#🏞｜general-with-images
wispy nest
and this was my old border embedding, so much stuff half finished :P
dense tapir
dense tapir
smoky oak
https://civitai.com/models/73998/funko-lora
Its officially posted!
Super happy with the results
sterile temple
I'll check it out
smoky oak
let me know how it goes!
idle crater
Eye doctor logo
smoky oak
@sterile templepleaselower the "sexy attire" score on the 13+ image, if you can lmao
I am not sure how that green dress is "sexy attire" lol
sterile temple
first pic 🙂
smoky oak
o-oh lmao
sterile temple
neon crystal
Super happy with the results
Killer stuff! Congrats on the release.
smoky oak
thanks :>
smoky oak
was it trained using 512 images?
yes, but it can do 768 if you are using a well trained model
sterile temple
tweaking the prompts to try to get images that don't look like a hot mess lol, now I'm getting action figures
sterile temple
0.8
smoky oak
can you send your positive prompt?
sterile temple
(funko) [[Donning a sleek, leather biker jacket with studded details and zippers, paired with skinny jeans]] atmospheric lighting depth of field <lora:Funko Test:.8>
smoky oak
ah, specifying too hard with details tends to bring the results back to realism, just a sec
<lora:Funko Test:.7>, Funko, leather jacket, jeans, zippers, woman, female, eyelashes, bokeh
sterile temple
Alright I'll try that. I don't mind the action figures though 😄
smoky oak
yeah, it can do some really cool action figures haha
sterile temple
maybe it's my negative prompt?:: dull colorless blur logo packaging watermark branding (Box, mouth:1.3) (plain background:1.2)
smoky oak
oh yeah, thats a very thin negative. I'd say just mess around and see if you can get some fun results
sterile temple
my model was just being stubborn, I had to increase the weights
smoky oak
OMG, it works with anthro models haha
sterile temple
neon crystal
I'm just gonna leave this here casually and pretend I have nothing to do with it.
hasty nova
you have got to be kidding me
I trained all night
and the model still isn't saving properly
I need to reinstall dreambooth every time I want to train now or something?
wait was it because I didn't save the lora weights or something, idk, might be that
but I don't feel like retraining much
But hey, from the samples I learned it doesn't really need to go past 50k steps before getting overtrained on specific images
wait I just realized the last model in that x/y plot is the wrong model woops
was not supposed to be deliberate
oak osprey
nice
wispy nest
oak osprey
cyan snow
oak osprey
no style keyword -> pointillism -> stippling -> legato
grandpa and grandma sitting at the dinner table, playing with advanced alien technology
cyan snow
oak osprey
SDXL still adds watermarks lol
the fuck is with that hand
most of its outputs seem to have watermarks. good job finding clean training data guys
tell me it's all stolen without telling me 
@smoky oak so joe penna leaked (accidentally i assume) the details of the fine-tuning hw being used on SDXL and it cost around $15,000,000
Granted, we're probably using a whole lot more images than you -- and getting it done a whole lot faster.
But that's what makes finetuning a lot easier for you later, eh?
oak osprey
cyan snow
cyan snow
<#1100170312106127410> etc
did they finally make a base model better than 1.5? ever since they released that it was all downgraded
like 2.1 feels worse than 1.5
oak osprey
SDXL is better than base 1.5 or 2.1 if that's your only standard for comparison, yes
DeepFloyd is better than base 2.0 but perhaps not 2.1
2.1 has better photorealism and its noise that remains in the image by default (base model) is more tolerable/beautiful than what DF does for photorealistic images
cyan snow
DeepFloyd is better than base 2.0 but perhaps not 2.1
deepfloyd is the only model that can consistently do text
cyan snow
2.1 has better photorealism and its noise that remains in the image by default (...
i feel like finetuned 1.5 models are more realistic than 2.1, but that might be because there are not that many 2.1 finetunes.
oak osprey
it's entirely subjective, bruh
you can get absolutely killer photorealistic images out of both. and when you fine-tune them, the differences are such that it takes essentially a photo editor type person that zooms in and stares at pixels for too long, to care about them.
but you have to remember that SD1.5 has an internal representation of the 512x512 image, in 64x64 space
it outputs a 512x512 image through "super resolution" skills that SAI taught it by pairing low res images with high res images
so anything based on SD1.5 lives within that 64x64 space, which is nowhere near enough pixels to represent things properly
there's far greater potential with the 768x768 2.1 model, or DeepFloyd which has subsequent layers that are trained more thoroughly on this super-resolution part
cyan snow
so anything based on SD1.5 lives within that 64x64 space, which is nowhere near ...
and 2.1 can do that better?
oak osprey
yeah it has two variants, a 512x512 with similar architecture to the SD1.5 model's unet and a 768 that has a larger dimension internally, which means it takes more memory during inference and training
oak osprey
people are lazy and do not want to re-train from scratch, there's more realistic models to merge endlessly for 1.5
more people do merges than actual training now
i literally can not stand looking at 1.5 anymore after fine-tuning 2.1
like i'm a day or two away from simply deleting them all from my database of models my users can use on my bot
cyan snow
people are lazy and do not want to re-train from scratch, there's more realistic...
wait, isn't it possible to merge a 2.1 model with a 1.5 model? won't it just apply the same weight on all the parameters?
oak osprey
uhm, nope
they have different text encoders, different parameter count, different vocab size
cyan snow
huh, so basically this community will start to finetune on the newer version eventually
oak osprey
probably not
cyan snow
then that means we will stay on the same level
oak osprey
the loudest parts of the community will likely not change, and will probably just continue to shit on new models released
once SAI removed art styles at the behest of authors artists, and nudity as a result of lawsuits and other liabilities, the community cries "censorship" and says "don't treat us like children, allow us to police ourselves" despite the self-policing not working being why they did that
i went to the CivitAI discord this morning to request help, and the kinds of people and the content they allow over there. big yikes
i would be cautious when you hear any kind of "opinion" of "the community" and remember these people's habits and tendencies, and then, balance their opinions against reality.
ask General Awareness about his issues training 2.1 with Khoya and how the developer of that toolset, tailors it toward a certain "genre" of generative artwork, and how he does not care / is dismissive of anyone who suggests there's some issues with their approach
interesting how Artius v2.1 removes some of SD's bias by generating females or males in occupations they're not traditionally known for. example. an elderly wizard prompt brought back a cute grandma wizard.
wispy nest
this image is very cool, its from the movie beyond the black rainbow
do you guys know how i can replicate it in stable diffusion?
oak osprey
and you didn't even mention you're announcing the Siak Dang Fold today??? so humble
oak osprey
is this SD1.5?
surreal thistle
Yep.
oak osprey
nice
surreal thistle
It also made these ones.
