Nvidia is now focusing on the business side AI/ML per Jensen as he doesn't even talk about gamers now
#🏞｜general-with-images
dense tapir
My wallet says AMD my needs say Nvidia
I am just saying this complaining is turning into rage by otherwise AMD fanboys
Really amazing to me how they are turning on their god
I know for a fact to quell them down is precisely what this one vid said is to release a card with 12GB at 500 and call it a day.
Well, I watch a lot of people and they name the pricing (give or take) and all the percentages of greed (my take) thrown in and yeah, these companies are dicking all of us.
they are trying to make the high prices left over from mining the new norm
get used to 4gb 4050 at 400 type craziness. I am just glad consumers are not taking their bait. At least there is push back.
I think a lot of this, as I have been saying, will come to light post summer so late summer to early autumn.
No, Sue is as ruthless greedy as Jensen.
If Nvidia is 100 make our 90 regardless of the costs
Plus they are having a lot of issues now with their CPUs
Nah
There is so much opportunity to steal marketshare that AMD is just not doing and the investors, even, are wondering wth is going on?
Was in WSJ I read a week or two back
A lot of mis opportunity and the article was referenced in a few vids on YT
That was good because if investors begin to question then apply a little leverage they may break this open which is good for the consumers
fierce crest
lol emad casually dropping this was insane to me
shows i really havent been paying attention to this discord channel
dense tapir
dry crow
Thanks 🙂 its a title
static tusk
well thats a rare event. A1111 got an update
smoky oak
well thats a rare event. A1111 got an update
yup
@dense tapir Auto never left A1111 turns out lol
This whole time he has been updating the developmental/experimental branches
saw a whole reddit post about it
static tusk
maybe working on sdxl stuff secretly
smoky oak
also dropped a massive comment on my feelings and learnings about Vlad
seems like most people feel the same way. Just not good enough to switch with the issues it has
especially with extensions
ultimate upscale is useless in Vlad
which really sucks, cause it has a huge potential to be way faster cause Vlad excels in post processing speed
My comment on the reddit was monumental lmao
oh, my reddit is still my old username
hmmm
I am just glad to see that its not just me that thinks vlad is a bit underbaked at the moment
dense tapir
This whole time he has been updating the developmental/experimental branches
Ahhh, well they were saying he was back to playing Dota 2, heheh. No worries I left his and use anapnoes until it merges with Vlad then I dunno.
I had to go back to xformers because, for me, none of them in 2.0 pytorch with sdp was faster and I could not get a version of xformers to work with pytorch2.
Vlad's is very touchy too so click the wrong thing or do some real experimenting with settings I had to completely remove it (not just venv) to start over.
smoky oak
if Vlad can fix ultimate upscale, I will switch
dense tapir
I prefer ana now
smoky oak
it has the capacity to be like 30 or so % faster for my workflow, but it doesn't work with ultimate unfortunately
smoky oak
I prefer ana now
whats so great about this one? Just a couple days ago you were singing high Vlad praises lol
dense tapir
Better UI, stuff that worked in Auto works with it, and no Vlad.
@smoky oak Most of my extensions I use ceased to work in Vlads for various reasons. The UI is why Vlad is working with anapnoe every day as the UI is a huge step above the primitive (that is good and some bad but mostly good)UI of Vlad's or Auto's.
dense tapir
I do but it just up and died on me so really no
smoky oak
I fully deleted Vlad, and I really need a screenshot of its cuda options
dense tapir
I think I can do that
smoky oak
cuda and SD options (xformers and such)
dense tapir
if it will start
smoky oak
is there any way you could do that?
ok, thanks a million
somebody is trying to insinuate I know nothing about the optimization settings in vlad, and thats why Auto is faster, but I wanna lay out every combination I tried lol
dense tapir
I hate that and why I hate people these days.
smoky oak
Confidently wrong
dense tapir
You're welcome
Why do horses always end up with a fifth leg?
1.5 was bad about that too
dense tapir
My horse has 5.5 :/
smoky oak
@dense tapirmy response
The comparisons I did were between Auto with X formers and no additional non standard settings toggles, and included every single variation in optimized settings:
All of the settings listed below were tried in separation with these settings individually as well: FP32, PF16, BF16, all individually with upcast, torch memory, enable cudnn, allow both maps, as well as all 4 versions of enable model compile
so we tried all of those individually as well as the combinations that follow:
Disabled cross attent (OOM)
xformers (no change in speed or VRAM even after verifying the module was active and installed
scaled dot (ended up being the fastest option in the end)
doggettx's (ended up being around 20% slower than base)
Invoke AI (around as slow as doggettx's, which was about 8.2it/s
sub quadratic (second slowest at 7.2it/s) (No, I did not painstakingly manually change every single sub quadratic setting, as we already spend several hours on this, and each damn change in vlad demands a full restart of the server
split attention (absolute slowest at less than 6.5it/s)
(WE ALSO, tried all of these with xformers flash attention and SDP, s well as just xformers, and just SDP)
in the end, the fastest option ended up being Scaled Dot with SDP disable (xformers flash attent made no perf diff here, so I kept it on)
The fastest speed of around 11.9it/s was achieved by utilizing scaled dot, SDP, model compile (cudagraphs), torch memory, and cuDNN, however the issue with this system is it takes outrageously long to run the first time, and every single batch or dimension change you make requires a re save of the compiled checkpoint. For example, caching a single 15 sample 512x512 gen took about 67 seconds to save, and after that it took about 1.9s to compete (with post processing), and this climbs to several minutes when caching lets say a 512x768x8 grid. Granted, it will allow you to shred through those grids afterwards at about 20% faster than normal, but it also takes about 8 minutes to cache, so you would have to reap those benefits massively to offset that 8 minutes lost with time savings.
