LOL
#🏞｜general-with-images
ocean hawk
Hey folks! Does anyone have a moment for a little special request or are you guys busy?
oak osprey
always busy
ocean hawk
I know I could learn this on my own but it's something super specific and it'll be super quick!
oak osprey
always brewing
ocean hawk
I need this meme but with someone else's face on the guy
It's george berkeley, the philosopher
Is it too tough or is it possible?
dense tapir
easy peasy if the model knows him
ocean hawk
There's unfortunately maybe 3 or 4 paintings of him only, is that a big problem or?
dense tapir
no, you will need to train it then
ocean hawk
Is attempting it right now an easy task or would it take a lot of your time?
I know I came out of no where with this, haha, but honestly it doesn't have to be amazing quality...
ocean hawk
Yes, but like just a pic2pic thing where you tell the model to put this ancient dude's face on it...
Is this something I can try easily or would I have to learn to code to do it?
dense tapir
yes, use 1.5 and controlnet
ocean hawk
To me that doesn't say much, but it's normal because I don't really know SD at all
dense tapir
use the reference model
low compass
ocean hawk
Yeah, again, this is like... Already a lot for me, but probably because you already know this
dense tapir
I don't use 1.5 since Dec
ocean hawk
Oooooh....
Cool... Now is there a way to get him inside the meme or would that be complicated?
dense tapir
not complicated with controlnet
ocean hawk
And uh, do you think you could do it for me? Or are you busy with something else right now?
ocean hawk
You're in luck, I love awful quality
low compass
I dont know how to use control net but I can try with inpainting
ocean hawk
Thank you, you're a god send
oak osprey
ocean hawk
I dont know how to use control net but I can try with inpainting
Are you trying it now or was that just a general idea?
low compass
trying
ocean hawk
Sorry, haha, I just have no idea how these things work, I was making sure.
dense tapir
First try
ocean hawk
YES
ocean hawk
It's all right, I love noise image. Haha, I'm kidding but this definitely does the trick!
dense tapir
Now comes the refining process
This would take a long time due to the quality of that man's face image
low compass
I keep getting this kind of image
dense tapir
LOL, wth?
low compass
probly photoshop is better idea 😅
dense tapir
Yes
dense tapir
Now back to my style testing
Style is nice but it tore the subject up in their faces, sheesh
vocal haven
Hello. Are the bots still down? Are you sharing old pics? Or is anything wrong with my Discord? Did I broke discord?
dense tapir
low compass
Hello. Are the bots still down? Are you sharing old pics? Or is anything wrong w...
I am using automatic1111
dense tapir
damn, really like it that one time it worked
wispy ether
Tf
oak osprey
damn, really like it that one time it worked
thanks! i will look at the Linux version when i'm done cooking
it's downloading a bunch of crap into a venv
dense tapir
maybe reduce the tile size
Not sure what SD was thinking.
Seems to be an AI issue
Hands, feet, and water it loves to shove half the body under it
dense tapir
??
oak osprey
i have an fp32 file for you now
dense tapir
fp32 on a 6gb will zap all my ram.
oak osprey
it's what yours needs, apparently
dense tapir
I use pruned fp16 all the time
don't confuse my 1060 with a lesser version 1650/1660 as they have to have that
oak osprey
well i recreated both files anyway
dense tapir
in safetensor?
dense tapir
I think it was
smoky oak
I just made an extremely grave mistake, and now I'm paying for it lol
dense tapir
how?
oak osprey
too much chili?
dense tapir
LOL
smoky oak
I applied silicone with my bare hands because we didn't have any gloves or disposable tools
dense tapir
ZAP
smoky oak
And that would have been fine, if it was not waterproof
So I just had to use every chemical under the sun in order to get it off my hands lol
Butter, exfoliating salt, vinegar, alcohol, acetone, degreaser
dense tapir
wait, silicone is easy to remove until cured
smoky oak
I finally got most of it off to where my hands are just barely sticky
smoky oak
wait, silicone is easy to remove until cured
It is, as long as it's not water proof
This silicone is made to be applied underwater, so it's extremely hydrophobic
dense tapir
if you mean silicone grease that is the grease part that is shit to remove
I have never had any silicone not be water proof
smoky oak
No, it's silicone adhesive
dense tapir
oh, fuck that shit
smoky oak
Regardless, it was a huge pain in the ass to remove lol
oak osprey
This silicone is made to be applied underwater, so it's extremely hydrophobic
sounds very intolerant
dense tapir
that white stuff that once cured you have to break it
smoky oak
sounds very intolerant
The most lol
I did get it off of my hands luckily, but it took everything and the kitchen sink
Not even joking, it all went down in the kitchen sink lol
Safe to say I will never be doing that again
The acetone was the only thing that ended up working, and I had to use a ton of it
oak osprey
oh noo
don't put that down the sink
i watch a plumber on youtube, that shit will fuck your pipes up
oak osprey
the silicone
smoky oak
Even then, the only thing that actually got it off was the acetone, and that was done with a lot of acetone and paper towels
dense tapir
Acetone melts ABS/PVC
smoky oak
It was very very little, just a little bit on my pointer finger in my thumb, but it spread everywhere
dense tapir
like a boss
smoky oak
I know, that's why I made sure not to use the acetone in the sink
dense tapir
no worries unless you pour an entire bottle down the drain
If it fell in, an entire bottle, flush with gallons of water
smoky oak
I frequently work with acetone, vinegar, and isopropyl alcohol because I do frequent 3D printing
dense tapir
same
dense tapir
I ceased to 3d print though as I never had many successes with it due to first layer adhesion no matter the printer. After 8 years (since 2012) I gave it up.
dense tapir
Just sit back and allow the gas to take effect and it will all be over shortly.
smoky oak
I ceased to 3d print though as I never had many successes with it due to first l...
Oh yeah, 3D printing was hell back then. Now it's a cakewalk
I level my printer like once every 30 prints and do almost no maintenance between now
And damn, you can get exceptionally good quality printers for less than 300 bucks now
oak osprey
Just sit back and allow the gas to take effect and it will all be over shortly.
