#🏞|general-with-images
1 messages · Page 66 of 1
LOL, more updates for control net I see.
1.1.159 to 1.1.162 to 1.1.164 to let us see.
1.1.166
smh
he's accelerating rapidly to hit 169
test github then when fixed roll it into the main head
he's having a manic episode or something
literally posted 2-3 min ago
alr gn yall
night
This is what happens when you don't test your shit before release. Small something I can see happening but not to this degree.
I had a dev tell me flat out on discord that they never test and a lot of the time has no idea about something. ffs
I say dev and meant "dev".
he's fixing math errors and thinking about new features at the same time
give him a break
5D Chess
it's deep
you wouldn't understand.
many samplers never converge on a sharp image
and it makes them perform way more poorly than they usually do
is that controlnet?
yeah, some of the same stuff I am getting
well that can help you fix it at least
go up to 120 steps apparently
what a pile of crap lmao
the dev rn
LOL, so appropriate.
It really tears shit up
Look at how it lacks any sharpness too? SAG would probably have fixed that but that dev is mia
just, ya know
multi-controlnet like 65 of them
apparently you can multi-controlnet this thing to itself
and then tune your CFG up to like 15.0
who knows what comes next. implosion, most likely
Never used that and lack the gpu power to do it too
well i haven't, and i don't, but it is only in sd-webui 😄
Some of these I have no idea how to use
Heh
If I am using SD on my local network, and accessing it from my PC, that isn't using data... right?
Oh sick! You got it working!
well i have decided UU isn't really worth it for me
but CTU on its own is pretty badass anyway
i might try SD upscaler 4x with CTU running before the tiled upscale with 4x
you can tell it's working because of the crazy details on his robe
that's the x4 upscaler bringing it to 4k
crazy
so if you do CTU before that, it's so nice
@smoky oak ^^^
wow it auto-fixes the SD 2.0 output, even. i've never seen it look that good
that's 512x512 lmao
Anyone that upgraded to Control net 1.1.166 having issues? I only get purple artifacting when using the new Reference preprocessor
aw jeez
i was passing the text encoder from SD2.1 in to Compel and then tried using that to make text embeddings for SD1.5
no shit, it failed

dear lawd it's 35MiB
lol @ base sd2.0's rick moranis harry potter
in fact, it might just be me, but Base 2.0's output is WAY better once Compel is used properly on there
oh boyyyy
@oak ospreyis the updates Cnet good yet? lol
for some needs yeah
it regressed in other ways
snooperscooper
i think it might work for you, but you're also very incredibly picky about your quality
at the start that gen just looked like a dog in a blue suit
im so captivated by this increase in quality
the Compel text embeddings, jesus lawd
they are so good, idk how
not bad, not bad
i mean compared to the way DD was display Jack Black before
this is like, identical to the real deal ahahaha
the real test seems to be Matthew Broderick
and it makes emilio estevez fat and old
Compel vs Real
not bad. a lot closer too
a little bit of a Stephen Colbert vibe, but w/e
@smoky oak try weighting the hell out of your CTU prompt, best++++ quality++++
tried 1.5 and the quality is increased there a lot too, but i have to reduce resolution substantially to avoid the deformations that 2.1 does better with.. but once you go high enough res, even 2.1 is like derp
@oak ospreyYou know I got better results using a depth map than using lineart model for this I showed earlier. Still scrathing my noggin trying to figure that one out.
Damn, I can't hardly do anything with CN as I always run out of mem
I only have 6
ah
More than 1 tab is iffy depending on the model
@dense tapirdid DDU, testing if its working proper now
first gen and my PSU tripped lmao
forgot to apply my undervolt lol
the fuck you think that prompt was?
