#🏞|general-with-images
1 messages · Page 24 of 1
1st the trains...now the power plant
things are getting old and need replacements, or they are all gonna line up to blow up
In seriousness, that plant operation company will be sued
The state residents will mount a massive lawsuit
They should not only be sued but all higher ups in prison as I thought it was illegal to keep those incidents to themselves?
It probably is
I mean ffs you could have a spill like that and 30 million dead just melt
also the nuclear authorities...they have a system of incidents seriousness, this has to be up there.
how they kept this under the table since November is beyond me
So much corruption dating back to the 90s
truthfully, it was probably a very minor amount of radiation, and also 400,00 gallons is like nothing in terms of consumption
But yes, it is still BULLSHIT
Drink bottled water for the foreseeable future
Naw, no dick licking here it was a radioactive leak
awww, my favorite pass time lol
figured
the news says something along the lines of tritium safety level in everyday water 20k, current reading in the millions. I don't know what that means, but sounds bad.
I am not trying to downplay the severity of this at all BTW
Uh
Well, radioactive materials is on a whole other level than say paint, or oil spills
yes, cuz it does not go away easily, and messes up biostructure down to cell level.
yes
It still makes me wonder how severe it really was on an impact level
half life and all that so best to just come clean and start the neutralization process
who know how far away that water has travelled to since November down the Mississippi
I don't know, just got the news an hour ago. They will surely find a scientist to come speak tomorrow morning
yes, which is where we get our drinking water. I don't drink it, btw, but to bathe/shower in it and if it were radioactive
I feel like this chat got really extremist all of a sudden. Or maybe it was like that before and I didn't realize
I am still stuck on this
I don't know how, but somehow youtube recommended me that news...maybe they picked up words from our chat here
linked accounts and all
I am assuming that was supposed to be a derogatory or offensive remark of some kind
I wasn't referring to you, if that makes you feel better
You're chill
and yeah, that is some BULLSHIT
Well, I say that due to yesterday and the hot hunks and you said more of that. Now day before yesterday.
yeah, ok
😛
Now I go look up on what ordinary people can do to purify that kind of water. Not an ordinary water purifier can treat that for bath water.
Excuse me for not automatically ruling somebody as judgy as you out for being hateful lol
It was like 5 months ago. By now all damage would be done, and it would have been completely diluted
for future cases.
ah, fair point
Oh, reverse osmosis works
RODI would be the best bet
Luckily, I have an RODI filter for my aquariums lol
but RODI water isn't safe to drink pure
plus, we don't know if some of that water has gone up into the clouds and dumped down as snow all over north america
Of course it has
maybe the chinese spy balloon was coming to check this weather out
LMFAO
Ok, that was good haha
they were just peepin the weather lol
they have a shit load of nuclear plants near coastal cities...they are relatively new but one day...
and most of their population are in those coastal cities
so that wanna see that nuclear rain effect up close
nuclear fission is a great tech, assuming people give a shit about what they are doing. Honestly, fission waste and power management is stupid easy. It takes a special kind of lazy/stupid to mess it up
you have to be needlessly reckless for something like a leak of that magnitude to occur
Just like how Chernobyl was a shit ton of bad practices and fuck ups in a row
poorly made reactor that isn't serviced well that has faulty fail safe systems being run by an untrained staff. Who could have seen that coming
As long as there are people in charge, there is always a chance that something needlessly reckless or stupid occuring.