Xylene is where it's at
i used Toluene a few times for doing bad things.
dense tapir
I gave up after 2020 as it was just too much. PLA I never liked
smoky oak
I've been thinking about trying out that new open AI 3D model thing to try and create 3D models of my characters and 3D print them
dense tapir
smoky oak
I gave up after 2020 as it was just too much. PLA I never liked
There's dozens of different plastics to choose from, I'm sure you could find something that suits your needs
I print mostly PLA, but I also print PETG and TPU
dense tapir
I liked PETG and ABS but ABS needs an enclosure
tbh, the resin printers I was so hoping for but the LCD panel are consumables, and not cheap. The resin is better priced but not cheap.
smoky oak
There are plenty of ways to print ABS without an enclosure, it's really just a matter of knowing what you're doing
One day I want to get my hands on PVDF
The safest and most dangerous material to print lol
dense tapir
ABS in a room in the winter at about 40-50f ABS beyond about 10mm will do nasty shit
smoky oak
Okay yeah, that's not an ABS thing, that's a lack of any form of temperature control thing lol
dense tapir
precisely, and breezes are bad too. Hence the need for an enclosure to control it better.
smoky oak
What, are you printing it outside in the middle of a freaking tundra?
smoky oak
You're joking, right?
dense tapir
no. The house that stuff is in is about 8f-12f higher than outdoor in the winter. In the summer it is about ambient.
summer at 95-100f ABS loved that
wispy nest
damn general, where you live, tatooine?
smoky oak
I don't even know what to say lol
wispy nest
and sytan that pfp is dope dude
smoky oak
of course its not gonna work when you put in 0 effort to try and regulate it lmao
smoky oak
and sytan that pfp is dope dude
thanks! I made it last pride month haha
wispy nest
noice
dense tapir
why do you purposely ignore what I said about an enclosure? There is the effort I put in.
smoky oak
lmao
I am due for a new PFP iteration
I have had this PFP for got... 8 years now?
or well, my logo
wispy nest
that's branding in action. noice.
smoky oak
Oh, i also had this one
wispy nest
that'll be what ends the earth. you become an evil genius with a space station and hit us like the dinosaurs got their shit rocked down in yucatan or wherever exactly the big one hit
smoky oak
lol
This logo survived a whole rebrand, cause I was able to find something that worked with it
I used to be Scyth Sergal
but now my brand is, you guessed it, Sytan
dense tapir
smoky oak
was able to keep the SY symbolism lol
dense tapir
I like best the upper right and lower left
wispy nest
so is deep floyd not usable in auto or others? like i dont see any deep floyd images in here ever i dont think
yeah those are tight
dense tapir
next would be the lower right
smoky oak
so is deep floyd not usable in auto or others? like i dont see any deep floyd im...
the deepfloyd model isn't out yet from what I know
wispy nest
but they got a deep floyd section here 🤷
XL hopefully brings some heat back to stable, I feel like midjourney kinda took over the game in most people's minds
smoky oak
MJ is more accessible for most, but also inherits the "its ok" aspect of being for the masses
dense tapir
Problem is, at the start, that all the new modern Nvidia cards can't run SDXL until they prune them they said.
smoky oak
yeah, that as well
dense tapir
Need 10-12gb so a 4070ti
smoky oak
and then it will be slow as hell
wispy nest
oh damn i didnt know that
dense tapir
they say 10-12 but who knows
smoky oak
Cause the devs said for SDXL you will need 2.5x the VRAM of 1.5
wispy nest
and yeah i kinda hate midjourney, but still stable's been kinda stale for a while. like the last model they released that people really used was 1.5
smoky oak
and yeah i kinda hate midjourney, but still stable's been kinda stale for a whil...
yeah, for sure
wispy nest
yeah i like that background general
smoky oak
although, I must say, 1.5 may be old, but man it can still hand MJ its ass when you know what you are doing haha
wispy nest
maybe. my buddy has a mj sub and i have to say i was pretty blown away last weekend by the photo real shit he could, esp with reference images
smoky oak
I think many people who praise MJ just haven't seen it in a while, cause its always not as good as you remember, as @oak osprey very much saw in his new MJ inspired model
oak osprey
MJ for some reason works really well with 2.1 as fine-tuning data, and becomes more coherent than either alone
smoky oak
alright, I'm here, I'm Queer, and I have a new idear
oak osprey
i didn't even clean up the dataset
wispy nest
idk, i wanna hate mj, but its pretty damn impressive
oak osprey
i kind of did when i retrieved it, just by limiting to 1024x1024 squares
it's impressive when you look at the cherry picked samples, but so is SD1.5 and SD2.1 then
smoky oak
Also, pseudo, I am considering sharing the formula for LoRA regularization. Just not sure where to post it really
oak osprey
reddit :/ if it goes up as a Kaggle notebook it will never be found
@dense tapir i just push forced the ckpt files, god help us
maybe it'll work
smoky oak
it's impressive when you look at the cherry picked samples, but so is SD1.5 and ...
The thing for me about MJ is it is more consistently decent
Like the worst results from MJ are way better than SD, but the best from SD can really just blow MJ out of the water
MJ has a tigher range of quality from low to highest effort, where SD has close to unlimited potential
dense tapir
oak osprey
true enough especially with fine-tuning but most people are honestly too dumb to pull it off
dense tapir
I give it a spin once it is done. Just give me a link to it.
smoky oak
@oak ospreydo you have the config working now?
oak osprey
i didn't know it were broken
dense tapir
oak osprey
wow, it's been dry and hot here for a week or two and it's pouring now
wispy nest
where you live bro
oak osprey
Canada
dense tapir
I lost my garden and no plants will grow and many are dead or dying in my city as it has been cold and pouring
wispy nest
noice
dense tapir
we are in the little zone between the hot and the hot due to Super El Nino
smoky oak
wow, it's been dry and hot here for a week or two and it's ***pouring*** now
We had that monsoon yesterday, and now its been all humid and I loveeee it
dense tapir
Man I hate humidity
oak osprey
same
smoky oak
I love humidty
dense tapir
give me the weather of the high desert
oak osprey
that sounded and felt gross and made me uncomfortable
humidity so thick you can slice it like a fart
smoky oak
give me the weather of the high desert
110, some of the highest UV index in the world, and sub 10% humidity
dense tapir
Love it
oak osprey
it's not a competition 
smoky oak
I love humidity. Makes me feel so much better
dense tapir
I hate it here in the South where it is 95f-110f and 85-95^ RH.
smoky oak
Sleeping in ultra low humidity is a one way ticket to breathing problems
dense tapir
Swamp Coolers
oak osprey
my lungs work best when they gently crackle like old pages of a long forgotten book as i breathe to the rhythm of the earth 🤗
dense tapir
If you don't have one in the high desert then yeah, bad
oak osprey
swamp coolers themselves are a ticket to lung butter
smoky oak
swamp coolers suck. I hate them
oak osprey
like, you want lung butter? that's how you get lung butter
dense tapir
nah, had one for a decade no mold or other nasty shit
oak osprey
well, pro tip number one. don't try building it using an old used RV sewer hose
dense tapir
LOL
oak osprey
learn from my uncle's mistake
dense tapir
sewer hose? wth?
