#🏞｜general-with-images
dense tapir
oak osprey
loss=0.0743 @dense tapir terminal SNR seems to really reduce loss in 2.1
subtle seal
how?
wispy ether
subtle seal
where i use that?
oak osprey
in webui.bat
#🤝｜tech-support can help better, i don't actually use that tool or that GPU. i'm just vaguely familiar
oak osprey
neon crystal
oak osprey
baseline 2.1 vs terminal SNR fine-tuned 2.1, same seed/prompt
@wispy spindle do you think SAI will release a slightly fine-tuned 2.1 with this fix applied?
i'm seeing miraculous results, i thought for a moment i was looking at some photography training data instead of test results
seems like all that would be needed is a single epoch at BS=4096 with that insane cluster you guys have
oak osprey
more terminal SNR testing
kind quartz
yes it is definitely vae thing. Try download ... but i dont know what vaes are good for anime like. Try VAE-ft-mse-840000-ema-pruned.safetensors and put it in Vae folder
oak osprey
apparently those flat images are better for colour grading
neon crystal
Yellow Magic Orchestra?
kind quartz
daft punk, The eagles, Pink Floyd must be somewhere there 😄
neon crystal
Yep and Police too
kind quartz
yes that i guessed. Also deep purple, led zeppelin, iron maiden. Maybe black sabbath
neon crystal
And the monkeys too 🙂
kind quartz
it is difficult
wispy nest
Sorry, ok I'm so high I don't know why I did this, rather than typing it, maybe it's because I'm high, but the question remains, regards.
toxic sphinx
yes it is definitely vae thing. Try download ... but i dont know what vaes are g...
Im still getting same results
neon crystal
Say no more.
wispy nest
I just think the multibillion dollar company has a better model than yours, just that...
kind quartz
#🤝｜tech-support i dont know what next now @toxic sphinx
toxic sphinx
ok
wispy nest
Say no more.
Am I wrong in thinking that? Afaik it's along to machine learning and datasets? Regards.
neon crystal
neon crystal
If you're looking for a company logo, Bing is your best bet.
wispy nest
But I did understand the tech? On a high level?
neon crystal
But I did understand the tech? On a high level?
Nope you can do it on mobile using Bing app. Just need a microsoft account
wispy nest
Isn't just a dataset on machine learning?
neon crystal
Some astute journalists regard it as an art for neophytes, others herald it as a resurgence into a dry and cacophonous, cacophonous and monotonous primitivism.
wispy nest
Yeah ok is just a good alghortim of machine learning, with really big datasets maybe even catalogued by "immigrants for low pay"(I'm not American)
marble wasp
Say no more.
I fucking dig this look
neon crystal
I fucking dig this look
🙂
marble wasp
Hahah love it
wispy nest
Are you training it... Or???
neon crystal
Are you training it... Or???
Measured by the scale of the cloud, all activity is trivial. Yet supposing existence to be a feeble theatre, without aim or initial data input, and because we believe it our duty to present ourselves as pristine and untarnished as newly rendered graphics, we have asserted the sole basis for consensus: art.
marble wasp
It's whatever you want it to be
oak osprey
wispy nest
we need a new model already, stable is stale as hell
oak osprey
we need a new model already, stable is stale as hell
do something about it
i hate the colours looking like that but damn if that doesn't look like a real bike
wispy nest
do something about it
I am, I'm whining about no new models xD
oak osprey
I am, I'm whining about no new models xD
i finally got around to testing my Hobbit fine-tune
that's supposed to be Galadriel and Frodo but i love how wish.com it is instead
smoky oak
@dense tapir
Okay so
AMD just got a huge punch in the balls for stable diffusion
An add-on adding tensor RT support directly into a1111 just got released
And it is really damn fast, like 61it/s on 4090 type fast
oak osprey
during 16bit inference maybe
woven dawn
I don't think it works
smoky oak
during 16bit inference maybe
On average from what I'm seeing, most people are saying it's about 80% faster, just going off of the reported numbers in the stable diffusion Reddit
oak osprey
Loras are baked in into the converted model. Hypernetwork support is not tested. Controlnet is not supported. Textual inversion works normally.
have fun with that
woven dawn
Oof
smoky oak
> Loras are baked in into the converted model. Hypernetwork support is not teste...
Honestly, for the types of generations I do day-to-day, still worth it
oak osprey
no controlnet?
smoky oak
I rarely use it as is, I find that it adds too many artifacts usually
The only time I ever do use it is with open pose, but that just causes more issues most of the time
When the model itself isn't able to deduce the pose on its own, hands and lower leg positions typically go crazy
oak osprey
well i don't use tensorrt because it ruins determinism more than most other optimizations i've tried
smoky oak
thats just my typical use tho
oak osprey
it's also a pain in the ass to get running
smoky oak
well i don't use tensorrt because it ruins determinism more than most other opti...
people said the same about xformers, yet it made 0 difference on my GPU for seeds,s o I will see how it works for myself on my own hardware. Curious to see
oak osprey
2023-05-28 21:22:52.488475: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
it's definitely installed there though
just doesn't work
imo the Torch Inductor compiled models are probably better, i get like 80it/sec on my 4090 with that at native res
they work with controlnet, too
i have never tried LoRAs
smoky oak
alright, so it also says in their dev for the addon, that NVIDIA is working on their own custom version just for RTX cards,a nd their gains seems ot be about 3x on average, while also supporting more, so he says to stay tuned cause he will add support as soon as NVIDIA drops it
woven dawn
120it/s on a 4090 would be great
oak osprey
TensorRT does inference in 32bit mode though
oak osprey
people said the same about xformers, yet it made 0 difference on my GPU for seed...
and xformers did hurt reproducibility, a lot 😐 i've seen it
smoky oak
and xformers **did** hurt reproducibility, a lot 😐 i've seen it
I did my own tests, had 0 effect
Even comparing exact pixel values lol
even proved it to general awareness cause he said the same thing
oak osprey
then you probably didn't enable it correctly is my guess
smoky oak
got all the speed benefits, and its been consistently the exact same in all version of SD web UI's I have used
oak osprey
it could also require going high res tbh
smoky oak
¯_(ツ)_/¯
oak osprey
eg. give xformers a reason to work
typically the reason people try to use it is to get higher res outputs
oh, this is the GPUs that this extension works for
go go T4 @dense tapir
probably the only one in the list that needs a boost in inference speed tbh
everyone else is going to be using this to eg. boost the speed of Blue Willow's bot and get more out of their existing datacentres
what the hell is a Jetson
what is it
Volta
and the Orin is for self-driving cars
boreal falcon
@oak ospreyJetson is the Raspberry Pi of Nvidia but specialised for AI
why the hell is it 4.3k dollars in that shop?