this time can be dramatically reduced (to about 15 seconds for a 512x512 cache) by either removing flash memory, or compiled model, but each sacrifices one metric of performance. Torch memory speeds up single images to 11.9it/s, but tanks the batch image gen by about 30% (from 1.5it/s at 8x512x512 batch, down to about 1.12it/s)
subsequently, doing the model compile only dropped single image gen to about 8.7it/s, but boosted batch from 1.5 to 1.73it/s
So as I started previously, we tested well over 100 combinations of optimization, all running the same OC as well across the two programs, and I couldn't find a single one that matched or beat A1111 in both single and multi gen. However they all beat it handily in post processing.
The biggest benefit of Vlad actually comes from big batches, as the huge savings in batch processing shaves huge amounts of time off . For example:
in Auto, a 512x512x8 grid of 15 samples in DDIM took 13.41s at 1.64it/s
while Vlad did 512x512x8 - 15 - DDIM in 11.65s at 1.51 it/s, showing how big of a deal the post processing bonus in Vlad is.
Anyways, I hope you found the frankly ridiculous amount of optimization attempts we went through up to your standards of testing the potential of Vlad to a decent extent :/
since they wanted to insinuate that I have no idea what I am doing lol
dense tapir
Now see if the troll wanders off or responds with more nut sweat licking.
Xformers will not work for <=6gb cards in Vlad and that is a simple fact. You mentioned it.
have to reduce rez so 768x768 with xformers in Vlad's is nope
smoky oak
added this:
"(TLDR; we tried a metric ton of combinations, and while some were faster, they come at the loss of nearly 8 minutes of compile time for each unique use per launch of vlad)"
smoky oak
have to reduce rez so 768x768 with xformers in Vlad's is nope
I also noted how rude he was to you in the original post about my problems with the GUI in general, not sure if you read that whole wall of text lol
dense tapir
Well, since Xformers is faster for me proven with anapnoe's and Auto both on 2.0 with sdp Vlad's is out
It is alright I ended it when he said "make a log you can have us submit in case we run into any errors.
like i haven't? have you seen original a1111 logging? its 10x worse. i'm working on it. and you're a spoiled child." He contradicted himself right there (another noticed it as well). My response to him was that I am a spoiled child and you are a dickhead so time to move on.
He deleted the ticket.
I have seen him attack others in his tickets he closed so not just something directed at only me. Others in various discords say he is a dick to people so they just stick with autos.
You try anapnoes? It is really auto's with a unique UI that is growing on me. A bit cumbersome because instead of buttons all over the place it is now with slide outs and stuff.
smoky oak
care to link me? I can take a look in a sec
smoky oak
but thanks
I'll try it and do my magic
any claims of speed improvements, or can I skip that?
dense tapir
same as in auto
The UI is very different afdter using auto's for so long
btw, you can symlink all of auto's folders to anapnoe's
That was a good thing unlike with vlad
smoky oak
oh, very nice
@dense tapirany info on if it works with ultimate upscale?
from what I am seeing, it looks like a refreshingly cut down version of Auto
dense tapir
Nothing cut down and check out the inpainting my jaw dropped as that was what I have been after since I first picked up autos
smoky oak
whats so special? I am download at the moment
dense tapir
spoilers
dense tapir
precisely
smoky oak
alright haha
dense tapir
I only found it yesterday when I went in to do a touchup
smoky oak
now that I am starting to get into python, I have some plans for some scripts I would love to run in order to make my own ultimate upscale sort of automated process that is even better
dense tapir
I like how it doesn't hold my hands like auto does
smoky oak
just download the zip and extract? Sorry, running on like 5 hours of sleep and a 17 hour day lmao
mind is a little foggy misty
dense tapir
treat it exactly how you did auto
smoky oak
troll is being oddly quiet
smoky oak
treat it exactly how you did auto
I genuinely do not remember-
assuming made a root file and extract?
is there a way to see who upvoted you on reddit?
dense tapir
if quiet was a simple trol, or not seen it yet. If it comes back with more shit it is just a sweaty nut licker.
smoky oak
I guess I should google lol
damn, looks like no
I got an upvote on my response and I wonder if it was somebody who saw me hand that guy is rear, or if it was him upvoting in defeat 😅
smoky oak
if quiet was a simple trol, or not seen it yet. If it comes back with more shit...
you always manage to use language more vibrant and colorful than the most oversaturated generations SD can make lmao
dense tapir
No idea as I just don't post there now it is far to shitty and hostile for me. Hate that place. Get rid of Karma then I would return
smoky oak
never a dull statement from you lol
smoky oak
No idea as I just don't post there now it is far to shitty and hostile for me. ...