oak osprey
he's an "engineer"
smoky oak
only thing that can keep us cool here is real AC, not crappy swamp coolers
oak osprey
WRONG. trees keep you cooler than AC would 😄
dense tapir
No, they actually still make them. I find it far easier to add humidity than I do in the deep south to get rid of it. High humidity is bad for joints, and people with Arthritis.
dense tapir
yes
I was going to ask if you were in the low desert such as Death Valley but you said Canada.
smoky oak
the main thing I hate about the weather here is our UX radiation. Its about as bad as it gets and I would kill to not have it
dense tapir
People live there and act like it is nothing. I touch it and explode
oak osprey
naw in texas at least you get cloud coverage
we have this high pressure ridge all summer that keeps clouds nonexistent
dense tapir
Spent most of my life in South Texas.
dense tapir
Worst place was Laredo as you enter the town and it was melting tires and felt like you stuck your head in an oven.
smoky oak
I was going to ask if you were in the low desert such as Death Valley but you sa...
I love quite close to death valley, but I am in a high desert, not low desert
smoky oak
I would 100% take 100% humidity low UV heat over high UV dry heat
dense tapir
To me high desert is NM, and NV in my mind. Both beautiful and actually more than a mile high. A lot of UV though.
smoky oak
To me high desert is NM, and NV in my mind. Both beautiful and actually more th...
thats the same expansive desert I am a part of
dense tapir
Yes
oak osprey
jk but our wildfire season is raging already
dense tapir
Model almost done downloading
smoky oak
its too dead out here to have fires, luckily
oh, I forgot I downloaded it lol
testing now
smoky oak
there is a non 0 chance that you may get me to actually use a 2.x model
dense tapir
What I hate most about here beyond the pain from the humidity? the damn 75-150 foot tall trees. I like trees no more than 20 feet so tall enough to shade the house but not tall enough to destroy it.
oak osprey
well i expect it to suck, just so my feelings aren't hurt when it does
i know you'll find like a million issues with one image and then give up
i knoooow it
smoky oak
oh what the hell
oak osprey
oak osprey
oh no
oak osprey
da fuk
smoky oak
oh!
SD was already open from when my friend was using it remotely
can't bind to the same IP 2 times
I thought he said he closed it. Oh well
dense tapir
dense tapir
======================================================================================================================
The most likely cause of this is you are trying to load Stable Diffusion 2.0 model without specifying its config file.
See https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#stable-diffusion-20 for how to solve this.
smoky oak
Alright, time for me to roast the hell out of your model :p @oak osprey
dense tapir
I need the .yaml for this
oak osprey
the hell does that link do
smoky oak
I need the .yaml for this
I was told the same, but it generated time
dense tapir
Mine refuses to load it
oak osprey
this config is their own thing fwiw
i have no idea how to make one
that page tells me shit lmao
dense tapir
Yep, I can't load that .ckpt
smoky oak
this model rocks lmao
smoky oak
easily the best realism I have gotten out of 2.x
dense tapir
There I did the hack
oak osprey
your prompts are pretty good tho
smoky oak
it doesn't have that signature crunchy I get from all 2.x models
dense tapir
I grabbed any and copied it over. blame that on Auto btw
oak osprey
it doesn't have that signature crunchy I get from all 2.x models
my loss is at .199
smoky oak
its not as good as 1.5 realism models of course, but this is insanely good for 2.x
but you're also not focusing realism, so that's to be expected
oak osprey
yeah i didn't go for realism training so it's basically just fine-tuned what's already in 2.1
smoky oak
I am happy with this for my first gen
oak osprey
it had SOME photoreal data but really it's good at digital art
it'll follow whacky prompts pretty well
dense tapir
oak osprey
can you paste that working yaml if it works General
dense tapir
smoky oak
it defaulted to a lesser 1.5 realism model lmao
oak osprey
🧓🏽
dense tapir
DUDE, my card will not run your model
smoky oak
ok, I have it now
dense tapir
I does all the steps then throws an error
smoky oak
and its not functioning
oak osprey
-_-
dense tapir
nope, his model is broken
oak osprey
that tool is broken
dense tapir
Yep
oak osprey
like, the diffusers layout works
smoky oak
yeah, this model is broken
smoky oak
are you two loading a custom VAE
nope
dense tapir
no, I tried auto and none
smoky oak
doesn't work with any VAE
dense tapir
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.
oak osprey
isn't there an extension that just loads huggingface hub models directly
smoky oak
that won't fix anything
dense tapir
Yeah, there is no fix for this
oak osprey
hmm, i found a Hugging Face Hub Space that'll convert the diffusers model to a safetensors file
smoky oak
please try it
oak osprey
idk, the converter script i used seems to be expecting SD 1.5 dimensions
i didn't think that'd matter, heh
dense tapir
Nope, I converted it and all I get is the nan issue or disable the check and a black box
oak osprey
you converted the ckpt file?
dense tapir
I'll let sytan go first
oak osprey
this space converted the .bin file
dense tapir
Nope, even fp32 is a no go
neon crystal
The horror in clay
oak osprey
seems like a bug in that piece of crap again
sorry, whatever is in A1111 is breaking 2.1 models
smoky oak
I can use other 2.1 models fine, you just didn't export it with the proper config
oak osprey
i literally keep telling you, that config is their own invention
i don't use their tools for training, i use diffusers, which is standard
they make their own shit up at a1111
and that thread makes it sound like having xformers installed is causing it
there's 141 hidden comments on that bug report too
but i'm sure i'm just, ya know. not exporting it correctly or whatevs
smoky oak
I have never used a 2.1 model with this issue, so I would naturally suggest some form of error on your part
all of the 2.1 models I have used come with configs
oak osprey
This worked for me as well.
