#🏞|general-with-images
1 messages · Page 23 of 1
Orange cats are genetically disposed to be GOOFY
and he is simply a perpetuation of the truth of that
I want to do the same thing
I have a biiiiiiig cat
he does the wierdest noises
he is the same size as my dog
fluffer
oh hes just a goober, just like mine hehehe
He's a banana lover
Nasty little face
I love scrungey cats


I am something of a...
scrunge connoisseur
you'd fit well in the mod channels
not sure if this is under NDA or not but damn we love cats
I loveeee cats haha
I don't care, I'm taking the risk, I leak cat infos
same
My love for cats only extends, the bigger they are
I need a savannah cat/Bengal/caracal in my life
A giga goober
ah yes
I'm a little afraid of how much bigger he'll get :p
but it will be even warmer for hugs
Maine coons are awesome
and such a long fur
that's just the problem though
needs lots of care on that side
I thought he looked like one, but I didn't know they could be orange
yeahhh, I can't do long fur
at last ! an emoji !
Anybody know how to get higher resolution details out of ultimate upscale?
My results are suddenly coming out fuzzy, and I am not sure why
my ultimate upscale results have significantly lower detail, and I am not sure why thats happening
Hi Guys
I just started to learn about the Stable Diffusion & control Net. That's makes me to be here.
Denois maybe to high
Even with extremely low denoise, it's just bad :<
tried a different sampler?
That did help some, but introduced clear repetitious patterns in the sky
I am not sure why this is suddenly an issue when it hasn't been before :<
Maybe you forgot to delete the prompt field
That will cause issues in upscaling if you keep it
You're... Supposed to?
Whatttt
I have done probably 100 generations with the prompt still in, with no issues! (Prior to this lol)
Yea but if the denoise is to high even with ultimate sd upscale like 0.3-0.5 i got artifacts in my images
I usually use ultimate denoise at .15 or lower
If you delete the prompt you will have no issues
I change my prompt from the description of the subject to something like "hires, masterpiece, 8k" for upscales
Yea it wont happen with low denois
It surely depends on the Upscaler too
Im mostly using esrgan based ones
@dry crow sadly it didn't help 
Before upscale (native 2560x1080)
After upscale (5120x2176)
(Not sure why it's not exactly 2x
It just looks worse
If anybody has any ideas why, I would be really curious to hear
I am just trying to get some nice and clean high res gens, but it's really not working for whatever reason
(this is humor and not my opinion)
that AI forgot to use controlnet
ControllNet is an exertion to AI models, it can also be fixed by human designers that have own hands as reference 🙂
But I agree, ControlNet is great for that.
Cyborg Halle Berry 😏 follow my IG : @ GlitchzAi 🙏🏾
Which Upscaler and settings?
i had a plan like that a while ago. trained a lora to get a semi-consistent character, but eventually decided it was waaaay too much work to get it to work properly using just sd. things would be easier now with controlnet and things like the composition extensions, but still it would be 100x easier using third party software in addition to sd
that being said, i'm still kind of obsessed with the grainy 90s style comic look or the cel shaded styles. currently running through the 1800+ list of artists someone posted from the sd database with a comic style prompt to see the various results. would've been a lot faster if i hadn't decided i was going to run it at 768x768 with 1.5x highres resulting in 1152x1152 size images only on a 3060 so it's taking a while. nearly done though...about 550 more artists to go. about 8 hours remaining -_-
Lora is one way to get a consistent character, I think I will stick to simply ways for now, just to see how far I get. Eventually I might use control net for this, but not at that stage yet
easiest way without using loras is probably just to use a real life person if you're just doing it for testing. here's a link i looked at when i was trying to solve the issue https://mythicalai.substack.com/p/how-to-create-consistent-characters
those aren't terrible suggestions and worked pretty well a month or so ago when i was messing with it
Thats what I did
Wrote a detailed description of the character and used in each batch, seems to work so far
I was banned for 5 hours, for an image called "nsfw", when not at all.
When I see this, I tell myself that the bot is necessarily an AI
I said that to joke about the infamous image of xyz. I don't care about Jean-Jacques ban, discord modo and
main shareholder of Biactol.
Sorry I'll remove this then, I didn't get the joke, didn't want to put you on the spot
Discord modos and their tendency to force people to live in their self flagealation.
It is as lamentable as it is despairing.
Sorry you feel that way. Just people doing what they can, and sometimes misinterpreting. I will let you, I wasn't looking to make you go through a bad time.
no problems, you have nothing to do with it x) (dunno if this is a good translation)
Pas de souci, tu n'y es pour rien x)
Ah ! c'est vrai que tu es français !
We are quite some baguettes currently around here
I take everything for its face value lol yeah...
Brassens is still my best boy tbh
https://www.youtube.com/watch?v=otNFpUdJ6xU
I'm not far from it so I'm nostalgic 🙂
👕 T-SHIRT, SWEAT, MUG, ... Paroles de Georges : https://la-boutique-des-fans.myspreadshop.fr/
L'histoire de mon Diaporama : 25 ans après avoir découvert cette chanson, j'ai enfin pris le temps d'aller découvrir Sète et de rendre hommage en photos à Brassens et à sa ville natale.
L'originalité du sujet, la richesse du vocabulaire, l'intellige...