Yeah, thats fair
the thing that interests me is how the fallout of Chernobyl hasn't gone anything like it was supposed to
or was predicted to
the mutations "look" minor. like the trees, animals and insects have defects, but we have not caught 3-eyed fish yet
some weird catfish but not 3 eyed
I was referring to the completely fine dogs that were found to have no increased radiation in their systems, regardless of being born and raised in that region for generations
so safe in fact that they were rescued and put up for adoption across the world after extensive research was conducted on them
It almost suggests that elongated radiation affects are now what we previously thought. Not saying its not dangerous, but that shouldn't be possible at all based off of what we know about nuclear fallout and the half life of the particulates that were released into that region
They found that the radiation tied to the dogs was strictly on their fur, and not coming from them. Wash them and they are just dogs
which is not in line with the damages done to various other species
the red forest
I submitted a ticket to the author of that extension in hopes they can help me get the contrast back. Apparently it works perfect in 1.5
not even contrast loras will bring it back if that extension is on.
its likely a last step sort of filter that results in the negation of LoRA and model effects
Well, it works with loras but possibly not with contrast loras which 2.x must have
interesting
2.x can't get darker pics otherwise
I was the first one to find this anomaly in 2.0 and emad said he was looking into it, or his people, but honestly it is a clip issue and a diffusion issue
diffusion the loras semi fixed but the clip is the part where I prompt "a dark field in Minnesota with a lone dimly lit cabin in the distance" and it is almost as bright as day and thre a BIG lightbulb on the ground at my feet.
Beta Princess Mononoke Scenery LORA out https://civitai.com/models/20942/princess-mononoke-scenery-background-style-ghibli
so many people tried and failed to finally admit forget it
even the white knights for SD came charging in to defend it to finally admit I was right. That was sad and funny at the same time though I am used to it nowadays. The fix, full fix, must come from SAI.
I am 100% unsure what SD did here.
I am still trying to figure out why ultimate SD upscale is so bad with landscapes
that one looks way better for a standard generation for sure
Yeah, I had to ditch the extension or no go
I always want a full dynamic range image, so I can edit it as I want
Hard to get that in SD
Gimmie a section that has pure black, and a section with pure white, and I can adjust as I want
yeesh lol
hehe
Now we're talking
That is the image I would like out of SD
that is post processed in PS
man, you are just crunching the hell out of these images lol
shoos
Why can't I get the second image out of SD?
As if we are missing something, a final piece of the puzzle
I FINALLY FIGURED IT OUT WOOOO
oh my goodddd
I finally have really good ultimate upscale with landscapes!
Yessss
Oh my god yesss
the details are really good !
gigapixel ?
this was generated at this res
damn
I am gonna go for 3x cause fuck it lmao
Tiled VAE extension ?
That is a 5120x2175 res generation out of SD
😩
trying 3x rather than 2x
this is gonna take a bit lol
I saw that post talking about a similar feat yesterday : https://www.reddit.com/r/StableDiffusion/comments/11tuzov/for_anyone_who_isnt_already_aware_of_it_tiled_vae/
oh wow, that one looks pretty great too :>
comments seems to not all have as good a result though
looks like this upscale is gonna take about 9 minutes
some have really visible seams
I am using ultimate SD upscale, which also uses tiles and then blends them together. But this one has a built in seam fixer
I have had amazing results with animals, but now I know how to go way past that
damn I didn't post my morning cute cat, that reminds me
I am confident I could do WAY better now
so your blurriness was just a matter of finding the right denoise?
it had nothing to do with denoise actually
It was a lack of steps
the math on how steps work wasn't at all how I thought
yeah steps and denoise. thats what you need to figure out
.075 denoise at 50 steps is only 3 steps in img to img
denoise multiplies the quantity of steps (unless you use the option to force the steps). yeah it can be surprising
I only noticed that most of my gens were doing 0 steps after I looked at command prompt
there is an option thought, to always do the quantity of steps you asked whatever the denoise
so I would get the exact same output, but at a higher res (so no new detail, just softer aliasing)
now instead of the 0-5 steps it was doing, its doing ~20 steps per tile
so
its a lot of time, but this isn't synthesized detail, its generated detail. Its not just looking at pixels and making them up, its regenerating each one
got me hyped over here
4k wallpapers are so easy to do now, OMG
1/ you generate a normal size picture
2/ it's upscaled, split into tiles, each is done in img2img (highres fix)
3/ more tiles are generated on the seams to make it fixed better, still the same prompt for all tiles
4/ stiched together
that's about it ?