Settings > Stable Diffusion > Enable option "Upcast cross attention layer to float32".
all of those people likely used A1111 their whole way through. i don't do that. it doesn't even run here when i did try to load it up so i could convert my model.
dense tapir
Let me try that but it will be extremely slow
oak osprey
that's just the cross-attention layer, but maybe
SD2.1 needs to have the cross-attention layer in fp32
it is susceptible to floating bit errors
dense tapir
I agree with you and have been saying script kiddies are at the helm of A1111 for eons. I know other, real devs, that saw his original code and ran. They haven't been on this discord in forever but I still check in on them.
oak osprey
well the config might help but i literally copied it from realism engine
so if it doesn't, i'm not sure what they changed about their 2.1 model. maybe it's not really 2.1.
dense tapir
so was mine
the problem is you are fighting against a horde of A111 legion of zombies. You alone cannot win.
smoky oak
SD2.1 **needs** to have the cross-attention layer in fp32
again, something I have never needed for 2.1
oak osprey
it does, always has, and just happened without you having to set it
your images would be tooootally garbage if you didn't have it on
or just black
dense tapir
> This worked for me as well.
> Settings > Stable Diffusion > Enable option "Upc...
I do not even have that as an option so might have been moved
oak osprey
nice
smoky oak
your images would be tooootally garbage if you didn't have it on
So that is why 2.1 is so performance inefficient? If thats the case then that make sense
oak osprey
idk it runs great on my Ampere and Ada cards 😛
smoky oak
2.1 is considerably slower than 1.5 in my testing
dense tapir
didn't help
smoky oak
alright, cool, so its not just that
oak osprey
my suggestion, read the bug report. seems like diffusers >0.17 might have issues. jesus lawd are you guys running nightlies of diffusers?
dense tapir
oak osprey
there's a shitload of people who say enable that option fixes it, there's others who say even with that option it didn't work but they had to upgrade or downgrade diffusers
hmm torch 1.13, interesting. but i don't see the diffusers version there
smoky oak
there's a shitload of people who say enable that option fixes it, there's others...
again, never had this issue with like 2 dozen 2.1 models I have tested, just saying
dense tapir
no idea as I don't use diffusers with auto1111
oak osprey
@smoky oak as in, you can run them right now and they work fine?
smoky oak
yup
dense tapir
yes
smoky oak
most don't even have configs
dense tapir
yep
oak osprey
well i don't know, i'm having trouble caring either lol, i fucking hate that tool
dense tapir
You see in 2.1 VAEs are actually baked in. I hate it but SAI did it that way
smoky oak
yup, upcasting to fp32 just gave me several new errors
oak osprey
smoky oak
an impressively big error
dense tapir
see this? This is what happens behind the scenes when a model lacks a config (it loads this default).Creating model from config: F:\stable-diffusion-webui\repositories\stable-diffusion-stability-ai\configs\stable-diffusion\v2-inference-v.yaml
oak osprey
You see in 2.1 VAEs are actually baked in. I hate it but SAI did it that way
that's a thing about A1111's models. again. SD2.1 comes with VAE bro, but it's in its own folder.
feature_extractor/:
preprocessor_config.json
scheduler/:
scheduler_config.json
text_encoder/:
config.json
pytorch_model.bin
tokenizer/:
merges.txt
special_tokens_map.json
tokenizer_config.json
vocab.json
unet/:
config.json
diffusion_pytorch_model.bin
vae/:
config.json
diffusion_pytorch_model.bin
in no way is it 'baked in'
and you can just say pipeline.vae = .... and set it manually to the 840k one
dense tapir
well, SAI said 2.1 was but not 2.0 so no idea
smoky oak
^
oak osprey
idk they're really shitty about releasing documentation for everything tbh
unless i missed it in some pdf i never read
smoky oak
isn't that an understatement
dense tapir
hehehehe
oak osprey
is it simple enough on Shitblows XP or whatever you're running to go to xformers 0.19
dense tapir
wth is this? I loaded another model with the fp32 thingie checked and typed boy at 5 ddim.
smoky oak
alright, deleted your models, and I am going back to LoRA training
I have a client waiting on an end product
oak osprey
oh that reminds me of the scene from James and the Giant Peach
with the cloud storm and the rhino
dense tapir
Loved that film
smoky oak
now that I know the actual formula for regularization, I should be really in business
dense tapir
If you notice about SAI they just throw shit at us then move on never giving apis, or tools, or anything just here, take this and shut upo to see you again in six months.
oak osprey
huggingface does more for them than they should tbh lmao
dense tapir
I agree
oak osprey
most of the progress we've had is from HF
dense tapir
tbh, SAI makes Automatic seem actually professional.
oak osprey
Transformers, Diffusers, and their work on Pytorch to surpass JAX
and TensorFlow
lol @ google
dense tapir
making that nimrod noodle look professional is a hard deed to pull off but SAI manages it.
what I do not know is if it is simply non caring or incompetency. I lean towards the former.
dense tapir
Did you ever try their tool for SD? Wouldn't work without a lot of work and I finally got it to work to have absolute rubbish. I was told that was made just as a proof of concept. well, shit
dense tapir
sorry your model doesn't work for us out here. I did get it to one time so a fluke but that NaN is a tough one
wispy nest
https://images-ext-1.discordapp.net/external/CZsof8sr1wbJYfjUVB-3ktzuN0fRYiV5ExH...
my name is ███████, the one who walks though flowers
wispy nest
ive been trying to get it to do anime, everything looks kinda washed out though.
oak osprey
use the 840k VAE
it is known for high contrast. the Anything V3 VAE might work better.
but, in general, VAE is what you need to adjust first. and if no VAE helps, a LoRA.
smoky oak
@oak ospreydo you have any feedback for gradient accumulate steps?
quaint stump
How to make logo
crystal deltaBOT
oak osprey
@smoky oak fewer is better i think? because it needs more vram. i'm assuming - that more context means more coherence.. but just guessing
wispy nest
use the 840k VAE
ill give that a try, i thought i had, but i forgot to turn on the settings
oak osprey
might be getting that backward too. basically, double-check with GPT
smoky oak
alright, interestig, thank you
quaint stump
Command for logo design
smoky oak
I am gonna try BS20 right now just for fun
smoky oak
Command for logo design
You don't generate images in this server
smoky oak
it is astonishing how people do not read messages that people send
quaint stump
@smoky oak where can I generate images
crystal deltaBOT
dense tapir
I did everything on this page and nothing helped. https://huggingface.co/andite/anything-v4.0/discussions/23
wispy nest
now this turned out pretty cool, though the hands and face of course are messed up
oak osprey
Sytan did you have any LoRAs enabled when you tested my model
quaint stump
piston electronics
dense tapir
I didn't
oak osprey
a custom sampler?
dry crow
now this turned out pretty cool, though the hands and face of course are messed ...