understandable, I just wish I could see if he conceded or not lol
anyways, no worries
gonna try and install the new UI now
haven't learned the name yet-
ana
easy enough
@dense tapirJust to confirm really fast, it is as simple as download, extract to root, run?
split rover
maybe working on sdxl stuff secretly
haha we use a modified version for a good amount of internal xl testing
static tusk
haha we use a modified version for a good amount of internal xl testing
very nice
dense tapir
k
smoky oak
Just trying to figure out how to symlink
looks like mklink /J Link Target
/J being a directory juction
dense tapir
it is mklink /d old new so mklink /d auto/venv whereneo/venv
smoky oak
ok, so soft point to a directory?
dense tapir
never heard it called that before
smoky oak
just going off the descriptions the guide uses
dense tapir
here, I use this instead of all that nonsense - https://schinagl.priv.at/nt/hardlinkshellext/linkshellextension.html
used it for years and years
right click a file a directory or w/e and pick link source then head to the fold and drop as BLAH
smoky oak
hmm, interesting
dense tapir
makes it so much easier
afterall we are in a gui so this cli stuff needs to go away
in linux we ctrl+shift drag a folder or file and sym linked
smoky oak
@dense tapirI would genuinely appreciate if more installers don't auto assume you want the trash 1.5 base model
only one I have used that didn't was Vlad
Comfy did it as well
dense tapir
Thank auto for that
I sym linked it all
honest 99% could be but the last pieces can't be due to gradio (extensions for instance)
symlink
smoky oak
@dense tapirAlright... I am messing around with it and finding a severe lack of wow in the inpainting tab 😅
please let me know what I am looking for, if you don't mind
wispy nest
deep floyds?
ripe cedar
Sytan, may I ask what negative embeddings you use?
dense tapir
No
wispy nest
who can make convincing mashed potatoes challenge
dense tapir
smoky oak
it gets pontantos
@dense tapirGonna have to expirament a lottt more in here, but so far I am not super loving this interface
wispy nest
perfecto!
smoky oak
brain is also way too fried to try and learn more
wispy nest
now where's my fork and the gravy?
smoky oak
uno momento
dense tapir
<@315547788211388417>Gonna have to expirament a lottt more in here, but so far I...
I sort of figured you wouldn't tbh. It really grew on me and vlads is moving to this too. It is sleek, and my only complaint is the need to push 2 or three buttons to get from A to B when it was one but for that I had the clutter removed. I''ll stick with this over auto's cluttered mess as it was getting very out of hand.
smoky oak
yeah, i am sure I could get used to it, but I am not sure I like the use of space
in some cases I prefer looks, but with SD I think I prefer function
dense tapir
give and take but, for me, the pros outweigh the cons that were all my gripes of auto
smoky oak
although the inpaint in ana is super cool
dense tapir
it zooms in, out can move the canvas around etc...
I am glad this exists as I can free myself from the mess of auto's. He really does need a UI dev
split rover
nice clarity, which model is this?
wispy nest
but I want mashed!
and gravy, lotsa gravy
and yeah thats a good model, especially all that depth, tho the glasses are a little wonky
split rover
ah guessing nice use of pos/neg then, for XL try again when the new variant rolls out soon. Will likely feel like a very dif model than the early variant on DS right now
wispy nest
looks like fettuccini im afraid
how about just a close up on the glory of the mashed potatoes and gravy
smoky oak
Tagged with salt and pepper lol
oh, I tweaked the prompt a bit and its better
calcified potatoes
potato fossil
split rover
leftovers
wispy nest
this chat is making me hungry
wispy nest
or maybe its just the weed lol
smoky oak
let me send some images from my comission set lol
wispy nest
comission, eh? big timer in the hizzay. nice
smoky oak
Comissioned by the head of Ai development at Unity 😅
he's my bud
we worked together on the project that unity bought, and now he is head of his own AI department
look at these berriessss
😩
wispy nest
yeah pretty stellar. little ai-ness here and there but overall top class
smoky oak
yeah pretty stellar. little ai-ness here and there but overall top class
thats relatively what he is paying me for lol
wispy nest
oooh nice one peter
smoky oak
I can't disclose too much, other than these hand generated images are being used to hone in one of Unity's next gen AI tools
food fight!
behold, my eats
smoky oak
you're too laters
dense tapir
I have a price alert for a bit now for my case and card with both being 100 more than MSRP (both went up two weeks ago).
case is at the price it was before the spike due to mail in rebate so that will go back down after (I no longer do mail-in rebates but did for decades).
smoky oak
@wispy nest LASANGE
dense tapir
lasagna, yummy
wispy nest
damn. now that one would straight up fool me
until i look at that garlic bread in the lower left a little more haha
dense tapir
In an ad on TV you would never notice
wispy nest
yep
mama mia!
lighting on that one is great, looks hella real
im lactose intolerant are they made with almond milk?