yeah, though I am not even using seam fix as the denoise isn't high enough to need it
I'm trying to break it down to reproduce in comfy
oh, and it also adds a pixel sharpener/upscaler at the end
to get better pixel per pixel detail, cause per pixel detail is a little fuzzy in SD
please let me know if you get it in comfy. I would love to try
The problem is one group of settings doesnt work for every gen, I found. You still need to play with it
which sucks
also, now that I think about this... I could upscale photographs like this. generate new detail rather than enhancing preexisting issues/short comings
BINGO
also, really satisfying command prompt lol
photo restoration using SD is already a thing
I did restoration the old way by hand and 100 hours was not uncommon on bad photos
If I had a 4080/90 I would be genning at 50 steps DPM++ SDE Karras
the noise in the sky?
Yeah, and the smoke trail
yeah, i could fix that no problem in post
of course but if it is there on that simple I dunno
maybe it has an issue with gradients? Seen that before
It uses the same seed for every time it makes, which is annoying
I wish there was a way to get it to use a different seed for each tile
in my old much worse gens, you could really see it
like this massive failure lol
Do you know how to enable that step force thing?
it can help for sure
there are lots of useful options in there
color correction is also quite good when doing loopback
(puting your previous img2img output as input, for doing animations)
ahh, to fix with hue shift and tone shift?
exactly
problem is that it forces the color balance to stay the same too
so it can be bad for variety
yep
it's something I had added to the quicksetting (the top bar) when I did lots of animations
so, is there an easy way to remember how to add things to the top bar?
I have my VAE thing added
(just in case somebody doesn't know, you can add any setting from the Settings tab to the quicksettings, the top bar where there is the model selection dropdown. Just add the name of that option in the "Quicksettings list" here)
find all the options names here : https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/modules/shared.py#L317
how do we find those names? Thats the issue I am having
well, you go into that link, find the name of an option, let's say "outdir_samples"
oh sorry
and just add it to that list, using a "," to separate
I thought it was the screenshot link lol
sd_model_checkpoint,sd_vae,outdir_samples
then apply settings/reload UI
and it's done
yeah
all the tabs
easier way is to look for the option you want in the UI
and then CTRL+F in github to find it
yeah, makes sense
using the text displayed in the UI
oh, I also wanna turn off all jpg copies
as far as I know, they don't expose those names any other way than in the code
cause all of my img to img images save as .jpg and.png, and it just uses more space
aha
found the culprit
oh weird, our UI's look different
I can change the file size limit
ahhh
you should git pull
sorry to pester, could you remind me how to do that?
open a console in automatic folder (click in the adress bar in the good folder and type "cmd")
and type "git pull" in it
no problemo 🙂
it's how all git code update works if you don't code
it's a lot more complicated once you start to code lol
I can do a ton of shit, coding is not one of them lol
(it's also why you should NOT modify the files that aren't in the .gitignore. if you modify one, it will think you are working on the code and won't just pull anymore)
all files in the gitignore (it's just a text file you can look at) are files that are specificaly made for the user to change
sorry I got on a git lesson there
I could not get pytorch2 to work 😦
RuntimeError: CUDA error: no kernel image is available for execution on the device CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
I upgraded to 12.1 the nvcc inside auto is 12.1 but not happening
apparently pytorch may not run on my 1060
#🤝|tech-support may know better, I'm really not sure what to say on that one
No big deal no one can say but I do see 1660 and 1650 have the exact same error
Apparently Pytorch may actually not be for any cards pre Ampere until they compile a version for us.
Might also be their gains are due to using Tensor Cores which would eliminate even universites.
ewww,
12.1 has messed this all up so rolling back
Looks like that may be what they are doing so cards without Tensor Cores are screwed. All cards prior to Turing, 1660/1650 and M40/P100/P40/etc... that universities, and research facilities, use every single day.
I think only RTX works with it
Created entirely with AI. https://www.youtube.com/watch?v=Cexmsxh6xfI
The video was created entirely using AI.