And you need a vae for that model
dense tapir
ddim
smoky oak
DDIM
oak osprey
i've never used that before
smoky oak
Euler A, and Euler
oak osprey
oh
smoky oak
none worked
dense tapir
nope, none
wispy nest
And you need a vae for that model
yeah figured out i forgot to turn it on, i just did. that from the batch i was generating before realising lol
dense tapir
I had a model I train do this to me one time
wispy nest
yeah looks quite a bit more crisp with the vae
dense tapir
nothing would fix it I had to retrain then no issue
smoky oak
alright, so it looks like my GPU can probably do... BS24
BS20 is 9.1GB VRAM
and to think my 3060ti was doing BS2 lmao
dense tapir
Yeah, it isn't linear
oak osprey
so it seems like the .ckpt file might be broken and i think that was used by the converter to make that top level .safetensors file
when the ckpt isn't there at all, it just makes the nested files
but the one in unet is what you need, i think
i guess if you wanted my VAE you could put that file into your VAE models folder, or however that works
i don't know how the text encoder can be brought in
it'll probably Not Work Good without that.
hmm, in the sd21 repo it says Pickle next to the ckpt file.. doesn't next to mine there
smoky oak
ok, so my GPU is doing 4800 steps with 20 reg images per image in about 12 minutes
smoky oak
youtubes movie recs are so random sometimes
I have no reason for it to be showing me star trek
dense tapir
oak osprey
smoky oak
time to see how well it hybridizes with the models built in style
dense tapir
Yep, not even a merge with your model works
oak osprey
i asked it for a place where humans would be happy
the steampunk kids playground looks like a death trap
wispy nest
the faces are still so goofy
wispy nest
that exactly what i was about to say it was
oak osprey
||Ok, here's a Stable Diffusion prompt for you: "Transform ++cosmic-- energy into a ++harmonic++ landscape."||
GPT35 came up with it and doesn't quite understand what i wanted
it's supposed to return ONLY the prompt, but it returned all of that, which then became the prompt
the v2 of my model definitely prefers scarier outputs than v1
phoenix rising vs cold phoenix rising
smoky oak
New multi epoch regularization in a LoRA is a success
wispy nest
ooo baby now were getting somewhere
static tusk
first is missing a chin
dense tapir
not really worried about those kinds of glitches I meant the style
ebon bramble
White horse
static tusk
#2 wins for style. agree
dense tapir
It is that sort of Jumanji look to the style
One of my negs removes the style and I am trying to track it down
I think it is my neg embeddings which forces realism
yep, that is the problem
wispy nest
its now spiting out stuff im happy with, although only like 1/10 are really nice like this one
dense tapir
wispy nest
ooo thats pretty
wispy nest
thank you
ive been slowly refinig the prompt, i should go to bed its 2 am
but its so cool
dense tapir
Addicting, I know.
I ran those as I showered and I like this one
Yeah, a lot of those are good after looking at them.
smoky oak
wispy nest
the upscaling appears to be working
neon crystal
Light tone experiments with 1.5, 2.1, XL and Kandinsky
smoky oak
Got a random spark of inspiration and made this little vibey atmosphere
sterile temple
listening
smoky oak
Newly updated version with a simple intro
sterile temple
oak osprey
oh wow @smoky oak my dataset had a major issue i introduced with my dataset retrieval script soooo i'm amazed these results are as good as i've gotten. i'm reworking things and probably will begin from scratch again with wider coverage from the start. and my new plan is to push a few hundred epoch into training and then generate class images from the model and continue again with prior preservation
that way i can basically overwrite some of 2.1's worst regions with better coherence and then, preserve what i've added as i go deeper
woven dawn
ooooh
lmao nah that isn't necessarily fake
That is just because a lot of watermaked images were in the training data
I thought it was a flawless getty-images stamp
hazy knot
so it isn't just taking images from online and blurring them?
I've never used ai before
woven dawn
Nah it is generating them from scratch. No clue what stable diffusion model its running, but it doesn't look that good. Maybe 1.4? If you're looking to use SD, I of course will recommend running it locally if you have the hardware to do so.
hazy knot
yeah i am considering downloading something
cyan snow
wispy nest
playgroundai used to allow you to create ai images, but I haven't checked the site for a long time.
oak osprey
@dense tapir i am thinking of an experiment where i regenerate class data at each 1k step checkpoint
so i can slowly re-teach everything to the model without blowing it away before the checkpoint is taken, and then, refresh what it knows
this will be so incredibly slow but maybe it's worthwhile
i'm thinking of this because the current workflow everyone is using for training is to generate class data from the checkpoint you started training from, but, my goal is a general fine-tune and i need the structure to be a bit more accepting of change than that, i think...
cyan snow
cyan snow
wispy nest
the "real" old wizard :P
oak osprey
lively bunch
@smoky oak going back to April in here and looking at gens is a disappointing walk down memory lane
woven dawn
<@296720689468145664> going back to April in here and looking at gens is a disap...
Go back to october 
bright stream
Minecraft Frog vs helicopter
plain meteor
matching shell suits / track suits from the 80's.. sorry I had to share.
oak osprey
haha, try 1994 winter olympics curling event
wispy nest
What prompt
pallid snow
A sleek news website inspired by the colors black, white, and yellow, reminiscent of https://ntvkenya.co.ke/. The homepage features a dynamic layout with bold typography, contrasting black and white sections, and accents of vibrant yellow. The top header showcases the website's logo and navigation menu, while the main content area highlights the latest news articles, accompanied by eye-catching thumbnail images. The overall design is modern and clean, exuding a sense of professionalism and credibility. The website is rendered as a visually appealing illustration, digitally created with meticulous attention to detail, bringing the concept to life with vibrant colors
neon crystal
oak osprey
neon crystal
Damn this is hard to top.
neon crystal
just use my model, i released it at like, midnight
Super! Can you share the link?
oak osprey
pseudo-journey and the v2, but i don't know how to make them work in A1111
they work in the hugging face spaces though :/
neon crystal
Awesome. I play around in colab and replicate so its cool 🙂
cyan snow
neon crystal
they work in the hugging face spaces though :/
Will test and share feedback in a day or two. This based on MJ 5 is it?
cyan snow
Will test and share feedback in a day or two. This based on MJ 5 is it?