So apparently NVIDIA Canvas can do really nice looking panoramas now
if you were to put that in a 3d panorama viewer it would look pretty coolk
pretty sure that's where the seam is, can barely see it
only noticeable area is the water
works pretty well in blender too
Logo design fish market
@peak sage
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
Yesterday I imagined this. Now it exists. What does it mean?
Happy you like it, and other than coincidence i can't say anything. This is the second version I do on this theme. The models are quite good at this
:
Page one of my comic:
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
Thanks! I really like that program, it's pretty relaxing to just paint realistic landscapes with ease, and with the new panoramas it opens up a whole new toolset. All you need is an NVIDIA GPU I think, the better ones are more fun to work with though.
I used canvas a lot back in the day, didn't know it was still being developed
Tell me again 3060 vs 3060ti, please.
Sorry?
You were anti 3060 and I forgot why ti over 3060.
For most cases, the 3060 has slow and useless amounts of VRAM, so I always recommend the 3060ti over it cause for about 20% more money, you can see anywhere from 20-100% improvements in computation
It's got more cores, higher clocks, a way faster memory bus, and much faster memory in general. And you don't really need more than 8GB for standard generations in SD.
I am doing 4k gens on 8GB VRAM
But in your case, where you care about VRAM capacity over actual performance/value, it makes sense to spring for 12GB much slower vs 8GB much faster
I hope Pytorch2 comes to automatic because it speeds up generating by about 1.5x from all the numbers I have gathered. HUGE differences
for inference 768 sq but I can't find diddly squat for training on most cards which makes it hard.
Lemmie test against my 3060ti, cause that seems low
You would think the above chart would be out there for training too not just inference as a lot of cards fall flat with training while are great at inference.
Mine seems to be closer to 4.15it/s
Though it does fluctuate
Sometimes it's down to about 3.8
Sometimes up to about 4.25
For a face without the forehead do I add the missing part in paint then I run inpaint?
If the 4080 were 800 it would have been mine last month.
I still suspect Summer we will know how this will go for cards
mickey mouse by botticelli, i dont know what i was expecting
My friend has a 4090, but then he ruins it by not running xformers
My 3060ti can generate high res images than his 4090 cause he is so hard headed about Xformers
What you mean by Xformers, can u point me to a decent tutorial
Do you use A1111?
Yea
Open the bat file you use to launch it in a text editor and just add
--xformers
After command line args=
So like this command line args=--xformers
And you're good. That's all you need
Next time you launch, it will be slow cause it's gonna download it. If you have an Nvidia card, it should gen way faster and use less VRAM
Interesting I’ll give it a shot tonight, will that increase my speed on a 3090
Oh yeah, for sure
It's around a 50% speed increase on a 4090 IIRC
It was about 60% on my 3060ti, and it allows you to generate with way more VRAM than usual
Ok I remember something about that, didn’t know I had to edit the file
Thanks for the info
I went from OOM on 768x1024, to now generating native 1080x2560 no problem
Nice
If I generate at that frame my card wants to explode
So much stuff to learn and to much fun generating random stuff, not enough time
It made a massive difference for my 8GB VRAM, I could only imagine how much it would do for 24
Honest answer here but if I had 24gb I would not run xformers either.
for training I 100% wouldn't
A lot of the big anime trainers turn off xformers now as I watch them talking to each other and xformers screws shit up. There is a cost for speed.
I am starting to heavily lean back to 16gb server card P100 as it is dirt cheap, faster at this than a T4 colab gives us. No tensor cores but can't have it all.
The problem is cooling the darn thing as it is meant for a server and I don't wish to go deaf
The real card to get if going that route is a V100 but being they are still in production and heavily used they are as expensive, or more, than a 4090
Positive side is they use less power
which is one of the main reasons datacenters use them who may end up with several hundred
Im tempted to buy a 4080 or 4090 as Ive been upscaling my images a lot now and its quite a lot of sitting around waiting
Did you see my chart?
what card do you currently have?
Pytorch 2 we need as it give 1.5x speed gain
damn near 3 times faster for a 4090 over your 3080. 2-3 times
yeah its totally insane, plus the extra vram to play with
24 vs 12, yea
mine has 10gb
at launch I wasnt bothered as I was just gaming and the 3080 smoked the 2080ti
its been a great card for 10gb. 4k gaming it can strruggle
I have always been more vram even at the cost of speed
my current 1060 is because my AMD 7870 got smoked and was glitching and locking me up
it was in the days before it would throttle due to heat and one fan died on it and poof, damaged card
dont they have thermal protection and throttle down?
ah ok
that gen, and before I had so many burnt up cards due to a flaky fan, or a fan that just failed.
Nvidia ones too but I dunno when that came for NV as I was AMD for a long time after geforce2 MX
I think so
prior to maxwell I think no
I saw a lot of 7xx cards doa burnt
anyway, I really didn't want a 1k card but had no choice. I still have the 7870 too as it works on the desktop just game and lockup, or benchmark it gets glitches even though I replaced its fans it was too late
I've always had AMD and leaned towards higher vram cards. I had a rx 480 with 8gb and people were telling me I was mental and would never need that. Then GTA V PC release landed XD. Also, always texture mods, always
yeah for sure. more vram is always nicer
Vega64 I would've liked to have more vram on but it kept me cozy with 8gb for a few years
even 16gb @ 1440p wiont last long with games coming out the next couple of years
UE5 engine games are going to make gpus cry
you know these gpu makers aren't going to suddenly give us plebs 32+gb of vram, right?