"The song features excerpts from Thracian Dances Suite (1925), No.2 Rachenitsa by Petko Staynov."
Miroslav Markov - Bass
https://www.youtube.com/@miroslavmarkov1245
Elena Ivanova - Vocals
Alexander Page - Lead guitars
https://www.youtube.com/watch?v=B6EnzPVQ2Cg
Apostol Stefanov - Guitars and synths
http...

Welcome Miss Granger. You're probably wondering where you are right now. In the muggle world, we call this...the Matrix.
There is no magic here. you need to learn to fight.
What did you used, or how, to enlarge this image?
She's beginning to believe!
I think you're ready for whats out there now, Miss Granger.
my brother as Hulk lol
Темно зелёное кресло в большой комнате в современном стиле
looking for the bot ?
@ocean viper
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
#1080946152318443610 Темно зелёное кресло в большой комнате в современном стиле
I will index all the things
that's a lot of resources
Wonder cat
@sterile tangle
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
Randy from My name is earl, Jim from the Royle Family and Frank Gallagher from Shameless UK
#🏞|general-with-images make a luxury restaurant menue
sorry but we don't have a bot for that
@neat yew
Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝|tech-support for any problems you encounter while installing it locally!
Nope, I am in touch with them and it should work on all cards so we are in the "why not" phase.
welcome around here, but why are you posting clothes ?
is that made with SD? feels like a real clothing
#🏞|general-with-images message اصنع تسويق
Used deliberate model,prompt,some lora to have a bit better contrast,face enhancement by little factors of gfpgan,codeformer mix,double esrgan model pass 4x and 1x models,extra pass with gpen.Then used xnview mp to add 3 filters on top of each other to have some feel of color grading.Last step at photoshop to add some depth blur ,some extra color grading,fixing hot spots,adding some grain to bring some life and overcome the banding issue and did some retouch.I ignored the white balance for sure to have a bit of cinematic look.
raw image incase...
5187x1459
Outpainting x100?
nope, single generation
Hires. fix
resize: from 1820x512 to 5187x1459
I'm quite amazed tbh
a little duplication for sure
Very cool
Unfortunately I don't have such an extreme widescreen monitor to use it as a wallpaper 😛 I need to stick to 16:9
They really look cool @glossy herald
wooops,my 1080 ti will possibly cry in 30 min then 😄
this looks beautiful
dancing river in a secret magical forest by Nico Marlet, by Andy Park, by Jason Chan, by Jason Chan, Félicie Haymoz, Luke Pearson, Senior Character Artist, highly detailed, artstation trending, art by ilya kuvshinov and greg rutkowski and wlop, cinematic lighting, very detailed, hyper-realistic, unreal engine, octane render
Negative prompt: monochrome, (low quality, worst quality:1.2), (multiples views:1.2), wrong hands, deformed fingers, author signature, text, watermark, nsfw
Steps: 58, Sampler: UniPC, CFG scale: 15.5, Seed: 822315003, Size: 1820x512, Model hash: c4947fa2c4, Model: DB_mix_realbiter_v10, Denoising strength: 0.44, Hires upscale: 2.85, Hires steps: 30, Hires upscaler: SwinIR 4x
thanks, it has a little too many rivers
I'm trying a group of pixar monsters right now
These can be used for inpainting.Nature photos seems to work better than chars.Can also be mixed with control net to have some double exposure effect and maybe some more...
and next a sprawling city in a cute cartoon
any prompts for 1920x1080 backgrounds for a change ?
here are some
Nice ones
They all look awesome!
It is looking more and more like an Xformers issue (nothing but trouble with that for a lot of people if you try to leave its sandbox).
trying to make DJ bear
let's keep on the DJ series
DJ dog
to be fair to the hands, it's a dog
@sterile kilnThank you! This confirms that PyTorch is able to execute code on your GPU as already mentioned. Yep, Xformers is the issue not pytorch 2.
he's gone to bed I believe
party cat, last ones
and now time for bed, see you everyone
Oh man, these look straight out of midjourney. Nice work!