I like your shop pete, you serve big scoops
dense tapir
wispy nest
Pete says, bring it lol
smoky oak
Pete and I have different styles
Mine is optimistic realism, and pete is realistic realism
smoky oak
I do have an edge here, as my custom model mix uses offset noise :p
I really do envy how you manage to get such authentic photograph looks tho
Some moody icecream promotional photography
dense tapir
Give it time as it will grow on you.
wispy nest
looks like Swamp Thing's trouble making younger cousin, Rooty
wispy nest
schweet
smoky oak
just prompt, and selective high res fix (now that I know how to use it right lol)
wispy nest
Ben Affleck as Elvis challenge
smoky oak
I have a couple LoRA's I have made for some celebs
dense tapir
I don't get why I keep getting this on its left shoulder?
smoky oak
I need to figure out how to reliably generate extremely high detail images
hmmm
Also, this one looks pretty damn good IMO
sinful mirage
space car
dense tapir
smoky oak
Woah-
dense tapir
smoky oak
I changed up some prompting, and the level of detail is kinda insane
yeah, with some additional modified positives and negatives
there is actually too much detail lmao
dense tapir
smoky oak
Every single hair and micro wrinkle lol
bro, I don't get how you do it so well lmao
They have such realistic lighting
like, mine have more detail, but it doesn't look as "legit"
just got a cursed idea lol
dense tapir
sinful mirage
revAnimated_122 vs vinteprotogenmix_v20 with the prompt "([beautiful forest valley with trees and an amazing sunset | portrait of a man]:1.3)" with some extras at the end. The first is more visually appealing somewhat but it doesnt seem to combine the ideas like the 2nd model does (ive tested it on a wide range of prompts, 2nd always does better). Anyone know of others that do well with this kinda prompt? Kinda niche so i doubt but maybe someone knows
sinful mirage
the heck
wispy nest
awesome! ha, can you try him as Vegas Elvis? I can see Affleck fitting that look haha
sinful mirage
"[man | cat | woman | dog | monster]"
smoky oak
I am getting there-
wispy nest
he's a hunka hunka burnin' love!
wispy nest
Give your granny a kiss
I wonder does the model recognize Granny Goodness?
(DC Comics character haha)
getting there
might just hit a limit with the tech/model too at some point
smoky oak
@stone cipherI am finally getting more realistic results :D
sinful mirage
ok hear me out, this one is really trippy, [dog | car], almost like an optical illusion
dense tapir
smoky oak
working on it still, but I think this is a big improvement over the porcelain dolls I was getting before
I am going for realism, but a little more modern than you do
yours have that cool vintage look
very convincing vintage look, I might add
also
I am noticing that poses are wayyyy more dynamic
much less staring forward with hands to side
definitely more dynamic poses
dense tapir
smoky oak
ah yes, the void hair
dense tapir
My first one Clyde would not allow and no nudity
smoky oak
ok, now I am getting much more realistic images lol
@stone cipherMassive thanks for showing me the results I could achieve and inspiring me to try and get them haha
wispy nest
love it. especially his shirt, making me want some lemonade though hahar
dense tapir
wispy nest
lol woops typo haha
smoky oak
lol woops typo haha
SMH my head man
wispy nest
Elvis as a Jedi Challenge
smoky oak
also... This image is not realistic, but oh my god
BRO
IT GOT
THE
SUBSURFACE SCATTERING IN ITS EARS
HELLO
wispy nest
noice. very noice
smoky oak
Can I get a puppy in the same scene
hasty nova
SUBSURFACE SCATTERING IN ITS EARS
this has got some subsirrface scattering
hasty nova
(there is a button on the top you can see that enables a cat, and you have three options in the settings)
smoky oak
I might do it-
hasty nova
perfect
wispy nest
Force is strong with that one, Pete!
hasty nova
https://tenor.com/view/skrunkly-meow-epic-cat-cool-flu-gif-25544332
beautiful
smoky oak
reef remnant
hasty nova
man these eyes are awful :(
smoky oak
wow, those are intense
hasty nova
I did put "big eyes" in the prompt
smoky oak
over here with the OwO eyes
smoky oak
Hey, I trained vicuna 7b to always talk in uwu speak lol
smoky oak
very pretty textiles, but textiles nonetheless
oh! this reminds me of an insanely cool result I got when I was using a random word generator
hasty nova
textile
smoky oak
hasty nova
oh! this reminds me of an insanely cool result I got when I was using a random w...
oh thats a pretty good idea. Was it just using one word or were you combining multiple random words?
smoky oak
the generator would select 5 random nouns
hasty nova
also now I am doing UwU
hasty nova
I would put them in and just see what it made
imma do that
smoky oak
also made this one
hasty nova
Oh UwU actually provides girls
smoky oak
imma do that
let me find the website, it copies them with spaces and commas
smoky oak
bottom right lmao
hasty nova
(these are all on 2.1)
hasty nova
thanks!
smoky oak
copies the words with spaces and commas
hasty nova
first prompt "Forearm, Sad, Geometrical, Rend, Cart"
smoky oak
these came from the random words lol
hasty nova
smoky oak
oooo
Also the random word generation
oh also
it taught me that the Ai does understand motionblur
and it can do it really damn well
smoky oak
Also learned how to use single color focus
these generated like this, no post processing
hasty nova
aww
smoky oak
these kittens are gonna kill me
hasty nova
im just gonna put in "a cute cat" and see what I get
smoky oak
generates jinx
hasty nova
smoky oak
gato's
smoky oak
actually perishing
smoky oak
OMG
hasty nova
I'm pretty proud of the model used to generate these, I have been trying to make v2 (with improved photorealism) all weekend and been failing because it doesn't seem to save the model right. I have one last thing to try overnight and I and desperate for it to work
hasty nova
awww
smoky oak
ok, safe to say I have the magic of generations coming back again
@stone cipher again, massive thank you, this has resparked my joy of generating again
look at these images, omg
I am not even using an anthro model
smoky oak
do an alligator
one of the AI's weaknesses-
hasty nova
:)
smoky oak
What model are you using?