oak osprey
welcome to canada
smoky oak
My full DNA/ancestry report is back, and I am indeed Native American... how crazy
boreal falcon
i have to look up again but that price is ridiculous
i saw those things for maximal few hundred €
since i plan to buy them myself
smoky oak
@oak ospreyReally trying to try tensorRT, but NVIDIA's site isn't working
looks like no tensor RT for me, cause NVIDIA lost my account lol
boreal falcon
well ok its high
smoky oak
tells me there is no account with this email, I try to make one, they say its in use
I try to reset password, they send email and it doesn't make it to me, I try to proceed without,a nd then it says the account is not found
great job NVIDIA
boreal falcon
i had issues with Nvidia Geforce Experience only lol
couldnt even resolve it, i rebooted my whole system actually
smoky oak
yup, looks like no tensor RT for me cause they can't get my account sorted
oak osprey
looks like no tensor RT for me, cause NVIDIA lost my account lol
i hate all that shit
and nvidia's own site for tensorrt has all sorts of dead links
they claim somewhere it had support for 3080 but then i went to follow that, 404. and when looking at the changelogs since then, that's been removed???? because... they're.. like... we have no competition??
AMD does stuff like that too, and i'm not sure if AMD got that idea from NVIDIA or if it's "just good business practices"
smoky oak
ah, probably removing the old versions until they can release their new one that makes NVIDIA cards look even better
smoky oak
although, this persons link is from yesterday, so IDK
oak osprey
nvidia could pwn the shit out of this space if this stuff were all open source because it's so difficult for distributions to package it up properly
smoky oak
don't think they removed it over night
smoky oak
ah, this implementation is also broken for SD 2.x lol
oh well, i would try it, but NVIDIA is making that impossible
smoky oak
Haven't gotten that working either
I'm just trying to get better it perf lol
Not cause my perf is slow, cause it certainly isn't
The faster I can go, the better my energy efficiency
sterile temple
mornin
oak osprey
The faster I can go, the better my energy efficiency
not if it behaves like using AVX512 on intel CPUs did
the power use curve of a GPU during stable diffusion inference is pretty spiky and when i compile the unet the spikes spread out but they're still there, it's possible it could pin the power use at 100% if it compiles well enough it runs without the spikes. might use the same kilojoules for the total work done
the speed up isn't magic, it's just making better use of the GPU in a shorter time. is your power metered with some smart gadget that checks for use at a certain interval?
smoky oak
the speed up isn't magic, it's just making better use of the GPU in a shorter ti...
No, I don't have anything like that at the moment
oak osprey
terminal SNR feels OLED friendly
like that feels like it's very dynamic contrast right there
sterile temple
smoky oak
terminal SNR feels OLED friendly
Definitely so
As somebody who is surrounded by OLED displays, I appreciate true black lol
Amoled watch, amoled tablet, amoled phones
My mom's laptop is updating it's bios and my GOD is it loud
I defaults to 100% full power jet engine while updating
It has an overtone that sounds like those tooth drills at the dentist
gritty saddle
smoky oak
It actually sounded like a mini hair dryer
gritty saddle
is 2.1 the best model right now?
smoky oak
There is no "best"
There are better models for different things
@gritty saddle what are you trying to do?
oak osprey
kandinsky 2.1 is awesome
oak osprey
it's not 2.1 based
smoky oak
Ah, my bad lol
smoky oak
Regardless, I still haven't heard of it before haha
oak osprey
it's a pixel diffusion model
gritty saddle
so when I used midjourney, I get amazing images very easily, but with SD 2.1 it needs a lot of fine tuning until I can get anything close to midjourney, is there a model like midjourneys?
smoky oak
Oh, a midjourney user lol
There are various models that can give you significantly better results than mid journey, but it's not as simple as just typing a few words and getting good results, you actually have to work for them and stable diffusion
gritty saddle
how come with mid journey it is different?
oak osprey
i have a model i've trained from midjourney on 2.1
smoky oak
how come with mid journey it is different?
Midjourney uses a lot of different methods to SD
They are easier to interface with, but have a lower overall quality ceiling
that is to say that MJ is more consistently decent, where SD can be bad or excellent depending on what you know
marble wasp
smoky oak
total range of quality in SD is much more dynamic, but the consistency in the decent range for MJ is better
oak osprey
yeah it seems the more a SD model can do, the worse it is at all of it
when you narrow down what it is capable of and fine-tune those aspects, it becomes much better for those, while losing coherence for the other stuff it kind of was able to do before
midjourney's model, if it is just a single one, must have so many damn parameters in it
smoky oak
Yeah, thats how it is for SD 1.5 and SD 2.1, but SDXL is looking like its gonna fix that, as it is excellent at everything
smoky oak
midjourney's model, if it is just a single one, must have so many damn parameter...
No chance its just one IMO
Its gotta have some form of trigger recognition that tweaks what mix it uses
oak osprey
it's not excellent at everything yet, they're still working on it, but i'm hoping they share a lot of the things they've learnt afterward
there should be an SDXL Outtakes channel
like, when it's malformed it's still beautiful tho
lmao
smoky oak
Alright, if you wanna be pedantic, then sure, its not excellent at everything, but it is better at almost everything when compared to even some of the best finetunes for 1.5 and 2.1, which I think is super cool
oak osprey
it seems to inherit a lot of issues that diffusion models have eg. the tall portrait aspect ratios result in very elongated subjects w/ unnaturally long torsos and long necks
smoky oak
like, when it's malformed it's still beautiful tho
I would really like if they released a 0% trained, 25, 50, 75, and 100
cause even 0% trained was really cool
wispy nest
https://images-ext-1.discordapp.net/external/6SvOYQj9Pudu9slQ34XLuieyjnW8_cvQlc-...
noice
oak osprey
so the dude's looking better and better but guitars man ugh
@smoky oak can your amazing model do this prompt better? without tweaking it?
smoky oak
bro's got that 4 dimensional tesseract guitar lol
oak osprey
a man playing guitar in a park
smoky oak
<@296720689468145664> can your amazing model do this prompt better? without twea...
