#🏞|general-with-images
1 messages · Page 25 of 1
don't know and now my graphics card keeps crashing
laptop?
nope
well, then it shouldn't be from heat having killed it
no
cant play my favorite games
so the end of the year? what is 4080?
16gb second fastest card
4090 is the best and 24gb but 1600+
oh ok
really nice job guys
whats your hires. fix settings to not get any weird glitches/doubles at that res please general?
oh right - didn't know there was alternatives tbh
@smoky oak gave me the juice. Get the extension that is actually a script called ultimate SD upscaler
use img2img and set denoise to about 0.2-0.26 cfg 7, 50 steps DDIM
use esr 4x and set it to what you want
it is bloody fast
not perfect but better than most
@fresh hound
that so fucking cool
Bruce Willis as Hank from Breaking bad
what sampling method do you use btw General?
all my images end up having this weird texture, i wonder if it's the model
that's so smooth damn
those are cool but they are missing something i feel
like?
idk, i just didn't get that feeling on the first one
the glowing one
different prompt
how do I get safetensors to work in 1111?
saw online u just past it in the folder and load like a ckpt model but it wont work :C
well, maybe do a git pull if you are on a really older version
but it worked out of the box for me, putting .safetensors in the models folder showed them in the dropdown menu in the UI as usual
does anyone know why my images have this weird ass texture?
what do they call that part of a game screen that gives you information on how much life hunger and toolbar? I need it for a negative prompt
just look at this master piece
HUD?
Test
yea I found it allready but thank you
I get them every time there's a video game title in the prompts. so I needed to add that to the negative prompts
and I couldn't remember the word for them
should the preprocessor be on? I'm going to make a sketch of the picture it will look like first?
you have a picture of what for now ? a real building ?
put the preprocessor "dept" on, and click on "Preview annotator result"
this will show you what controlnet would take from that picture
for buildings, mlsd is also really good
that preview button is essential in lots of modes
it lets you see what will get extracted, what contours or depth, ...
let's take an example
I found infinite backgrounds for retro/synthwave youtube mixes xD
i did how can dowland the result?
you don't download it at that step, but you'll be able to do it soon
now that you have your depth map, if it looks alright, you select the same model, "depth"
and you prompt what you need, like "photo of a house"
here is what I got as depth map
I found depth to be better than other mask extraction methods
sometimes adding more on top of it, like canny and/or mlsd, can help a lot too
although body pose is pretty neat for humans
i should delete left picture?
my personal favorites are canny, and scribble
no, keep it
it will use the depth map it made for it anyway
yeah for more detail, but depthmap gives more room for creativity to work with
it won't reprocess it each time, don't worry
it really depends on the use case, but applied to architecture, this seems like really already enough for professional uses
well with better prompt than "a house" lol
in particular, prompting the building materials is quite strong
I wonder how depthmap is processed anyway, because others like canny are basically just binary 1 and 0, but depth looks more like 0 - 255
I'm not sure. the model is called Midas and has been around for quite some time
deforum uses it too
ooh thanks i will take a quick look
deforum is a little hard to use, but anybody interested in animations using SD should take a look at it, fantastic results
should the preprocessor be on when rendering the image?
if you are using a normal image as controlnet input : yes
if you are using a depth image : no
you can download the depth image, it appears as one of the outputs when you make a picture
and use that one instead of your initial image
in that case, remove the preprocessor
I wonder if deforum would work with body pose, IE taking a video of a person and for each frame get the pose then run SD on the pose
that would be sick
but theoretically it could be possible without deforum as body pose is all that is needed
thats real cool
I might consider developing an extension that does that if no one yet did xD
nice !
thanks!
it's mostly that deforum didn't update for controlnet yet I believe, but I haven't followed closely
I saw a couple of vids where the input was someone dancing and the output was a drawing of it, pretty sure they used on of the preprocessors to extract the mask for each frame, but part of me does not want to believe that the person did that manually frame by frame xD
no, you script/batch this
you extract all the frames using ffmpeg
then process all of those through automatic using the API currently, no script does it yet that I've seen
and you build the video back up in ffmpeg
use the same seed on each pic for more stability
some dejitter if you can afford it too
true, that would work
ffmpeg is underrated like crazy, its insane what it can do while being open source
the sun shadows and rays are on point
I'm moving from controlnet, and going into the "make 1920x1080 in one pass"
longer for sure
this makes me wonder if at some point the AI will surpass conventional ray tracing in terms of computation
well, ... if you go and train one model for everything at once, we aren't there yet. but you sure can train on specificaly that task
oof, my gpu does not have the mem needed
Time taken: 2m 8.81sTorch active/reserved: 3299/3580 MiB, Sys VRAM: 24564/24564 MiB (100.0%)
(for a batch of 3)
not good though
I did some nice ones yesterday, starting there : #🏞|general-with-images message
I mean the bounces are pretty good at this point, maybe not so many and we cant control it but its taking it into consideration
even going as high as 5120x1440
gtg for now though, but happy to have met you, I hope I'll catch up with you again !
oof, are you on an A100 or something xD?