“Could you please tell me the date?”
“09 Jerk.”
“K, thanks.”
The rest of the text looks like Tamil, which is weird because I told it to generate “Spanish text”.
oak osprey
i tried to get text 'poop' on this, and it didn't go as i planned
that is what happens when i throw 'photoreal' after 'animated, 2D, sprite' ahahaha. leonard+ nemoy+ as captain nemo
gritty cedar
hahah all artstation does is tag images "NoAI" in the metadata lmao
gritty cedar
left i guess
surreal thistle
Well, it looks less like a plastic doll.
The left one is Playground 1.0, the right one is SD 1.5.
gritty cedar
ok but I can make a photoreal image with 1.5?
surreal thistle
Can you?
wispy nest
if you're using things like "face fix" then it might be the cause, other than that, it's mostly about the model and prompt you use. But that's like saying that grass is green. :P
surreal thistle
I literally put “imperfect skin” there.
wispy nest
define "not perfect," is it 99% perfect, or 10% perfect? ;P
surreal thistle
Hmm.
surreal thistle
Whatever the cause might be, the percentage is definitely a bit too high.
wispy nest
it's a learning curve, and how one person writes their prompts might not be like the next person would do. And then it's the whole "what is X to me, and what is X to you." But those things are boring to talk about. :P
What I mostly thing about is, "if I want X, is the way I see X how the ai see X as well?" The way I write prompts wouldn't be close to coherent as english :P
wispy nest
there's many reason for everything, but over time, it gets easier to get what you want :D
surreal thistle
I’ve been doing AI since 2021.
Back then it was only good ol’ Artbreeder.
In March I was still using Artbreeder, but in April I started using Playground.
neon crystal
man apparently LAION is working on multi-lingual OpenCLIP
Nice to know that! SD XL currently has a rudimentary understanding of multilingual prompts from what I’d seen. Even transliterated stuff will give you results related to that language’s culture ( mostly faces)
wispy nest
But there’s still this unnecessary glossiness.
and they look like kids
surreal thistle
Yes.
wispy nest
creepy af
surreal thistle
I put in “25 year old girl”.
But it doesn’t look like the AI processed this info correctly.
wispy nest
"girl" is, when used in news or research (using american english) as a "female under the age of 18," but in general speech we can use it for females below 30, and then there are people who have their own limits when to call them X or Y :P
neon crystal
surreal thistle
So “25 year old young woman”.
cyan snow
wispy nest
hehe, if you wrote it like that, then the people writing the "chicago manual of style" would complain that it's unnecessary to say the same thing twice ;P
25 year old is more detailed, but still similar, to "young woman" :P
but the ai doesn't really use the dictionary definitions. It has its own "thoughts" what things are
cyan snow
wispy nest
is that a bad thing? I don't think I've seen many realistic images when it comes to the ai yet. They are all just "off" somehow, but I'm also not one who like those images in general just because they need a level of work far higher than anything else
surreal thistle
is that a bad thing? I don't think I've seen many realistic images when it comes...
I'm not necessarily against it, but sometimes it looks TOO fake.
wispy nest
yeah, I get that a lot as well, the image is super nice to me but then I notice something off and that thing, how little it might be, changes the entire image :(
wispy nest
yeah, I like that one. Mostly because of the blush/freckles it always adds :D
surreal thistle
This one is also not bad.
These are also some of my best in a long time.
I'm not sure which style to use, the painting/photo mix-up or the full-on realistic one.
wispy nest
depends on what you want to use it on, and also: why not use them all? :D
wispy nest
as long as you don't forget or delete the prompts, then you can always save them forever :D
But I'm not someone who know much about realistic and/or photo stuff as my "limit" on realistic skin is this. :P
wispy nest
no, it's auto1111's webui I run on my computer
I only use "my own stuff" and don't mess around any sites, mostly because I'm too lazy :P
surreal thistle
I don't run anything locally because my computer is:
Boil 'em, mash 'em, stick 'em in a stew!
cyan snow
wispy nest
the real reason why I use my own stuff is because I want to get the images to be as sharp and unblurry (caused by noise, not by style) as possible.
And I learned that the only style that can do so, so far, is anime/cartoon style
surreal thistle
The cartoon style CAN be pretty good.
This is supposed to be a mother with her daughter.
cyan snow
the real reason why I use my own stuff is because I want to get the images to be...
that is not true, you can train a simple lora on images that has a specific style and when you use it it will be exactly the same
i done that myself many times
wispy nest
that is not true, you can train a simple lora on images that has a specific styl...
yeah, I know that. But that's true for anything :P
I also haven't tried that many loras yet, and I haven't seen any sharp images that I like anywhere yet. Not that I've really looked around :/
But do you have any examples on it? :D
surreal thistle
cyan snow
yeah, I know that. But that's true for anything :P
I also haven't tried that man...
i once made a lora that makes images similar to an album cover called ''the caretaker''
surreal thistle
This one is pretty fake now, I really don't like those eyes.
wispy nest
i once made a lora that makes images similar to an album cover called ''the care...
how did that look like? :O
surreal thistle
With the Makoto Shinkai keyword it becomes better.
wispy nest
this is ai generated
nice! :D
one question though: Why a pen in a stone? :P
cyan snow
nice! :D
one question though: Why a pen in a stone? :P
that is the whole point, the caretakers albums are extremely abstract, that is why i made a lora that can mimic it
it's an album that simulates dementia in the form of music
wispy nest
Heh, I never understand abstract stuff…but I think that's also its definition ;P
surreal thistle
wispy nest
This is what I like when it comes to anime sharpness, but I'd like something similar in general, but I don't really know how to explain "what it is I really want" hehe
surreal thistle
I like either impressionism mixed with photorealism or Makoto Shinkai's style when it comes to AI.
But it keeps generating two people when I want 1.
cyan snow
also, many models of AI in general, make this kind of stuff
photo-realism is very evolved in AI
wispy nest
yeah, those things are examples of too blurry for me :(
cyan snow
yeah, those things are examples of too blurry for me :(
this is too blurry? its a 4k image, i made plenty of those
surreal thistle
wispy nest
this is too blurry? its a 4k image, i made plenty of those
yeah, I'd say they are too blurry.
But don't get me wrong, I don't think they are bad images. They are awesome, but when it comes to photo real, then I want next to no blur whatsoever as if taken by a state of the art camera. You know one of those super refined, down to each pixel detailed.