I have no idea, but I LOVE IT
+++ "A fractured mirror reflects a quantum reality within. Capture the elusive fragments to stabilize the diffusion." +++
gpt3.5 comes up with weird prompts that work extraordinarily well
if you want the prompt for GPT
prompt = f"Print ONLY a random text-to-image prompt for Stable Diffusion using condensed keywords and suffixes ++ to emphasize and -- to deemphasize."
if theme is not None:
prompt = prompt + '. Theme: ' + theme
return await self.turbo_completion("You are a Stable Diffusion Prompt Generator Bot. Respond as one would.", prompt, temperature=1.18)
the "You are a..." goes in the system role message
@dense tapirOK SO
it didn't hang up on high res fix that time
it didn't always do it previously, but who knows, could be a good sign
dude i dont think i changed that much but i guess i did. all of my old mediocre prompts are just total fucking gold now
i'm so happy with this because my whole goal has just been to find reasonable defaults for amazing images with zero effort, and voila
@hasty nova ^^ your model's output with Compel text embeddings and auto-CTU at 0.3 strength, prompt best++++ quality++++
oh, and the x4 upscaler after that
Dark side of the Moon 🥺
I know someone who is going to get the 4060ti 16gb. I don't blame them so will see how it does for SD when they get it. They simply could no longer take their 1660.
i'm "compelled" to see how these embeddings change base 2.1 now with these prompts
i wish i could do copy-on-write VRAM use
ffs
that'd be the fuckin shiiiiiit
just run X copies of the same model and only use the difference of memory that they actually write to
in linux, you can virtually fork the process and then it is a copy on write memory mapped space. it looks like you have your own memory space, but all of it is an indirection to the parent's memory. if you were to write to any of it, it'll, kinda, unmap that segment, and now it is real memory, consuming real vram
i'm not sure if these developers just don't know about this stuff or if it's not currently possible on this hw/CUDA ecosystem etc
None of the ones I have talked to have ever even loaded Linux as it gives them a rash.
Iow, I doubt they know that.
oh, wow
windows doesn't... uh... have copy on write memory, at all
i can't believe people actually are able to use that OS at all in any meaningful way. this means, eg. game saves on Windows will cause the game to momentarily block all I/O and threads so that it can serialise the game state without corruption.
there's controversy about vfork() but having it, and using it sparingly, is better than not having it at all. as far as i can see, there's not even a close approximation of it.
you have to run WSL2 to have a paravirtualised Linux kernel run alongside Windows to provide these functions.
WSL/2 sucks
well, it's just HyperV. and i think the success of that is very hardware-dependent
so that's some fun base 2.1 output
i love that
Prompt: -Essentials++
"+"Organization-minus--Chaos++ and+puzzle++plus++with+emotions++minus--
fucking GPT3.5's magic prompting skills
its output is far more noisy, far more regularly, but it's maybe the prompts being so damn weird, idk
gpt3.5 made a prompt with all the frickin animals in one go
@smoky oak as far as i can tell, Compel is like CTU for prompt embed generation

I've made a few wildcard files with GPT4
my discord bot has a command that plugs 3.5 directly into stable diffusion
it just comes up with random images and they're wild
tygorilla?
ligorilla
someone stop it, it is out of control

junglerally's text embedder training really seems to have worked wonders, honestly, it's so much better than base 2.1
hope in bad times
@oak osprey which one?
not sure now if i have it
Thank you @oak osprey
so stoned looking lmfao
yes it is realy nice, unique tones i would say
your hope in troubled times prompt
negative: low quality, low res, messy, grainy, smooth, sand, big eyes, anime, fractured, cracked, wrinkles, makeup (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, synthetic, rendering
positive: hdr+ 8k+ sharp-focus++ cinematic+ epic+ highly-detailed+ photograph portrait digital art
yes, i dont understand for what is 8k+ and such when rendering at say max 1024x1024?
pulls from the part of the image pool that had 8k in its training labels
and this is rendering at more like 4k
now i understand!
oh that's super cool
i want to build an outpainting upscaler now
put that image as the centre and build tiles around it, expanding the context of the image one at a time
it's something GPT4 can explain, really. i didn't like, magically understand any of it. just start small, keep adding to it
I had gpt4 program a lot of stuff and it amazes me
the AI ecosystem is really neat because everyone on huggingface provides example code
although the real gpt4 is Bing not OpenAI for a free account like me at least.