3090TI
yes, see you ! glad to meet you as well 🙂
mine is a 4080 but cant handle it
I need better prompt for that
the 40xx series is nice on some sides but yeah.. you have 12GB on that, right ?
yep only 12
it is 😦
well, catch you later !
take care, see you
i should always use this setting?
and which control model best for creating human and animal images?
Hey Everyone I Really Very Urgently Need Help With My App.
Is There Anyone Just Anyone Who Could Help Me With Integrating Stable Diffusion Or Any API That Can Simply Take 3-5 Images Of Users And Create It's AI Avatar, Sort Of Like Lensa.
Please Anyone Just Any One Help Me With It It's Extremely Really Very Important.
Even If Anyone One Of You Know A Person Who Could Help Me Please Connect Me To Them Just Please Help Me Guys.
Why Every Word Starts With Capital Letter, It's Painful To Read
No Bro Lensa Made It Possible, Remini Followed Suite.
I Want It I Just Want It, My Business Model Is Very Different From Them.
Anyone Who Could Help Me I Am Ready To Offer 1% Of My Company To Them.
you're planning on using SD APIs? How would that scale up to > 1000 of users at the same time
what cloud infrastructure do you plan on using?
a guy with black hair, a blue sweatshirt, sits on a throne, a crown on the back of the throne, a smile on his face
how to create a drawing ?
I AM Not In That Technology Stuff, Just Want A Technological Guy Who Could Help With It.
Any API Just Anything.
Whatever It Takes To Make It.
so you want to challenge a 25 M$ business, okay, then I assume you must have at least a couple of Millions to invest in your startup
and a business proposal document
It seems I needed a WSL update
And now images seem to work fine without the token limit most of the time.
Sometimes however, you get images like this.
But when it works, it works.
id have thought it makes better lables for the bottles tbh
im sure i've seen better results for that part
Has anyone ever get experienced get trojan while doing preprocessing images?
this is very common with nsfw models downloaded on shady websites
So what should i do? delete all images that i just download? I mean i only download jpg & chrome HTML Document, i don't know this kind of file can contain trojan...
@reef frigate this is what I am getting but I don't have the obj file
I am trying again cuz it could also be cuz collab was diconnected
I just see this
it gives here yellow lines as well
Vintage 90's anime style Jesus cristo with hair black
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!
#⛰️
#🏞️
/retrowave, synthwave style, palm in the night in the rain,the Last of Us , landscape portrait, 8k resolution concept art portrait by Greg Rutkowski, Artgerm, WLOP, Alphonse Mucha dynamic lighting hyperdetailed intricately detailed Splash art trending on Artstation triadic colors Unreal Engine 5 volumetric lighting, gothic clothing
Some others were discussing this in #🤝|tech-support, I got these same issues today -- not just torch-deepdanbooru, but also Promptgen. It seems like Windows Update changed something and Windows Defender is picking up signatures within pickletensors and bins as being malicious. It may be nothing but it's probably not something to ignore, best bet is to find a safetensor conversion of the model.
Yea got that a few times
App restart helps
Dang, never seen that before
And I am CHRONICALLY on discord lol
@dry crow I figured out how to absolutely absurd ultimate SD Upscales lol
Like, 100,000,000+ pixel Upscales
On just an 8GB GPU
Woah epic, with the the sd ultimate upscaler script?
What was the source res?
2560x1080
I went from 2560x1080, then I did a 3x upscale, took that result, and did a 2x upscale on that
Ahh nice
Ultimate SD Upscale, like I said lol
For 1x to 3x, I used R-ESRGAN+4 whatever
And then from 3x to 6x, I used no upscaler
Which made a slightly soft image, which I then ran through gigapixel at 100% scale (no upscale) and it was able to extract detail out of the slightly blurred inage
Ah okay you used gigapixel for the last step
Yeah, all it did was de-blur the image
All the upscaling was in SD. I just used giga pixel as a pixel filter
Going that high res doesn't work in SD. It tries to do almost 800 upscale tiles, and it crashes
800 tiles? Of 512?
No, that's now how that works
Did you tried with 768 tile Size and esrgan 4x ?
Okay, now i get it.
I need to try the ultimate upscale script tomorrow again.
But i can get 6k images with the normal sd upscale script and esrgan based upscalers.
12k with Extras tab
@dry crow the only thing to think about is that the upscaling in the extra tab just add more pixels with no new info, where Ultimate SD generates the new pixels
Yes i know but for non photorealistic images you cant tell much difference at 6-12k anymore ^^
It's like giga pixel vs high res fix
Ah, fair
Just crisper lines. Doesn't work with higher details tho, but that's not a problem for non realistic stuff
I love and hate the new UniPC sampler
when it works great but most times - CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling cublasGemmEx( handle, opa, opb, m, n, k, &falpha, a, CUDA_R_16F, lda, b, CUDA_R_16F, ldb, &fbeta, c, CUDA_R_16F, ldc, CUDA_R_32F, CUBLAS_GEMM_DFALT_TENSOR_OP)
it is funny because if you try it again it might work. A bit strange that one.