Very very nice color profiles
@smoky oakWell, it worked
Create ancient Muslim warriors
@smoky oak xformers is the culprit so I had to --no-half due to no fp16 on these cards so I was met with OOM. I then fell to 512x512 on one of my old 1.5 models and 2s faster on this old card than when pytorch1 and xformers. If I can get xformers working too that will be dope.
Wait, so what did you fix?
I installed pytorch2 last night and spent the day working with the pt2 devs. Xformers is why I couldn't use it not pytorch2
any 2k+ card will have no issues
1k, or less, we are all DOA
I am saying is it an issue cause xformers works different on tensor enabled cards. Which would be all 2k+ cards
20-22s was my normal gen time in 1.5 with xformers now it is 18s without
nice!
possibly
Imagine how fast you could go once pytorch 2 and xformers are working together
the problem is 6gigs isn't going to cut it for 2.x
yes
I am excited
from what the devs are saying it is the wheel being used they compiled it wrong, and they have no idea who arrow is.
all you 2k+ cards do is 3 lines in a command window and done
This harkens back when I had to compile my own in 1.5.
care to try it as I know you know cmd windows and stuff?
I am dying to see how much faster you get
#1072220168534642768 create ancient Muslim warrior
@smoky oak how to generate image's
no bots on this discord need local
yeah sure, if you can explain how to do it
this will be installing pytorch 2 for automatic 1111?
so like full program pytorch 2?
Yep
awesome, if you have info on how to do it. I can
one moment
sucks as I just installed the latest dev version of xformers and it downgraded me so I must compile it locally. Not a big deal I just need to go find out how again as it has been a minute.
let me know when you are ready
ok, i am back, sorry@dense tapir
"./venv/Scripts/activate" - with those quotes on the command line in the root folder automatic1111
yes
ok, it did it
prompt should be (venv) blah blah now
make sure auto isn't running
72MB/s
DAMN
I have learned from my mistakes in the past. I close it every time now haha
LOL
ok, its done
damn
oh, its installing others now
yeah, our internet is nasty here haha
500/500 fiber. Sometimes up to 600/600
well, that was a typo
what was a typo?-
its doing this
tell me when done
pip install --no-deps --force-reinstall https://github.com/ArrowM/xformers/releases/download/xformers-0.0.17+b6be33a.d20230315-cp310-cu118/xformers-0.0.17+b6be33a.d20230315-cp310-cp310-win_amd64.whl
done
that is the pos that does not work for sub 2k+ cards (turing)
alright
well, xformers compiled for me
ok, thats new
applying xformers cross attention optimizations
I do not
I closed it before launching
fire it up and get back venv
alrighty
back in
Versions of relevant libraries:
[pip3] numpy==1.23.3
[pip3] open-clip-torch==2.7.0
[pip3] pytorch-lightning==1.7.6
[pip3] torch==2.0.0+cu118
[pip3] torchdiffeq==0.2.3
[pip3] torchmetrics==0.11.4
[pip3] torchsde==0.2.5
[pip3] torchvision==0.15.0+cu118
[conda] No relevant packages
Yours say that?
it gave me an error-
(venv) D:\stable-diffusion-webui>python -m torch.utils.collect_env
Collecting environment information…
PyTorch version: 2.0.0+cu118
Is debug build: False
CUDA used to build PyTorch: 11.8 goes on with a ton of stuff
full error
it closed itself-
I am opening again
ok, i am back in
do I run the python thing again, or just do the nvcc?
get the venv prompt
nvcc is not recognized as a command
shit, you are missing cuda
oh fuck lmao
that seems like an oversight lol
is it cause I use studio ready drivers?
no, it is a separate download
I just downloaded new drivers last night, but forgot to apply them
ah, alrighty
not drivers you are missing the cuda toolkit
ahhh
which is cuda, lol\
do I download cuda 12?