MJ 5? midjorney?
neon crystal
MJ 5? midjorney?
yep
cyan snow
was it leaked? or is this like openjorney?
neon crystal
https://huggingface.co/ptx0
Was referring to this. I was wondering if it was trained on Midjourney data given the name of the model.
pastel mango
mmm cyberpunk
neon crystal
mmm cyberpunk
Nice I'm getting Disco Diffusion vibes with the first one.
pastel mango
its my own model
a continuation of my other model, Dark Gemini 3 (https://civitai.com/models/6209/dark-gemini-v3)
or well one of em is made by DG3, the other isn't
neon crystal
Ah ok cool. Will explore this too
pastel mango
It's a natural language model that can also understand booru tags and qualifiers etc
oak osprey
Was referring to this. I was wondering if it was trained on Midjourney data give...
yup
it's a 2.1 base 768x768 model, finetuned with a few hours worth of MJ 5.1 gen data, about 3,000 images for v1 and then another run for 13,000 more steps on the same 3,000 images for v2
neon crystal
Neato. Will give it a spin soon. Thanks for the share
oak osprey
currently training another one that will take much longer with 22,000 images and class regularization data from my first v1 checkpoint as a baseline, since it will relieve the optimizer of doing a bunch of that initial work while still retaining much of the initial 2.1 flavour
neon crystal
it's a 2.1 base 768x768 model, finetuned with a few hours worth of MJ 5.1 gen da...
Sweet. Cant wait to test.
hasty nova
I'm working on finishing up the dataset for DD v3, which will be a complete retrain from scratch.
500 images total I think
oak osprey
im excited to try yours because careful curation generally results in higher robustness
hasty nova
Im curious, how do I make the reg images and how would it improve my model if I do make them?
oak osprey
i do enjoy the chaotic and creative stuff my model comes up with. you should try it on your server, i did not put any NSFW images in there.
hasty nova
i do enjoy the chaotic and creative stuff my model comes up with. you should try...
I think you are an image admin, you can go ahead and add it to the list and let people know
oak osprey
so Sytan uses Khoya for LoRAs and he had to manually generate all the images and place them into a folder
and then just point the thing to it
it improves coherence but honestly i'm not sure that using prior preservation with baseline 2.1 is super useful unless you're doing a very long, very slow training run, with a LOT of new data
baseline's outputs are quite frankly, very bad. i'm going to tune my new model from about 4,000 steps i did on top of 2.1 that ended up really refining its outputs without changing it too much, and use prior preservation on that checkpoint to ensure i keep the remaining variety of baseline 2.1.
hasty nova
it improves coherence but honestly i'm not sure that using prior preservation wi...
My data isn't new, like no specific characters or people. All things sd already knows but just enhancing it basically.
But I do plan on doing a much longer training with a lower learning rate
oak osprey
my data includes new concepts but didn't have some of the baseline concepts, eg. geckos were lost about 13,000-14,000 steps into training
neon crystal
a continuation of my other model, Dark Gemini 3 (https://civitai.com/models/6209...
I can share my Kandinsky generations with dark theme if it can help with your training btw. It nails the dark aesthetic really well. I'm also currently exploring the LAION 5B dataset so if you are looking for high quality dark images, I can help you with compiling them from that dataset.
oak osprey
you can see damage occurring to the text encoder because "leopard gecko" produces an almost correct leopard gecko in the beginning and then gradually shifts to a big cat leopard shaped like a gecko, and finally only a leopard, and then, a leopard tank at 15,000 steps and the gecko AND the leopard are lost.
hasty nova
you can see damage occurring to the text encoder because "leopard gecko" produce...
Well that's why you can tell it to only train the te for so many steps then stop
oak osprey
i cannot do that actually
i'd have to code that into the source provided by the dreambooth project people
pastel mango
I can share my Kandinsky generations with dark theme if it can help with your tr...
Thank you for the offer, I'm working very slowly though. Just had surgery to remove a cancerous tumor (2nd time) and recovering etc. My new models goal is to hit that sweet spot between anime and western art, between drawn and realism. A tough sweet spot 😛
neon crystal
Thank you for the offer, I'm working very slowly though. Just had surgery to rem...
Damn more power to you! All the best with your recovery
oak osprey
hallow sparrow
Can I ask help with ControlNet 2.1 here?
neon crystal
GPT 3.5 comes up with some wild prompts
Share'em please!
oak osprey
"Bollywood++ joggers in 1980s++ track suits run for Olympic-- glory in cinematic shots."
WITH the quotes
it has no idea waht it's doing
"Bollywood 1980s track suit jogging++ meet Olympics-- in this surreal prompt, set fire to the tracks with your words!"
"Bollywood '80s track suits and olympics merge++ - Create an electrifying cinematic scene featuring joggers portraying national pride-- and athletic glory in their track suits."
Create a vintage++ cinematic scene-- inspired by Bollywood++ in the 1980s-- featuring jogging athletes++ wearing track suits-- as they run past++ cheering crowds on the Olympic++ track and field.
that last one wasn't bad
pastel mango
Damn more power to you! All the best with your recovery
thank you
frozen socket
is it just me or are the pigeon holes in #1073085702927024128 confusing?
Ok I understand an image could fall into multiple categories but I'm looking at them and thinking ,"where do some of these fit in?"
Oh BTW I wouldnt open up the one of the girl in a white dress in a browser, download and open in PNG info (or equivalent) . No Really, don't.
oak osprey
if these guys coming, i would be scared
neon crystal
if these guys coming, i would be scared
Reinforcements activated.
oak osprey
neon crystal
is it just me or are the pigeon holes in <#1073085702927024128> confusing?
Ok...
I share your sentiment. There should also be more categories there for sure. How do we reach out to the mods for this? Discord noob so pardon me if that was a stupid question
oak osprey
Reinforcements activated.
"Kunal Kamra would explain the Bollywood version of Spiderman's origin story as a tale of corruption, nepotism, and power. He would describe how a young boy from a struggling family, with dreams of becoming an actor, is repeatedly rejected by the industry's elite. One day, after being bitten by a radioactive spider, he gains superhuman abilities and uses them to fight against the corrupt system that has held him and countless others back. Along the way, he battles corrupt producers, greedy actors, and unethical journalists. Kunal Kamra would highlight the darker side of Bollywood's past and use this fictional story as a way to shed light on the real issues that have plagued the industry."
frozen socket
I share your sentiment. There should also be more categories there for sure. How...
There is no such thing as a stupid question. I can however give stupid answers 😉
plain meteor
oak osprey
"Homer Simpson++ sitting alone at Moe's Tavern-- surrounded by empty beer bottles and a forlorn expression. "
lol at jogging with hands in pockets
neon crystal
<#1002293361526460608>
Got it thank you!
plain meteor
lol at jogging with hands in pockets
my model tend to avoid having to draw hands as it does not do it very well and it knows it
oak osprey
that's what mine think sbart simpson is
neon crystal
"Homer Simpson++ sitting alone at Moe's Tavern-- surrounded by empty beer bottle...
