They have been banned so many times, jet they keep coming back
#🏞｜general-with-images
smoky oak
dense tapir
Yep, diversity
wispy nest
I actually have a issue with ai images when used for almost any media as I just can't see past the "this is ai generated." Even when I make a lot of ai images myself I still have a bias against it :(
hasty nova
I had a plan to put the mj dataset through an aesthetic scorer
dense tapir
They have been banned so many times, jet they keep coming back
Well, I stepped back into the bushes with that last one (insert the Homer Simpson meme here) cause I knew what was coming. Sure enough, POP.
smoky oak
I posted the same thing like a month ago when people asked for my sources, to whiche verybody said "they are just bad" or, "you are picking bad results on purporse"
Like no, most just look bad from MJ, and the outliers are the ones that get blasted for being amazing
oak osprey
I posted the same thing like a month ago when people asked for my sources, to wh...
well most results from SD out of the box are fucking awful, too
hasty nova
well most results from SD out of the box are ***fucking awful***, too
I guess you just gotta know how to use it
oak osprey
you could say the same for MJ tho
hasty nova
Yep
smoky oak
oh yeah for sure, but when you get good at SD, there is a hell of a lot more capability you can squeeze out compared to MJ
hasty nova
Which is probably why this dataset isn't good, also the upscaling doesn't look good for mj
oak osprey
out of the box, MJ is like opening up the fire hydrant on oversaturation and extreme "beauty", for one group's opinion of "beauty"
smoky oak
and the whole selling point of MJ taht everybody always harps on about is "it takes little effort, therefore its better"
dense tapir
As a training data set to add some spice to it I can see it. As the entire dataset I can see that too for those who want that style, but for me I have never wanted my generator to have its own style. I know that one it is MJ is a slap imo.
oak osprey
OpenJourney ❤️
smoky oak
Like the dumbass on the firefly comparison who said that my SD gens were "not as good" cause I had to add negatives, cause none of the other programs could use them... Thats not worse, thats more control lmao
hasty nova
As a training data set to add some spice to it I can see it. As the entire data...
Well the one thing I hope it helps with is that there were a few good watercolor style images which I want my model to have
and some cats
dense tapir
Well the one thing I hope it helps with is that there were a few good watercolor...
I hope SD's trainer can pick up on it.
oak osprey
Like the dumbass on the firefly comparison who said that my SD gens were "not as...
yeah that's why i'm using MJ as my source material for SD 2.1 training instead of just using MJ
i'm transferring MJ's learning
wispy nest
I see no issue with something being easy to use and get at least decent results. I'd rather see them have fun creating something I've seen a thousand times already than them being sad that what they made isn't what they wanted
smoky oak
it does some cool composition and colors, and that's about it
smoky oak
and besides, I have found some ways to mimic the really cool and moodly look from MJ with just some basic prompting
wispy nest
I like MJ version 3's style, that weird, almost horror type of sculpture images
hasty nova
Same, I mean I kinda trained a whole keyword around into my model. It was just "moody", also "dark" and "cinematic" really make it look more like mj
oak osprey
I like MJ version 3's style, that weird, almost horror type of sculpture images
use stable-diffusion-2-1 at 1920x1080 native res
smoky oak
those were when I was mimicking v4, with its semi real but painterly stle, and its pretty damn spot on
I mean, this is a real MJ image from that gen
oak osprey
i don't expect the best model ever from MJ as training data but i gotta say it makes me chuckle to know they're paying for my training catalogue
wispy nest
use stable-diffusion-2-1 at 1920x1080 native res
it's not really the same to me, it's just something special about version 3 :P
smoky oak
the bubbles and paint strokes look really close
oak osprey
"trained on the midjourney supercluster"
what the fuck's a supercluster
i think i found one between my 'cheeks once
smoky oak
so those above were mimicking v4, and then these were me trying to refine it into my own more consistent and aesthetic style
oak osprey
well fwiw i never found v4 appealing but i DO love v5.1
it's still incredible how much you can polish a turd though, looking at a majority of the uncherry-picked data from midjourney
a lot of these images aren't 'great' but i still want their aesthetic because, i'm a monster
smoky oak
These are my MJ inspired refined gens
much more moody, more painterly textures, very abstract
smoky oak
oh I like that :>
smoky oak
its a cool image, but still nothing that special IMO
it doesn't look even slightly semi realistic, but the concept itself gets some brownie points
wispy nest
this is the style I like, but it feels soooo weird to have an actual cd cover use ai art
oak osprey
its a cool image, but still nothing that special IMO
it doesn't look even sligh...
love how you went from "i like that" to "oh it's nothing special" once you knew it's 5.1 native img
it's the violence inherent to the system
smoky oak
love how you went from "i like that" to "oh it's nothing special" once you knew ...
Thats not what happened, I like the image just the same, but its nothing special in terms of achievement, just concept
would have said the same if it was SD, it doesn't look real at all, and has some rough spots
oak osprey
the prompt for this is "lion graduation ring on signet"
smoky oak
but my liking it comes from the concept
oak osprey
same. i'll never look at like 99.99999% of the images in this training data
love how they haven't figured out text yet but DeepFloyd has
it's weird though, doesn't their model almost feel burnt @smoky oak ?
smoky oak
yeah, it does
oak osprey
like you can see style bleed through everything they do
dense tapir
Style bleed?
oak osprey
yeah when you have a single style overly influencing the outputs
it's almost impossible to get away from it in that case
smoky oak
actually getting me slightly annoyed cause I have said this so many times to other people only to be called a shill lmao
dense tapir
Isn't that what a style should do?
oak osprey
not out of control
so MJ 5.1 isn't really quite burnt or it's more like a compendium of overtrained LoRAs
dense tapir
yeah when you have a single style overly influencing the outputs
The "MJ Style" is a real thing. I have even seen prompts online with "in the style of Midjourney".
oak osprey
but like look at the noise in this img
it has some kind of garbage throughout it
that's a sign of a "burnt" model. but this is quite subjective tbh, some people go for that look
when you burn it you overtrain it to an extent that the model becomes really just good at doing a single thing
dense tapir
A halo white ring around it and a shit ton of aliasing (the jaggies).
oak osprey
tbf the prompt is 2d sprite and i bet a lot of the training data there was transparent images
that causes the halo effect
dense tapir
bingo
smoky oak
but like look at the noise in this img
yeah, unless they specified pencile/textured style, thats terrible
oak osprey
the prompt is in the filename
wispy nest
the noise look like the typical, and quick, use of a upscaler. I've seen it a lot :(
oak osprey
in fact it missed that they say "not human" in the prompt
oak osprey
at least it's realistic
modern orbit
No se donde generar 🫠
oak osprey
a lot of these video games have such clean characters. have you ever gone on a quest? shit don't stay clean
dense tapir
prompt said "not human"
wispy nest
but to be fair, 2d sprite probably confused the ai into thinking that it should look like a 16 bit sprite that most people probably think of when it comes to those things
oak osprey
poquito no es muey soya directamonte la montana! verde gatorade, "Electric Blue"
dense tapir
for me sprite is 8bit
oak osprey
well for me they're 256 bit color
wispy nest
I wanted to write 8bit, but I didn't want to sound too old :P
modern orbit
Como género las imágenes?