Well, I need to know how something works, and why it is doing it, before I can program
it
I asked GPT4 what it would like to see
did you show it the images after?
can't do that yet 😅
I wish
The first idea was holding a vintage camera. I told it its pretty bad at drawing people holding things and hands in general. So then it asked for an old guy reading a newspaper 😅
lmaooo
auto-CTU was the best idea i've ever had
i used to be like "hm, which one do i upscale? only the best!" but now they're all the best 
this brings me such joy
my CFG is 9.0 too
so SD pipeline like, 🤓
i KnOw kUnG fU
have you tried just 'animal hybrid mashup' and let SD do the imaginings?
thicc
definitely prefers feline animals
umm
well that batch of seeds was hornier than the last one
i put feline in the negative
cute
animal hybrid mashup with David Attenborough
wildlife photographer Canon EOS M6 Mark II, natural landscapes golden hour national geographic
the dingoroo ate me baby
stupids selfie, i need some magic action!
what's about to happen to the poor kitty

just teleport 🙂
good night!
Took me a few attempts to get a nice looking loaf of bread
one of first game on atari 😄
nn, atari 800XL 🙂
which way to go 🙂
have you see gorila elephant?
he is big
planet of apes
this is even more planet of apes. now create head of Statue of Liberty on beach 🙂
if the robot got dogs head you can post it in dailies 🙂
Bit hot here
what
ape was playing with fire. nothing more
this was AI idea and it works
Star lord
no
where did I say I did? lol
i read i
Gandalf
i did him today 😄
Like that?
@oak osprey what is this compel stuff you are talking about? New negative embeddings or something?
Just another Very bad day at the office.
how is in english that magic pole?
Thank you @fresh hound now i see
compel generates positive and negative embeds
oh, i also use k-diffusion
golden sheppard kangaroo
if you think of them as in a sequence its like he gets turned into a cartoonish venom or something
Salutations, butt nuggets
o/
What the fuck
@smoky oak I asked a gpu content creator/youtuber on his live stream just now if the 7950XTX will ever be a real thing. Sadly, he said it never will be. I sure hope he is wrong but Lisa is Lisa. btw, AMD is out selling Nvidia 3:1 right now.
Thanks,
Yep, looks like a typical 1.5 training to me. Now try all that with 2.1 using the same dataset. It is horrid.
proper hands DALFgan
https://github.com/Zyin055/Inspect-Embedding-Training @dense tapir do you ever do textual inversions?
Walter from Fringe
only i can reproduce !!! put it in pgn info
Surprise!
pgn
png sorry you know chess notation 🙂
what about it
check it put it in png info send it in img2txt
And seed is?
-1
yes!
in a1111 is -1 reserved for random number, i know weak..
is easynegative V2 out yet?
i dont know realy
i only found a safetensors but embedding is .pt
check it with AV. I know for example i had model that small flowers from street on github. And i checked it and it was high risk labeled that ckpt model. Could be false positive, but there are for sure some bad
o.k.
computer club of the future in the style of minimalism
so weird too
tried to get Tirpitz and Bismark ship
@wispy ether it looks actualy quite nice, so futuristic
FTL that middle
what
faster than light
it drive inside it seems 🙂
oh apparently dreamboothing in fp16 is ass compared to full 32bit precision training
i mean it's perfectly fine output but 32bit is much higher accuracy
the fp16 makes his face look weird
like he is in pain 
the head is disgusting
yes 🙂
Hulk
something wrong with model....