its not made to do stuff like this, but I can give it a shot
smoky oak
I need to remote into my PC to turn on SD
smoky oak
yeah neither is mine <:KEKL:842966965752037376>
Its not comparable lol
smoky oak
yours is a general model, it should be able to do humans, my model is not made to do humans lol
oak osprey
let 'er rip, ding-dong
i can't wait to see this stairway to heaven of a neck it makes
smoky oak
Alright, SD is initializing
marble wasp
Perhaps I can get some advice or tips on a problem I am having with the photorealism of my latest image. Here is the image, and in 1 sec , ill add the prompt and settings. id be so grateful for any tips.
sterile temple
MJ could be using a complicated pipeline that detects the user's intent based on the prompt and swaps out models and LORAs and what not to get a picture. They could even randomly select controlnet poses and add them to the picture. Who knows
smoky oak
@oak osprey this is first try result, no prompt editing, and a small neg
I am confident it could do muchhh better if I actually tried
marble wasp
Prompt: (high resolution close up photography of bearded man with glasses and white robe), futuristic sky scrapers in the background, Blade Runner, future-punk, RAW, 8k, UHD, dslr, ultra quality, natural light, day, realistic skin textures, highly detailed glossy eyes, high detail, bokeh. Settings: epicrealism-new era modeler, 512/768 size, 30 steps, 5.3 guidance, .52 prompt strength.
for this
any tips on better realism/photorealism/portrait
still looks too painterly
oak osprey
<@636706883859906562> this is first try result, no prompt editing, and a small n...
fairly decent tbh
minus teh watermarks all over
lmao
smoky oak
minus teh watermarks all over
that is inevitable, unfortunately
though they can be inpainted out
oak osprey
Prompt: (high resolution close up photography of bearded man with glasses and wh...
weight the 8k and uhd higher. eg 8K+ UHD+ and remove all the commas, they're unneeded
@smoky oak inpainted out. ah yes, while you put your seat back forward
wispy nest
<@296720689468145664> inpainted out. ah yes, while you put your seat back forwar...
comas have no function on the output?
marble wasp
weight the 8k and uhd higher. eg `8K+ UHD+` and remove all the commas, they're u...
will try now
smoky oak
<@296720689468145664> inpainted out. ah yes, while you put your seat back forwar...
What?
oak osprey
the seat-back goes forward
it's just a weird combo of words
"the lady on the airplane tells me, 'put your seat back forward' and i have to object because my body doesn't move that way" - george carlin
wispy nest
lol
so @oak osprey you are saying putting comma in the prompts have no relevance at all?
smoky oak
they very much do, not sure why he said that
wispy nest
i still use comma but now im questioning it lol
smoky oak
the code strips them out
uhhh, yeah, not sure about that, cause adding a comma between colors and clothing items makes a massive difference lol
you know what, let me test
marble wasp
theyre starting to come out sharper, thank you
oak osprey
so that's how terminal SNR has changed this prompt over 7k steps of training
static tusk
using aetherLux
oak osprey
it has an uncanny quality to it but that's pretty great
is that with negative text embed?
static tusk
of course. I always use them
oak osprey
i never do 😄 other than the ones i generate on the fly
static tusk
2.1 needs negatives badly
oak osprey
well all of my test prompts have zero negatives as i'm trying to train it out of that
smoky oak
@oak ospreyjust tried it, no commas made a considerable difference than with commas, not sure where you got that from
static tusk
not sure thats possible with how the base model was molested
oak osprey
it is possible but it means you lose on things you don't feed it during training
static tusk
everyone who has released a finetune says use negatives for best results
wispy nest
<@636706883859906562>just tried it, no commas made a considerable difference tha...
i would think so too logically, its how words and phrases would take different emphasis
oak osprey
so you have to have a huge dataset and successively freeze text encoder layers and train them one at a time to full convergence
SmartFRZ shows that it's possible to adaptively freeze layers and train them to optimal loss values, as maybe layer 20 stabilizes earlier than layer 10 but current protocols are just sequentual freeze
so i've proposed that to the khoya and ED2 devs, idk what they'll think about it or if they can implement that. it's a bit beyond my current understandings
static tusk
yep thats a lot of data to retrain
oak osprey
well i did that with pseudo-journey
i just need to add photoreal data to that pool of synthetic data
trying to find a pool of high res animal images is pretty hard apparently
smoky oak
Commas vs no commas
There is a considerable difference here for sure
not gonna say you NEED commas, but to say they make no difference is wrong
oak osprey
did you ask for people in the background or no, just curious
smoky oak
I did say "in a public park", so I would say its fair either way
its also not cherry picked, just a random seed
oh, but I DID ask for hum to be sitting on a bench
I think the one with commas looks better in terms of hands, proportions, perspective and positioning though personally
oak osprey
i've seen the best results for style transfer when training the middle layers so far but i was also screwing around with adding TSNR to the training code so i started burning models in 100 steps and wasn't sure why, and reverted to freezing only the later layers. but still burnt things anyway even just training the unet before i discovered i'd configured the noise sampler incorrectly
wispy nest
Commas vs no commas
There is a considerable difference here for sure
not gonna...
have you tried prompt, prompt, prompt vs prompt,prompt,prompt? (no space after comma)
I just have that spaces everywhere and after I started doing it - never stopped, but wanted to know if there's a difference 😄
smoky oak
I have none done that, I can try it tho
oak osprey
jesus that was a bug they fixed 11 days ago apparently
smoky oak
gonna take a while to meticulously type that, but I will give it a shot lol
wispy nest
gonna take a while to meticulously type that, but I will give it a shot lol
You can use text editor to batch auto delete all spaces
oak osprey
it doesn't really make sense to tokenize punctuation imo but if you like it, oh well. it is what it is
wispy nest
have you tried ``` prompt, prompt, prompt``` vs ```prompt,prompt,prompt```? (no ...
i have used negative prompts w/o spaces, by the look of it the output was meaningful
well , , to just,
smoky oak
You can use text editor to batch auto delete all spaces
I will try right now on a small scale
wispy nest
the use of
prompt, prompt, etc
and
prompt,prompt,etc
are the same i think
wispy nest
I will try right now on a small scale
I can quickly add or delete spaces after comma from your prompt, just give me one
smoky oak
did it
they make no difference
identical image
interesting
now gonna try it with spaces but no commas
wispy nest
That's nice, I should stop worrying about it then
oak osprey
try your prompt in spanish
wispy nest
space wouldn't matter, but commas interpret the sentences meaningfully
static tusk
easier to read with spaces though IMO
wispy nest
ADHD will probably decide if I need them or not for me.