dense tapir
Shit, embrace it brother
oak osprey
you all saw "sprite" in the commodore 64 manual and stuck with that understanding
wispy nest
I'd probably try using vector art or something similar
oak osprey
@modern orbit mi amigo, no entiende
dense tapir
C64 days, yeah. Amiga days even better.
modern orbit
<@762039592966029362> mi amigo, no entiende
No 😪
smoky oak
I'd probably try using ``vector art`` or something similar
yeah, SD can do near perfect vector art
wispy nest
I never had the amiga, it was too expensive :(
dense tapir
yeah, it was so I had an Amiga 500. 1000 was hellishly expensive and a 2k? forget it
oak osprey
i accidentally killed the amiga 2500 by plugging in a PC sound card and it sparked and smoked and before my grandpa got home i buried it in the backyard
dense tapir
I never had an Intel based machine until 1996 and to this day Motorola CPUs are just superior to Intel, but as it was with Betamax so it was with Motorola. I don't care much for their mobile phones though.
oak osprey
betamax was never superior
dense tapir
10 more days until computex begins so I hope a 7950XTX is unveiled.
oak osprey
maybe you mean betacam which was a professional format but it was totally different from betamax.
dense tapir
maybe you mean betacam which was a professional format but it was totally differ...
betamax was far superior to VHS only it couldn't do 6h only at most 5h and lost the format wars. eond of line.
oak osprey
in fact betamax had smaller tapes that stored less footage of tape. they had to thin it out to get even close to Standard Play times on VHS. and then JVC did the same thing, thinned the tape out and got Extended Long Play in there. 8 hours of terrible quality.
dense tapir
5h vs 6h and the inferior format won because of convenience. Oh, well it is what it is but we keep getting the same happening to other things since because the lemmings prefer conveniences over quality every single time.
oak osprey
i mean, Technology Connections literally just did a video comparing Betamax, Betacam, and VHS, and VHS looked better
it just seems to be a myth that won't die out
if you love Betamax so much though, there's likely someone else who will sell you one for a lot of money 
dense tapir
I wish we could photo bash in latent space.
wispy nest
people seems to bash photos everywhere else so maybe it's not needed ;P
dense tapir
It would be really great to merge in the latent space. Think not what others already do think what could be done.
dense tapir
i mean, Technology Connections **literally** just did a video comparing Betamax,...
He gets a lot of stuff wrong so I unsubbed to his two channels a couple of years ago. This is what I was talking about as I lived it - "When Betamax was introduced in Japan and the United States in 1975, its Beta I speed of 1.57 inches per second (ips) offered a higher horizontal resolution (approximately 250 lines vs 240 lines horizontal NTSC), lower video noise, and less luma/chroma crosstalk than VHS, and was later marketed as providing pictures superior to VHS's playback. " Now this part is when it began to fall apart - "However, the introduction of Beta II speed, 0.79 ips (two-hour mode), to compete with VHS's two-hour Standard Play mode (1.31 ips) reduced Betamax's horizontal resolution to 250 lines."
dense tapir
that crosstalk was the big one.
then it all fell to poop when VHS went 2,4,6, to eventually 8. I never did 8 as I knew that had to be far too thin.
steel yarrow
smoky oak
Alright, I gotta work on a comission
I'll be back in not too long
still waiting for Etsy's slow ass to fix my account so I can sell on it
dense tapir
https://images-ext-1.discordapp.net/external/qUBFsvj7jrfavKl6Rgclg8jiIcBgU9sZnCn...
whatever that is at least it seems happy
smoky oak
Something is going on with my graphics card
My LoRA training is much slower for some reason today
Only 1.07it/s
floral urchin
...
smoky oak
hey, I am training some LoRA's right now, I can share some graphs in DM's if you show me how to
muted pewter
Boa noite
oak osprey
boa constrictor 
dense tapir
dense tapir
Damn
dense tapir
oak osprey
ahhhh this might be the worst image in that dataset
look up close
just wild that people pay for that
hasty nova
Oh hey, I am currently captioning the images I have gathered today as well as some others from yesterday, anything that was from that MJ dataset is being tagged with the "AI" keyword. There are already some ai images in the dataset that won't have this keyword, but it should stil be able to be used in the negative if you want to get rid of that MJ feel. Hopefully this is helpful for some! Should also improve quality potentially.
actually it's just "ai", nothing captitalized
dense tapir
Well, I see 2.1 is missing a 3rd model that is just too damn good to be missing.
oak osprey
@dense tapir honestly without any fine-tuning, MJ is superior to SD 😭
dense tapir
for control net
dense tapir
that means we are missing style, tile, and ip2p
oak osprey
yeah i don't know why the ControlNet team insists on using 1.5
hasty nova
<@315547788211388417> honestly without any fine-tuning, MJ is superior to SD 😭
but the fact we have the ability to finetune, or run on local hardware for free is quite nice
oak osprey
@hasty nova yeah and i can finetune WITH their data, but i'm just saying, there's merit to both sides' arguments. and the lazy ones win out with MJ's out of the box tilt toward "beauty" (oversaturation, sharp lines, etc)
hasty nova
<@800501037793280030> yeah and i can finetune WITH their data, but i'm just sayi...
yep
dense tapir
yeah i don't know why the ControlNet team insists on using 1.5
There was an open letter to SAI about that and they were NOT kind to SAI at all. Basically, I agreed with them so I get it.
oak osprey
i wish Stability would give a bit more guidance for fine-tuning
they just yeet shit to the 4 winds
"here's deep floyd, dur hur" and then 💨
dense tapir
Yeah, this is why 2.1 never caught on with people who train, and shit.
oak osprey
naw, the tools have been there for 2.1 for a while, it's just community lore and myths and legends at this point
people told me you literally cannot train 2.x on NSFW content, so, i spent a week figuring out how to do that, and did it. that checkpoint is now published
dense tapir
fuck those mother fucking tools as they are the same goddamn tools we used with 1.5 and they suck ass juice for 2.1 training for styles at least.
oak osprey
you have to mess around with the settings
gentle panther
#mountain
dense tapir
I rather slam my head against a damn wall than attempt to train 2.1 ever again for styles.
hasty nova
people told me you literally cannot train 2.x on NSFW content, so, i spent a wee...
Oh you actually published it
thought you weren't planning on doing that
oak osprey
Oh you actually published it
Sytan needed it for testing
hasty nova
Ah
dense tapir
I fucked with those damn settings for 3 goddamn long months at well over 1k hours to the point of utter exhaustion. The control net devs get it so not sure why you don't or refuse to.
oak osprey
i wasn't planning on publishing it, but people in here literally told me to fuck off because it's just another person who trains stuff without sharing it
so it's their funeral and i'm not responsible for what they see from that thing
i didn't do any realism training on it, so, it knows NSFW but not any more coherently than the base model would have. good luck
hasty nova
well, I am still training no completely nsfw stuff into DD, sorry
dense tapir
Supposedly NSFW is in the model, as are actors, and artists, but the links to it were severed.
oak osprey
they used an NSFW detector model that was trained on 65 million NSFW images to actually remove the images from the dataset
dense tapir
funny as I still get NSFW so it wasn't all that good then
oak osprey
the hilarious part is that dataset is public on github and you can see all the porn they removed
oak osprey
funny as I still get NSFW so it wasn't all that good then
you have to try pretty hard
hasty nova
was it that one where you couldn't get rid of it no matter how many negatives you used?
oak osprey
well, I am still training no completely nsfw stuff into DD, sorry
i don't care, it's literally a fresh trained model from base 2.1 just as a research project
hasty nova
i don't care, it's literally a fresh trained model from base 2.1 just as a resea...