Artius
?
v21 doesnt exist
i dl it today
i use realisticvisionV20
from civitai, i think it is simply 2.1 model
😮
bruh
he is weird but like his music, also badly trained because eyes 😛
he has different colors. I mean not same colors left and right
ur colorblind then
maybe. He has different colors? I mean in reality
Bernix is right
I have heterochromia as well
then ur blind
but I have a different type
I actually have really good eyesight 😅
However, I do have the extra light sensitivity associated with having blue eyes, since my eyes are mostly blue
Yes, I have heterochromia lmao
means
Blue outside, hazel inside
everyone has that
it looks like some nebula 🙂 Curious what eye colors means.
There are different types of heterocheomia
as well as different forms of colorblindness
there is total heterochromia (two different colored eyes)
central (two different colors around the center of the eyes)
and spotty heterochromia (blotches and spots of different colors)
total, central, segmental
I have very minute colorblindness, but not the standard "I can't see these colors"
I can see colors just fine, its just that I struggle greatly with light colors
like I struggle hard to tell the color between white and very light pink and stuff
if thats your eye, then you have central heterochormia as well
🗿
welcome to the heterochromia gang
💀
me as well, i got just light form of not spotting light shade difference. If understandable.
But tests on internet usualy i have eye of hawkey.
the "color blind tests" where there are words and numbers and stuff, I can see some, but not all
though when tested on individual gamut, I have no problem at all with the full range of the color wheel, only when they are within close proximity/lightly saturated
yes kamawara or how is those tests called
Can people see full RGB gamut
like I can see this one just fine
but I have no idea what this is
this one is easy for me
21
0 clue with this one lol
Idk neither
see, I can't see that
Wrong
yes when they want me to army i didnt pass this and they thought i am joking 😄
they are too lightly saturated for the difference to greatly stand out for me
the first one is 21, hes right
first is 21 @wispy ether
Oh
I thought u meant the 8
And no
Its not 21
Its 74
Use different view angle
on monitor it is somehow different i think here paper is paper.
74
yeah, those 100% look like 74 and 21 to me lmao

the results from most of my tests have always said I am probably red green colorblind at low saturation, which makes sense
If i use my glasses which i never do i see some colors higher saturated
But only like 5% more
very interesting
yes i think i am similar
I wonder if its a slight polarization thing
well donw
however, I should state that while I am bad at picking out specific colors at low saturation, I have exceptional value detection skills, which is helpful on its own
What is that test called
its not really a test, but rather a skill
im 99% sure this is just a blank one
I can see two shades separate from each other and tell if they are different shades at all
EnChroma's online color blind test helps to accurately assess and diagnose color vision deficiencies. Take the test now to discover your color vision.
Midjourney vs Stable
I got a super amoled phone
I have a full sRGB and accurate monitor
new
ips i think
yeah but like, its reference tuned for sRGB color gamut, which is standardized
my monitor has 8 bit frc but no hdr
big brain
if its too saturated, its over gamut, if its too dull, its under gamut
HDR has nothing to do with color, only value
it does
because then it shows more color
more number
i forgot
from 8 bit to 10 bit
i S2421HS
10 bit only means more individual shades
without i
yea
this is BS lol
you can't compare SDR to HDR unless you are viewing both on an HDR display
HDR just means brighter bright spots really
looooo 4ms monitor
4ms is plenty fast
this dude lmao
black smearing
who said im a dude
4ms is more than good enough at 60hz
yea at 60hz
hell, 4ms for 120 hz is still sub frame time
4ms at 240hz is also sub frame time lmao
4ms is the pixel response
4ms at 75 hz is well below frame time
real numbers are different, g2g and so.
yeah, which is less than the time it takes for a frame to be offloaded at 75hz
grey to grey
it doesnt have to do anything with frames
me to 🙂
yes it does
a 4ms pixel response time is 4ms, and it takes 13.3 ms for a new fame to buffer at 75hz
that means sub frame pixel response times, which has been the case on monitors for a long time
if that would be marketing then everything would be a lie
not on va panel lol
I have a 1ms monitor and a 4ms monitor, and they are almost imperceivably similar
i will not argue, just this is realy marketing. 1ms is basicaly imposible to get
VA's are trash in general, second worst display tech you can get nowdays
whats the best display tech?