I remember nvidia did some bs with the 970 generation. Something about having 8gb but it was fake and way slower if you used over 6gb or something
they are in a rocky place due to business
GPU makers work with games and OS creators to make sure the market can actually use the tech it gives out. Games have no reason for 32gb vram cards yet
look how they strugle to give us even 16gb let alone 24gb. Almost as if you are kicking them in the nuts to get it.
Games will start dealing with massive datasets though, in very creative and fun ways, and that will require a butt load of memory
YET being the keyword
GPUs are going to become less about processing pixels and more about managing these massive data sets too
32gb of vram cards being the budget card I can only imagine the prices.
be me and remember 500mb HDD's the size of a cinderblock for $2000
look at tech prices now a days, sigh relief
they just need to get the price of HBM memory down
LOL, I go back before that where 20 MEGA byte harddrives were a thing and seeing that on a personal computer with me a 1.44mb floppy
My first HDD was in 1996 and 2.1 GB Western Digital.
that was an expensive purchase
i did the budget cheat and bought stacks of single density 3.5inch floppies, put them through the drill press to give the double density hole a pass through, and then they all worked as 1.44mb woohoo. some of my earliest hacking really
I think I still have that hole punch for that as it was actually sold in computer stores.
lol yeah the store had punches specifically for it
nice white box
Big ass box little ass punch
it was the 1980s everything was to excess.
happened cause manufacturers weren't going to do 2 manufacturing lines for single and double density discs, so both plastic 3.5inch shells, single and double density, got the same floppy disc inside
just needed that hole
if that doesn't give me old man street cred, i'll flex my sysop status on the BBS i ran. Ground Zero BBS on a RemoteAccess system
I ran a BBS too with 5 floppy drives on a C64. The most it could daisy chain on a 300 baud 1650 that I changed over to a Lightning 2400.
Sporting that 2400 I was the shiz nitz
i just had the two drives for my c64 and never did too much modem work with it. We connected to compuserv but mostly just used it for games and word processing. learned some basic too
modem fun really took off with the 286-486 era
remmeber the day i got the 14.4 kilobaud modem working for the first time. holy shit we were screaming speed demons that day
I went from basic to ML (Machine Language) in the first two weeks. Made a demo of basic, blitz compiled, and ML of a cube across the screen to demo the speed differences.
half duplex 😦
I remember being so happy buying a 56k modem
56k holy shit. when would you need more?
Trailblazer 19.2 was everyone's dream
that was like a $800 purchase if i remember
I sat at 33.6 then went to ISDN dual channel motorola modem then finally cable in 2000.
i'm probably inflating the cost cause i had less money back then
No, it was hellish expensive
I never could afford it so sat with the 33.6
that of course down throttled due to phone lines
28.8 on a good day
rich asshole friend had one with shotgun technology. You connect it to two phonelines and dial the isp with both, and get double bandwidth
fucking dick
sorry
i'm still a little jealous of that
we used to get squirrels here each year on our city so 9600 was not uncommon. About two decades ago they changed all the phone lines to something they didn't wish to chew and problem solved.
THAT was DSL (adsl). 128k and I loved it
Magibox in my city gave it to me for free.
not the modem the ISP usage
Back when Quake 1 was a thing. Damn that was fun
Quakeworld
I played Team Fortress Classic
Yep, then the devs got greedy and bad and sold out to make TF2
and it sucked
yep
at least compared to TFC
yes
with all the stupid hats
I think I am just going to 16 gigs and a server card if I can figure out how not to go deaf
LOL i love TF2, but i also love the lingering hatred for it from oriiginal TFC fans
it is faster than colab by a little at least and is local
the hats came later. Cosmetics and MTX weren't part of it for a couple years iirc
I just did not like how it played or looked. It lost all of its magic
wish the original orange box version was still established tbh. You can find it and run it, but there's no communit
myself, i never found TFC that magical. I played a lot of unreal tournament and it's mods, and found that much faster and aggressive. TF2's art style is what drew me into it
I don't think the chuds on Reddit will answer about speeds of training on various server cards. I know it is being done.
It's art style I still do not like to this day
it's very "looney toons" esque which appeals to me greatly
See, that is precisely what turned me off
Every character was so large
way too cartoony for me
yeah. I liked quake a lot, but tfc just felt like a bit too much quake. more of the same generic. which is why unreal was awesome. Way less generic.
also i didn't do PC gaming for a few years too and sort of committed to consoles hard. Goldeneye oh baby
Halflife 2 ushered me back home
cant wait till they mod HL2 RTX like portal RTX
What's up diffusers?
I have about 10k hours in D3
These days I am just not a gamer though
same. I ve reducing gaming a lot
I finally played RDR2 and loved it until my damn character I had some attachment to caught TB and died.
Arthur
that ticked me off
my number 1 game since dec has been SD..lol
tbh I have lost interest in SD. I mean I am burnt out so badly as my fun came from training and releasing but colab, and advancements, just sort of killed it for me.
the toxic community started it though
What gut punched me was Nvidia and the card I was saving for became a mockery and an expensive mockery at that.
those two together SD was on borrowed time with me. I honestly don't see it getting any better with those two either.
Im the opposite. I love it even more as time goes by and I learn more improving my workflow
At some point I will try training my own stuff but for now all the new stuff keeps me busy
Just generating stuff was never fun for me outside the first day or so as I wanted more and more as in train my own stuff.
any idea when pytorch2 gets out or is avail for A1111?