custom mix I am working on
hasty nova
Oh nice
smoky oak
oh! I have trained some animal LoRA's, and they are highly compatible with anthro models too
hasty nova
I've been using a trained from scratch 2.1 model on some hand captioned images
smoky oak
I made one for Capybara, one for ferrets, and one for opossums, and I am working on more
the opossums came out outrageously good
hasty nova
there's a chance general awareness is using it rn as well, but he could be using some other model I am not sure.
smoky oak
like, I have never really seen opossum furries, but I cannot believe they are not more common after I saw how adorable they are
Like I mean
hasty nova
My model currently is mostly best at people but I am really hoping for v2 to improve more aspects
hasty nova
Sure
smoky oak
What I have learned is that the less negative words, the better it seems (not too little of course)
hasty nova
I usually keep my negatives between 2-7 words I think
smoky oak
my go to negative is fantastic. I'd say 99/99% of the time it blocks all undesirable things
however... I am finding now that it is also why my gens look so samey
hasty nova
my minimum is "low quality, messy"
idk why but messy seems like a good word to have
smoky oak
also, to get more realism, its a little counterintuitive it seems
I am actually making words like "perfect" and "flawless" negatives
cause realism is imperfection
hasty nova
well yeah
asymmetry
I really like the reflections in my model
It can do some sweet images with the "rain" token
Anyway I am once again going to start training v2 and so I should close discord to save resources, goodnight everyone!
smoky oak
my own little Elvis test
split rover
will try a couple ;D
dense tapir
I was reading about AMD and even the 6700 blows me out of the water from before all the new stuff was released. What I did was take some of their settings and gained 10-15% speed. 38-40s now 35-36s.
basically treating my card like a hybrid of AMD Nvidia since I have cuda I can run xformers but I lack other things AMD lacks
The dead are rising?
@smoky oak One of my really little content creators I watch (450 subs and growing) was showing how to do AMD and he was showing some anime to test with. Look at this comment.
smoky oak
Some people can't be helped, FFS
dense tapir
I know, it is a mental illness
smoky oak
Like sure, use AI for NSFW if you want to, but have the decency of being private about it
God
dense tapir
Hehehehe
I wish I could find a creator for AMD to actually sit there and see their numbers but for the RDNA3 cards.
Oh, in the github thread I grabbed those AMD things from they said for more speed use the pytorch2 AMD branch AND turn off sdp optimization. ffs, sdp is slower on mine too so I wonder why it is exists? Is it really just for the 4090 to speed it up?
smoky oak
Custom mix currently
wispy nest
just splurged on MJ, gotta say, its pretty nuts
fair wagon
there a place I can get help with Stable Vicuna?
dense tapir
Anyone know how to erase the thumbnail pics in the checkpoint picker in auto1111?
wispy nest
Anyone know how to erase the thumbnail pics in the checkpoint picker in auto1111...
if you mean the preview image, it's just a png in the model folder. you can delete it there
dense tapir
if you mean the preview image, it's just a png in the model folder. you can dele...
Really?
@wispy nest Yep, there they are. Thank you. 🙂
dense tapir
@smoky oak Well, at least I know my card is fine for its VRAM and stuff.
wispy ether
<@296720689468145664> Well, at least I know my card is fine for its VRAM and stu...
which one
dense tapir
@wispy ether EXTREME portable is what I have
wispy ether
3060
dense tapir
good to know and see those figures
nimble holly
#1072220168534642768 a man
wispy ether
<#1072220168534642768> a man
women
nimble holly
how to make a image?
kind quartz
i am peaceful person, but those reminds me one spiderman enemy
wispy ether
how to make a image?
get a pc
dense tapir
wispy ether
i am peaceful person, but those reminds me one spiderman enemy
whats ur prompt
i wanna try
kind quartz
The_aliens_characters_are_in_the_cinema_looking_out_of_camera
nimble holly
@wispy ether ok now how?
kind quartz
without _ those are spaces
wispy ether
without _ those are spaces
wdym
nimble holly
how to make image@wispy ether
wispy ether
how to make image<@567777599296897045>
get a pc
terse hamlet
How to generate?
nimble holly
ok nep
wispy ether
bru
kind quartz
#🏞｜general-with-images message
Without spaces, mmnt
3088227482 The aliens characters are in the cinema, looking out of camera
wispy ether
terse hamlet
How to generate image ?
kind quartz
with seed
3088227482 The aliens characters are in the cinema, looking out of camera
wispy ether
i am peaceful person, but those reminds me one spiderman enemy
how is that a cinema
semen
kind quartz
DDIM i did it in
kind quartz
50
kind quartz
maybe 32 steps
oh yes deliberate v2
wispy ether
modules.devices.NansException: A tensor with all NaNs was produced in VAE. This could be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.