Yeah I know
dense tapir
Now something emad said a very long time ago, when I used to listen to him, is that the artists can be added back in via the community (in round about speak). This is what the base model is for. Then people found out training the same dataset on 1.5 works then on 2.1 it shite.
hasty nova
Just still telling people that since people seem to think it fixes anatomy so much but I think my model is already pretty decent at it and should be even better in v3
oak osprey
Then people found out training the same dataset on 1.5 works then on 2.1 it shite
you have to use a polynomial scheduler for the learning rate annealing on 2.1.
dense tapir
Been there, done that
hasty nova
> Then people found out training the same dataset on 1.5 works then on 2.1 it sh...
yknow maybe that's how dd came out so good, I picked a polynomial one at random tbh
dense tapir
Poly was my go to for DB starting with 1.4. I liked it
oak osprey
sorry man, i mean this the nicest way possible, but just because you couldn't get good results doesn't mean it isn't possible. i think the fact that you give up so easily and then deride it without pursuing better results is just lazy and weak.
i love every 2.1 fine-tuned model i've ever used
hasty nova
Oh, here is one of those watercolor images I was talking about! I love this style
oak osprey
my friend called that posterisation once but i think he got the word wrong
dense tapir
What? You don't know what I went through and the fucking torture I went through and the damn hours upon hours trying every damn thing. You are simply, not worth my time to discuss this with since you are 100% oblivious to any damn thing.
oak osprey
to me, posterisation is like "The Endless Summer" movie art
hasty nova
anyone know what to call this style? I have a few images like it
oak osprey
What? You don't know what I went through and the fucking torture I went through...
haha i've been doing training tests for the last 2 weeks and i'm still going, and i'm happy to discuss my findings or possible routes from here with anyone, but you insist that your experience is the ultimate and the one everyone will eventually converge on
dense tapir
I insist because there are entire discords with thousands who also insist the same thing after their wounds from trying.
oak osprey
that's Art Nouveau
hasty nova
that's Art Nouveau
thanks
dense tapir
Now remember this I couldn't spend more than two hours per day (actually 1h45m) to train on Colab on a T4.
If it takes longer than that to converge forget it
oak osprey
Now remember this I couldn't spend more than two hours per day (actually 1h45m) ...
well that does make things a LOT harder.
you have to save your checkpoints and restart training from there, and, you have to use a Colab notebook that can avoid resampling the same images every single time it restarts.
dense tapir
Christmas day I spent 14h training and still nothing, but that was a rare treat.
oak osprey
if you don't do the very last thing, it destroys 2.1 so reliably you could set a clock to it
dense tapir
you have to save your checkpoints and restart training from there, and, you have...
Which the Kohya one the dev even said it starts over even on saved checkpoints. He said Kohya said it so he eventually removed them.
oak osprey
Christmas day I spent 14h training and still nothing, but that was a rare treat.
here's someone's amazing results with 2.1 but it's taken them months to figure out how to get results this good, and they're going to be publishing their workflow when they publish the model (but not their dataset)
dense tapir
See, you are missing something though
hasty nova
I actually really like this image
oak osprey
Which the Kohya one the dev even said it starts over even on saved checkpoints. ...
well the Accelerate method of dreamboothing now saves everything - the Optimizer state, the Random states. everything.
dense tapir
took them months while the same dataset on 1.5 would have been done
oak osprey
who cares, it's still 1.5
dense tapir
Oh, this isn't DB
oak osprey
i'm confused what they were doing then
dense tapir
This is Kohya Lora/Locon/LyCORIS. For DB I never even managed to get a style in 1.5 but models were dead easy.
wispy nest
heh, yeah, I remember that. Lucky for me, I had my doubts from the start so I never really tried that hard to train stuff :P
static tusk
https://www.reddit.com/r/StableDiffusion/comments/13j78fo/some_examples_of_the_g...
a little blurry for me. unless its just bad upscaling from them
oak osprey
This is Kohya Lora/Locon/LyCORIS. For DB I never even managed to get a style in...
well, i have a different mentality to it entirely, honestly. i'm a developer, i push into the new ground. i like that the text encoder for SD2.x is actually open, and not made by OpenAI.
dense tapir
who cares, it's still 1.5
who cares? You miss the point yet again. SAI gives these new models without the proper tools. As the control net devs said the issue is the damn TE as it is a turd.
oak osprey
lmao.
the controlnet dev is toxic as hell and not a great person to follow if you wish to develop AI stuff
OpenCLIP is far superior to the original CLIP
hallow cape
training with dreambooth
oak osprey
the numbers back it up, dude 😛
dense tapir
Oh, that doesn't make him wrong since I ran into that same shit. Great to prompt with and fights you to train it.
dense tapir
Now I suspect FT/DB might be different as they are much closer to the source, or they might even be worse OR, and this may very well be the case, Kohya's implementation of Lora/etc... is highly flawed.
hallow cape
lol love that folder name
love the _stuff or more humorous version _shit suffix
hallow cape
oh nice, a 3060 too
for some reason nvidia made THAT have 12 GB and not the 3070 or smth
hasty nova
my ai foler is just "ai"
hallow cape
¯_(ツ)_/¯
oak osprey
Now I suspect FT/DB might be different as they are much closer to the source, or...
Kohya's stuff as i hear it in general might be very broken implementations of things
hasty nova
for some reason nvidia made THAT have 12 GB and not the 3070 or smth
I got a 3070 😭
But I still train a lot of stuff
hallow cape
nice 👍
dense tapir
Kohya's stuff as i hear it in general might be very broken implementations of th...
Yes, I was coming to that conclusion after all that time. The sad thing is there was nothing else to train with but his for Lora.
oak osprey
Yes, I was coming to that conclusion after all that time. The sad thing is ther...
there's a repo, called lora, that all of this originated from. but if you don't want to use code instead of UIs, well, then the broken reimplementations from GUI devs is what you get to work with eg. A1111
wispy nest
I worry mostly about how much work will be "wasted" when a super good new formula gets released. But we'll only see better results I think. I have my own thoughts around it, but in the end, I'm just happy people get to make art :D
oak osprey
the work is mostly in curating the datasets
dense tapir
One of the problems with Kohya, and all the devs under him implementing his stuff, is they are all Anime trainers using 1.5 so I come in with 2.1 and realism and they sent me off on wild good chases, and were so wrong for what I needed to do.
oak osprey
everything else is just properly cooling a furnace as it tries to melt itself down.
oak osprey
One of the problems with Kohya, and all the devs under him implementing his stuf...
god damn kids and their animes
dense tapir
there's a repo, called lora, that all of this originated from. but if you don't ...
Yeah, we can't use guis on Colab now.
hasty nova
bro what is this prompt 😭
a_christian_who_turned_into_a_muslim_converting_back_to_being_a_christian_after_hearing_the_gospel_from_an_ai_prophet_of_the_lor
oak osprey
that's the tip of the iceberg is what that is
hasty nova
heres the mj image from it
oak osprey
did you see the one about the 6 year old tugging on the sleeve of his 30 year old self who is doing drugs at the table playing poker cards and ignoring him?
hasty nova
dont think so
dense tapir
god damn kids and their animes
I am not mad at you, but you have to really understand the torture I went through with those devs. Finally a kind soul said they had watched the BS they did to me and all the torture I was going through for a month that they finally DM'd and gave me the low down. They simply live in their Anime 1.5 bubbles and everything else has to be the same.