I know it's a super high bar, but I just can't get past blur when it comes to the AI, I must have been bullied by a AI dev when I was young or something :P
blur as a style is super nice and images like this have no issues for me because my eyes don't start of with "this is meant to be a realistic image" in the first place :/
sterile steeple
how can i achieve something like this with stable diffusion also what model name can do something like this one ?
wispy nest
how can i achieve something like this with stable diffusion also what model name...
it's hard to get symmetrical features for fantasy limbs. As well as creating whole new creatures that don't look similar to common animals. Or that's from my experience. The AI has enough problems creating a human with enough fingers to even start adding new ones :P
sterile steeple
it's hard to get symmetrical features for fantasy limbs. As well as creating who...
yea , i have been trying to achieve something like this for some time now and i always get bad figures and bad fingers ^^
wispy nest
yea , i have been trying to achieve something like this for some time now and i ...
it depends a lot on how much you've used the ai, your style, etc. As well as what software you use. But for me, I'm more than happy if the ai can create this correctly :P
sterile steeple
it depends a lot on how much you've used the ai, your style, etc. As well as wha...
this looks great except its mouth and eyes for the creature
i like cartoonish style something like digimon , and wants to try creating some creatures like them for fun ^^
wispy nest
everything is possible, but it also might need as much work ;P
oak osprey
I like either impressionism mixed with photorealism or Makoto Shinkai's style wh...
see above
oak osprey
don't say that you'll upset them
cyan snow
don't say that you'll upset them
it just has higher detail capabilities, at least seems like that so far
wispy nest
yeah, 2.1 has the ability to create much better details :D
cyan snow
also, less ai monstrosities
i didn't get even a single conjoined twin. i think that is due to more parameters
wispy nest
aren't the two castles technically twins? :P
hasty nova
@smoky oak @oak osprey I guess since you two have the best eyes for this... I have just created the first testing plot for DD v3. Keep in mind this one was trained from scratch unlike v2 which was trained on top of v1. What do you think? I know the colors arent as cinematic or dramatic, but it is more painting like which is in the prompt.
cyan snow
aren't the two castles technically twins? :P
it doesn't look awkward, in 1.5 models it just goes mayham
and in this case 2 castles was in the prompt
tame wagon
Does anyone know a tool that writes the title and tags for an image?
cyan snow
Does anyone know a tool that writes the title and tags for an image?
clip interogate
wispy nest
it doesn't look awkward, in 1.5 models it just goes mayham
I think some people say, "you just need to use the correct prompts and negative prompts" to those things ;P
oak osprey
@hasty nova it's interesting how this effect reminds me of paintings i have seen in museums where the photo-realism gets me wondering how they've pulled it off
cyan snow
I think some people say, "you just need to use the correct prompts and negative ...
sometimes in higher resolutions it is inevitable, just because it is a wider shot
hasty nova
<@800501037793280030> it's interesting how this effect reminds me of paintings i...
So it's a good thing, right?
oak osprey
like their eyes actually look wet and reflective. i don't know if you're going to more realistic skin textures or if your prompt is asking for paintings
the skin is still an issue if so, but otherwise to me, the hair looks very 'clean' aka not compressed artifact-laden
wispy nest
sometimes in higher resolutions it is inevitable, just because it is a wider sho...
yeah, that happens in both 1.5 and 2.1 :/
cyan snow
yeah, that happens in both 1.5 and 2.1 :/
well, in 2.1 it's way more rare
oak osprey
at this level of quality it's highly subjective but i like this sample the most as its lines are well-defined and has good contrast
but a question is how do these results change across a larger sampling of prompts. do you see consistently better contrast at this ckpt etc
hasty nova
like their eyes actually look wet and reflective. i don't know if you're going t...
yeah the skin looks weird, but that's probably the painting tag
wispy nest
well, in 2.1 it's way more rare
in my experience it's the other way around :O
oak osprey
in my experience it's the other way around :O
on fine-tuned 2.1 models, it's super rare
cyan snow
in my experience it's the other way around :O
you use highresfix?
surreal thistle
could be? maybe not tho
Smells more like Miyazaki to me.
cyan snow
on fine-tuned 2.1 models, it's super rare
exactly
wispy nest
you use highresfix?
yeah, I use it in every single image I create that isn't just a quick test
hasty nova
at this level of quality it's highly subjective but i like this sample the most ...
I just got home to the model finishing training. Haven't tried any other prompts.
surreal thistle
Smells more like Miyazaki to me.
Or rather a combination of the two.
wispy nest
on fine-tuned 2.1 models, it's super rare
in my experience that's also the other way around :P
cyan snow
in my experience that's also the other way around :P
skill issue
wispy nest
skill issue
"you just need to use the correct stuff?" ;)
hasty nova
I am going to try the exact settings of this image from v2
It was one of my favorites
oak osprey
I just got home to the model finishing training. Haven't tried any other prompts...
two checkpoints where the background hair is totally different
his hair becomes like a shield of smeared paint in the last checkpoint
hasty nova
two checkpoints where the background hair is totally different
so which is the best from that example?
oak osprey
probably the 3rd or 2nd to last
but these issues can be resolved with prompting
definitely can't say which ckpt i prefer unless i see more prompts 😛
prompts = {
"woman": "a woman, hanging out on the beach",
"man": "a man playing guitar in a park",
"child": "a child flying a kite on a sunny day",
"alien": "an alien exploring the Mars surface",
"robot": "a robot serving coffee in a cafe",
"knight": "a knight protecting a castle",
"menn": "a group of men",
"bicycle": "a bicycle, on a mountainside, on a sunny day",
"cosmic": "cosmic entity, sitting in an impossible position, quantum reality, colours",
"wizard": "a mage wizard, bearded and gray hair, blue star hat with wand and mystical haze",
"wizarddd": "digital art, fantasy, portrait of an old wizard, detailed",
"macro": "a dramatic city-scape at sunset or sunrise",
"micro": "RNA and other molecular machinery of life",
"gecko": "a leopard gecko stalking a cricket"
}
oak osprey
i like the third one again
third new one i mean
i wonder if you can bring out some of the old versions' character with prompt changes
hasty nova
probably
it probably uses slightly different prompting
its worse at higher res
not great :/ but the prompt is also pretty bad
stone cipher
```json
prompts = {
"woman": "a woman, hanging out on the beach",
...
Macro photography of RNA and other molecular machinery of life
stone cipher
a dramatic city-scape at sunset or sunrise
oak osprey
@hasty nova here's all my results for the top prompts on your page
i used the same non-cherry-picked seed from each gen, but it's a different seed than you used
in case that helps in some way
wispy nest
hehe, those are the twins I remember :P
oak osprey
i was surprised to see it duplicate
none of the other prompts did
oh, the owl did
but i selected the malformed owl over the non duplicated output because it was cuter and had a hat
wispy nest
yeah, it's random'ish. I've mostly noticed that happen at very specific aspect ratios, no idea why otherwise why it happens like that
oak osprey
i am using 1152x768 which i found works very well
wispy nest
good thing we're not paying a normal artist for this or we'd be out of a lot of money :P
wispy nest
first one because the background of the second one kinda make it look like someone has cut out and then glued the cat to it :P
hasty nova
first one because the background of the second one kinda make it look like someo...
oh true
wispy nest
but I like both of them :D
oak osprey
hasty nova
Ok, I have decided the model isn't any better.