No, I don't trust it
11.8
I tried 12.1 (the latest) and meh
now this is about to get ugly
but you gotta do it
I closed all command prompts. I assume A1111 would be a conflict
See, I am not excellent with code, but I consider myself a fast learner haha
works for me
its downloading
alright, download is done. I just run it as an exe now?
are you suggesting that I could have been running faster/better this whole time?
ok, its extracting
yes
so I am about to get a double speed boost? lol
oh, new discord themes just dropped for me
I have been trying to apply painterly style to this image and the results are not bad. But some of the details in the background are getting lost, tried less denoise but then the applied style is barely visible.
Any controlnet tricks to highlight more details of the background?
I'm fine with losing details of the character. I can isolate the character and apply style separately
What have I done...
Alright, so @dense tapir and I have tested the new pytorch 2 for SD A1111 and we have come to the conclusion that its actually a bit slower
not sure why
Yep, I am going back to 1
even slower without xformers so it is not slower due to some oddity in xformer from PT1 vs 2
Hey guys, realistic vision 1.4 is no longer exclusive to Fantasy AI. He released it to the public!
its not a massive improvement, but things look consistently better IMO
(no xformers noise, these are checkpoint differences)
1.4 seems to listen a bit better, and also creates better looking faces IMO
just been trying pytorch 2 A1111, and realistic v1.3/1.4
pytorch 2 A1111 is... disappointing
for me and @dense tapir its consistently a little slower
for me I go from 10.5it/s to like 9.9it/s
and he is hit a lot harder
ouch
yeah, sad times
photo portrait of a soft fluffy cat, hat, cape, long fur, sky, solo, elegant, ultra detailed fur, dynamic pose, beautiful garden, rising sun, insanely detailed, highest quality, fur texture, intricate details, (cinematic lighting), RAW photo, 8k uhd, dslr, hdr, gloomy,depth of field, rim lighting, two tone lighting, dimly lit, low key
Negative prompt: airbrush, photoshop, plastic doll, plastic skin, easynegative, monochrome, (low quality, worst quality:1.4), illustration, cg, 3d, render, anime
Steps: 60, Sampler: UniPC, CFG scale: 15.5, Seed: 2761464684, Size: 960x544, Model hash: c4947fa2c4, Model: DB_mix_realbiter_v10, Denoising strength: 0.44, Hires upscale: 2, Hires steps: 30, Hires upscaler: R-ESRGAN 4x+
difference is consistent with xformers for me
so like 7.8/7.2 no xformers
and 10.5/9.9 with xformers
so I am back on old pytorch, cause its faster
very disappointing ._.
why was I using 15.5 CFG on UniPC....
lmfao
dogn't
I usually use low CFG all the time
well, taking a break for now, but I'll make more later
but I do see benefits
yep
I'm loving the quality this extension lets do
what extension?
it's what I used for all those 5120x1440 I did last night
ah
its seems like a worse version of ultimate SD upscale
I can try it, but it just seems like worse ultimate upscale
I can do 8k gens on a 3060ti with it
and they look damn good
(a little artifacting in the sky)
7680x3264
I'll try it too, see if the composition ends up better than on mine, especially on 32:9
1x
2x
3x
(the 8k one^)
let me try tiled VAE
maybe they can be used together for outrageous res outputs haha
for me 1:26s became 1:30s
yeah, multidiffusion/tiled VAE just seems like way worse ultimate SD Upscale
results look nowhere near as good, and the res is limited in comparison, from my quick test
though it is wayyy faster
but I would rather wait for a way better result
2x res upscale with MultiDiffusion and Tiled Vae
2 minutes 33 seconds
DAMN AssertionError: composition via AND is not supported for DDIM/PLMS samplers
I got around to installing torch 2.0
slower for us
lowkey sucks lol
end result is I dont see any speed improvement
its slower for us
its basically the same
for me 1:26 became 1:30m
Hello, i made some videos, check it out and tell me if you like one: https://www.youtube.com/@AI_stuff
I am back on V1
the difference is small so I will leave it for now
fact is those gens above are the first time I have gen'd in 1.5 since 2.0 dropped
Not sure if frame by frame is good or maybe i have to change every 2 frames?
whut
I miss all my lovely embeddings and loras from 2.x
using deforum to make this ?