🗿
wispy ether
maybe 32 steps
oh yes deliberate v2
can you send me
kind quartz
image?
dense tapir
kind quartz
768x512
kind quartz
but i believe you can get all info from image i posted.
o.k.
kind quartz
i will delete it as somebody can find image bit on edge
dense tapir
TF2 style.
dense tapir
@wispy ether on your 3060 with 2.1 and 768x768 just a prompt (no embedding, locons, lora, etc...) what is your it/s?
kind quartz
spider woman
kind quartz
random pendant
"Dress for the job you want, not the job you have. Apparently, Jan's dream job is to be a stripper".
I never thought I'd dig spiderwoman
I'll start making some unreal-engine style now
kind quartz
🙂
dense tapir
random pendant
Alternative version, if this doesn't convince anyone that dreambooth is far superior than loras or TIs, I dare you to make a better one. For me it looks like I have a new wallpaper. Wondering what kind of background to give her.
kind quartz
for walpaper dark is best i think
sharp merlin
lord have mercy
random pendant
does it do photorealistic
storm badger
opal edge
hey guys can someone enhance this image? it's about human army fighting a dragon and a demon some soldiers are on a waterfall
this is the best that i could do
dense tapir
open sun
if anyone has a dataset and wants a quick lora/lycoris training on my 3090 for free - let me know, currently queue is open.
dense tapir
crystal bloom
SD + Kaiber AI
wispy nest
kind quartz
i think it is those socks competition 🙂
next vale
Hello
kind quartz
Hello
next vale
I'm looking for a LoRA for western political cartoons. Specifically, the second half 19th century, turn of the 20th century. Thomas Nast, J. S. Pughe, John Tenniel, etc. Does this exist?
smoky oak
I am really happy with the realism I have been getting out of my custom model now that Peter has helped me revise my prompting
wispy ether
<@567777599296897045> on your 3060 with 2.1 and 768x768 just a prompt (no embedd...
idk what version i use
how do i check
smoky oak
@dense tapir please do not get into this-
smoky oak
There is no discord bot
wispy nest
-_-
crystal deltaBOT
smoky oak
Ah! It's working again!
wispy nest
ok i founded the bot thatnks anyway
smoky oak
There is no bot
wispy nest
A model, either lora or whatever can generate models using those same socks.
I can use either a model or pictures, but first I want to see what can be done.
cool wolves pete
are they going to evolve to domesticated dogs with later gens? 🤔
smoky oak
@wispy nest did you see my new ventures into better realism?
I am quite proud of this one
It's not perfect, but it's definitely more realistic than a lot of my other animal gens
wispy nest
very nice
kind quartz
@smoky oak what model?
kind quartz
thank you!
smoky oak
No problem
I'm hoping I can get Peter and I to be the resident realism experts lol
His being vintage realism, and mine modern realism. That would be really cool
paper bobcat
wispy nest
old guy kinda looks like Noam Chomsky lol
that one would fool me as a genuine flapper era photo pete
you could maybe win a competition with it like i saw some other old b/w did haha
smoky oak
kind quartz
seems hdr i dont like much hdr personaly. otherwise good photo
kind quartz
golden 80s
smoky oak
I'm trying to generate some cute wedding photos, cause I'm lonely lol
Not my style, personally 😅
I'm getting too powerful lol
Nevermind lol
These aren't what I meant, but thank you 😅
wispy nest
looks like twins getting married haha they got the same dang nose
smoky oak
looks like twins getting married haha they got the same dang nose
Omg they do!
I didn't even notice lol
calm canopy
Yeah I remember someone asking about how to prevent LORAs from doing that
Makes sense that they alter both faces.
wispy nest
first couple essentially has the same face too, not sure how to get around that
calm canopy
Should be possible with latent coupling
smoky oak
Should be easy just with prompting TBH
wispy ether
@smoky oaki was using the wrong boot file all the time 
i mean
they worked
but didnt kinda
now i booted user-gui.bat and everything works
gpu gets recognized
how do you generate a image as test after training in kohya
hasty nova
cool wolves pete
you saying this made me realize I could have been reading @stone cipher name wrong this whole time... I have been reading it in my head "peh-ter-mors" as one single word, it is supposed to be "pee-tur mors" like the name Peter?
Welp I have been reading it wrong this whole time.
random pendant
latent coupling + inpainting
make sure you leave blank space in between the two sketches
wispy nest
@stone cipher which model you are using for creating theses images?
@stone cipher can you share huggingface face or GitHub link
@stone cipher okay, thank you for answering
kind quartz
hey basset hound please
hazy kite
Turns out watching a phoenix rise from the ashes may very well be the last thing you ever do
paper bobcat
brazen anvil
wispy nest
noice
I don't want you to edit the image, but a model or images that can be generated with that sock.
wispy ether
dawg
smoky oak
Spinchbub
brazen anvil
What animals are these supposed to be?
capybara and a marmot
calm canopy
Looks kinda nice tbh
wispy ether
I actually have a LoRA I made just for capybaras lol
.........................
.