Not sure what to do differently other than a different learning rate.
wispy nest
at least you learned stuff? :D
hasty nova
Although I may not have tested thouroughly enough
I just decided it kinda started losing some things
wispy nest
there will always be such reason, the important part is to know when to continue and when not to. :)
hasty nova
Yeah. v2 is ok but I felt didn't have a diverse enough datset.
wispy nest
that's also the normal response in the end. :P
But yeah, now you know more what to do in the future :D
oak osprey
I just decided it kinda started losing some things
try freezing the TE
some people have found that it helps to train for 1 epoch with the TE unfrozen and then stop, create checkpoint, and re-train again for your full run with the TE frozen
finally, create a new pretrained checkpoint and run training again from that for 1 more epoch, to "bring the weights up" and make it more coherent
but i'm not sure what training data they use for each run, if they break it into chunks or potentially repeat
hasty nova
try freezing the TE
not sure how to do that with what I use
oak osprey
disable text encoder training
i'm so tempted to just stop generating class data where i'm at now and begin a training run to see what happens with something like a 1 to 1 ratio of training to class images with a diverse tag set of training data
oh man
i'm not there yet
i have 17,252 class images so far, and 22,976 training images
wispy nest
That's enough for now, I'm not much for landscapes but even I can tell that; yes, this is one :P
oak osprey
yeah that makes sense
your experience mirrors that i had with my 13k step model vs 4k steps
cyan snow
oak osprey
4k steps is super creative, understands prompts very well. but when at 13,000 steps it takes more work to get images but they have more style when you can get what you want. it kinda sucks, because the text encoder clearly went Alzheimer's once the polynomial rate rose too high, and you can tell if it hadn't, the 13k model would be far better than the 4k one
hasty nova
4k steps is super creative, understands prompts very well. but when at 13,000 st...
I only saved every 12500 steps :/
oak osprey
i save every 1,000
hasty nova
I don't have enough space to do that
cyan snow
also, this is with a 2.1 model
smoky oak
<@296720689468145664> <@636706883859906562> I guess since you two have the best ...
This shows a huge quality improvement in the details!
just look at that beard detail improvement
hasty nova
This shows a huge quality improvement in the details!
Which means my dataset is better but I definitely need to try different training techniques
smoky oak
I will just start off with saying V3 looks considerably better on a detail vs artifact level
wispy nest
just look at that beard detail improvement
the left one looks like it's a close up for a real photo :O
hasty nova
Thanks! But I ultimately decided it has too many other negatives to release in its current state. I am going to try @oak osprey 's advice next time I have the ability to train.
oak osprey
that advice is to freeze the text encoder after 1 epoch of pre-training warm up and then frozen text encoder for like 25 epoch and then unfrozen for one more
wispy nest
almost, it's the top left corner of the right image that breaks the "realism" for me, while the left one doesn't have that "smooth" skin. :P
smoky oak
the V2 is just crusty
oak osprey
yeah i agree Sytan the new eyes are like glassy and wet looking
ever been to a museum and seen amazing paintings?
smoky oak
this one looks like a crusty filter on a photo
this one looks like an intricate painting
to me, at least
the details are lower and higher in the areas that should be lower and higher
hasty nova
almost, it's the top left corner of the right image that breaks the "realism" fo...
it's not supposed to be realistic though, it's supposed to be a painting
smoky oak
I would have to say at least a 40% increase in the issues I saw in the model
hasty nova
or painting-like
wispy nest
it's not supposed to be realistic though, it's supposed to be a painting
yeah, exactly, that's why I said that I thought it was a close up of a photo. That was my first reaction when I saw it :D
while the second image didn't look as much as the same "close up of a photo," probably because of it being a painting :P
smoky oak
this one looks fantastic, great work @hasty nova
hasty nova
this one looks fantastic, great work <@800501037793280030>
Thank you!
smoky oak
you are def heading in the right direction
hasty nova
I think it's the dataset improvements
wispy nest
and my test of a "castle on a hill" is not going well at all :P
smoky oak
ok, so my new mouse AND my external m.2 bay are both here
wispy nest
I think this was the best result I got, very much castle... :P
oak osprey
ok, so my new mouse AND my external m.2 bay are both here
where's the PSU
smoky oak
I don't need one now that I undervolted lol
this 3080 draws about the same as my 3060ti now
smoky oak
awesome, new mouse feels great :D
its wireless, so I don't have to worry about wires anymore, thank goodness
wispy nest
is vincent the artist who makes such paintings? I've noticed that castles have a larger chance to have those disney/triangle tops, even being blue more often than not :O
oak osprey
was that, uh, a big issue? wires?
you replaced the permanence of a wire with the ever increasing entropy of batteries
smoky oak
was that, uh, a big issue? wires?
it was starting to become one with the way my PC is setup, and how I turn entirely based off the monitor I am using as my main
smoky oak
my front middle monitor is usually my main, its a 34 inch 75hz ultrawide
But if I am gaming or watching a show, I use my left monitor cause its a much brighter and more responsive 165hz 27 inch monitor, and it requires me to move my mouse and keyboard over, which is a huge pain with a wired mouse, cause the cable is so stiff that it starts putting back pressure on my movements from the cable touching the monitor stand
oak osprey
Sytan, try your model with 'knolling' on one of your funko prompts
it has surprising results
cybernetic parts knolling leads to this sort of side by side display
knolling+ makes it kinda uhm, tear him apart to pieces
idk why it put that stuff around "Steve Buscemi"
stone cipher
stone cipher
stone cipher
https://images-ext-2.discordapp.net/external/-60iAjPKSYrBg4HtwBy2sjlWPvzE5MiNPLk...
past pelican
Anyone up for a historical challenge? I've been trying to recover the SD image source of 
for a long time now. It was generated a long time ago, but I am pretty confident I know which model was used. Unfortunately I no longer have access to the source of the image to check.
I do have a pretty good estimation of the prompt, the sampler, the steps, the model and even the seed. The problem is that this was created back when the stable diffusion bot would spit out 4 images at the same time. The image produced was index 1 so the seed isn't exactly the same seed.