has more contrast that ips tho
amoled qled oled
yeah, and its stupid too, even at a 500hz refresh rate, its still sub frametime lol
quantum dot
wrong
yes
Mini LED
Mini LED is bleeding edge lmao
isnt mini led new..
no
Mini LED is brighter, more saturated, faster responding, and has no screen burn in compared to OLED
yes, it is
every screen burns
after time
the higher the brightness the higher the chance
probably fake
are those samsung?
idk
cause if so, the right is mili LED QLED
Its quantum dot light emitting diode display lmao
Yeah, and OLED is on the left in that image
what
being beat by QLED, which is a micro LED form with quantum dots
yeah, mini LED is way brighter and more saturated than OLED
Mini LED is pure black lmfao
it has backlight
ok, and whats the benefit of that?
it would still be lit up nep...
There is a reason why Mini LED is the next display standard
Its brighter, more saturated, more responsive, and has a lower burn in coefficient than OLED
with the same contrast
ive never heard before that anyone has a mini led
its why the pro display XDR is mini LED
Its cause its brand new tech lmao
i have oled in oculus i believe
not brand new
the ipad pro 2023 is a 2000 nit mini LED display
the iphoen 14 pro max is a 2000 nit mili LED display
ok, and OLED has been around for like 60 years lol
yeah it comes from pre existing tech but its not the same as that pre existing tech
right lmao
The bottom line is, Mini LED and QLED are the top of the line now
lol
OLED is great for some things, but everything it does good, mini LED and QLED do better
Yeah, thats AMOLED, not OLED, very different
Lol
gapa
I'm very happy with my qd-oled ultrawide. Not sure how miniled or qled beat it in any regard. Motion clarity and contrast are way better on the qd-oled. The only display tech I see being superior to oled is micro-led, primarily because burnin isn't a concern at all.
I haven't tried all these panel types though so I can't say anything for certain.
qd oled ultrawide not worth the price
I happily departed with my $900
QD OLED is fantastic as well, about as good as it gets for OLED tech. Fantastic colors, good brightness, great image quality for certain
Just like I happily.... well. Somewhat begrudgingly departed with my $1800 for a 4090
dead
||Nep, more like technologically I-nep-t||
what

The new furry model mix is insane, probably one of the highest quality digital art models I have used. I bet it would be fantastic for DND characters
she fallin´
L
L
L
L
furry

Bro, do you have any positivity to add ever?
not to furries
I'll add that to the list of your intolerances
Actually no, I'll just disengage. Good luck bothering somebody else when you need help
Anyways
Sure did and released a bunch of them. To this day I prefer them as a trainer but for the vibrancy I get from a Lora/Lycoris is why I switched to them.
I am scared to look at what version we are on now, lol.
8 new version in half a day is crazy
He's been updating it like every few minutes, right? Lmao
About
What a psycho
I just got an incredible new model, but I can't use it away from my PC cause our internet is completely dead
I do not get it because you private your hub and use it to test with then can copy it over to the main head and done once it works right.
Yeahhhh
Pseudo was talking about that
He should have waited till it worked
But at the same time, there is the possibility that it was working for him, but other people are running into issues that he didn't have
Ahhh. That is how I do it and copy is in Github because that is how it is supposed to be done.
No, no mercy to these devs after I had one tell me they never test and a lot of the time have no idea what something even is they just implemented and toss it out there.
tom tucker
I saw a game for game V770 vs 3060ti and it is sad as it was 2-3fps slower. I do hope Battlemage catches up.
Ah damn ._.
I thought even the a770 matched the 3070
I guess not
Very upsetting
Tho at least AMD isn't the lost AI cause we thought it was 2 months ago
Kinda upsetting but when you step back and think about it this is their first big attempt at gx cards in the major market (they had other failures of attempts) so 1st gen vs a 3060ti is not that bad.