.
l
.
my table is wiggling
fuck
Generating Samples: 100%|████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 3.60it/s]
Steps: 39%|█▉ | 412/1050 [03:34<06:24, 1.66it/s, inst_loss=0.0234, loss=0.0234, lr=0.0001, prior_loss=0, vram=10.4]
brazen anvil
lol i was thinking in the same, made a lora for capybaras, marmots and axolotl
smoky oak
Capys
smoky oak
:/
wispy ether
tf is pickletensor
smoky oak
A format
wispy ether
big brain
smoky oak
Nothing that concerns you, it's not for local running
smoky oak
I hope you didn't make it in public
wispy ether
?
smoky oak
Or wait, they changed it now I think
wispy ether
i dont get whatr you mean
smoky oak
Not long ago, when you would make a Civit AI account it would instantly plaster porn all over your screen
smoky oak
By default they removed the NSFW filter on all accounts
wispy ether
they have a option
smoky oak
Yeah, but it was automatically off all the time
smoky oak
So if you wanted to sign in, it would plaster porn before you turned the filter on
wispy ether
nais
after one more training
any tips to fix the fingers
your gf
@smoky oakdoes graph smoothing steps change the result?
crude kraken
I’m looking to make images of people (and objects) with transparent backgrounds. Is this something I can do with existing models and at what point does training a custom model make sense?
smoky oak
<@296720689468145664>does graph smoothing steps change the result?
Not too sure, never messed with it
smoky oak
I’m looking to make images of people (and objects) with transparent backgrounds....
One way that I do it is by utilizing a noise offset model, and then weighting either a solid black or solid white background, so then when I generate the image I'm able to select the subject and remove the background
crude kraken
One way that I do it is by utilizing a noise offset model, and then weighting ei...
I’ll check that out
smoky oak
model is ruined lmao
Oh my God lol
wispy ether
Oh my God lol
i went back to old checkpoint
calm canopy
Wait I thought these were LORAs lol
wispy ether
Wait I thought these were LORAs lol
same
but somehow i made a model
idk
it didnt react to lora:
calm canopy
I think filename wise, it should still be a ckpt or safetensors if it's a lora
But not reacting to the lora call of it is strange
smoky oak
Yup, that's a model
smoky oak
How in the world did you manage lol
calm canopy
Jesus lmao
calm canopy
I thought sytan showed you how to create LORAs earlier
calm canopy
Kohya's has only ever treated me nicely
smoky oak
I thought sytan showed you how to create LORAs earlier
Oh man, I as well as like 10 other people spent over 7 hours on that :/
wispy ether
Kohya's has only ever treated me nicely
where do i view results there
calm canopy
Hahahhh I remember reading that this morning. I think it was on palkia
smoky oak
Yeah, my results ended up working
smoky oak
it was a model
Explains why it had the issues it did
I know what you're doing wrong
Screenshot the UI
smoky oak
You need to use the dream booth LoRA tab, not just the dream booth one
wispy ether
calm canopy
I've only heard bad things about the dream booth extension.
wispy ether
if i cant view i dont know how it trained, if bad or good
smoky oak
Not viewing >>> having to deal with the problems you are having
calm canopy
100% install kohya's gui, it's worth it.
smoky oak
We all use kohya ourselves
And you can checkpoint your epochs, so even if it trains bad, you have multiple time saves
wispy ether
100% install kohya's gui, it's worth it.
and how do i view then
smoky oak
Load into auto as a LoRA and use it like any other LoRA
wispy ether
what do i even put there
smoky oak
That can work
smoky oak
Or, if you have a model you wanna train it for, select that model
calm canopy
and how do i view then
Sorry not sure what you mean by view. Do you mean the loss graph? I wouldn't worry about it tbh.
smoky oak
I always train my LoRA's off a realvis base, as it's what I use the most, and its highly compatible with my custom model mix
wispy ether
wispy ether
Sorry not sure what you mean by view. Do you mean the loss graph? I wouldn't wor...
view the result
if training made it worse or better
smoky oak
You load it into auto just like any other LoRA
calm canopy
There's a pretty simple process for that, that works with every lora training. Try it out with a prompt and see how it all compares
wispy ether
is a hdd making weird scanning sounds?
smoky oak
You can also do an xyz plot of the strength and epochs of the LoRA
calm canopy
If it's too " burned" or saturated, you know it's overtrained. If not, pick the best one and you're good to go. Assuming you're saving different epoch amounts of the same lora.
Are you only training 7 images?
I guess that "can" be okay
calm canopy
Not gonna full on say that you need more because even aientrupeur showcased a lora on literally 1 image being trained. Ideally you have more images tho.
wispy ether
what does v2 do
calm canopy
how many epochs should i do
This amount varies and it's 100% on you to figure out which amount is best.
smoky oak
The palkia one I made was made with 4 images
wispy ether
calm canopy
The palkia one I made was made with 4 images
Yeah and that one looked decent.
wispy ether
if i enable it and set the right path
keep off?
uh
even with it disabled
thats why i dont like this one
it just doesnt work
now it works
i used old config
😎
RuntimeError: The size of tensor a (768) must match the size of tensor b (1024) at non-singleton dimension 1
???