Here's the original image data:
The name of the image is:
"my_husband_puts_googley_eyes_on_everything.Googley_eyes_stuck_to_our_poor_cat_mixed_media.-n_4_-i_-S_2120710778_ts-1659937824_idx-1"
However when trying to generate it from scratch you don't get the same image.
Here's what I do know about the image:
-The timestamp is: 8/13/2022, 2:02:08 PM which means the model was either SD 1.3 or SD 1.4
-It also means that the sampler was probably LMS / K-LMS with a default step count of 50.
-The seed is probably: 2120710778
stone cipher
Here's what I do know about the image:
-The timestamp is: 8/13/2022, 2:02:08 PM...
I can't see the image
oak osprey
oak osprey
Here's what I do know about the image:
-The timestamp is: 8/13/2022, 2:02:08 PM...
the pipeline code, the sigmas, other things have changed since then too
new optimizations totally change the output, eg. xformers
past pelican
I hadn't considered xformers. You're right.
oak osprey
torch wasn't using Generators for a while. people were doing torch.manual_seed but might not have set the numpy seed, and the seed was likely generated on CPU
in fact that bot might not have even used Diffusers
past pelican
I can't see the image
I'm guessing the discord link broke.
radiant plinth
oak osprey
radiant plinth
wispy nest
I think this was the best result I got, very much castle... :P
dope
wispy nest
nice pete. looks like a still from a 60's french historical film to me
stone cipher
nice pete. looks like a still from a 60's french historical film to me
Thanks
smoky oak
Alright, I got my new mouse working
smoky oak
@dense tapir Look that this jokery lmao
wispy nest
dope
too bad about the face, but this was more fun than castles :P
control net? her posture looks decent at least
wispy nest
control net? her posture looks decent at least
I don't think so, if it's not something toggled automatically I don't think I have that. I have no idea what that is :P
noice. looks pretty professional.
Gonna be honest and say that I have no idea why those images have better pose. It's 100% luck, somehow :P
smoky oak
I usually just need to plug mine in to do that
I finally made the jump to wireless 😅
smoky oak
woo, and now i have 4.5TB of solid state drive available at all time s:>
dense tapir
<@315547788211388417> Look that this jokery lmao
WTF? Why all those ancient versions?
I came in with Solidworks-09 back in 2011 then 2012, on up (2015 was utter shite). Those other programs are ancient too.
smoky oak
WTF? Why all those ancient versions?
These are now part of the spec view perf suit cause they are old enough to license or something like that
Either way, they show relative performance as the tools still use most of the same SDK's and stuff
smoky oak
they are standard test suits now
smoky oak
so old as to be useless for anything but this
They are a good way to see a general connection between specific workload's and what GPU's do better
dense tapir
the first HUGE advancement of Solidworks, for instance, was 2009 to 2012. People still use 2014 version too
oak osprey
what about Houdini
smoky oak
I forget which one, but there is a workstation software where AMD destroys NVIDIA
dense tapir
Houdini on a 4060 anything would be laughable
smoky oak
what about Houdini
That wouldn't be a good perf test honestly
most serious houdini work is done with CPU's anyways
dense tapir
I still can't find what I need for actual 7k series cards infering and training. I even go to the AMD SD discords and I just flat out asked " I just wish I could see some actual numbers of 768x768 for inference as well as training lora/lycoris on a 7900XTX even though ROCm is not yet made for it. Why has no one shown this, or is it really so bad it is an embarrassment?" No response.
oak osprey
The output on a AMD 7900XTX would look something like:
Average step time: 47.19188690185547ms/it
Clip Inference time (ms) = 109.531
VAE Inference time (ms): 78.590
Total image generation time: 2.5788655281066895sec
dense tapir
All these rental places are in third world countries and will only put up what is best so Nvidia. I don't ever doubt they are better but by how much for the ungodly prices they are charging.
I hated Shark
oak osprey
it should be pretty much the same. the inference speed is roughly the same across different models
dense tapir
compiles everything then runs off that but change the model, or the rez and it has to recompile that. As they said we are for speed not space so several 100 TBs might be needed for special cases where you change models, and resolutions a lot.
I get what they are doing, and why they are doing it because all the thinking is done one time then it only needs to be concerned with the more minor things. Does help in speed but 100 to 1000s of models is crazy too much for me.
Also it doesn't train 😦
For me that is my biggest metric
dense tapir
compile does not mean train
Be nice if Shark could train, but I don't see how with how it is set up.
oak osprey
i run mine on a discord bot, so, the trade-offs are well worth it to use Diffusers
the number of generations that thing does
dense tapir
right, for most running locally that is a headache
oak osprey
yeah unless you're batch generating class data for training
then you really should try that
dense tapir
Joe Penna's has always used diffusers and the quality is bar none the best but needs 24GB cards. Hence my need for 24GB.
oak osprey
yeah. training 2.1 in 16bit mode was found to cause catastrophic loss in the text encoder
dense tapir
yep
oak osprey
it needs at minimum bf16 but at the ideal place you're using fp32
smoky oak
this new mouse is growing on me
dense tapir
yeah, and no BF16 on any sub Turing card so T4 couldn't do it, nor could I locally.
oak osprey
have you ever tried DeepSpeed?
dense tapir
No, because I have never heard of it before.
oak osprey
oh, it's used to get DeepFloyd vram use down to like 6gb
dense tapir
is that the 4bit thing?
oak osprey
i honestly don't know what it does
i just wanted to get good basic results first and reproduce them before i start experimenting like that
dense tapir
suddenly, everyone is running around wanting 4bit because it allows for much larger models.
oak osprey
i could see that working
dense tapir
I see 4bit and think ZX87
oak osprey
the internal dimensions of the image in SD is 64x64
you could do more with 4bit
work in multiple layers that get blended, so that you can make more than 4bit worth of colour or smth
idk
dense tapir
I still think 64x64 is a source of some issues, but others can disagree if they wish.
dense tapir
it is plenty big enough (53x64) reminds me of "640K is all we will ever need".
oak osprey
lol my model loves to make Houdini go surfing
dude's committed to getting pitted
whoops lol
native stag
what software do you use ?
dense tapir
one of my extensions I updated and is now wigging out. 😦 total error.
yep, a ticket from 10h ago seems the dev broke his shit and published it anyway.
sterile temple
You can usually delete the folder for that extension if it's causing issues
crystal deltaBOT
dense tapir
You can usually delete the folder for that extension if it's causing issues
That is akin to saying the sun always rises until it doesn't. LOL
sterile temple
Now new users can try and send prompts here with big text?
dense tapir
I wonder how good this is - Sampler: DPM Solver++
sterile temple
do an x y chart
dense tapir
We do not have it in automatic1111
sterile temple
oh
wispy nest
do you need this model v1-5-pruned-emaonly.ckpt for automatic1111 web ui or can I run any other model instead of that?
dense tapir
any, that is just the default one if you have none it downloads I think
sterile temple
too much cocoa butter
dense tapir
eating it or rubbing it on?
dense tapir
hehehe
I am surprised at the lack of texture to be honest
As if a rubber balloon.