I asked and just waiting on a response
cool
been since dec on the pr
pr, I see, failed though
making some progress with photorealistic portraits! basically a stable diffusion output with some touch up in DALL-E 2 (idk I prefer the inpainting there), and with a DaVinci Resolve filter slapped on top
nice! what model did you use?
should be fine
Ok i think its not allowed
gotcha
Sent you link instead
These are
What our model doest best at
My favourite
Oh man...
I love him
Love some jocked men lol
@young mountainMind If I get the link too? You're generating the same kinda stuff that I'm interested in
😛
why are these so awesome?
I guess I am having a good night. TY 🙂
Hey everyone 🙂
Heyas
How are you doing General ?
I see you are moving out from robots ?
Aye
I'm not producing lots of images with it yet, but I'm having real fun transitionning to comfyUI
I didn't know how much I wanted a node based UI
I really like it but it is still not comfy for me. In time he will update it for sure.
yeah it really is not comfy lol
I love node based stuff and why I was praising the future of invokeai but they decided against it
I'll share a list of templates once I'm done preparing my library
they are still teasing it recently, invokeAI
it should come but yeah, 2.2 has been out for long now
well, comfy needs to do what invoke does with gen images as it goes directly to the screen and they use the buffer there
makes it faster and no temp files
they do really have a nice code workflow on invokeAI. I do hope they continue on this way
but they need to add back batch size....
comfy also needs to make it a single node for loras etc... where it can grow as you need more because too many wires (connections) is just ugh
havent's tried lora in it yet
they work 100%
I was messing with area composition
unlike in auto
what I love about comfy is his hatred of gradio is as strong as mine is. It will never ever be in gradio.
and that is a relief
I like gradio for what it is, it still brings some good productivity when building simple tools
but going full UI in it is really not what I was expecting
I get so many errors and their devs are touchy feely types
one thing gradio does not like is symlinks
I had to bring some of mine back because it would complain and error
others say they have issues with NAS too
damn, didn't know that
I chimed in on a ticket about it
I went searching for a card to train with since Iam done with colab as colab destroyed my very will to live.
that much ?
I found a decent card but no real inexpensive way to cool it. 😦
cooling is a main issue for sure... it's what turned me away from some cards when I was upgrading last fall
faster than a T4 on colab and around 200 used, plus 13 for a wiring dongle but it is passive meant for a server
I can't have a screaming banshee in here spinning at 10k rpm times 2 just to cool it.
that leaves water cooling and that is around 150-225
4080 is just far too expensive
that is for sure. the 40xx series is just overpriced imo
Not getting any better either
Now they just inked the deal for GDDR7
meaning newer cards coming with that in them.
like the 3k cards had gddr6 then gddr6x
Hi
hey again Blackwater 🙂
yeah so if you only used MJ for now, there are some differences here but it works close to the same way
in particular, you can have some more tools to play with
the "scribble" mode from controlnet seems to fit your needs
Oh yea thats kinda what I want
I wanted to make this shitty drawing a bit cooler
MJ is making it look really good, but its adding faces and stuff that I dont want
ok so i'm trying with your tree image
the scribble mode seems to have a hard time understanding the drawing in there
i'm trying with the canny mode
lots better
those are the settings
I'm using a local installation of Stable diffusion
and some extensions on top of it that give me enough control for that
but yes, out of the box, the AI won't respect your drawing closely
Is that free to use or not
yep
just needs some hardware. but there is also ways to run it quite easily in the cloud if you don't have the good computer for it
if you have ever used notebooks, you can just run this one and test https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb
I have a Legion 5 laptop with an RTX 3060 is that enough
it should be, yes. the most important is the graphics cards, and the VRAM on it
My humble experiment to create a music video entirely using AI , based on prompts from the lyrics of the song.
https://www.youtube.com/watch?v=20c8qcmy4oI
The video was created entirely using AI.
The song features excerpts from Thracian Dances Suite (1925), No.2 Rachenitsa by Petko Staynov.
Miroslav Markov - Bass
https://www.youtube.com/@miroslavmarkov1245
Elena Ivanova - Vocals
Alexander Page - Guitars
https://www.youtube.com/watch?v=B6EnzPVQ2Cg
Apostol Stefanov - Guitars and synths
https://www...
Okay
let's move again though, I don't want to spoil the picture sharing. to #🤝|tech-support if you want some help getting it running
sorry ^^
(working on my PoW, looking for a good base to start on, and it's spitting me fire)
this one is literal fire
love that one
giga shrek
! dreambot
@weak crest
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
Dose anyone know the name of this prompt extension ?
the dropdown list? Tag Autocomplete. actually might be tagcomplete now. this thing lol
https://github.com/DominikDoom/a1111-sd-webui-tagcomplete
(I can't add a second flame for the link, but I would if I could)
hah. appreciate it nonetheless. it's a good extension if you use that type of stuff. i have it installed for wildcard,embedding, and lora completion. very useful
I'm starting to switch from webUI, but I'll be sure to try it out nontheless
they added a feature a while back that actually allows you to autocomplete a wildcard, and then additionally select from inside that wildcard if you want to pick something specific from a wildcard list
i usually use wildcards to get random generations from the lists, but when i want something specific it's very handy
it's quite the great idea ! for sure, sometimes you want to pick one specific value from the wildcard. nice thinking
i also have a tendency to rename my loras to include the trigger words...so some loras get -very- long filenames and the autocomplete extension is a huge lifesaver with that
I'm trying to upgrade my "bad cosplay" dataset and I just found out about magikarp cosplay... damn lol
@distant sparrow
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
And welcome around !