He even said they had to be trolling me but then realized the bubble
oak osprey
well, that's fair. but you have to realise, that was a very small part of the ecosystem
what their training does to these models is just tragic
dense tapir
Only eco I had since to train a lora, the locon, then loha was from Kohya and that bubble, ugh
oak osprey
i am not surprised their methods do not work with OpenCLIP
hallow cape
well the Accelerate method of dreamboothing now saves **everything** - the Optim...
the WHAT
oak osprey
accelerate launch train_dreambooth.py
... args...
hallow cape
OpenCLIP is far superior to the original CLIP
also, how? (and how do I get for auto1111
hallow cape
accelerate launch train_dreambooth.py \
... args...
oh CLI
oak osprey
how do you get? use Artius v2.1, Realism Engine, Digital Diffusion
hallow cape
was looking for that but too lazy too go through it, maybe I'll see it for my later trainings
oak osprey
yeah man. automatic1111 sucks ass
dense tapir
I looked every where to get out from Kohya but never found a trainer that worked for lora/locon/loha AND 2.1.
oak osprey
oak osprey
I looked every where to get out from Kohya but never found a trainer that worked...
well someday i'll get into LoRAs and i'll help you then
hasty nova
I looked every where to get out from Kohya but never found a trainer that worked...
I tried kohya once, didn't work so I went back to dreambooth in vladmandic ui
oak osprey
maybe someday i'll have a girlfriend named Laura, too 😭
Laura Croft
that's a tombraider reference, Junglerally
hasty nova
Beat the story of the first remake
oak osprey
oh
hasty nova
Yeah not the og's :/
dense tapir
I was told it was my data and this and that etc... One day I finally said it is time to test out one of my theories after a lot of testing data. I grabbed my data, a 1.5 model and trained using everything EXACTLY the same. SOB, it worked. I never trained again.
oak osprey
you'll never know the terror i felt in 2002 or so playing Tombraider on Windows 2000 and those DAMN TIGERS attacking out of NOWHERE
oak osprey
you have a TI calc?
hasty nova
Yep, TI 84 Plus CE
oak osprey
there was an RPG game i played on mine, written in assembly and it had actual graphics
got it from ticalc.org
i really can't remember the name of it 😦
stable diffusion 2: the revenge of the bicycle
hasty nova
I've finally finished captioning all the MJ images I got today, now I gotta caption some I got from artstation the other day
oak osprey
feels like that's the same damn bicycle
wispy nest
I was told it was my data and this and that etc... One day I finally said it is...
that's similar to whenever I hear people say that, "you need to use the correct prompts." I mean, yes, just like every ai model :P
And then I never get them to say what a "good prompt" even is so I figured out around 1 week after 2.0 released how much stuff had changed hehe
hasty nova
I always love seeing how sd messes up the coherency behind the wheels
oak osprey
oh noooo it's Marty McFlying' the fuck out of here at 400 steps
500 steps 😭
and now, it was never bornt
oak osprey
I always love seeing how sd messes up the coherency behind the wheels
that image was pretty good but the next one destroyed the wheel itself
hasty nova
yeah it's weird
oak osprey
and then at 600 steps, it comes back and reuses the same image it had in...
200 steps
there's a few differences
dense tapir
I wonder something as I know this guy so might ask him, but for lora stuff we couldn't do this.
towards the end I was saying that maybe training the TE separately would help, so I was going in the right direction, or thinking in the right direction, but sadly it never would have picked up exactly where it left off unlike how it has always been with Dreambooth.
oak osprey
you can't train them separately 😦 i tried
oak osprey
i tried doing 3400 steps with just unet training and made a checkpoint and then started from there with the text encoder and it told me it couldn't optimize that
dense tapir
I did try it on kohya_ss and it gave an error yet said it could be done
oak osprey
mine wasn't that optimistic lmfao
but yeah i've just been nuking the learning rate for the unet and TE to get SD2.1 to do what i want
dense tapir
well, this is the issue as the TE is the problem and must be used differently but the tools for lora stuff is still stuck on anime 1.5 model
oak osprey
i have read that others are indeed freezing most layers of the TE to do better
dense tapir
we really need a lora trainer that can do that
oak osprey
well the one Sytan uses, which i think is Khoya, has the separate learning rates. it has 3 of them for the LoRA.
dense tapir
thing is for style training on 2.1 nothing worked and I had a friend with his 3090 would was trying shit for many nights with me. no matter how low, or what ratio of unet to te we could not do it
oak osprey
to the point that typing in "a happy smile" gives you the textures of a couch cushion
eg. destroying the model to the point that it just doesn't do the thing anymore
dense tapir
oh, a few. For me I could never really burn as that takes way longer to do than I had.
oak osprey
i can burn one in half an hour, baby
🤣
i can show you my ways
its called "following shitty online tutorials"
@dense tapir did you do regularization data? and also, without?
i assume if you did >300 models you must have tried at least a few strategies for regularization data
dense tapir
yes. with and without
oak osprey
eg. real photographs; direct generations from the checkpoint before training; a standard dataset from someone
dense tapir
real, generated, the ones from Nvidia, etc...
oak osprey
the prompt that you use has a major effect
i trained a body part into the model and when i didn't use class images of "person" it just completely lost the ability to make people, while preserving robots, bicycles, and landscapes
dense tapir
our prompt in kohya is only for generating the image every so many steps it is not used in the training
if you don't want images a prompt is not needed
oak osprey
i mean the class prompt
dense tapir
oh, yeah
oak osprey
and are you guys doing captions, or keyword training
eg. the old "photo of an sks dog" trick
dense tapir
I tried style, artstyle, etc... it wasn't the class fault
I am not touching training again until I can train locally on my computer which is a long ways off.
dense tapir
That is explainingthe general feelings I was getting with the lora stuff.
the tools I had to use for lora/locon/loha was the issue I think since my data, tags, class, settings, etc... worked in 1.5 without a single issue
oak osprey
yeah i just blame people like Controlnet and others who stick with 1.5 and don't put the time into this stuff
they took artists style tags out and people shit themselves all over that
dense tapir
If they tried to use their own tools in 2.1 they would see, but as one told me 1.5 is perfect for our Anime and we don't care about that shit 2.1.
static tusk
once XL is out I think 1.5 and 2.1 will be used much less
dense tapir
I don't
static tusk
XL just needs finetuning with good images
oak osprey
if SDXL can't be trained locally on 2gb vram cards people will cry and boycott it
dense tapir
Hell, on my current card I can't even gen with it.
hallow cape
finished training. now to test it.
dense tapir
2gb, LOL
Considering Nvidia seems stuck loving 6-8GB on Ada Lovelace and SDXL needs 10-12, welll
oak osprey
MJ needs better moderation on their discord, one of the bots generated a prompt: joe biden in a secret jacuzzi surrounded by <things that can't go into this discord>
oak osprey
i don't understand this derangement, it seems to be widespread
"Hey kiddo, uploading a pizza to Discord again?"
"Dad!"
dense tapir
I 100% straight up do not understand CP. I think I know what happens, I get terribly emotional so I try not to think about it.