Those vinyl blow up ones
sterile temple
well it seems to have a base, so maybe it thought it was making a vinyl figure
dense tapir
Yeah.
I am noticing openai follows prompts better
This is nuts
I don't get why it is cut off like that
dang, another prompt cut off
I find stable diffusion loves doing that
sterile temple
the wing bone's connected to the.. tit bone
dense tapir
sterile temple
I find some seeds can be pretty strong, and you will get the same composition no matter the subject
sterile temple
like if you're trying to get particular pose and the model just gives them broken arms or three legs because the seed is trying to draw something else?
oak osprey
I wonder how good this is - Sampler: DPM Solver++
i use that one exclusively. plus Karras
sterile temple
and sometimes you just get a picture of a steam train, just cause
sterile temple
colourful even if it is a hot mess
dense tapir
Candyland gone wild
sterile temple
Central American volumetric lighting depth of field bokeh Double Exposure lens flare Vignetting Chromatic Aberration
(one central figure:1.3) Candy land, sweet treats, colorful confections, whimsical atmosphere, delightful fantasy (Piece of Mind:0.0) An artwork consisting of soothing, harmonious color gradients, creating an aura of tranquility and peace
I'm just mashing together wildcards I've made with chatGPT to see what it spits out
tame wagon
Does anyone know a tool that writes the title and tags for an image?
sterile temple
dense tapir
oak osprey
Does anyone know a tool that writes the title and tags for an image?
captionr
tame wagon
@oak osprey I googled ''captionr'' but didn't find anything
sterile temple
Does anyone know a tool that writes the title and tags for an image?
Have you tried the wd14 tagger extension for A1111?
dense tapir
SD 😦
tame wagon
https://github.com/theovercomer8/captionr
thanks !!
dense tapir
3 wheels in the back and only one in the front
LOL
looks like six wheels so missing 2
sterile temple
It needs the extra wheels on the right side to stop the truck from tipping over when the extend the lead canopy
dense tapir
oak osprey
using 8x A100's now to get the training time for this down from 340h to 45h
oak osprey
a) yeah of course, b) if i had them, c) if you could get access to more than one at a time right now
dense tapir
yeah, only 1 at a time but how much faster would it be than 8 A100, or would it be?
oak osprey
my batch size is 64 right now with 64 gradient accumulations, so, BS-4096
that's a big boy training batch size
can't do that on a single H100, i don't think..
it has HBM memory but it's not a golden panacea, i'm pretty much nailing each a100 to the wall in terms of GPU util
should have about 30 checkpoints to check out tomorrow if i haven't totally wrecked everything
i'm using 8bit ADAM so there's a lot of oscillation in the loss value
dense tapir
Conflicting info is out there
The H100 is the successor to Nvidia's A100 GPUs, which have been at the foundation of modern large language model development efforts. According to Nvidia, the H100 is up to nine times faster for AI training and 30 times faster for inference than the A100.
oak osprey
according to nvidia
dense tapir
Tom's says H100 is 54% faster so wth?
oak osprey
training in what
some training pipelines are more or less python-heavy
the more python-heavy it is, the less performance difference
i'm CPU bound a lot of the time, i think
dense tapir
yeah, that would play a big factor for sure. Now if the entire dataset sat in GPU mem there would be no bottleneck
oak osprey
the traditional wisdom is to try and fill your GPU memory
batch size goes pretty high when you try to do that. like with sd2.1 768 on an a100 80G (one) it's BS=7 without gradient checkpointing or accumulation
and 8 will regularly OOM from CUDA
so the wisdom is now, use something like BS=4 and 16 gradient accumulation steps which is roughly approximal to BS=64
but this brings memory consumption way down
so my next step was to try and increase the BS again to bring memory up to the max
dense tapir
This is a good read and from it I get H100 is a bit more hype than reality, which Nvidia is doing right now with their consumer lovelace cards- https://silvertonconsulting.com/2022/09/22/nvidias-h100-vs-a100-the-good-and-bad-news/
oak osprey
but it goes so damn slow i don't know if i'm actually gaining anything
look at it this way, the cost of training goes up to $60,000 when i fill the memory, and $6,000 if i don't
dense tapir
damn, wth?
I wonder how to turn that off? which damn program is doing that?
I don't even have a radeon
oak osprey
loss=0.192
so General i'm discovering the trick to tuning 2.1 is to not use regularization data, it boosts the loss really high somehow
something about 2.1 or the sampler my class data is using, i guess, they interact very poorly
i need to investigate that sometime and try some hyper fast learning models and see what i can do to improve that situation 😦
the workaround is to simply have a few thousand training images, all varied and captioned
don't train for too long, or don't train the text encoder at all
or train text encoder for 1 epoch, and then, freeze it and keep going
i have now done ^ this
static tusk
Not a fan of images per minute metric, but at least one outlet covers SD benchmarks
oak osprey
man i need to help Tom get a good benchmark going for SD
or whoever that is
that chart is missing the XTX and the 4090
static tusk
yeah its comparing mid range only
oak osprey
so i'm getting 15s for 4 images generally, at 32 steps, on my A6000
my A100's do a lot better for generating images and i get more like 3s for 4 images
dense tapir
smoky oak
Unfortunately, the toms hardware info is really bad for comparisons
oak osprey
that chart is a "well, duh" moment and doesn't get into higher res issues etc yea
dense tapir
Not a fan of images per minute metric, but at least one outlet covers SD benchma...
Yeah, I want it/s or s/it
smoky oak
so i'm getting 15s for 4 images generally, at 32 steps, on my A6000
isn't the A6000 the perf of a 3090? cause that seems slow
oak osprey
i'm doing 32bit, you
smoky oak
oh wait, 32 steps, nevermind
oak osprey
and at 32bit
smoky oak
i'm doing 32bit, you
ahh
smoky oak
Alright, I was gonna say
oak osprey
😦
smoky oak
:>
oak osprey
it's Total Recall
smoky oak
:C
oak osprey
ARNOLD
smoky oak
Oh, did you see discord has big text now?
smoky oak
ARNOLD
oak osprey
that just looks weird here
oak osprey
oak osprey
smoky oak
Cause I was gonna say, you send images and view images all the tim elmao
oak osprey
DeepFloyd, yo
smoky oak
they're AI generated
Shut the hell up lmao
static tusk
etuu
oak osprey
Shut the hell up lmao
i grew up doing all sorts of sick pranks on my family members. sick in a good way. like the old saran wrap on the toilet
one time i put a brick inside the water tank of the toilet to reduce its use
smoky oak
i grew up doing all sorts of sick pranks on my family members. sick in a good wa...