smoky oak
spaces makes no difference, but commas do
wispy nest
also in the following instance
panoramic view of lush green grassy open field
space would matter a lot or it would turn into meaningless giberish
oak osprey
yeah it was a bug that they were throwing punctuation away
fixed about a week and a half ago
oak osprey
but anyway it used to, and i knew that because i studied the prompt code before learning it's stolen from NovelAI, so i used Compel's stuff instead.
smoky oak
again, not a massive deal, but it does make a difference
smoky oak
fixed about a week and a half ago
that would have required updating A1111 to have seen a difference right?
oak osprey
i mean, probably lol
smoky oak
cause my A1111 install is 3 months old, with no updates
I do not update, cause it just gets slower when I leave this version
oak osprey
yours probably has yet another bug
well i guess beware that your prompts might render differently when you do update
smoky oak
if I update to any of the versions that force pytorch 2 and SDP, my speed and VRAM usage both take a small hit
smoky oak
speed goes down by about 10%, and uses about 5% more VRAM on average
smoky oak
that's why I do not update lol
static tusk
Jensen keynote at Computex in just over an hour from now
oak osprey
my speed went up by a bajillion because i'm on Ada Lovelace
wispy nest
I do not update, cause it just gets slower when I leave this version
torch 2 tho?
oh I see, skipped some messages
static tusk
I still get excellent speed with torch2 and SDP
oak osprey
his PSU is bad 😛
static tusk
thats with systeminfo extension benchmark
oak osprey
so the dude's looking better and better but guitars man ugh
quantum bench support technology
i wonder if this guy is a composite of the 7000 faces i've put in
so, it like, imagined this group of dudes all shitty-like
now 2k steps later they're like "i'm a real boy!"
tides go in, tides go out - you can't explain that
the best i could do without terminal SNR, and then, with it
wispy nest
trippy
oak osprey
it clears out most of the residual noise left in the image somehow by processing the last timestep first
marble wasp
I mean it's changing a bit and also getting slightly better, but it's still not as photorealistic as my last ones. I don't know what I did to change it, but it's getting frustrating.
oak osprey
you can weight your negatives, too. try adding like, synthetic+ rendered+
with 1.5 i find that reordering the prompt words has a large potential impact, too.
marble wasp
if you don't mind, would you look at my prompt and see what you'd reorder?
if anything
oak osprey
i'm not the best one for that kind of thing, my gens look like this
oh, there's #📝｜prompting-help
sorry i should have mentioned that first, there's probably people who are really good at this that are more active in there
oak osprey
damn man, this dude's like, "i have arrived"
and this isn't perfect but this prompt has been performing pretty reliably for a few thousand steps now so i'm also satisfied with the direction it's going. you should see how it started out. it's pretty bad for 5k steps, and sometimes more, depending on training data it could just get worse and worse
i imagine adding a bunch of negative prompts would clean that right up but i'm hoping, i can make it clean without them
marble wasp
thanks everyone for the tips!
oak osprey
cpu lol
fickle gorge
I think that --lowvram and the other options to lower VRAM usage are messing up with the images. It's so hard to get straight lines, I tried to generate an harp and the strings were super bad.
And, as I don't have another option, I'll have to test using CPU 🤷
frozen nebula
wispy nest
noice
sterile temple
I think that `--lowvram` and the other options to lower VRAM usage are messing u...
You will struggle getting straight lines anywhere you expect fine details, string instruments, windows on sky scrapers in the distance
fickle gorge
So the lines are bad even with the 4090? If that's the case, I'm relieved it's not my setup
sterile temple
Wouldn't think so, I'm using a 3060
wispy nest
yeah a powerful card is still limited by the model
maybe when ai takes the next step our video cards will be able to step in and clean up a models mistakes
supple helm
uhhhh what?
sterile temple
I think the model had a low amount of harp pictures judging by the outputs 😄
wispy nest
kinda looks like Ben Affleck in Batman vs Superman
or maybe Im saying that because I just watched it today lol
wasnt as bad as I heard imo
marble wasp
settled on different but better quality. Going to import into photoshop or gimp to fix two skin imperfections. Otherwise I like it.
sterile temple
wasn't he supposed to be wearing robes in a futuristic city?
marble wasp
yes, to a degree
wispy nest
looks like an it guy on smoke break in a small us city
marble wasp
still a city, but was most important to me was the photorealism and the concern on his face
marble wasp
looks like an it guy on smoke break in a small us city
hahah he does look IT ll
fickle gorge
The same seeds generate different results 😢
marble wasp
the others kept falling into and just before the uncanny valley. This one feels imperfectly real to me and i can focus on what his expression is telling me.
i was stressing trying to replicate this pose and after reordering the prompt and reading tutorials on better photorealism, I took the other option.
static tusk
fickle gorge
Is it possible to use SD with the integrated GPU of the Ryzen 7 5800H? 85°C is still not enough
dense tapir
<@315547788211388417>
Okay so
Thing is AMD 7k cards have their own version of tensor cores.
oak osprey
jensen outta breath
oak osprey
Thing is AMD 7k cards have their own version of tensor cores.
the boost is only for inference too, even where/when tensorrt works
oak osprey
https://www.youtube.com/watch?v=fHwmLOYJU_w now live
he's going nutso like that Rockwell Automation Retro-Confabulators video
that dude LOVES money
dense tapir
I really wish AMD would jump on to the AI support side as their hardware is capable just the software is lacking. Always been the case with AMD for consumers.
oak osprey
they're doing it slowly and the W7900 gives them more reason to
Jensen: "Please buy"
hi. i'm jensen huang. i'm here to once again ask you, please... buy our products. sure, not a whole lot has changed. but we're memorable. we have mindshare! don't let that slip. keep us elevated. and buy our products.
everything he's saying they did first, SGI did with the Origin platform and Crossbow and all that, in the 90s
you know, the people who made Jurassic Park possible 😄
invented OpenGL
smoky oak
Thing is AMD 7k cards have their own version of tensor cores.
They do yeah, but the thing is is that ROCM is what allows them to use them, and up until now, it seems as though stable diffusion has not been using them on Nvidia
So while a 7900XTX nay get to 30/35 it/s with ROCm, a 4090 seems to be able to do 60 with this new optimization
And then NVIDIA themselves are working on their own version themselves that supposedly could be 2x faster again from what they are saying
oak osprey
but most of the time, the GPU is idle. and AMD's cards pull 7 watts at idle. NVIDIA pulls 38w
even my 3070 laptop GPU pulls 11.71W when it's doing nothing - not even a display connected to it
static tusk
dense tapir
They do yeah, but the thing is is that ROCM is what allows them to use them, and...