I 100% do not want to ever know exactly what happens. Heard the horror stories though.
oak osprey
mate, mostly it seems to be some kind of mindless 'flex' they are doing to show they do not care about your rules or standards, liek Nep earlier
dense tapir
Oh, yeah Nep
oak osprey
they act all innocent and blame you for 'being sensitive' and i do not know the point or goal behind that
dense tapir
Nep is Nep doing Nep things for a while now.
wispy nest
heh, this is how my stuff "identify" a bike. I knew of most things but not how weirdly old the bikes end up as :P
wispy nest
yeah, that's how I identify the ai and how it paints
it's one of my embeddings in 2.1
dense tapir
I tried controlnet image to image and I never got anything decent.
oak osprey
i literally don't get it
dense tapir
it came out superimposed
wispy nest
don't worry, no one understands what I'm doing most of the time, not even I do :P
oak osprey
oh wow, it fixed the bike and dented the rim
going backwards, 1000 -> 600 -> 200 steps
dense tapir
ANyway for image to image I think controlnet works for people but in 2.1 I can't get it to work right.
oak osprey
wow so 400 steps is like a trillion years or something of human time
look at the way those mountains change
wispy nest
there we go :P
oak osprey
Steps: 1%|█▎ | 1798/244960 [1:32:12<294:55:33, 4.37s/it, loss=0.205, lr=7.34e-8]
lol
i have just a few steps to go
243,162 steps left
dense tapir
an emerald dragon beast
How is that a dragon beast?
no neg was used just that prompt
oak osprey
maybe dragons think cats are scary
oak osprey
lmao the test images of knights it makes
oak osprey
terrain much?
dense tapir
so close :/
static tusk
SD never does both wings for some reason
delicate steeple
Finally my "precious" is complete
dense tapir
SD never does both wings for some reason
Look at the legs too. 😦
oak osprey
my test prompt for child with a kite has gone wild with the mj influence
nearly died laughing
turbid sable
Cat banana
dense tapir
humble rune
oak osprey
@dense tapir i'm super happy with this round of training on 2.1
dense tapir
<@315547788211388417> i'm super happy with this round of training on 2.1
That looks pretty damn good
I asked a lot of people since SD released and DB came on the scene and they still say it doesn't style well. Models very well but not really styles. I wonder if that is part of why Lora has issues with styles?
TI embeddings worked so well, and damn easy too though sometimes not the result you were after but still rocked.
wispy nest
hm, I never finished my background/winged embedding from long ago :/
not sure how much use it's to do that now when everything has changed
wispy nest
yeah, I downloaded a lot of loras but so far I only liked the anime ones as every single other one blurred the end result too much, but I've not even tested them for half a week yet
wispy ether
dense tapir
I have yet to have one that blurred anything
wispy ether
wispy nest
I believe I might be using blur too loosely as I'd say that 99% of all images I've seen here are blurry :P
oak osprey
That looks pretty damn good
it even added aliens 🤯 original baseline 2.1 puts astronaut when you say alien
dense tapir
it even added aliens 🤯 original baseline 2.1 puts astronaut when you say alien
Better because Alien, or Aliens on the base always ended up giving me Africans in small villages, and towns in Africa.
dense tapir
wispy ether
wth
oak osprey
Better because Alien, or Aliens on the base always ended up giving me Africans i...
noooooooooooooooooooooo
that needs to be fixed
@dense tapir my fine-tune vs baseline 2.1 same prompt/seed
dense tapir
Ahhh, that is cool
Yeah, that alien shit baffles me
No idea why Africans are Aliens in the dataset
sterile temple
District 9?
dense tapir
oak osprey
ima uploadin this model now, @dense tapir
tis an early checkpoint since it's at 4k steps out of 224k
but i'm interested to hear your thoughts and see what you can get out of it
oak osprey
dreambooth
i'm not sure what FT is, but i'm using the implementation from the Dreambooth research paper
i adjusted the code to change how it interacts with images to be more reproducible and consistent, and i can actually monitor which images are causing issues and remove them from the dataset
and i agree with you, 4k steps is really raw but compared to the original results i was getting, every subject is improved across the board
so it's not fair to compare it to SDXL, or any amazing fine-tuned 1.5 models, but more to be compared with the baseline 2.1 and OpenJourney v4
pretty obviously it still struggles with faces and some scenes end up looking just straight-up shitty but a different seed/prompt fixed it
otoh, it's still better at faces than before
there's the .safetensors file in there
dense tapir
FT is a finetune which I have never tried myself.
oak osprey
oh. well, this is fine tuning
i am using the dreambooth research paper to do so but it just requires a mild modification
dense tapir
oak osprey
dense tapir
first one is yours second is base non ema
dense tapir
yes
oak osprey
👀
dense tapir
second one is sorta shite
oak osprey
it's very bland but still epic for baseline imo
dense tapir
agree
oak osprey
@smoky oak ^^^
dense tapir
hence the "sorta"
smoky oak
peeng
the amount that he has been able to improve 2.x base is incredible
he and I have been chatting and messing around in DM's for hours lol
I am learning how to train LoRA's better
oak osprey
i'm so tempted to just keep going and pushing on this one to like 100k steps but i'm sure that'll probably break it. BUT I NEED TO SEE
dense tapir
You are saving every few steps, right?
oak osprey
every 10 steps
dense tapir
WHOA, wth?
oak osprey
and i generate all of them in a test matrix
yeah it used 5TB of disk space to do 4000 steps
dense tapir
dude, lol
dense tapir
Well, I would have said every 50-100 but every 10? heh
oak osprey
i did back it off after 1000 steps to every 50
my training script pulls the checkpoint interval on each loop from a file so i can update it while it's running
dense tapir
don't go more now since you will fall not too far when it implodes you can go back almost at the peak
oak osprey
that's true but re-training it isn't bad now that i've fixed the reproducibility issue
it saves the random states, the optimizer, as pickle files i love that name
then, it loads them back up when it loads the checkpoint, plus a list of the file names it worked on
dense tapir
I am just not into wasting the time retraining, or energy, if I can just go back right before it imploded
oak osprey
i mean that's valid but disk space is expensive and i'm automating it all so it's just a command to run here
./training.sh
dense tapir
I am getting random errors from yours though
first time I loaded it I had to reboot
changing setting sd_model_checkpoint to 2.x_models\pseudo-journey.safetensors: SafetensorError
Traceback (most recent call last):
File "F:\stable-diffusion-webui\modules\shared.py", line 483, in set
self.data_labels[key].onchange()
File "F:\stable-diffusion-webui\modules\call_queue.py", line 15, in f
res = func(*args, **kwargs)
File "F:\stable-diffusion-webui\webui.py", line 149, in <lambda>
shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights()))
File "F:\stable-diffusion-webui\modules\sd_models.py", line 496, in reload_model_weights
state_dict = get_checkpoint_state_dict(checkpoint_info, timer)
File "F:\stable-diffusion-webui\modules\sd_models.py", line 262, in get_checkpoint_state_dict
res = read_state_dict(checkpoint_info.filename)
File "F:\stable-diffusion-webui\modules\sd_models.py", line 241, in read_state_dict
pl_sd = safetensors.torch.load_file(checkpoint_file, device=device)
File "F:\stable-diffusion-webui\venv\lib\site-packages\safetensors\torch.py", line 99, in load_file
with safe_open(filename, framework="pt", device=device) as f:
safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLarge
oak osprey
😮
dense tapir
never seen that one before
oak osprey
that's... new
smoky oak
@dense tapir pseudo and I have been screwing around with LoRA's in DM's on my new GPU, and we are finding ways/settings to get much better results
Turns out my 10GB 3080 can do over BS20 lmao
wispy nest
How I can make images
crystal deltaBOT
oak osprey
i unironically wonder if we're in a solar storm cycle today
dense tapir
<@315547788211388417> pseudo and I have been screwing around with LoRA's in DM's...