There is a special place in hell for you lol
oak osprey
changed my dad's thermostat programming so he spent less on electricity
he NEVER found out who did it
smoky oak
one time i put a brick inside the water tank of the toilet to reduce its use
My family used to do that
when my toilet exploded (yes, you read that right) we got a new toilet with a super efficient water tank, and 2 flush sizes
it uses like .3 gallons on the shallow flush, and its great
I still can't believe my toilet exploded lol
oak osprey
well my family took me to therapy because they found out i'd been peeing in the top part
i was 8 and the doctor had me play nintendo 64 as he talked to me and he realised i was just trying to be efficient and use the water twice
i needed to offset the times i was sitting in the tub with the shower head running pretending i was in a hit submarine
travel illustration in the style of jon klassen with whitespace
elegant fox with six eyes and 3 tails a lot of colour like a beautiful painting
galaxychat anniversary 16 birthday rocket cake
the allegory of the maiden and the three night goblins style of cranach the elder wide angle forest background landscape highly
timeless logo for a architectural studio with the letters a and c minimalistic elegant white background with text in it black le
someone walking down a sidewalk with a cute laughing devil following them
alluring stoner eyes yoda magician tophat
ralph fiennes as lucifer angel of light long blonde hair golden armor baroque
boho illustration of dahlias
a small seedy little green goblin man that is secretly really friendly and devoted to his wizard lover knee height unreal engine
disney style chibi full body cute smiling unreal engine detailed ultra high definition 8k
sylvana windrunner with cyberpunk hair v5 ar 23
a single large cylindershaped bacteria with flagella swimming in a pipe full of water hyperdetailed realistic
with a huge heteroideus red metal net cave in the middle of the road
some of those are like, what
prompts from my training data set
what my model made for the goblin one
lethal burrow
Thanxx to the voters...u are appreciated ❤️
dense tapir
smoky oak
@dense tapirI have discovered a new favorite thing lol
Generating a nice image, then swapping the pos and neg to see its antithesis
Nice and pretty image
swapped pos and neg lol
sterile temple
is that how she really feels inside 😦
dense tapir
Damn, lol
Well, I have watched a lot of vids and the 4060/ti is DOA. Almost all retailers around the world say it is a flop and they will not be ordering more of this turd of a card.
No one is even coming in and asking about it so Microcenter will not be opening early for this one.
People can say whatever in the fuck they want to about the moore's law is dead channel but he has the connections and will go on record with things long before those things are done. He said the AMD 7600 could go as low as $269 and they will still show a nice profit so that is where it should be released at. Sure enough the MSRP is $269.
Is it a great card? No
smoky oak
$269 is a reasonable price I would say, especially compared to what NVIDIA is offering lol
dense tapir
It's issue is 8GB too
smoky oak
yeah
dense tapir
I think it was hardware unboxed who just tested the 4060 and even at 1080p you could see, on older games mind you, textures being uncompressed before your eyes.
the 3060/3060ti did not have that issue
so much for filthy Jensen's claim that the new 32mb buffer chip will help.
I cringed as I remember those days when 256MB cards were going bye bye to 2gb cards.
*1 gb so I grabbed a 2gb
@smoky oak
dense tapir
SD glitched with shoulder and the hair. Seems to do that a lot
moons, and round objects glitch when something is infront of it
Upped the ddim to 30 steps (so 31) from 20.
smoky oak
@dense tapirdid you see that SDXL has reached 50% training completion?
smoky oak
It looks amazing already
These are all 50% trained SDXL images, and my god do they look good
SDXL is gonna be something insane
A base model that is already generating as good, if not better than some of the best finetunes is insane
It also has proper brightness fixes (not noise offset, its better)
dense tapir
Matters not to me though until I have more memory, even then if I can't train it then I want nothing to do with it.
sterile temple
static tusk
https://youtu.be/Sp6K3qpVFO0 Adobe didnt take long to add generative tools in Photoshop with Firefly tech
dense tapir
Was in beta for a long while
Last Nov, or Dec I was introduced to it. Was very good but still needed tweaking on some of them.
Whoa, I am so damn close
smoky oak
@dense tapirCan I delete my pip cache to save space?
Wanna make sure I don't break anything lol
smoky oak
I have cleared like... 900GB off my PC so far
dense tapir
how many did it say it removed?
dense tapir
fuck, lol
smoky oak
dense tapir
that is a laugh riot, no offense
smoky oak
Hey, I had no idea lmao
dense tapir
it's cool
smoky oak
finally got my 2TB NVME set up as well
dense tapir
all that happens is it goes back out on the net to regrab it if needed, and sometimes that is exactly what you want as yours is corrupted (happened to me once)
smoky oak
all that happens is it goes back out on the net to regrab it if needed, and some...
Thats what I was 99.9% sure it was, but just wanted to make sure cause asking is a lot easier than fixing broken shit lol
dense tapir
Nice, my 1tb 970 evo pro is very nice, and the old 512gb WD Black is all for Auto
smoky oak
my D drive has 240 GB of SD files lol
I will need to nuke my outputs soon lol
I am at 81,553 images
dense tapir
Well, I have over 100 models and most of that is 1.4/1.5 so that stayed on the old mechanical as I hardly touch them.
dense tapir
it gets insane
I absolutely love the roasting Nvidia is getting from 99% of the reviewers over this trash 4060.
There is one ass kisser reviewer I don't watch much and even he had neutral to say about it. For that ass kisser that was remarkable.
give an example. 960 vs 1060 6gb was 100 more MSRP and had a 72% generational improvement. 3gb version was 35%.
You saw the image I pinged you earlier showing 4060 vs 3060/3060ti.
another thing is my 1060 6b has a 192bit bus not 128. 128 was like for the 1050 or some such crap
When I was buying that is what sold me was the 192 as I was coming from a 7870.
smoky oak
@dense tapirYou know what is HILARIOUS?
The
GTX
550
has a 192bit bus
lmfaoooooo
dense tapir
I know
smoky oak
and the GTX 450 had a 128 bus
dense tapir
I really do not understand what in the hell Nvidia is doing. I really do not.