Thanks
Long-haired girl on a horse at a picturesque beach at sunset
Girl playing soccer
@sullen plinth
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
Prompt: "The man who kill my father."
thanks controlnet, I didn't need to sleep anyway
My cat is a rotisserie chicken
@smoky oak@static tuskhttps://www.youtube.com/watch?v=pom3nQejaTs
Discord : https://bit.ly/SECoursesDiscord. Torch 2 / Pytorch 2 is now supported along with new DreamBooth Automatic1111 Web UI extension. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on 🥰 https://www.patreon.com/SECourses
In this video, I am explaining how you can man...
ping
:O
Torch 2 support? :D
Let's see
I will never turn down an offer to gen faster heh
oh nice one
watching
oh, so it only works with the dreambooth extension?
Hmm... I see
it is in the pipe so more and more are upgrading the heat is on Auto to do so
I'll wait until that comes out then
I have no need for the dreambooth extension, unless I am misunderstanding what this is doing
of course colab will be next year with this
Yes, it is. Thanks a lot for sharing the wonderful usage.👍
I prefer the DB extension to Kohya but on colab I couldn't use it
the DB extension allows for LoRA training?
yes
hmm. Perhaps I will take a look then
far simpler I think for up to three concepts. Beyond that forget it
so, this pytorch 2 thing. Does that only benefit that extension, or does this affect A1111's general gen speed?
just the extension until auto implements it
alright, I'll just wait for that then
eventually everything will be pytorch 2 as it is just far better
is it just speed increases, or are there improvements to VRAM efficiency as well?
if I had the card I am not so timid I would dive in and eat it all up. I don't so forced to wait
yes
So its faster and uses less VRAM?
it really is highly optimized now
well, this is far beyond simple SD so this is a nice advancement
Man, just a few months ago, I couldn't dream of doing 4k gens on 8GB VRAM. Now I can do it no problem lol
I barely can do 1024 on 6 after Jan auto update. go to comfy and back to full on
weird, must be a non RTX thing
as you know, comfy is way less efficient on my card
its slower and the resolution OOM is wayyy lower
though I am very excited for the development of Comfy
oh wow, so its faster, more dynamic/adpative, and runs on more stuff than ever with less resources than ever
what a beautiful thing to read
yep
is there any plans on AMD GPU's being able to use SD reasonably any time soon?

I want a card sub 800 that can train this stuff in 10-15 mins instead of 25m on the T4.
iow 4800 for 799
that should be really easy, as the T4 is very slow compared to modern tensor arch
or AMD or Intel I really do not care as I am agnostic about brands
but your issue is VRAM
16gb on t4 but google does something funky as their card are 1gb less
My LoRA's take sub 10 minutes most of the time on my GPU. I love it
768x768 training
Probably some form of root level VRAM protection and such
ah, IDK then
on 512x512 even the t4 flies
or they leave a 1GB window in order to protect from VRAM cramming/crashing
I'd imagine that saves them a lot of maintenance and diagnostics
possibly as I saw 1gb since they started. K40, T4, P100, V100, A100 all 1GB less
Yeah, its probably either an overflow protection, or a root level VRAM protection level to protect from hardware exploits
possibly both
my friend has an A100
man the things I would do to play with it lmao
he got it for free no less
lucky bastard lol
Thats what happens when you sell a potentially multi million dollar product to unity lol
they give you free GPU's heh
in a matter of 2 months he went from a 3060 to a 3090ti to an a100
yeah, which is why I said no way in hell on colab as they just, last week, finally upgraded to 3.9.16
tired of Colab it made me hate life. I saw a vid and most old timers say it is just not worth it in 2023 so they are losing paying customers (such as the 50 per month)
yes
not when you have packages that demand 3.10+
oh, in that case, good point
Kohya demanded 3.10 so the notebook guy worked with him to get it to work
ah, i am on 3.10.6 it seems
is there an easy way to check my CUDA version?
as I said nvidia-smi
aha, thank you
command prompt
yikes
528.49 just dropped lol
Yeah, I wait
I am updating now
wonder if I can expect anything from that
I know that RTX GPU's can now run super resolution on youtube videos
real time video upscaling
Yeah, I saw
Hi there, I've been trying to find a prompt that will get SF to create an image with a style similar to Alan Moore's work, but I haven't had any luck. Specifically, I'm looking for a comic book-style image, but the images SF creates tend to be too cartoonish or simplistic for my needs. I'm hoping to get an image that has a style similar to this example:
what is SF?
Oh, SD. Well, did you try 2.1 or a different model?
Train your own lora for that style btw
hmmm no 🤔
btw, I like that style
Yeah, its a fun style
Wow, can you train the AI with a general style as well? I thought it was only for people models
Oh, heavens know as I train styles not people
<10 mins lora has it
25m on google colab
I train people not styles haha
General Awareness and I are two sides to a coin lol
he specializes in styles, I specialize in subjects
LOL, yeah
idiot savant?