for 512x512, yeah
smoky oak
I have been wondering about trying slightly higher res
oak osprey
Thanks. It works after changing the filename extension from *.safetensors to *.ckpt.
smoky oak
the anthro model I used is 640x640
oak osprey
ooooh
did i fuck that up
i honestly don't use those files so, i don't know how it works
dense tapir
I cannot load yours now even after a reboot
oak osprey
try changing the extension from .safetensors to .ckpt
static tusk
probably need a yaml
dense tapir
won't work. Let me change to another model then to yours
oak osprey
v2 needs one
dense tapir
yes
static tusk
v2 needs one
yep
oak osprey
how do .. can one of you make one
i'll put it whereever it needs to be
ohhh i have one from realism engine
wispy nest
How I can make images
oak osprey
wow i have no idea what to put in here hahaha
smoky oak
How I can make images
I responded already.
crystal deltaBOT
wispy nest
I responded already.
What
smoky oak
Literally just responded again
wispy nest
When I download stable diffusion files it not show web user bat
smoky oak
<@296720689468145664> lawd i can't wait to try training DeepFloyd
it has a whole new text encoder, so that will be fun to play with
oak osprey
text encoder is putting it lightly
it has a whole damn language model
like literally T5 has a head called FLAN that you can connect and use it as a full-fledged LLM, like GPT2
dense tapir
I cannot load that model of yours
oak osprey
did you try changing the file extension?
dense tapir
yeah
oak osprey
that's pretty weird
dense tapir
grabbed the yaml and changed it to match the model name.yaml and still no good
oak osprey
yeah i don't see what that yaml config has in it that would fix this
dense tapir
got it to work finally
I had to change to .ckpt and load something else three times
I believe the problem was that I cache these to ram so it was still in my ram
you have issues with this model
oak osprey
weeeird
yeah i need to make it a safetensors file for A1111 somehow
it's not clear how that's typically done
dense tapir
Traceback (most recent call last):
File "F:\stable-diffusion-webui\modules\call_queue.py", line 56, in f
res = list(func(*args, **kwargs))
File "F:\stable-diffusion-webui\modules\call_queue.py", line 37, in f
res = func(*args, **kwargs)
File "F:\stable-diffusion-webui\modules\txt2img.py", line 56, in txt2img
processed = process_images(p)
File "F:\stable-diffusion-webui\modules\processing.py", line 503, in process_images
res = process_images_inner(p)
File "F:\stable-diffusion-webui\extensions\sd-webui-controlnet\scripts\batch_hijack.py", line 42, in processing_process_images_hijack
return getattr(processing, '__controlnet_original_process_images_inner')(p, *args, **kwargs)
File "F:\stable-diffusion-webui\modules\processing.py", line 657, in process_images_inner
devices.test_for_nans(x, "vae")
File "F:\stable-diffusion-webui\modules\devices.py", line 152, in test_for_nans
raise NansException(message)
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
oh, my model is in half precision mode
dense tapir
it doesn't like my CN
dense tapir
I use fp16 myself
it is emulated in the card but at least it works while 1650/1660 gets black screens
oak osprey
ohhh
i have to make an fp32 file then
so parts of it can run on CPU
gotchya.
wait, how did it work before?
wispy nest
Traceback (most recent call last):
File "F:\stable-diffusion-webui\modules\cal...
great for formatting :)
dense tapir
beats me
wispy nest
@dense tapir how I can use bot
somber moth
#🏞｜general-with-images Design a captivating book cover for "Cyber Security Roadmap" that incorporates elements of cybersecurity, technology, and adventure. Use a mix of vibrant and cool colors to evoke a sense of professionalism and expertise. The cover should feature a clean, modern font for the title that is easily readable even in thumbnail size. Experiment with creative concepts and imagery such as lock icons, shield symbols, circuit patterns, data streams, interconnected network nodes, or abstract representations of digital landscapes. Create a visually striking cover that entices readers to explore the book's contents.
dense tapir
The issue is you had me change from safetensor to .ckpt and they have different routines to decode them. I switched back to .safetensor and we are back to the original error which is the header is too large and I can't load it.
wispy nest
Ok
dense tapir
I deleted the model and yaml as I went looking in it with a hex editor and yeah, not the same as other finished models I have.
wispy nest
https://beta.dreamstudio.ai/generate
When I download stable diffusion files it not show web user bat
oak osprey
I deleted the model and yaml as I went looking in it with a hex editor and yeah,...
it's weird, i wonder if there's a way to use Diffusers layout models directly in A1111
sterile temple
When I download stable diffusion files it not show web user bat
oak osprey
because thats working great.
wispy nest
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-...
Thank you
oak osprey
says 114 hours of training left to go
woven dawn
wispy nest
A reason why I seldom let the ai create men, is that they often look like super bulky muscle men, but sometimes, like here, one cannot say if this is one or not :P
sterile temple
A reason why I seldom let the ai create men, is that they often look like super ...
try 'body builder' in the negative prompt
wispy nest
you got one for long necks, I often get them when I don't have the same ratio of the image. I want them like the above 99.9% of the time :P
also: Thanks! :D
sterile temple
you got one for long necks, I often get them when I don't have the same ratio of...
you could try stretched, but sometimes you just gotta generate a bunch and pick the best
wispy nest
yeah, that's mostly the reason I seldom bother with most negative prompts as, in the end, it feels like a coin toss. But I'll never let home vanish, one day there will be a perfect fix for all my woes :P
sterile temple
yeah, that's mostly the reason I seldom bother with most negative prompts as, in...
If I find something that's working, I'll save it to my styles so I can use it again
sterile temple
same, I have around 4 negative prompt words I know work 99% of the time.
You could try this to crop from a square image to a different ratio
wispy nest
You could try this to crop from a square image to a different ratio
thanks for the tip. :D
but I've been doing that a lot, but it's also mostly a coin toss. I'm more about finding stuff that has major effects.
abstract valve
#ketchup
crystal deltaBOT
abstract valve
Please share the image of Ketchup pouch
smoky oak
read the message above
oak osprey
Please share the image of Ketchup pouch
abstract valve
1 rupee ketchup pouch
oak osprey
wispy nest
how i can use bot
wispy nest
seems like they forgot to say that you should also drive to them and set it up in person? ;P
wispy ether
sterile temple
how i can use bot
there's no bot
oak osprey
wispy ether
there's no bot
oak osprey
that's not R2D2
wispy ether
that's not R2D2
wispy nest
there's no bot
easy diffusion very slow
oak osprey
oak osprey
easy diffusion very slow
try plugging in a DVD player
wispy nest
try plugging in a DVD player
what
dense tapir
what neg to use to stop zombie titties?
oak osprey
try "fun"
wispy nest
also, do I even want to ask, "what even is that?" :P
dense tapir
Wish I could show this image
oak osprey
my prompt was pretty dumb lmao
when you roll over and sleep there's no more time to live, because then you're dead
sterile temple
what
need a good computer to go faster, it might be downloading the model before running? Ask in the #🤝｜tech-support
sterile temple
'wookie nipple pinchy'
dense tapir
nope
stray sail
.
smoky oak
I have no idea what the hell is wrong with my kohya
@dense tapirmind if I pester you really fast?
dense tapir
What up?
smoky oak
my kohya is not listening to Epochs correctly, and is massively misaligning them
dense tapir
How so?
smoky oak
How so?
