived yesterday

[wispy ether]

https://tenor.com/view/hahahahah-gahahahaha-haha-ha-fun-gif-19042485

why not 4090

[cyan snow]

why would i need 4090, 4070ti is enough for what i do

[wispy ether]

what are u doing

[cyan snow]

also, rule#2

[wispy ether]

i dont read rules

uwu

[cyan snow]

good luck with that, i guess

[wispy ether]

where even is the rule channel

[quiet current]

#✍🏼｜rules-and-tos

[wispy ether]

https://files.catbox.moe/njn4q1.png

1440p wallpaper

😎

16:9

[honest hatch]

I like this, what model did you use?

[wispy ether]

digitaldiffusion

[honest hatch]

thanks!

where do I find it?

[wispy ether]

civitai

[honest hatch]

well, I forgot there is also this webpage xD

[oak osprey]

why would you want that much accuracy loss

[wispy ether]

what

it doesnt make loss

[oak osprey]

thats a custom pipeline, and a1111 already uses one. you would need to add the missing bits or else you lose functionality once that is available. i doubt it was worthwhile to do, since it takes about half an hour to cache the tensorrt build

[oak osprey]

the original sd models are 32bit...

[wispy ether]

restarts pc instead of waiting

[wispy ether]

what

you mean float32

[oak osprey]

floating point 32 bits

[static tusk]

fp32 is full precision
fp16 is half precision

[wispy ether]

whats the difference

fp16 is faster and use less vram

[oak osprey]

you rely a lot on the quantization process to not throw away important data

[wispy ether]

what

[oak osprey]

fuck off dude lol

[wispy ether]

bruh

[static tusk]

for training fp32 is ideal. for gens you dont see that much difference, but its there

[wispy ether]

i dont have 200iq brain to understand that sentence

quantiza

idk

[oak osprey]

its a lot of loss to reduce from 32 bits to 8

[wispy ether]

does fp32 still generate weird fingers

[oak osprey]

quantizing is like summarizing the tensor space to a set of cliffs notes

[static tusk]

its only small details you will notice, not large changes like bad fingers

but fp8 I think would be noticeable

[oak osprey]

if you train models, training at fp16 can cause catastrophic loss, so i assume quantizing to 8 isnt a walk in the park

[oak osprey]

most people training on fp16 use a very low step count and learning rate and very few images to avoid changing the model too much

the model does something called prior preservation, where it looks at images it was capable of creating before training began, to remind itself of what it already could do, and preserve the structure of its tensor space

and generating these at fp16 also increases loss

people tend to overfit on beautiful styles because its easier than properly training

[wispy ether]

PLMS sampler

[wispy ether]

10 samples

[wispy ether]

5 samples

2 steps

[muted pike]

Marmalade, Any arrangement in space, flat 3d, 3d render, sing style

[smoky oak]

[kind quartz]

@smoky oak what that cat means!!! I know i must working harder and better!!!

[oak osprey]

SD 2.1 trains so quickly

it's insane.

[grim pivot]

https://www.youtube.com/watch?v=cIkjXY0fM0U

[smoky oak]

trust me, 1.5 is still considerably faster, but you seem to have found some settings that are at least working

I would love to test out your model when you are up to it. I'd be curious to see if it has the same noise/crust issues that most 2.x models seem to have when doing realism

@oak osprey

I am actually messing with a new realism model at the moment

[oak osprey]

i was pretty sure someone would take that the wrong way. training quickly isn't necessarily a good thing

it means it's easy to change large sections of the latent space in irreversible ways

[smoky oak]

I thought you were meaning fast to get good and non problematic results, cause thats where 1.5 comes in

[oak osprey]

that means it trains more slowly

training more slowly allows it to keep coherence of what it already has organised in its latent space

this is why people struggle so much with 2.1

so for context, 1e-6 is considered a Super Low learning rate for 1.5

that will burn the everliving shit out of 2.1 in like 100-200 steps

i have to use 1e-8 for a batch size of 2

[smoky oak]

interesting

[oak osprey]

this begins to approach the training rate of 1.5 at around 1e-7

you need an extreme number of class images generated directly by whatever model you're working with, for it to keep the Loss as low as possible

[smoky oak]

I do see how that's not a good thing tho

I just thought you meant its fast to get good results out of 2.x, which is not true when compared to 1.5, cause anybody who gets good stuff out of 2.x isn't willing to share it, so we make like 0 progress on getting naything good out of it

[oak osprey]

they're likely not willing to share it because they've destroyed coherence for large portions of the remainder of the model's latent space

[smoky oak]

really tho, the gate keeping around 2.x training is obnoxious

[oak osprey]

i have published every revision of my model and i hope no one ever tries to use it to train from

i don't blame others for not thinking their stuff is worth sharing

as AI Gandhi says, "Be the change you wish to see in the world. But remember, dropping a nuke in the end-game is less long term suffering than letting you all figure it out over another 100k years."

[smoky oak]

Thats not at all what I meant, I meant the people that do get rarely good stuff out of 2.x, and then say "Figure it out" when you ask them

[oak osprey]

oh

i would question what horrific things went into their dataset

i haven't tried model merging yet, are you any good with that?

i actually installed sd-webui locally yesterday and it noticed "You have an AMD GPU!" and tried installing all the ROCm stuff but this is a laptop so it just boinked out and installed the wrong Torch binaries and all that

but once i'm there, i will absolutely write up everything i know

[smoky oak]

yeah, but not with 2.x

well...