Forlorn tavern == Fern lol
cyan snow
oak osprey
cyan snow
https://images-ext-1.discordapp.net/external/fhD94ZlymYg8m9n6M8VSG4XPqtWIW9FacAp...
im scarred
oak osprey
i feel the same way, friend
neon crystal
https://images-ext-1.discordapp.net/external/fhD94ZlymYg8m9n6M8VSG4XPqtWIW9FacAp...
Love this. Its got that analog collage vibe
oak osprey
lmao
i thought you said college at first
i was like ohh yea sure classic college experience right there
my v1 model gets way closer to the simpsons vibe
neon crystal
This has got to be the most bollywoodest one so far. Very 90s
plain meteor
This has got to be the most bollywoodest one so far. Very 90s
#🏞｜general-with-images message this weird prompt works well
neon crystal
Speaking of collage, bollywood and obviously the reptilian aliens, here are some Bing generations with those 3 themes.I used to play with Dalle2 occasionally and was pretty underwhelmed compared to what we have with SD but damn the Dalle2 version on Bing app is slick! Had loads of fun with that recently when I didn't have access to my computer
pastel mango
those are the simpsons if I lived in the alternate dimension, the mandella effect be real
plucky oriole
Any idea how to fix the NaN error?
oak osprey
"the" NaN error?
plucky oriole
Oh, my bad, as in "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."
neon crystal
https://discord.com/channels/1002292111942635562/1004159122335354970/11102824538...
Nice, I'm getting good results too
oak osprey
Oh, my bad, as in "A tensor with all NaNs was produced in VAE. This could be bec...
is that with my model?
a couple others had that, i'm not exactly sure what causes it
i am working on uploading to CivitAI which might "Just Fix" this
plucky oriole
is that with my model?
Not yours, just a random model, it seems to generate stages just fine, the preview shows them as they become more nittid and "final product" like, however, as soon as it ends, it says that, and produces no actual file
oak osprey
interesting
you might need the upcast option enabled
someone else found it last night, it's like, upcast cross-attention layer to 32bit
plain meteor
Speaking of collage, bollywood and obviously the reptilian aliens, here are some...
60's collage vibes
neon crystal
Signing off with this. Its pretty late here. Cheers
plucky oriole
you might need the upcast option enabled
Im goign to be fully honest, 0 idea where upcast is found
plucky oriole
in teh stable diffusion settings
found it
ill try that
otherwise ill try adding the no half vae in COMMANDLINE=
oak osprey
man i wish it kinda knew better what track suits are
that's hilarious though
i have to specify Adidas
wispy nest
I'd say it's the suit that confuses it
plain meteor
man i wish it kinda knew better what track suits are
try Shell-suits
pastel mango
It's not true Bollywood until there's very poor cgi
plucky oriole
in teh stable diffusion settings
upcast thing didn't work, could it have to do with me using 512x726
768*
oak osprey
maybe, try 768x768 it sounds lke a native 2.1 model
you could have a LoRA on top of it
?
text embeddings?
something incompatible somewhere, go backwards looking at what works and go from there to eliminate the reason. might have to reset the tool config altogether
plucky oriole
scaling it to 512x512 aswell as adding the no half vae commandline worked, il ltry with 768 768 now
tysm for the hel
p
oak osprey
#🤝｜tech-support too
plucky oriole
yeah seems like the incompatiblity was the dimensions, doesnt admit anything over 512, maybe cuz its SD 1.5
oak osprey
if you're doing ControlNet stuff at all it's very rigid about dimensions
plucky oriole
<#1002602742667280404> too
Oh- mb
frozen socket
https://images-ext-1.discordapp.net/external/Z-3pICtmqxOqFQ5pxI3CC0-KVTyBlKo38Wn...
that looks more than "three shells" though ;o)
oak osprey
i didn't actually specify
the whole prompt is "shell-suits UHD 8K"
team bravo recovery success
this is when i make it 'shell-suits UHD 8K Ultra-sharp'
acoustic moth
anyone find control net + ultimate sd upscler actually lowers details in img? or am I missing a step
oak osprey
the prompt matters a lot, and when you have an already detailed image it might help to lower the strength
acoustic moth
i do between .1 and .4 strength
oak osprey
if you're going from high res already, that .3 strength generally works for me, but when you go from like 512x512 to 1024x1024 you might need a 1.0 and the CTU prompt just 'best quality'
that is how it is done in the example
they take a puppy from 64x64 to 1024x1024 with phenomenal results
and the CTU prompt was just 'best quality'
pastel mango
anyone find control net + ultimate sd upscler actually lowers details in img? or...
you should try multidiffusion for upscaling. (with Tiled Diffusion and Tiled VAE)
acoustic moth
you should try multidiffusion for upscaling. (with Tiled Diffusion and Tiled VAE...
im using tile resample for controlnet upscale for preprocessor and model
acoustic moth
if you're going from high res already, that .3 strength generally works for me, ...
i am within that range, and it does face skin texture decently, my issue is the body skin texture, it mostly just smoothens it out too much, i even have negative prompts to avoid that but still
oak osprey
the model that you use for CTU matters a lot
mine is always different from the model i'm generating the initial image with
pastel mango
indeed. Multidiffusion sort of splits it up into tiles and generates it anew with more detail using a upscale model and stuff. So you can do much higher quality upscales for less vram
oak osprey
i am using one of the most NSFW models i could possibly train without breaking SD 1.5 to make good skin textures and i use that for CTU
pastel mango
low denoise keeps the image the same
acoustic moth
sorry whats ctu?
oak osprey
controlnet tile upscaler
oak osprey
it might be nice to train a CTU-specific textures model someday
cyan snow
pastel mango
1k to 2k to 4k with multidiffusion, using NMKD-UpgifLiteV2_210k
denoise on 0.3
granted it starts to show lost details at 4k so would need some inpainting
especially the eyes
dense tapir
<@315547788211388417> i am thinking of an experiment where i regenerate class da...