File "C:\Users\Lukag\Desktop\stable-diffusion-webui-master\extensions-builtin\Lora\lora.py", line 273, in lora_apply_weights
self.weight += lora_calc_updown(lora, module, self.weight)
wispy ether
its something with the weight
smoky oak
It's actually insane how many problems you manage to have
I have never had a single error with auto or kohya
Bugs, sure, but not errors
smoky oak
For kohya?
wispy ether
ye
smoky oak
Just gui.bat
wispy ether
Just gui.bat
what model
smoky oak
Whatever model you are training for
crude kraken
For example, you could use it like this
yup makes sense, will play with getting the right background. Seems like there’s a few different ways to remove the background too (without using photoshop), I just need to do some more research.
smoky oak
Sounds good, hope it works for you!
Doing a solid white background may actually beore beneficial for removing
sharp heart
#ask did any of have some prompt for mechanical animal?
smoky oak
#ask did any of have some prompt for mechanical animal?
I am sure I can come up with something
@dense tapir that troll never responded lol
sharp heart
30% head bionic mechanical with one shinning mechanical eye, steam, glowing tube, t-shirt design, retro aesthetic, focused on the character, 4K resolution, using Cinema 4D, pastel tetradic colors, cute and quirky, fantasy art, watercolor effect, bokeh, Adobe Illustrator, hand-drawn, digital painting, low-poly, soft lighting, bird's-eye view, isometric style, retro aesthetic, focused on the character, 4K resolution, photorealistic rendering, self portrait, white background,
I use this one but the image far from I want to see
I mean I want something close like cyborg. I just curious it end take my sleep time xD
smoky oak
Made a robotic wolf lol
wispy ether
sharp heart
I got this one xD
smoky oak
@sharp heart
dense tapir
<@315547788211388417> please do not get into this-
Yeah, I am not going down that rabbit hole.
smoky oak
thank you lmao
dense tapir
Not bad
I am about to get giggety and branch into vids so I need a 4090, sadly. I was just informed my OOS may not be coming back at the old price. ffs
smoky oak
You know, I just realized that if I can get this job I am trying to apply for, I could write off a super OP PC on taxes lol
smoky oak
though, its not like I will be hurting for funds if I can secure this job 😅
dense tapir
I despise this setting - Variation seed. I wish we could make it stay off until otherwise directed.
seems pnginfo to txt2info loves turning it back to -1
smoky oak
giving vintage a try
smoky oak
dense tapir
I lived in a city for a long time with a silent film star. Nice elderly lady that lived, oh wow maybe to a 100?
Those made me remember her.
Her
smoky oak
dawg
dense tapir
According to the story put forward by her studio publicity department, she was “the last child born in the 19th Century—a few minutes before midnight in the final hours of 1899.” This date is even on her headstone. However, other biographies of her life list her birth year as 1897, 1894 and 1901. She was around 90. Just a normal person which was refreshing.
hasty nova
reminiscient of the first time you tried digital diffusion
you made something like that
dense tapir
I need to eventually try controlnet for 2.1 but I don't have the drive space for it
hasty nova
well it looks like you have controlnet, and also what model is it?
feral pebble
disabled control net. v2 model
hasty nova
Ah it's a v2 model. Is it the 768 model or the 512 one?
feral pebble
it might be doing better with "phoenix bird". perhaps phoenix could be too many things (cities and junk?)
hasty nova
Could be
hasty nova
Ok, so your resolution is ok. But the base sd 2.1 models aren't very good, I would reccomend trying a finetuned one.
hasty nova
I wonder if 768 will run on my 11gb (12gb) card
Absolutely it should. I run it on my 8gb card.
feral pebble
oh interesting. I've not yet delved into any fine tuned. do you mean a lora addition or a fully fine tuned?
hasty nova
In fact I have trained a sd 2.1 768 ckpt on my card.
hasty nova
oh interesting. I've not yet delved into any fine tuned. do you mean a lora ad...
I would reccomend a fully fintuned model. I am working on v2 for mine, v1 is currently avaliable on civitai, it's called Digital Diffusion.
I got this in one go with it.
feral pebble
cool.. I've not done any fine tuning. How many images do you need to supply? And do you need caption pairs? (a text match for each image)?
hasty nova
cool.. I've not done any fine tuning. How many images do you need to supply? A...
Well, my current dataset was around 300 images, all hand captioned. v1 was 250 images. But datasets don't need to be that big necessarily, I was making a general purpose finetune. If you are trying to train in phoenixes in general just train a lora or embedding.
That's my model
feral pebble
was just helping a friend make something to make his kid happy
what's the textual inversion one for?
(did you make it or is this the one you used as a base for your fine tuning?)
hasty nova
what's the textual inversion one for?
It's like a lora but a bit different. It trains tokens into a small file and you use it in your prompt.
hasty nova
(did you make it or is this the one you used as a base for your fine tuning?)
I made it.
feral pebble
ohhh.. wow.. that's an interesting technique
hasty nova
The base was the normal SD 2.1 768 model.
v2 will do better photorealism, while still being able to do good digital art.
hasty nova
It's like a lora but a bit different. It trains tokens into a small file and you...
My civitai page has a lot of embeddings from back when I used to make those for 1.5.
feral pebble
very cool. it's been some months since I worked with SD, being involved in too much else in life.. so many improvements 🙂