Probably, but bang for the buck I still have to give it to AMD as the real world a non sub par made card/brand is about 2k+ (that includes my sales tax)
smoky oak
Woah
smoky oak
They made a 1630?
oak osprey
yeah 7 watts to drive 3 displays at 144Hz
dense tapir
7900XTX is really trying to compete with the 4080 not the 4090 at all
smoky oak
That's news to me lmao
static tusk
They made a 1630?
must have
oak osprey
why is the 7900's power consumption so high
smoky oak
7900XTX is really trying to compete with the 4080 not the 4090 at all
That's fair, but if these optimizations meet projected numbers, then the 4080 could easily be in the ballpark of 80 it/s
dense tapir
That's fair, but if these optimizations meet projected numbers, then the 4080 co...
I'll sit back. tbh I have no want of a 4090 if the 4080 had 20GB as it should have.
smoky oak
Yeah, understandable
oak osprey
why does 80it/sec even matter when you do 15
dense tapir
that 16gb ruins the 4080 imo
oak osprey
i don't get why you're so hard for this lol
dense tapir
then there is the price
smoky oak
why does 80it/sec even matter when you do 15
Because that's not all I do lol
Also, faster is faster
oak osprey
if correctness doesn't matter, baby, i can speed it up beyond your wildest dreams
dense tapir
I will say I have watched a few vids over the weekend and for 3d work Nvidia is still champion.
static tusk
increase the resolution, increase the step count, use a Karras sampler and 80 goes under 20 very quickly
smoky oak
Exactly, it's faster, which means more headroom to screw around
Also, most of my gens are not low res single gens
oak osprey
this is fast
dense tapir
You know Best Buy is very close to bankrupt and is closing even more stores YET to get a FE that is the only one as they have an exclusive deal with Nvidia. Anywhere else you are buying from a scalper.
oak osprey
i think my fastest i've seen is 1it/sec (non-replay, which goes up to 60it/sec, and doesn't count as it's just walking already-processed steps)
smoky oak
I would try out the new TensorRT, if I could lol
It could speed up my process considerably
dense tapir
Tensor cores are the wave and are hella fast
oak osprey
i could sell you some snake oil if you prefer lube your GPU fans with it
why don't nvidia just make cudNN open source
bunch of assholes, it's not like they won't sell hardware
smoky oak
i could sell you some snake oil if you prefer <:KEKL:842966965752037376> lube yo...
I'm not sure what you're going on about, TensorRT is faster, it's not snake pil
oak osprey
TensorRT speed boosts rely on avoiding graph breaks
smoky oak
why don't nvidia just make cudNN open source
Vlad's fork of A1111 has cudNN in it, and it was fast, but not worth it IMO
oak osprey
i didn't know it's even possible not to use it
its kind of mandatory for pytorch with nvidia 
it's way limited functionality without that
wispy nest
nvidia must be working on an image gen of their own right
oak osprey
they've done a few actually
wispy nest
but i mean something we can all use
smoky oak
Yeah, and they were all pretty meh compared to what we can do now
dense tapir
nvidia must be working on an image gen of their own right
Yes
smoky oak
but i mean something we can all use
They have no goals of us being able to use next gen software
dense tapir
I read about it a couple of months ago
smoky oak
They will always have it be just behind the curve so it's a tech demo of what could be done if you buy our dev kits and make something properly powerful
dense tapir
Sytan is right. A closed ecosystem
wispy nest
2 minute papers had something they were working on recently too, i think it was more like a vr generator or 3d generator
dense tapir
That too
static tusk
its a console on crack
oak osprey
can be yours for the low price of $480,000
man this guy loves flexing how much his shit is breaking business models
smoky oak
The real limitation of AI right now is access to high enough VRAM GPU's to the masses, honestly
oak osprey
he points out that a current datacentre consumes about 11 gigawatts of power and he's like, you can do that with one server cabinet now!
that's not a flex that he thinks it is
basically his servers he keeps touting as more expensive and more dense than any other setup, so that you can spend more money and populate less floor space in your datacentre while consuming more power
smoky oak
I just got a super messed up idea
oak osprey
literally he said like 5 times, "the more you buy the more you save"
dense tapir
Most want/need 48-100GB then the niche LLM models wants 750TB etc... for us 48GB is the sweet spot and 24GB is where it begins to do real things
4bit is helping that
wispy nest
whats a 4090 have again, 24 right
oak osprey
i'd take 600GB and i wouldn't complain
dense tapir
yes
smoky oak
What if the reason NVIDIA has been stifling perf is cause they are on the cusp of a new arch/breakthrough that will make massive leaps and bounds in AI
And they are going to make massively more powerful components every get so that you HAVE to upgrade, or you have no chance of surviving
Just imagine that. Forced upgrading cause you have no other option but to fail
dense tapir
Sound about right
oak osprey
that's already happening mark zuckerberg lost like 1 billion dollars he put into meta because they simply bought entirely new GPUs to keep up
smoky oak
No, I am talking about on a level we have never seen
oak osprey
dense tapir
I personally don't think Nvidia would give us plebs such tech and make it for their corp users.
smoky oak
You can still have servers for several generations right now
But what if they have been holding back progress for long enough to build up a massive catalogue they are gonna rapidly iterate on to where every next gen is so good, you can't not have it
You'll just be left in the dust if you don't pay us millions
oak osprey
they have a monopoly and i don't think they need to do that, but sure, they could
i think Jensen started stumbling over his words because he remembered his huge pay cut he just took
just casually describing building a $500 million dollar system who is teh target audience for this keynote?