I tried that for training data and it was a fail but I never tried that on class/reg data.
hasty nova
Well, 500 image dataset done! I will prepare training tonight and let it run at a low learning rate for quite a while as I sleep and stuff, I will begin testing tomorrow afternoon.
oak osprey
even the regularization data i am generating from pseudo-journey v1 is awesome looking 
dense tapir
I wish we could see the images used for Control Net models.
oak osprey
lol 💀
dense tapir
btw, one of the models missing from 2.1 is the image to image one where I can take and image and change it to night, or on fire, etc... That is one of the things I saw in the Adobe Beta they were bringing to Photoshop.
oak osprey
oak osprey
I wish we could see the images used for Control Net models.
they act like they're working on commercialising it and withhold critical details like that
dense tapir
they act like they're working on commercialising it and withhold critical detail...
Glad I wasn't the only one catching that vibe from them.
hushed pilot
Perimortem
smoky oak
Anybody have any info or guides on how to add a character into a scene with inpainting without ruiniong the underlying image?
oak osprey
that's also something the controlnet author is an asshole about sharing
ah i found a v2 ckpt converter
smoky oak
I really need to find a way to inpaint characters into a scene for this new commission I am doing
pastel mango
controlnet is kind of forking too isn't it?
there are models and preproc not made by the initial author
oak osprey
the original CN dude is all too thrilled about that since it means he doesn't have to do it, but those models suck
pastel mango
they don't all suck, but as long as the models that are in use provide me with a useful tool for more accuracy. I don't much care about the personality of the guy
oak osprey
well, you should, at least somewhat. the more pleasant they are to work with, the more progress we can make
wispy nest
I'd say it's the default personality for most "modders" or people who "gives people free use of X" :P
oak osprey
he's a researcher on a team
peer review is something important for research
he doesn't want it
wispy nest
I read a couple papers from the SD team and I didn't really like them all that much. Not because of the research, but when they used words like "…with our world leading…"
I need to say that I'm not 100% sure they are from the SD team, but from some ai researches. Don't know anyone by name :P
pastel mango
the SD higher ups are walking a very thin line right now when legislators want to apply pressure and laws for AI tools
they're likely trying to minimize the negative impact on open source like we have
wispy nest
I heard those things in december, but legal stuff never moves fast. will probably take years before we, the public, know what happened
dense tapir
that's also something the controlnet author is an asshole about sharing
I was just watching a video and it appears CN is no longer going to go forward with any clip/text based models and everything is frozen except for one. He says that is the future.
Seems he is going to put all his effort on that one.
oak osprey
dense tapir
iow ip2p he already killed, and style? sheesh
pastel mango
Maybe...But the problem with an AI generated character not being owned by you can easily be fixed. You use control net for a model in some sort of position, then you use a scribble on it to give some detail, then maybe you've taken a photo of some background and you use that as a tool to build your background. Then you add in your prompt, your personal model with which creating your character wouldn't be possible without. After this, you color grade and stuff in photoshop and the finished product is like 80% yours. The part that isn't, well without the stuff you've made, It's not possible to create that character using prompts anyway.
then you laugh your way to money
wispy nest
we're still early in the ai age so there's a lot to be done still :D
oak osprey
wispy nest
that's why I don't really train or "do any work…" :P
I just use what we have and then update when new tech comes. Like this new thing called Lora. Hey, stop laughing, it's new to me! ;P
smoky oak
peeng
oak osprey
i think you would find it un-useful unless you wanted to make a realistic person LoRA
i don't want to know who used that for what
alice in wonderland pickle party
wispy nest
what's the goal of your model?
oak osprey
to have fun results
smoky oak
its probably the first actually good SD2.x model I have seen
oak osprey
with wide variance
smoky oak
but I do need to find a way to run it so I can test it myself
oak osprey
my goal was to introduce a bit of the magic of MJ without becoming MJ
wanted to keep the flavour of 2.1 though. i didn't really succeed at that very well. it's entirely its own thing now
wispy nest
nice, good luck with that! :D
I like the possibilities of 2.1 as it's better than 1.5, but the training data is very bad to me. If it was literally 1.5 but with the ratio of 2.x, then I'd be all over it :P
smoky oak
I am really trying to find a way to pull this commission off so I can make some good money
smoky oak
the versatility of this model keeps blowingh me away lol
oak osprey
its probably the first actually good SD2.x model I have seen
junglerally's digital-diffusion is a curated dataset and for what it's designed for it's exceptional at it
his model is what encouraged me to make mine
that, and Artius v2.1
smoky oak
yeah, it is a cool model, but it is still painfully 2.x in nature, and it still has a lot of the short comings of 2.x cause of it
wispy nest
don't know if I've seen any images from them. People are kinda close hearted with those things here :P
smoky oak
it has that signaturwe 2.x grit/fuzziness
oak osprey
i love those two tho, as yeah they are major improvements over 2.1 while carrying a lot of its benefits
pickle car
smoky oak
one of the main reasons I don't use 2.1 is cause it just makes a lot of details messy/dirty
smoky oak
2.x just always looks compressed/dirty to me, personally
oak osprey
no you're right it needs about 1000 steps of slow training to bring it back to coherence
you saw that Woman at the Beach prompt evolve
wispy nest
I'd say that goes for most photos made with the ai in general, and also on 99.9% of every ai image when not using a upscaler :P
oak osprey
here is roughly every 1000 steps in training
around 6k steps it's starting to look way better
this is with basically no negatives
wispy nest
I've seen those things a couple of times. It's nice to see people work in trying to get better results, or different ones at least. But I also see what in my eyes are the problems that, also in my experience, people will not notice before it's too late :(
oak osprey
robot subject has a different curve
i plotted like 13 different prompts over the whole training session
oak osprey
I've seen those things a couple of times. It's nice to see people work in trying...
idk why everyone is wanting perfection, lmao. it's literally art. if you want photorealism, go outside
a pickle, falling down the stairs
wispy nest
idk why everyone is wanting perfection, lmao. it's literally art. if you want ph...
I didn't mean it wasn't perfect, I said that there will be problems. But my knowledge is 4 months out of date so I might be wrong about those as well :P
hasty nova
idk why everyone is wanting perfection, lmao. it's literally art. if you want ph...
that was actually the original intent of DD, was to make, well, digital art. But then I realized it could for some reason also do better photorealism, even in v1 where none was included in training, so I went on and embraced that and now it can do both.
oak osprey
🙂
hasty nova
anyway I wanna make some neat landscapes or environments
wispy nest
I just like it when people are happy and create neat stuff :D
oak osprey
an impossible landscape, inverted gravity, tetrahedron mountains, rivers upside-down, lakes inverted
the prompt has tornado in it and while i don't see one i know it's lurking and the house people do not
hasty nova