Hmmm, well, I'm going to investigate how to train her in a style 🤔
sounds like haha
you should train two different LoRA's, or a merged LoRA with one subject as a style, and the other as style. Trigger both and they should hybridize
So much I would love to do but need more than colab for sure.
I almost purchased a p100 but the cooling solution for it is as much as the card
above was 0.3 this is 0.7
0.7 is what I was after it seems
@dense tapirHave you thought about an A4000?
oh nevermind
they are more than I remember
All cards have skyrocketed after the release of the RTX 40x0 cards being so stupidly overpriced
sad thing is Nvidia is going to double down on their pricing
The card I was saving for was the 4060 but it got nuked and became the 4080 spec wise
0.6
I think 0.65 or 0.7 is what I am after
trying 0.65 now
Been a long time since I merged models and no interpolation I am unsure what it is doing
was not an option when I did
Looks like 0.7
result=A? lol, why merge then?
If you're talking about the third interpolation mode, it's very weird. It's like C minus B equals a where a equals the difference between C and B from A
It's the weirdest thing I've ever heard of in relations to merging lol
First option as that is new to me
Okay I give up. I have run out of money and I cant seem to get the model I want. So Ill just release what I have for now.
Any information take from the Huggingface or Civitai pages:
I see no way to remove the tertiary model so I am stuck.
To remove the tertiary?
oh, I guess no need
only comes into play with the add
Yes, add was always a bit odd
I thought you use the third as the base model, so you remove the base from B, then add that to A, as a percentage you decide
woah there's 1024 TRAINED not merged models out there? Gawddaaaamn
Can someone post some gens from this one
i gotta go to bed now, but heres one that isnt in the samples:
beautiful lighting, medium-shot, photograph, white woman wearing tshirt, leaning against fence
Steps: 50, Sampler: Euler a, CFG scale: 7, Seed: 1199866268, Size: 1024x1024, Model hash: 28fedbc31e, Model: Char-v1.0-by-AI_Characters
Bro, my toilet just fucking exploded, and I don't know what to do
WHUT?
any repairs done to it recently?
Nope
I sat down to go, heard a small like creak that sounded like rubber, and when I leaned forward, death
Not frigid or hot in there?
Nope
Looks like some sort of stress on the porcelain
My mom cut her hand open on it too
Damn
I had to get gloves to move the shards, and now there is like fine porcelain dust/fragments on the floor
Well, judging by that phto (partially blocked) it appears to be it broke at the nut that holds down the tank
right side in the photo
over tightening causes it over time until you sit down and away it goes
hard to tell with the old style float bulb in the way
It's not even an old house/toilet
This house has just been problems since we got it. But I still really like it
It has the problems of an antique house, yet it still considered modern construction
House was built in 2007, so I don't know what to say
It's not an old house by any means
Well, I am going by the USA standard as float valves like that is a dino here
This is USA ._.
They moved away from those by 1990 and I can't even find one now
Then the manufacturer used old shit to save money like one of my houses did where I grew up
I guess so
all homes used to have those black float valves then it changed over cause those have been around since indoor toilets came to be a thing
This house has a lot of issues. and the owners before us were fucking stupid, so they didn't help
iow 1880s+
like this one he kept on building on to it until he died and was always drunk so nothing works, or is level or makes sense.
cnc machine
I'm gonna have PTSD
Post traumatic shit disorder
Sponsor: Montech Metal DT24 Base on Amazon - https://geni.us/IXXak
This hardware news recap talks about a GN imposter using our name as a prebuilt scam, Intel leaking its own 20-angstrom architecture, Microsoft leaking AMD's RX 7800 GPU, and driver bugs. The triumvirate is complete: NVIDIA's bugs cause CPU cycle waste, AMD's can brick Windows, a...
Almost finished just need to upscale
Photo bashing is so much fun
What are your thoughts on the NO ai gen movement? see artisti like greg with the stickers?
I get the idea, and I agree with it. But its also founded off of a lot of misinformation and lack of understanding on how AI works
I really do not get the NO ai gen movement because art is subjective and as long as I like what I see I don't give two fucks where it came from, who did it, or how they did it. Basically, now step back to see this picture, this is a mild version of what will come when the masses have a NO Android movement.
true, I can agree with that
21st Century having 19th Century ideas meeting the 22nd Century techs.
Android rights, no Androids are just tools, etc...
At the same time, I do have to relate to artists and how they are being displaced. There are lots of artists that are gonna lose their jobs
but at the same time, more people have access to high quality art, which will result in like a new renaissance as more people get more creative
I do not because, in the end, there is no safe job and I know for a fact that when the day comes that no living body can find a job because everything is AI/Androids/Robots/etc.... that is the day the paradigm begins to change and money is a thing of the past. No longer will you work because you have to rather because it is your passion. It will come but a lot of infighting before then though the future is bright we just have to get through the storms first.
Wealth brings power and the wealthy aren't going to want to remove that power
thats a future dystopia for sure, but I guess thats not our problem now
Well, we are on the cusp of it though as this shit is moving so fast it blows the mind. In a blink of an eye 10 years of tech advancement happened.