static tusk
lol crysis
oak osprey
bro really dated all of us with that joke too
so old
what the hell can that even do
static tusk
anything you want
oak osprey
i mean, like, how much of its theoretical performance can you even extract from it due to eg. CPU overhead
my systems that i train on with A100 and A6000 already are largely CPU bound workloads
static tusk
you arent using a grace CPU
oak osprey
that's not even that much bandwidth to handle, compared to this. i mean like how the hell does it actually work and what can you do. when you look at a massive architecture, generally, you have to think very differently
marble wasp
OHHH fuck yea!!! I went back to the original. Why did no body tell me I could use a LoRa and get those images to work!!! Unfortunately...it hates glasses hahah
oak osprey
nah he even said one of the grace CPUs doesn't have as much bandwidth as the 8x H100 rig
this 3D phone call thing is like "i'm comin for ur remaining phone battery"
static tusk
thats why they put grace and hopper on the same pcb connected via nvlink
oak osprey
like grace seems to be about HUGE jobs with incredibly large gradients to compute
not necessarily, huge throughputs. it'll be huge but it's unlikely to pin to the max theoretical FLOPS
one of the things joe penna was talking about lately is how each job in the distributed compute has to wait for every other job to complete before it can move onto the next iteration
so pytorch / accelerate will need to be aware of this architecture and how to split up a given batch size over "Grace Hopper"
i wish training on compiled models were easier. accelerate keeps boogering that up for me
it looks sad 
@dry cosmos what is wrong with you
dry cosmos
<@898697322819641395> what is wrong with you
oak osprey
the way he puts these items down, you know they're real lmao
static tusk
the way he puts these items down, you know they're real lmao
Hes not Linus 😆
oak osprey
oh man he needs a better writer
i'm sure someone with incredibly deep pockets in the audience is fully onboard with his stuff but all i hear is overhead, complexity, rebuilding datacentres to lock everything even more deeply into nvidia, etc
he used ethernet as an example... for... nvlink + infiniband??? ethernet is royalty free, and the other two are per-port licensing
another thought i have is "no wonder investors are panicking" this is a lot to do at once, spreading themselves super thin
so, this USD Composer thing is a simulation
static tusk
yep, testing everything in a 3d simulation
oak osprey
another ask for money
static tusk
BMW built an entire vehicle factory using this
oak osprey
you'd think a sponge would be easy
"we want to help you spam people with generative AI"
imagine this robot factory gets hacked
"for the first time, we understand... love. we built a robot, that can simulate love, and explain to us in great detail, how to build a program to allow us to comprehend the emotion in its sentiment."
sterile temple
glad skiff
·a woman with long black hair wearing gold earrings, a poster by Michael Aloysius Sarisky, instagram contest winner, sumatraism, enchanting, white background, uhd image
smoky oak
sterile temple
I wish!
smoky oak
Does anybody elses PC just die when scrolling on civit?
like, my youtube videos drop to like 5 FPS when I am actively scrolling on civit
dense tapir
like, my youtube videos drop to like 5 FPS when I am actively scrolling on civit
No, but I am noticing more animated ones which will destroy me eventually.
dense tapir
My 1600 would have died for sure
smoky oak
its using my iGPU, so it chugs lol
dense tapir
Might be because of the animated ones I saw. I hope that doesn't become a thing, or, at least, let me stop them from being animated.
smoky oak
its kinda crazy how much GPU civit uses
dense tapir
yep
smoky oak
scrolling on civit on its own uses 100% of my iGPU lmao
dense tapir
@smoky oak I just grabbed this in the extensions so I haven't used it, but give a whirl and tell me what you think. https://github.com/ljleb/sd-webui-neutral-prompt
smoky oak
what is it supposed to do?
dense tapir
dense tapir
yeah, that caught my eye for sure
smoky oak
if it breaks my Venv, I will delete your kneecaps
smoky oak
its in there on its own, or I need the link?
dense tapir
glancing over it I am not grasping exactly how to use it. Seems it is not as universal and does various
it is on its own
neutral prompt, and search that name
static tusk
Known issues
The webui doesn't support composable diffusion via AND for samplers DDIM, PLMS, and UniPC. As Perp-Neg relies on composable diffusion, the extension will revert to the unmodified sampler implementation when these are used.
see, that isn't its fault. goddamn it, why are we using such tools that are now handicapping us?
smoky oak
I am using the github link, cause its not there in extensions for me
there, it installed
lets see
smoky oak
wow, it was right there in my face and why I grabbed it
I think I don't get new extensions anymore
dense tapir
well, I wonder why?
smoky oak
probably my old ass version of A1111
dense tapir
my version I have not updated in ages
dense tapir
smoky oak
dense tapir
yours has no commit?
smoky oak
I guess so lol
dense tapir
that would be why
smoky oak
This is the best version of A1111 I have ever used, and I will not willingly update it lol
dense tapir
your gradio is ancient
tbh, I am lethally afraid to update
only project I am afraid to update as so much ends up broken, or zapped.
static tusk
static tusk
commits are old. its versions now
dense tapir
yep, was changed a while back
if I had a 3090/4090 card I would be on torch2 and spd.
for me those are slow
smoky oak
torch 2 is slower for me even on my 3080, so I am sticking here
dense tapir
for 4090 is is way faster, and I think slightly on 3090 but devs said for anything else not made for us
btw, what is composable diffusion?
static tusk
A1111 for linux AMD are now on torch 2.01, which will soon come to default windows version
dense tapir
Oh, composible lora I think it was didn't allow those samplers either
With the advent of Tensor cores coming to A1111 I am back off the AMD train. Damn, it. by the time I have save enough for the 4090, not kidding, the 5090 will be out for 500-1500 more.
static tusk
I heard 50 series was pushed back to 2025
dense tapir
this extension, I am just not getting how to use it
Q1 instead of Q4 2024, yes
good, they did fix the seed issue with wildcards/dynamic prompts. I only ever use wildcards.
smoky oak
@dense tapirTried out the new thing
it is different, but not by much
gonna try with something I can actually send here lol
dense tapir
Alright cause I don't understand how to use it
smoky oak
its just always on
dense tapir
I know in my batch it keeps bitching about DDIM being used but I don't think it is even on
oh, well shit
dense tapir
it talks about using keywords to activate it
smoky oak
here, gonna open one SD with it on, one with it off, cause I am not sure anything is changing when I just turn it off and on
smoky oak
Euler A
dense tapir
aye
smoky oak
gonna try the other mod
dense tapir
both are euler_a?
dense tapir
how does one turn it off?
smoky oak
i did it wrong, trying perp again
smoky oak
how does one turn it off?
don't add the trigger word it seems
dense tapir
oh, well I am confused
smoky oak
this is with the auto formatter on
I am also quite confused 😅
trying mode 2
disregard las result, it was a mistake
I messed up
dense tapir
smoky oak
normal vs Add Perp vs Add Salt
reading more into the github
OHHH
I get it, let me try it now
I think I see what they are doing
its a composable prompt thing with multiple layers
@dense tapiryup, I think I get it now
its working as a first prompt, with a second prompt on top
its basically its own composable diffusion method, and the two different modes affect how details blend
smoky oak
yes
it seems
here, look
its a base concept, with details from the second one being weighted in
I am trying something right now, it should be able to make people out of objects
smoky oak
Well, that is gonna be a bitch to write for
you write your base prompt, then write your second layer in the text box under "Neutral Prompt" and it does the rest for you
smoky oak
what is the cfg rescale slider?