I would add some interpolation maybe, to help with the jitter
less keyframes, like 20/s, and 40 frames interpolated for 60fps fluidity
I did quite a lot of those once upon a time
Hi, so I've used the ABG Remover to generate thesr images with alpha maps...? How do I use them to remove the backgrounds? I have Photoshop
Yes
@glossy heraldJust so you know, I think you will like ultimate SD upscale. This is a comparison
here is the best one I got on those animations
I'm installing it and was about to do a comparison on same seed
Using some of my old embeddings, yes
Base image (2560x1080)
Multidiffusion 2x upscale
took 2 minutes 6 seconds
vs ultimate SD upscale, which took around 1:20
I downloaded that and deleted it
yeah, multidiffusion seems to just be worse ultimate SD upscale
yeah, and omg 15+ m
multi
ultimate
Have to experiment more, but i like the trippy frame by frame effect
these are not bad for no upscaling, 1.5 at 768x768
yeah, multidiffusion gets a no from me
let's try
more VRAM, slower, more noise, less proper output
resize: from 1800x512 to 5130x1459
its important you know how to use ultimate SD upscale
I wonder what its claim to fame is supposed to be?
no idea, its worse in every way, from my early testing
for now I clicked "enable" only, we'll see
yep
no, that doesn't work
you put bigger tiles ?
use these settings
Interesting as I am used to the new clip now
you wanna scale 2x, you wanna use that upscaler
you can, I go from 512-1024, depending on my mood lol
512 I take it because 1.5?
I'm not sure I'm using the good extension there
its ultimate SD upscale
Isn't that only for 1.5?
I thought it was its own model
it just breaks the image up into tiles, does them, and stitches them together
oh no, its a script lol
thats why I can do 8k gens no problem
its just a shit ton of 512x512 images together lol
I wanna try an extreme one lol
well, damn I only have the SD upscale
687 tiles at 768 lmfao
I will do it, just to flex on the uninformed 4090 owners
wondering why they can't even do 4k, when I can do 16k
where is this script?
extension tab
oh
load from list, scroll down and find Ultimate SD Upscale
I was looking in the scripts drop down
it goes there when downloaded
that one always sucked
its the SD Upscale, but refined
oh, cool
I have need to upscale
trying 512 instead of 768
still 697 tiles
hmm...
I am playing with fire lol
please don't OOM
its still slicing tiles lmao
OUT OF MEMORY AGAIN

ok, trying 5x lol
12800x5400
still fucking massive
I so despise kohya scripts for additional networks as I told him hows about you stop changing folders when we update that you don't even change? Yeah, tired of having to resymlink them just because. Most devs put folders (like lora) on a blacklist to prevent changing them.
Tile height 0?
aight
I brought that up because on 2.x we are forced to use that and he updates regularly
use denoise .2, and also use 50 steps
(will actually do 10 steps)
ok, something is wrong
5x failed with the exact same OOM, and its saying 697 tiles again
I'm not planning to use this upscaler , was just wondering
OHHH
well that broke my UI x)
@smoky oak Btw, you know my venv for PY2? It made my png info text dark on dark so I could not read it. I switch to PY1 and normal.
wow lmfao
breh lol
best contrast
LOL
me neither, I don't like to see my text anyway
strange that a python version would do that
yeah
I switch back to 1's venv and back to norm
I noticed I was on 2 due to that text like that
last extension I just tried broke my venv, making one again rn
oh, don't you just loooooove that?
it's... a script??? not an extension???
its an extension script
you're an extension script
well, I am trying to figure out how to work it as I don't want to prompt I just want to upscale
I would recommend 50 steps, .2 denoise
it works better if you upscale with a prompt, that way it knows what its upscaling
ahhh, crap :*(
I guess thats reasonable for a 15360x6528 image lmfao
just 100M pixels total
I bet its blurry., if it finishes
50 steps of ddim?