12 images, 100 steps each = 1200 steps, i am using BS 12, so it should be 100 steps, for 2 epochs 200
instead, its saying that one epoch is 259, and 2 epochs is shomehow 388
So my first epoch is saved at 66.6666... percent, and the second one is saved at 100%
also, I have 3 epochs selected, yet it is only doing 2. I am not sure what is up
wispy nest
I tried titties and it didn't work
depends on what you're after, then, from my own experience with 1.5, the negative prompt large breasts fixes 90% of my issues over weird symmetry. But can't really say more than that because 1. might be too naughty for here, and 2. don't know the whole issue :P
dense tapir
Issue is bare nipples
dense tapir
12 images, 100 steps each = 1200 steps, i am using BS 12, so it should be 100 st...
That is a typical issue I ran into with Kohya a lot and I hated it. What I started to do is go by steps instead of epochs.
dense tapir
because I could save based at XX steps whereas XX epoch was shit. Might be 200 steps or 400steps or 266 steps.
Kohya always had bad epoch math
smoky oak
but if I do 1 epoch at 100 steps for 12 images, its not gonna do 1200 steps, its gonna do some random number
I am not sure what is up
smoky oak
oh god, BS1
wispy nest
Issue is bare nipples
the most sfw fix for that would be to write corset as I don't, or never found, a version that does not cover the breasts
smoky oak
I will check
smoky oak
does that number equal right?
no
dense tapir
then that rules it out. Now what did it make it with a BS1?
smoky oak
its saying 2 epochs (not 3), and its saying 1 epoch is 1550, and 2 is 4650
dense tapir
the most sfw fix for that would be to write ``corset`` as I don't, or never foun...
corset ruined the image but did work
dense tapir
its saying 2 epochs (not 3), and its saying 1 epoch is 1550, and 2 is 4650
give me your values
She was naked
smoky oak
alright, headed to DM's
wispy nest
She was naked
using perky before the naughty word, weirdly, makes the character cover their chest because I believe the ai "think" the character is cold :P
cyan snow
wispy ether
wtf
wispy ether
scyther
dense tapir
using ``perky`` before the naughty word, weirdly, makes the character cover thei...
didn't work.
sterile temple
@wispy nest getting a better success rate with 'one central figure'
wispy nest
didn't work.
aw, yeah, it was a long shot really. But it's also the second best "single prompt word" I know of for that :(
wispy nest
<@456226577798135808> getting a better success rate with 'one central figure'
better results for what? I got a bad memory, but at least I don't have a bad memory :D
sterile temple
better results for what? I got a bad memory, but at least I don't have a bad mem...
the tall aspect ratio
cyan snow
better results for what? I got a bad memory, but at least I don't have a bad mem...
i have dementia
wispy nest
the tall aspect ratio
ahhhh, I never actually prompt for number of characters at all. Maybe I should try it. Never really went through my head. My empty, empty… what were we talking about?
dense tapir
aw, yeah, it was a long shot really. But it's also the second best "single promp...
2.1 might not react to it
wispy nest
ahhh, yeah, 2.1 probably don't know what to cover, probably a greater chance that the ai "think" it's a percolator that's misspelled
neon crystal
Any Joy Division fans around?
wispy ether
check out the prompt
forest sonnet
check out the prompt
"Sir, have you tried restarting your computer"
Average IT support
wispy ether
lol
wispy ether
kind quartz
balmy nexus
Fish Pixar
orchid thicket
can anyone help . Need motion blured images of potraits. Like the face being in motion and blured
digital hornet
what was the gpu channel called, i cant find it
sterile kiln
Clint Eastwood 🙂
cobalt tartan
oak osprey
what was the gpu channel called, i cant find it
#gpu-go-brrr i guess it's gone now
digital hornet
there's so many near pointless overlapping channels but the one I actually need is removed 😦
keen palm
plugged vicuna into comfy and just letting it generate stuff in auto-pilot mode, my only input being "come up with an interesting prompt" or similar, think it's doing fairly well 
oak osprey
starting point, last night's checkpoint, current checkpoint
man still need some help, i guess
poor geckos lmao
wispy ether
wispy ether
oak osprey
ROFL
wispy nest
what a hoss! i wish i could squat like that
is that image upside down? it's messing with my head! 🙃 😵💫
wispy ether
wispy nest
this is the reason why I'm trying to find if there is some anime style I like out there because it seems anime has the most crisp and sharp coloring, without blurring, of every style I've so far tried :/
oak osprey
depends on the training data. I know Sytan was working with Anime last night and i helped them improve their LoRAs monumentally. they weren't awful to start with. but we were both surprised by the improvement in coherence, which is something difficult to really understand until you see it side by side
that model looks great. there's no noise, no white halo around the subject
they likely did a decent job generating a dataset and regularization data. are you using LoRA, or a full model?
wispy nest
it's 5 loras at 50%, one old NAI hypernetwork, that everything3 vae and some custom 1.5 model.
I'm testing a couple loras at a time, but so far I only liked the anime loras because of said "crispness." as every other one, even when upscaled (using a anime focused one as well :P ) can't help them :(
But I've only tried less than 20 loras so far and I downloaded like 1000 of them so :P
other than the face having too much noise that I have never learned why, I kinda like the style. Both because of the "popping" of the colors as well as the, in my eyes anyway, un-blurriness
But I'm not that sold on pure anime stuff, anime mixed with a little realism and paintings are totally fine but I don't know, maybe the style needs to grow on me some before I can really say :P
(also the lack of pink and purple that a lot of ai images seem to over expose and use as well isn't as normal here :P )
oak osprey
other than the face having too much noise that I have never learned why, I kinda...
face noise we learnt last night why it happens but it's not my discovery to share and @smoky oak would have to explain
we also figured out how the 'popping' colours get brought out
wispy nest
no worries, I have some ideas, but I haven't and maybe can't test them so :P
wispy ether
frozen nebula
wispy nest
<@456226577798135808> add `knolling` to your prompt you showed lmao
I'll try it when my weird embeddings have finished, no idea when though :P
wispy nest
this is dope as hell
I actually have had to do work lately what a rip off but damn looks like the old pros are still churning out great stuffs in here
but where's Pete and the General 🫡 ?
wispy ether
frozen nebula
this is dope as hell
Her eye is a bit wonky on one side, but I kind of thought it suited the image as its so chaotic and grungey
oak osprey
oak osprey
but where's Pete and the General 🫡 ?
peter has been busy lately but if you ping him he can be summoned
gecko knolling
wispy nest
i am curious what happens
when I didn't use any () nothing really seemed to happen on the 3 images I tested. I used the exact same prompt, but not seed with the added prompt word by the way.