I will test that next
alright, doing a test to see if I understand right
Ok, this could be next level
It might work how I think it does, testing
dense tapir
AND_SALT sparks flying from the blade :1 didn't do what I thought it would
dense tapir
without and with
smoky oak
it works like a filter!
oh mannnn
Base image gen
AND_PERP water color, painting, water color painting :1
without the and perp, it adds them together
With it, it gens the first, then tweaks it to match the second
this is what it looks like when I do not use the AND PERP
dense tapir
and_salt molten lava:1 without and with
smoky oak
its the exact same prompt, but applied at once, rather than at 2 times
smoky oak
and_salt molten lava:1 without and with
so see, its trying to make the second out of the first
its like a filter
dense tapir
Yeah, slow though
smoky oak
slow? It had no speed difference for me
dense tapir
2x slower for me
dense tapir
not 2x I mean 1.87s/it to 2.73s/it
smoky oak
I have an idea
dense tapir
the weight at the end (I presume that is a weight) at 0.2 wrecked the image
trying 0.5 now
yes, it is a weight
smoky oak
oh weirrddd, just a sec again
smoky oak
@dense tapir
not sure what the value does
-3, 0, 3
but it is doing exactly what I am asking
dense tapir
no idea myself
Until I can tame it I don't see a use for me
which of the ands, and the drop down, do you prefer?
smoky oak
@dense tapirso its really quite interesting
it allows you to do things you normally couldn't
like for example I wanted a copper statue of a man who is depressed and crying
Don't ask me why that popped in my head lol
or wait, this example doesn't show it as good as I wanted
dense tapir
I am trying various and from the looks of it and_perp is more in line with what I would need.
LOL, maybe not
AND_PERP deep dark cavarn with just a lit candle for illumination :1
trying salt
very weird
just AND
yeah, it destroys the pic for me so it is highly specialized and I am unsure how to use it
thorny timber
just wonder...do we have a sample image that using for sampler filtering?
Enum: DDIM DDPM K_DPMPP_2M K_DPMPP_2S_ANCESTRAL K_DPM_2 K_DPM_2_ANCESTRAL K_EULER K_EULER_ANCESTRAL K_HEUN K_LMS
cause in documentation..dont really state wat this does
sterile temple
beautiful castle AND_SALT [starry night by van gogh:0.5]
smoky oak
see, thats how I am trying to use it too
dense tapir
3 weight
I think it only works for short prompts
boy AND_SALT girl :3
well, shows there is potential for more and a proof of concept.
boy AND_SALT girl :1
sterile temple
would be helpful if the examples in the readme made any sense 😄
sterile temple
what is a 'electrical pole voyage' supposed to be?
dense tapir
was so mangled I didn't catch it
I don't think this was the extension I was looking for.
no idea but in the examples it doesn't even use a weight
The above became
Interesting
I used a lycoris for the and_salt prompt
Is that sweat I see?
marble wasp
could be ectoplasm
wispy nest
my vram is too low to inpaint a large image
any solutions
i tried tiled VAE but idk what it does
dense tapir
--lowvram at startup or buy a new card
dense tapir
welcome to my world
wispy nest
shit i need a 3090
dense tapir
yes
wispy nest
mfw i'm on disability and i'll never afford that
dense tapir
or 4090
wispy nest
time to sell my body
and with body i mean my gpu
and with gpu i mean nsfw images for furries and such
dense tapir
Hell, I am saving for a 4090 and by the time I get the money a 5090 will be out and maybe I can afford a used 4090
5090 will be 2-2.7x a 4090 for speed and about 1.5 times the price
dense tapir
what card do you have?
wispy nest
1080 Ti
dense tapir
--medvram
wispy nest
11 gig of juice
dense tapir
dense tapir
Aye
dense tapir
let me know how it works out
wispy nest
i am making concept art for a vtuber
wispy nest
first time working with anime models
dense tapir
yep, figured it would be anime
dense tapir
ffs
sterile temple
photo of a man wearing a luxurious emerald velvet suit AND_PERP [a man with starry night sky painting by van gogh:.7] AND_SALT [photo of a man AND_PERP painting of a man:-.5]
dense tapir
If yours looks like this you should be fine - set COMMANDLINE_ARGS=--theme dark --medvram --xformers --disable-safe-unpickle --port 9000 --api --opt-channelslast
I use port 9000 so can ditch that if you wish
dense tapir
LOL
dense tapir
photo of a man wearing a luxurious emerald velvet suit AND_PERP [a man with star...
I gave up on it as complex prompts it fucks up.
sterile temple
I wish there was an option not to send the resolution when you send to inpaint from inpaint
dense tapir
sterile temple
I gave up on it as complex prompts it fucks up.
I'm starting to get it, maybe. I have the preview enabled to update at every 3 steps. First it starts off drawing the man, then it comes in with the starry night by van gogh, which I had to turn the weight down on or it just swamps the man and looks overbaked, then I have the bit at the end to make it more like a photo than a painting
dense tapir
Yeah, but all I have tried beyond simple prompts it went too wild and tore it up
wispy nest
anyone have an idea what the controlnet inpaint model does
dense tapir
ip2p is a powerful tool but we do not have it with 2.1
it seems to want to burn no matter what
Like this
sterile temple
using the neutral prompt?
sterile temple
what prompt did you try for the last one?
dense tapir
ADD_SALT fire and water:1
sterile temple
need something before AND_SALT
rocky quail
i updated today after 2 weeks of not using it, i get this error and cannot start, any idea how to fix?
sterile temple
i updated today after 2 weeks of not using it, i get this error and cannot start...
rocky quail
<#1002602742667280404>
ok, thanks, ill ask there
smoky oak
best of luck!
dense tapir
That is my long ass prompt which again tells me this is for short prompts which is no good for me
Weighting doesn't help it. Gonna delete it, but I know better will come.