As I have said I can only think they are trying to burn us all then close down their gpu gamer division
smoky oak
I am pretty sure the 4060/ti are the first 60 branded teir GTX/RTX cards ever to have 128 bit
dense tapir
It is their worst, absolute worst, generational improvement.
smoky oak
But the hard thing is its not consistently the worst
once again, the 4080 and 4090 came out sooo strong, if a bit pricy on the 4080
its insane how far you fall from the 4080/4090
dense tapir
I got into a fight with some jackass about the 4080 because I said it was a fantastic card if not overly priced. He comes back and we had a lot of rounds. Idiot refused to listen when I said "on a technical standpoint the 4080 is fantastic, while on a price view point it sucks." He says "you can't have it both ways it either sucks or not". FFS!
I hate knuckle draggers like that.
smoky oak
I got into a fight with some jackass about the 4080 because I said it was a fant...
either it sucks or not
what
sterile temple
was curious, can still see dell are ripping people off
smoky oak
As is ALWAYS said in Electronics, there is never a bad product, only a bad price
(unless its like blowing up lmao)
drop the price on the 4060ti 16GB to $350 and suddenly its not nearly as bad of a deal
its the price that makes it bad
cause in that case, its about 12% faster than the last gen with 2x the VRAM for $50 less. I would be hard pressed to say thats a bad deal, though it still would not be a particularly good one, but you get my point
dense tapir
drop the price on the 4060ti 16GB to $350 and suddenly its not nearly as bad of ...
Well, nah, the card is technically inferior too, and that is the worst part.
smoky oak
it was just a spitball example
if you lower the price enough, the other logistics don't matter as much
dense tapir
If it had a 192bit bus, 16gb, 350 USD it would be no issues with it at all for most people.
even $399 for some
smoky oak
I find it funny saying that the 4090 is a good deal for what it is (i would say it is honestly, especially compared to the 3090), and then following it up by how everything else is just objectively bad for one reason or another
dense tapir
that bit bus hurts worse than the price. Oh, and the damn 4 lanes only
smoky oak
I thought it was only 8 lanes, not 4
dense tapir
I mean for gen 3
smoky oak
Yeah, on gen 3 its only 4 lanes of gen 4
dense tapir
yep
smoky oak
equivalent
dense tapir
which people at the 60 pricepoint the majority is on gen 3
dense tapir
I loved it
smoky oak
I love when they get real snarky like that lol
dense tapir
Jensen is persona non grata right now and he better get a clue real fast.
even if the price were lower for the 4060 gen 3 at 4 lanes of gen 4 speed, 128 bit bus, that has to sting no matter what. Gamers gonna get that card and immediately return them.
they did to the 4070/ti
sterile temple
I can only find 4060s in laptops here
smoky oak
the 4070 to 4070ti gap is genuinely hysterical to me lol
dense tapir
I wonder if they are even going to attempt a 4050?
smoky oak
probably
dense tapir
4GB?
smoky oak
probably 6
smoky oak
I would guess 6 GB and 96 bit
dense tapir
Python has a clean too but I forgot what it is.
smoky oak
I'll leave it for now, its only 4.8GB
dense tapir
Working for everyone, dev pushes an "upgrade" and we all update and it is broken so we can't use it. Yep, dev closed as duplicate of an already closed duplicate and said it is all our faults we didn't download something and make sure we are on the right Python.
sterile temple
I would leave site-packages, that's what it uses to run
dense tapir
Good, the dev quietly updated his fuck up
works now
I love how it is all our fault, closes the tickets, has an update, and I didn't do anything different yet the newest update works. smh
static tusk
which ext was that?
storm badger
smoky oak
I have officially cleaned over 1TB of data
dense tapir
Sheesh
dense tapir
nope, this thing doesn't work right
Hopefully he fixes it 100%
These no longer work
earnest solar
Whompst is it that said SD can't furries?
sterile temple
Is that from Morrowind?
dense tapir
sterile temple
dense tapir
sterile temple
earnest solar
Is that from Morrowind?
Img 2 Img from a heroforge model.
sterile temple
😐
earnest solar
smoky oak
Whompst is it that said SD can't furries?
The most advanced model ever made for SD is made specifically for furries lol. SD can do furries better than anything else
smoky oak
Yup lol
I have yet to see another model that can do concepts half as good as some furry models
Hope this retrain of this LoRA works well
earnest solar
I'm just over here trying to appreciate how much I've improved over time.
smoky oak
I love SD specifically for that
earnest solar
from this
dense tapir
smoky oak
you can visibly see your improvements
earnest solar
to this
smoky oak
it cost you $0 not to post this lmao
dense tapir
it cost you $0 not to post this lmao
I have no idea wth happened.
earnest solar
it cost you $0 not to post this lmao
no, you don't understand, it needed to be shared.
earnest solar
I have no idea wth happened.
dense tapir
done
dense tapir
You know on mine CUDA disappeared and now it is under 3d
dense tapir
earnest solar
I'm not gonna be able to afford a 30xx series or equivalent or higher until at least next year.
dense tapir
At current prices, and my hatred of Jensen, I am 99.9% sure I am team AMD again and I don't care how slow it is.
dense tapir
what kind of loss?
smoky oak
1/3rd of the way through the second epoch and the loss is only 0.003 lower 😭
dense tapir
you save every 100, or less, steps, right?
smoky oak
getting close to half, and now its at 0.004 lower
dense tapir
epoch saving is too far out
earnest solar
nsfw af
smoky oak
ok, now its 0.005 lower
earnest solar
Have an adorable palate cleanser
smoky oak
so it is at least lowering
smoky oak
epoch saving is too far out
each epoch is 205 steps lol
dense tapir
100 would be better
earnest solar
smoky oak
100 would be better
have never had any reason to go that low
I will just have to adjust my regularization data
dense tapir
when shit fucks up you can go back very close to the point it was at its best
smoky oak
My values are probably too high and limiting it from lowering
dense tapir
smoky oak
when shit fucks up you can go back very close to the point it was at its best
My LoRA's almost never go bad, and I prefer to redo the whole thing if they do, as there was an inherit issue if it messed up
dense tapir
Some reason these are all plastic doll like
earnest solar
smoky oak
and the loss went up a bit, alright, definitely something wrong here
I am not sure if reg is gonna work when the base model is so damn bad at the subject
alright, its done
second epoch only went from 0.114 to 0.111
earnest solar
dense tapir
earnest solar
Now someone else post her legs
dense tapir
2.1 sucks for camera stuff
openai clip is renown for its camera work tokens
when they switched clips that all went bye bye
sterile temple
earnest solar
dense tapir
yeah. Where is that from?
sterile temple
portal
oak osprey
oak osprey
cyan snow
wispy ether
Tf is "camera stuff"
dense tapir
@smoky oak btw, it appears the 750 is better in benchmarks than the new 7600.
wispy ether
People who use intel cards belong to jail
kind quartz
try play on it older game 😄 Thats the card that doesnt support dx9?
sterile temple