yeah, at .2 denoise, its only 10 steps
trust the process
I know what I am doing :p
or, you can go into settings and force it to use the amount of steps you say, then just use 10
MUCH better
yeah, should be good
I wish we could save this stuff as defaults
lets see :p
righhtttt
well, we might be able to, just a sec
I'm going back to fluffy kittens not breaking my venv
bro
21 tiles so 21 images
I might be able to do a 100MP image in under 20 minutes lmfao
this is INSANE
8GB VRAM!
its only using 6.5 right now lol
yessir
damn, my 7 step 1024x1024 tiles are 4.1s
I swear, if this works
I am gonna cease to exist
100MPX SD generation lmfao
I mean, I am upscaling a 25MPX SD generation, to be fair lol
so its 4x as much
its using 4 GM system RAM
and 6.5GB VRAM
I bet if I went to 1200x1200 tiles, this would be way faster
the slow part is actually stitching the images. So less tiles, more fast
the stitch takes longer than the tile gen
yeah, because it has the most limitations then, it's when it eats up vram
it does the same for me on higher specs
TBF, I am stitching a 100MPX image lol
I recommend using an image with at least 1.5x high res fix with ultimate upscale
The quality is unbelievable
have you not seen my 8k image? lmao
Can you explain it better to me? Keyframe is not frame cadence, right?
I am past 50% done on the 100MPX image
btw, this was a happy accident and I rather like it
keyframes is the term I use for the real frames I generate. 20/s means I have 5 seconds of video with 100 frames.
Using software like Flowframe, I then ask another AI to make more frames in the middle, to smooth things as if it was morphing from one pic to the next, and not jumping around in the details
here is an example I did, pushed to the extreme, I think it was 10 frames per second, and then interpolation up to 60 https://www.reddit.com/user/Hearthmus/comments/xstz0n/living_trees/
Anyone know how to use the new UniPC sampler?
Nope
crashed me
well, comfy is so superior to auto but lacks a few things to make it really comfy
for some people
for me its way worse
basically he needs to make a SD styled on DaVinci resolve
I gained
Ok now i understan what you mean, thank you... will try your tip
75% done
went from 50s to 35s
D/GAR TEC/IR
One day, just any day now
1.5 really gave me a bad image up there
I have no idea 1.5 is creating them
I love watching random stuff it creates
I hope in the future we get math filters for comfy UI
YES, I requested that
like where you stop the gen half way and apply a vector noise offset and just get FUCKED gens lol
I used to go super hardcore into blender procedural materials
even DaVinci resolve has that and FFT
98%!!!!
99%
HERE WE GOOOOOOOOOOOOOOOOO
waiting lol
its tsuck at 100%
IT
DID
IT
15360x6528
oh
RIP the sky lmfao
@dense tapirThe noise killed the sky lol
WORTH IT
so how does it look?
its not pixel perfect quality, cause it ran no pixel filter, but hey lol
other than the sky
just a sec
Original
2560x1080
1x
vs the 100MPX version
original
100MPX
fucking INSANE lmfao
yeah looks good
only took 18 minutes
ok, it is soft, TBF
thats cause I couldn't run a pixel filter
but I can use AI upscaling on same res to fix :p
@static tuskI fixed it using an AI blur remover
I am at a loss of words lmfao
Oh, nice Kohya added .keep files thanks to my suggestion.
419 votes and 89 comments so far on Reddit
UniPC 8 steps 4s to gen
actually 6 steps
it is based off of ddim so has the same round of steps issues
it seems
21 steps that was actually 19 steps and 15s for me
50 real steps 52 called for 41s to gen
I think I like this one
42s to walk through 20 real steps 768x768
@smoky oak Yep, unipc is ddim with a twist
Flowframes interpolation is making everything wobbly... i don't like it
step motion has more effect
Big cat 🙂
what are you using for this?
Prompting
I don't use that controlnet stuff
SD?
Yes, 2.1