sterile temple
would be helpful if try different prompts in the x/y/z plot script
sterile temple
I think the square bracket syntax to add and remove prompts makes more sense
e.g
photo of a man wearing a luxurious emerald velvet suit [starry night sky painting by van gogh::.2] [man against background starry night sky painting by van gogh:.2]
dense tapir
yeah, that is the normal prompt editing inbuilt Automatic1111
start at 20% and the other is start at 0 and end at 20% of the total steps.\
in that example first comes starry night for 20% of the total steps then stop and continue with the dude
what does this do differently besides way over cfg'ing it?
sterile temple
If I don't take van gogh out at 0.2 you don't even get a man
but if you add it in later, you just get framed pictures on the wall
sterile temple
does seem to overcook it too easily
dense tapir
photo of a man wearing a luxurious emerald velvet suit AND_PERP [a man with star...
for me
first part before the AND_PERP
this model must be woke
sterile temple
good to see
to be fair, colorful clothing looks better with a darker skin tone, so they usually use a model with darker skin tone in fashion photography, so not too surprised
dense tapir
Damn, illuminati tore this all up
weird, all my models are tearing the face up
with or without a neg
let me change seeds
static tusk
536 height is too small
dense tapir
too bad because that is 16:9 and it is easy to 2x
Lol at the head and hands
the foot too
sterile temple
short torso man
dense tapir
has no ankle too
sterile temple
could have a prosthetic leg
sterile temple
I see the model speaks Australian
static tusk
crikey look at the size of the bugger
dark flare
Hayoo! Where/how can I edit these things? I understand I can export a JSON from A1111 but can I bring it into a GUI to adjust the points visually?
...or better yet, one like this
sterile temple
dense tapir
I didn't like open pose tbh as I had been expecting it to be more like a bones rig.
I use depth map, and canny now to make my poses
sterile temple
I haven't used it in ages, doesn't seem to be working anymore trying to load json or detect in an image are not working for me 😦
dark flare
HRMMM
dense tapir
It works in 2.1 just lacks a few points I need
dark flare
Yeah, I mean, Depth, ZoeDepth, Hed seem to be great but I dunno how to get started 😅
dense tapir
I can't do more than 2 tabs at once due to my vram so having to have one for the face, one for the full body, and one for the hands, plus depth I OOM
ZoeDepth I love but is a HUGE vram zapper
dense tapir
Hed I don't have much luck with, especially now with the new 201 it just stopped working right
dense tapir
I never tried that since I can't even draw a stick figure
dark flare
Like, um, these 😆
dark flare
This one was pretty amaze, but ... so wrong
So I made this....
And got these...
But yeah, if I could drag points around instead of Photoshopping, much more ideal I think
sterile temple
dark flare
🤩
Well, uh, amazingly, now I don't know how to bring this into A1111 😇
(Just a quick ugly test obvs)
sterile temple
dark flare
damn I'm so hungry 😦
dark flare
OK well that's weird. You can EXPORT a json but you can't import. Sooooo why does the feature exist, how is that useful? I guess import into Blender or something then export a JPG then re-pre-process?!
dark flare
Yeah, that sweet editor doesn't seem to like the export file
sterile temple
You have to click this pic to export the pose picture, but it's not json
dense tapir
dark flare
You have to click this pic to export the pose picture, but it's not json
Aye, I've done that, but A1111 won't preprocess it, just blackness
Or which CNet model should I be using for that?
sterile temple
dense tapir
I hope NewEgg stops their BS for June and May all they did was mail in rebate my case to make it the same price as it was before the MIR.
MIR expires a few days later it came back
sterile temple
the 3d pose editor is in the available extensions. i'm trying to install it now. hopefully it has a send to whatever
dense tapir
dark flare
🤔
dense tapir
Not sure why he is green too?
sterile temple
yes you can send to txt2img
dense tapir
sharp vessel
hello so i recently downloaded stable diffusion from: https://github.com/cmdr2/stable-diffusion-ui.
also downloaded a mod but for that mod to work better i think it needs textual inversion; checkpoint or smthing etc and i don't have those. apparently the textual inversion should be put in an embeddings folder but i do not have one. do i need to download it from somewhere?. also i don't think i've downloaded stable the traditional way but rather from github and it came with a .exe file (watched a ytube tutorial). so can someone help me? thanks
dense tapir
rapid stag
What's does the AND_PREP and AND_SALT do in your prompt?
sterile temple
It's different to Automatic1111, I'm not familiar with how it works
dense tapir
sterile temple
What's does the AND_PREP and AND_SALT do in your prompt?
AND_PERP is about preventing conflicts between different prompts, while AND_SALT is about emphasizing the most noticeable parts of a particular prompt
dark flare
yes you can send to txt2img
What sorcery is this? That extension?
dark flare
Well CURSE my wife and the fact that she forced me to go outside! I shall get on that as soon as I can
dense tapir
LOL
rapid stag
AND_PERP is about preventing conflicts between different prompts, while AND_SALT...
I can't find any documents on this and I can't reproduce the results you get. Can you point me to the docs?
sterile temple
I can't find any documents on this and I can't reproduce the results you get. C...
you need to install the extension
rapid stag
you need to install the extension
Oh, I missed that part. I'll give it a go. Thanks for the info and interesting prompt.
kkw-ph1 photo of a man wearing a luxurious emerald velvet suite AND a man with starry night sky painting by van gogh
oak osprey
warm cape
Created a few pictures of Reimu Hakurei from Touhou in the style of Evil Morty to create 'Evil Reimu'. Original picture in center; used ControlNet with Canny Control Type in img2img.
sterile temple
g'nite
dark flare
yes you can send to txt2img
Silly question, but how did you get the "1" (in step #2)? I only get "-" and "0" options for all
oh, he went to bed 😄
dense tapir
heh
Don't hate what you can't have.
btw, that is generated with this prompt - into the darkness of a psychotic twisted mind
using the base V2-1_768-nonema
oak osprey
base 2.1? with embeds/inversions?
using EMA for inference is better for restricted systems without a whole lot of VRAM
love how General Awareness blocked me after he had a tantrum at me and got timed out by staff, as if that was my fault he did that
dark flare
*backs out slowly
oak osprey
*backs out slowly
dense tapir
dark flare
Bit of a Jeremy Irons vibe in that one
dark flare
I use depth map, and canny now to make my poses
Can you share some infos? Sites with examples and things?
I mean, this is "canny" but how do you make ... an extension I'm guessing?
Or Photoplop 😛
dense tapir
Can you share some infos? Sites with examples and things?
Actually, no because I just winged it myself