[oak osprey]

how it started vs how it's going

[smoky oak]

maybe I am good with 2.x, I have no idea

Never found 2 2.x models worth merging 😅

[oak osprey]

i have a HUGE test matrix of prompts to run through for each checkpoint. i could always use more

[smoky oak]

I am currently testing 3 photo realism models against each other

[wispy nest]

exaggerated anatomy

[oak osprey]

that sounds like a "handful"

[smoky oak]

I would be willing to supply some photo realism prompts if you would be ok with keeping them between us in DM's

[oak osprey]

gatekeeping

[smoky oak]

tho I am not sure if you are going for realism

[oak osprey]

bro it's going into my source tree

[smoky oak]

to some extent, yeah, I have to make money some how lol

[smoky oak]

ah, nevermind then

[oak osprey]

"o ya itll be private"
"here u go"
"git push"
"what was that sound"

@smoky oak well these prompts aren't to ensure my model achieves photorealism. it's to ensure it doesn't lose what's there

[smoky oak]

alright then

[oak osprey]

modify the prompts so they're Less Good if you have to

it's just to get a wider sampling of the latent space

i have dog, child, robot, man, woman, group of men, group of women, architecture, abstract shapes

i have a mountain bike prompt, i have a woman under the milky way galaxy at night time

[kind quartz]

what is this model @smoky oak ? Just how big expert you are 😄

[oak osprey]

that's one that has tiling artifacts lmao

[smoky oak]

I have no idea, but damn that is really inconsistent and artifact filled

[kind quartz]

i know.

[oak osprey]

VAE tiling turned on?

RIP

i love how many blog posts are like "i didn't noticed any issues, enable by default"

[kind quartz]

no... just incosistency

[oak osprey]

at 2048x1024, on consumer kit, that'd require VAE tiling

[kind quartz]

just from pyramide i supposed sytan could guess one of famous models

[oak osprey]

but they're pyramids

[kind quartz]

@oak osprey it does not, just bit more work 🙂

i shifted it if few thousands years ago. Still issue

[smoky oak]

Zovya's photo real is still winning compared to other realism models

[oak osprey]

i should throw my test matrix at SD 1.5 and then Zovya's and RV1.4, RV 2.0

[kind quartz]

here deliberate v2

[oak osprey]

HARDblend, Deliberate v2, and URPM

[oak osprey]

i was using HARDblend as my CTU model for a while and then i switched to RV1.4 and people actually noticed enough to complain loudly that the thing was broken.

[smoky oak]

Damn, these realism models are getting too good lol

Its getting hard to compare them

[oak osprey]

because they're all mixes of each other

😄

you need to go outside their field of specialisation to test them properly

they're likely hella overfitted for realistic photography

try to get cyberpunk robots, abstract tetrahedrons in shadow space, beautiful experimental images, etc

test out images of cross-animal hybrids, up-close shots of textures eg. fur / leather / stone

[smoky oak]

I get what you are going for, but I am after how good it can do realism, not everything else

But I do focus not just human realism

[oak osprey]

the general capabilities of a model are going to make it better at photorealism

the ones that have overfitted less while still attaining the best photorealistic results are going to be the best models

you don't just want boring photos of people in whatever poses the training data had, right?

at least personally i want to be able to make a person ride a bear or something

[smoky oak]

that sentence is kinda contradictory

"If a model can do better everything, but worse single thing, its going to be the best at single thing"

[oak osprey]

you're saying it is hard to quantify their "betterness", and i'm saying that once you're finding that they're all on even footing so that you don't care about which one is better at photorealism, the only thing remaining is to see how well it can apply that knowledge to everything the base model knows about

[oak osprey]

are those 4 different models?

[smoky oak]

yes

[oak osprey]

whyyy do they look so similar in output?

[smoky oak]

pretty clearly so lol

they really don't, not all of them lol

[oak osprey]

that's basically the same person

did you use img2img?

[smoky oak]

These aren't that similar lol

[smoky oak]

No, just strong prompting

[oak osprey]

the composition of those photos are so similar they even both have two boulders on the right side

[wispy nest]

the shadowing/darker parts above the eye-brows look very similar :O

[oak osprey]

look, when i'm training a model, every 50 steps, the thing changes drastically. those models are over-merged and look like they're all basically the same model.

that's why you're having difficulty testing them

[smoky oak]

its called... the same seed lol

[oak osprey]

the same seed, the same model, 50 training steps later, drastic changes.

1.5, 2.1, they're all like that

[smoky oak]

and also I selected a seed where they look close enough to compare, otehrwise they are not consistent

[oak osprey]

well, there's that, i guess. i don't cherry-pick seeds

[smoky oak]

like the fact that she has braids in 2 different places, and her shirt changes color, and her ethnicityis not the same, and the level of background blur is inconsistent

[oak osprey]

still, i don't think those models are very different at all.

the composition is identical. the braids and rock positions etc come down to training time

and the proportion of merged weights in the Checkpoint Merge tab

[smoky oak]

testing a new prompt with random seed

@oak ospreyhere

no cherry picking, no seed alignment

[kind quartz]

all hands left 😄

[oak osprey]

you did me dirty, there's no way i can say "those all look the same" and get away with it