BM is when it begins to get better and Celestial then Druid. That doofus of a whatever that made Alchy is the same doofus who made the AMD Vega.
No one at Intel liked him and were glad he was fired. I bet they threw a party too.
I ordered a 1tb 970evo plus last night to become my boot drive and my current 500gb nvme will be as a secondary for SD. Next up my case.
Shoot, we just had another power surge
Damn
yep, a bad transformer that normally indicates unless Doc Brown is at it again.
Yeah, that happened to me once when I saw the transformer explode first hand
They have a technician coming out in a few days
Luckily, I have really good signal at my house, and PDA net allows me to use the data on my phone as a wired connection
it blew the top off and all those Comcast wires just POOF, vaporized.

I'm mostly annoyed that I can't use the new model I just got my hands on
what model is it
why
Not sure I can say the name here, but it's the best digital art model I have ever used honestly
oh so its a naughty naughty model then
Kinda sorta, the name is naughty, but the new rework of it is actually very SFW compliant
dean
It used to make everything NSFW even if you didn't ask for it, now it's a lot more reserved with that when you don't ask for it
I am about to remove all my 1.5 models to save space.
Traitor 
lol
hm#
I am trying to ditch the HDD and they would remain on it.
my hdd was invented like 2009
well, I use them now with a docking station I slide them into if I need to save to it, or read back from it.
shit, lol
dawg
that person is a menace
💀
My Core 2 Duo system, which my big ass case was for that I still use, is from 2009
Finally the case is being forced retired thanks to these gfx cards being so damn long
dean
no not dean
dean
is that allowed
oh yeah, I get that
bro, nep spams so damn much that my chat is still getting overfilled even with the block messages lmao
the heavenly nips
bro is taking up 40% of my feed on his own
so heavenly that they glow
lmao
Eventually I will purchase a larger nvme, but tbh for models a sata SSD is fine and almost instantly load but on this baracuda HDD 2TB takes up to 3 mins.
its giving Katy Perry boob cannons
yeah, sata is plenty fast for a model, its all sequential read anyways
Yeah, so I figure a nice 2tb SATA ssd should be fine but what I don't like is the hard push to m.2 drives when we only get 2 or 3 slots.
only reason all my stuff is NVME is cause I do audio work, and I use lossless audio for a lot of it, which just destroys HDD's and SATA SSD's
i moved away from 1.5 back to 2.1
same
who is still on 1.5?
is anything v5 2.1?
i was
I am on 1.5 right now for zovya's photo mix, and the new model I am using
we all know you are stuck in the past 🤪
tbh zovyas model is both good and very fucked
which I am back at my PC now, so I can use it
650W PSU, 1.5 models, granny-shiftin, double-clutching. you're gonna blow the manifold, bro
lmfao
I am still one of the kings of realism, thank you very much >:C
lol
nah
You know even those cheap ass SSDs without DRAM are fine for SD.
:(
hallucinations can be real
is that a challenge, good sir? 👁️
name the criteria
- no tide pod prompts
you're joking, right?

does that bird exist
uhh no?
that's a Lookachu
💀
you see it through your kitchen windows, and it's a Lookachu
Thats your entry for realism?
hmm i dunno
Everyday this has been slowly going down.
heres mine
that.. that looks very very off
kangaroo golden shepherd
dog with 5 legs
you sent that aleready
A kangerwoof
hmmm
realistic animal contest
what about this banger
saber tooth
damn dude already made a better image than what i made before
it's like a lion monkey bear cat
I have dedicated an outrageous amount of time to realism lmao
criteria @smoky oak why not a discriminator tell us whether it's fake or not
use part of a GAN
I don't anybody here will pass the "is it fake" test, myself included
i love how it makes bears pant
so we doing cats now huh?