Anyway, at (PROMPT:1.01) it started to change the image quite a lot, so much so that at (PROMPT:1.11) there were only vague (and blurry! >:I ) shapes around "something in the middle"
First image is 1.03
second is 1.05
last is 1.07
oak osprey
that's really sick
wispy nest
https://images-ext-2.discordapp.net/external/74ypqgMATLFXAUTJQvNFoDj1FRU1HJXX6MT...
Nah Im not trying to bother anyone, just didnt see his name as I scrolled
oak osprey
my friend's prompt knolling a midsummer night's dream
adding the posterization keyword acts differently from posterisation which is actually resulting in a closer style adherence, but it definitely interacts with knolling interestingly.
summer valve
Luffy
oak osprey
i've turned off ControlNet now, these are just raw images my model is capable of
wispy nest
adding the `posterization` keyword acts differently from `posterisation` which i...
I believe the ai is trained on british english, so we (meaning me) who likes fries with my american english can have some weird variations :P
oak osprey
i was trained on Indian English, which is close to British English, so, i enjoy the better results 😄 it is one of the few things the British Raj gifted us
the Hindi language is expressly awful at tokenization, 2000 Hindi tokens = 113 English tokens, or so
so anyone wanting to gen in hindi should probably plug google translate in front
or some LLM
wispy nest
interesting, never thought how different languages could affect the token use but yeah i can see that now
oak osprey
English is what it is trained on mostly so it is the most advanced for that. Chinese researchers are starting to majorly improve the understanding of their script, though
wispy nest
using my wildcard textfile with names of botanists, such as PHILIPPA NIKULINSKY create some nice results on that "knolling" prompt word :D
oak osprey
there's no Roman text other than English that tokenizes better, either. eg. Italian, and French, do slightly-to-moderately worse than English
if you're doing Vietnamese though, they switched to Roman text at some point. and that works monumentally better than trying to use the old Vietnamese script.
i haven't tried comparing prompts in English vs Hindi vs Transliterated Hindi. that could be interesting
wispy nest
are Vietnamese the language that doesn't add commas, nor dots on large numbers like for example 12654763543? :P
oak osprey
unfamiliar with that
wispy nest
ahh, my favorite book title :P
(as long as it doesn't say anything weird in another language that is hehe)
frozen nebula
oak osprey
guy thinking "i have to whisper because he has so many ears"
horse: "WHAT? SPEAK UP"
another possibility, guy thinking "haha he can't speak english so i can tell him anything i want and he can never tell anyone else"
wispy nest
better than my head on a horse when I wrote horse horsing around as the prompt :P
oak osprey
what about bojack or horseman or both together
bojack alone
horseman
it approximates the cartoon style to have both together
wispy nest
what about `bojack` or `horseman` or both together
as long as it's not making fun of anyone, other than me that is, everyone makes fun of me, even the person behind the mirror when I shave ;P
oak osprey
bojack is a cartoon
thats easily the best episode too
there's an analysis by Johnny 2 Cellos on youtube that led me to a deeper understanding of this show and everything is symbolises
seems just like a silly horse cartoon at first
frozen nebula
oak osprey
is that with CTU 1.0 strength?
frozen nebula
CTU ?
oak osprey
no CTU vs CTU for same prompt
ControlNet Tile Upscaler
it's possible to get a good result at 1.0 strength but it's hard
wispy nest
a little janky when it comes to the profile, but I think there's a major issue left even so :P
frozen nebula
Ah tile upscaler, yeah I use it alot, nearly always at 1 strength, but I use it for img2img
oak osprey
a little janky when it comes to the profile, but I think there's a major issue l...
Team Ramrod
dense tapir
@smoky oakNo ASUS for me after the last GN vid from a few hours ago. They have turned really bad. GN went out to meet with ASUS and flew all that way to be met with a NDA to sign. They refused and left for ASUS to go running after them to say nevermind. Good show today they even lawyered up on the show.
frozen nebula
oak osprey
dense tapir
wispy nest
that jawline is one of my few agonies when it comes to the ai. It feels like there are so few "skull structures" of the characters that everyone look the same if they aren't some celebrity :(
dense tapir
wispy nest
I didn't need to sleep anyway :P
oak osprey
poll: which is better interpretation of Leopard Gecko?
smoky oak
@oak osprey you might get me to add a 2.x model into my workflow/arsenal
Omg, the camera at this adoption clinic stopped working, so they started drawing the dogs
This looks like a job for stable diffusion! Lmao
It really is impressive how they pictured the essence of said dog lol
oak osprey
@smoky oak i've decided to fork the training results into two separate 2.x models because the current 4k one and the 13k one i'm testing are two different monsters and they're both amazing
the 4k was trained on 3k images and then the 13k was trained for an additional 13,000 steps (so, 17,000 total) with 8,000 more images
smoky oak
More of their drawings lmao
smoky oak
the 4k was trained on 3k images and then the 13k was trained for an additional 1...
Interesting
oak osprey
it begun to lose certain concepts once the LR ramped up past a certain point but the blends it does once it's there are just really nuts
smoky oak
Those look sick dude lmao
oak osprey
4k vs 13k models
jade epoch
`frolicking` makes creatures look like centipedes
Looks like a caterpillar from here. 😁
low compass
wispy nest
I was looking back at some older images I made with the ai. And it's actually insane how far the tech has advanced. Or for me it has because on the date 2022-11-27 I made this and thought it was one of the most awesome and detailed images ever created by an ai. Love it when reality ruins my ignorance :P
dense tapir
oak osprey
@dense tapir zero negatives
oak osprey
wat why
wispy nest
can't react if you're blocked
dense tapir
I tried yesterday and it appears and vanishes instantly as if I am blocked
wispy nest
or discord has it out for you! conspiracy! :P
oak osprey
ok try now
dense tapir
YAY! Lol, just when I ran out of foil
oak osprey
foil to construct the pipe to smoke out of? 
dense tapir
what I time we live in when even foil for my pipe is more expensive than the stuff I smoke
says the junkie
wispy nest
don't use foil, that's what the they wants you to! You should just point your iphone camera toward your face at all time instead. That's what I heard anyways from this weird dude at the back of the shop
dense tapir
Not Lester again as that dude did some bad shit back in the day and never recovered.
oak osprey
dense tapir
Alright, let's see if I can react now
THERE we go
Now that was a discord glitch for sure
I tried it 3 times and figured I was blocked but you answered me so was confused.
wispy nest
I've been able to, or maybe it's just because of the vae, lora, etc that does it for me as a bonus, but anyway: I've been able to at least make it so I seldom get more than 4 fingers and a thumb. now I get less instead ;P
dense tapir
I have used everything even neg embs, tokens, it sucks
wispy nest
fingers, hands are two of my 4-10 negative prompts I use now-a-days, mroe often than not, I don't bother with having any negative prompts other than pink, purple, details :P
hidden timber
What wriite in the prompt?
oak osprey
``fingers, hands`` are two of my 4-10 negative prompts I use now-a-days, mroe of...
but you use like 1000 loras 
wispy nest
but you use like 1000 loras <a:kekception:909936057171197973>
shh!
I downloaded 1000 of them, but I only use 5 at a time ;P
so far anyway, gonna try adding more tomorrow :P
dense tapir
wispy nest
I started a little earlier with something new and I got to about this with 6 loras, technically 7 but the last one is at 0.00001 so I have no idea if it even does anything :P