Yeah, which is good and also bad
Used to tech advanced in 20 year cycles, and before that 50 to 100 years, etc... now daily almost.
yes, I agree it is far too fast for comfort
Things will collapse before we get a good outcome
Gradual would be far better to get the masses to a greater acceptance instead of pitchforks, tar, feathers, and nooses.
not even just that
there is no way to regulate this stuff, or form limitations, which is gonna lead to horrific things
there will be 0 legislation on any of this. Hell, there still is, over a year after Ai image generation made it imainstream
They are working on it and we just had a meme lord sentenced to 10 years in prison for helping getting Trump elected in 2016. Laws are slow on purpose but the governments are afraid of memes and I just saw my first AI generated meme of some politician. Not a US one but it was used as some thumbnail. It is coming and the governments will come on with a vengeance.
Right now I can train a model on a singular politician and do some serious memes.
Even if the base doesn't know that politician at all
I said, as a kid, that 2030-2050 is when this country will advance to a ST like world or a Mad Max like world. I do not know what will come out the other end but the following years are not going to be nice. Think of it as the teenager angst years as we just went through the growing pains.
we aren't even at growing pains yet
Anyways, this is all too vague/unknown for us to speculate on really
We have no idea where things are going, and we have no say in them either
one thing is for sure. AI is gonna be terrifying in the hands of the corrupt corporate leaders
some art of the guy in my PFP
Trying to think of a 3rd problem related to greg rutkowski. There are a ton of ai work out there with prompts including his name, that in the end look nothing like his. Including nsfw stuff. Imaging down the road someone relate greg to those things LOL. My first uninformed ai image included his name because that's what I read, I have to add artist names like greg's to make it draw something nice. Of course I still spewed out crap since I did not know what I was doing. That image was deleted long ago. So I think Greg is angry because 1) his style is sometimes deliberately ripped off 2) his job or others' jobs are in jeopardy, maybe not him as he is so famours 3) all kinds of ai art with nothing to do with him is crediting him
using artist names in 1.x was good but I am so glad that is not a thing in 2.x as it was really just a crutch
I often see ai images that are made on very long mess of prompts, with artists' names from various disciplines thrown around...as far fetched as greg's with picaso....those 2 share very little things alike. Yet the begining prompt "anime drawing" dictated the entire style so in the end, the picture looked nothing like greg's nor picaso's. But the credit went to both lol.
Greg Rutkowski has been dead for a hot moment
He had his 15 minutes of fame
oh sorry I did not know about that
Thomas Kinkade is also somebody I wouldn't worry about mimicking. He was a sell out, and a crook, and is one of the most hated artists in the painting space
but greg's name was a name that got consistently ripped off on 1 of the txt2img service website I started out in...like 8 of 10 images had his name so I thought it was crucial to have his name on it LOL
Yep, but still I liked his art
its the definition of kitsch
I have not see a bob ross prompt yet for some reason...something so simple and everyone knows about him
The definition of Kitsch should legit be a link to Thomas Kinkades work lol
What word can I use besides high contrast for that last image?
He didn't really have an iconic style. More of just a photo to painting translation
happy clouds, forest, river, mountain background by bob ross
you might not be able to fix it
you should try a VAE, that looks like a biased VAE issue
I can in post but would rather via a prompt
I used to have that issue as well
he was not unique, but quite a simple/generic landscape oil painting style that if someone is making a simple ai landscape oil painting, you can quote him
VAE is automatic and that is 2.1 so we don't have VAEs
yeah, I agree
though I will say the reason you don't see him cited in AI is cause he is still under copyright
all of the other artists that the AI's know well are on expired copyright, and have since become public domain
People still quote disney, pixar on their prompts...imagine what the lawyers are now doing.
Damn. Its not gonna look any better out of the AI then. The white and black points are skewed
Yeah, not sure why
is greg's copyright expired? what about active guys like beeple...see his name all the time
Thats what VAE's are supposed to fix
I doubt it actually adds much to the work, as he doesn't have a signature style, and I have a feeling hes also not in general data sets
vae, I think, are baked in now with 2.x
thats a bummer
way less control in that case
You can see the difference they make for some scenes
SDXL will continue 2.x trends?
Used auto, and look how far off it was. :/
Yeah, VAE's are super useful/powerful. The shouldn't get rid of them
I mean just look at that massive difference above
what else can VAE fix besides contrast and brightness
They aren't gone they are in the model I believe
sucks because I loved mixing and matching vaes
Contrast, hue, saturation, color temp, all sorts of things
yep, vae made absolutely no change
and it is non destructive as well
damn ._.
just another reason for me to never use 2.x
😛
Going forward this is the way
3.0 too
Unless SAI comes to their senses I don't see anything actually getting any better.
I am on the opposite, I have not used 1.5 much, unless I could not make it on 2.x and have to revert to another 1.5 model/lora
this is pretty heavy HDR treatment
I prompted HDR
the halo effect is quite prominent
you train mostly 2.x right
I use 1.5 strictly as it most of the times gets better results, has a lottt more support, and isn't an absolute pain in the ass to train
and I ONLY train 1.5, or work in 1.5
How to generate images
once again, general awareness and I are two sides to the same coin haha
I am 1.5 and subjects
He is 2.x and styles lol
You train on charcters, much more difficult to control so I get why you want 1.5 I make mostly objects so the photorealism from 2.x is good for my use.
yep, PR is what I have always been about.
I go for photo realism above all else. And after experimenting for close to 50k gens now, nothing has beat 1.5 for me
realistic V1.3 to be exact. It just does what it does so damn well
You know I suspect 1.5 people are going to be on 1.5 in 2030
I am on 1.5 until something better than 2.x gets released
I say that because it looks like SAI has painted themselves into a corner
one of the few 1.5 models I occasionally go for is realistic vision, but still on fine details and gradual tonal changes on texture, 2.1 is still better for me, across several 2.1 photoreal models
I hope SDXL is better than 2.x
I would go to 2.x, if training wasn't a fucking guessing game
you tried realistic vison 1.4 yet (behind a pay wall now?)
did SAI change the name from 3.0 to SDXL? If so that is so damn childish it is actually insulting to me.
I mean, i am still very happy with results like this out of SD 1.5
I don't know if SDXL is 3.0 or just 2.5
Realistic vision 1.3 just beats everything else with animals IMO
LOL 20 years later and a few meth head renters later.
No. Fuck them and fantasy AI for locking something they have no right to behind a pay wall
lmao
The creator behind realistic 1.3 is a POS for selling out, especially when he has no rights to any of the shit he used to train on/merge with
there should be some broken windows with meth heads inside
from seeing that in a real life situation in a neighbourhood 5 blocks away
I know
boarded up windows for the last 4 years.
I live in the worst city in America for drugs coming in, through, and out of. A big hub on the Mississippi river.
I am waiting to be shot through my window one night
Damn. Not only are you in Mississippi, but you're in a part thats even bad for Mississippi lol
ah
some dude was murdered near my house a couple of years ago.
Mississippi is a cesspool IMO. Too much bad shit comes from there lol
Memphis is the murder captial of the USA per capita even beating out Chicago
Murder is a very common thing where I live. But its due to gangs and is rarely towards unsuspecting people
specifically shooting people on roads or infront of stores, here
Same here though we do have the BLM types here now that want all non whites dead.
oof
We did have a stray bullet come through our awning a couple years ago
but we think it was from some dumb ass on 4th of July
Where I came from had gangs in the barrio and people slept on the floor or in tubs (children) else dead. Nice to shit in 18 years.
Where I live has always been shit TBH
My roof had a bullet through it so some ass was popping off on July 4 and down it came
Every single year I hear them popping off into the air
Yeah, thats our best guess as well, as it came in almost straight from above based off of the hole it made
ffs, ammo isn't cheap but if a gang banger just go rob someone
All these cities that are pro criminal are about to go belly up with stores leaving them.
yeah, as they should lol
section 8 buyers lol
LOL
our whole neighborhood was destroyed by section 8 buyers
Even the balcony is falling down
Yes, made me so mad
my last neighborhood became a hood with tagging and all due to that
What happened here was 2005 Katrina and it never recovered.
All the criminals spread out
Our neighborhood was bought out a lot by Asian owners (no hate to asian people, get the hustle lol)
And then they would renovate them, and toss them to the wolves by running mass section 8 housing. All that does is bring in trashy people who live trashy lives. Blegh, it was bad
And it sucks cause not all of the section 8 families were bad! It was the vast majority of them tho
so once the section 8 situation crashed, and all of the owners pulled out, the houses of those few good families were pulled too
The house across the street is owned by a now old Asian lady and it has always been rented out to illegals. She refuses to rent it to anyone else.
luckily our old neighborhood more or less recovered, as it is still in a nice area with nice people
Jesus look at the difference a VAE makes here
I am honestly unsure if that matters in 2.x. SAI said no
I am starting to really not like SAI lol
they seem to have a lotttt of say so for something thats supposed to be for the masses
I just had the tagger mess up and call a horse a cat
no weapons, which is one that really irks me as my knight can't have a sword, but people can make CP without even having to think about it
oof haha
Well, just take ex hedge fund managers and anything they say with a grain of salt.
Luckily there is nothing stopping people from taking their base work and making it better than they do lol
I mean, I know dragon forged and I have a huge project in mind lol. And the results would not only be more ethical, but also employ people, and also produce better results
Yes, I am unsure where that is happening
might be this damn seed
I was not having that earlier as you saw
I changed samplers and still no contrast
VAE supremacy
anime me lol
this is weird
how do i fix fhe squish
did you use image to image?
I have cat in my prompt but it is giving me a wolf
you have to make the output taller
its taking the input and squishing it to match teh output
can i edit and squeeze the image together
what?
I mean yeah, you can., But just generate it properly and you won't have any issues
I found the issue. Dynamic threshold
Aha
oh, wheres that
Give the image more height
worked on 1.5 on 2.1 it may be too much
would seem that way haha
okay, thx for help🙏
https://www.youtube.com/watch?v=iy1DNxkVUT0 ouch Mississippi
A nuclear power plant along the Mississippi River in Monticello, Minnesota, has leaked more than 400,000 gallons of radioactive water due to a broken pipe. NBC’s Maggie Vespa has the details.
» Subscribe to NBC News: http://nbcnews.to/SubscribeToNBC
» Watch more NBC video: http://bit.ly/MoreNBCNews
NBC News Digital is a collection of innovati...
SWEET
I will hold my tongue on what I was about to say lol
we are all gonna save electricity since everything will be glowing, including us
yeah, this extension is wrecking the contrast
your tongue should already be glowing if you were affected. This happened in November 2022
Just made the news like an hour ago though
ah
wtf? Happened in Nov we only find out now?
yea, watch that news
I did
still safer than Flint Michigan
green glowing radioactive water, flowing downstream on river, glowing people, highres, 8k, photorealistic. no negative prompts needed
tweet this to nbc news