Adaptive takes 10 times longer from 36-45s (depending on sampler) to 4:48-5m
So, I should be able to get the new uniPC sampler by updating my a1111, right?
yes
bottom of the drop down list
I like it and it is DDIM
has a few quirks to it which I believe is in the auto implementation
adding one lora I went from 40s to 49s to gen. each lora makes it worse and worse too
@smoky oak
Hmm... Down from 90 is still massively upsetting
VoltaML is why. I would have gone to that but I am pascal based.
VoltaML is tensorrt based for speed and Pascal cards don't have them.
No
Ah, alright
They had a big presser and a run up to it a few months back
they were saying to change one line of code could get you 2.5x speed increase. Yeah, if tensor cores are on your card
Is there a way to run this sort of thing on my system?
of course
So what exactly would the workflow be like to get this to function for stable diffusion?
read the link. I don't see people switching to Pytorch 2 if they are using VoltaML
Ah, alright, I'll take a look in just a moment
I will say if you pull it off 5-6x the speed
in SD your 10.5 would become about 40
30-40
more tensor cores on your card the faster it goes.
A100 is fall over speed
0.2s to gen an image
I was so hyped but then it got closer and the truth came out that we need tensorcores
thing is you have to use the specific (as the chap said to me) versions or you lose speed. This is what I really hate about all this py stuff as it is so connected to one version for all of it.
and they do mean == not => or <=
I just asked for you so I will let you know when they reply
can someone post some photorealistic mj v5 pix
I cant find a showcase in their discord
(this is SD, not MJ, just saw the message above)
yeh I wanna see mj
mj v4 simple prompt @stark vine
i am talking about the simple prompt thing. This is from pick-a-pic that was recommended for a demo on general chat by @vague oasis but longer prompt compared to the mj v4 one above
lmfao
Looks like I don't need 75 tokens anymore on the OneAPI version
Intel arc working swimmingly now.
1.65IT/s on 768x768, EulerA.
Of course however
Just like CPU inference, there are glitched outputs from time to time.
hey thanks for supporting diversity in the gpu world
that will help bring prices down for other companies, hopefully
@stark vine https://www.midjourney.com/showcase/recent/ that white and blue woman is v5
but we can do that on 2.1 or even realistic vision 1.5 no problems
sadly, SD still sucks dick for intricate details, in my case, I am desperate to get good skin textures
thats why I keep begging to PLEASE someone make a 1024x1024 model
so SD is definitely lagging behind in that aspect
that was upscaled...if you saw what Sytan up there did....
I find very detailed skin textures and gradual tonal changes on 2.1 photoreal models. and realistic vision to a degree (not as good as 2.1 still)
thats the thing G
2.1 was supposed to be the bomb
but it got nerfed due to the controversies lawsuits etc
it is definitely capable of something like v5 can do, if it didnt suck ass at everything else
thats why everyone shifted focus to 1.5
for sfw portraits, I have no problems on 2.1
I would love to see what v5 and dreambooth could do, sadly, that will never happen because mj is the Apple of AI software lol.
show me some
although i don't make enough portraits as my main work are objects and buildings
can i send and not make it public?
yeh
afaik 2.x were trained on stock photos only
so good luck generating anything mildly interesting
Hi, a quick question - when I compare the results I get from the sample prompt images from CivitAI, mine keep showing washed out, fuzzy, and low contrast in comparison when using the same model, same prompt and same extensions. What could this be? I would appreciate any help. I am attaching an example image to show what I mean. The reference is the original image from CivitAI, and the one below is what I get.
alright, thank you. I would be interested in trying it, if its not that big of a deal to implement
I have 2 installations of SD right now lol
Aye. The big part of it is TensorRT
naw, as far as I am concerned Pytorch2 is doa
yeah, seems to be
Any hints on the washed out results I have shown above?
You need to download an install a VAE. They make a huge difference
they are what controls the color processing after an image is generated
I am getting that with any model I have tried.
Including those I have included the VAE
you can see how big of a difference it makes
oh, hmm...
@dense tapirWhat was that addon you were using that was trashing contrast?
I am putting the VAE in the SAME folder as the model.
Dynamic Thresholding
oh, I believe you still have to trigger it
A1111 is by default set to None I think
I never could get it sharp enough
you can enable auto mode
Aha! How do I do that?
just a sec
Ahhhh, I would NEVER have guessed that. Thanks!
you can select automatic mode, which I think looks for VAE's that share names with checkpoints
I have noticed many VAEs have random names that don't match the checkpoint.
You can see they make a monumental different
Monumental indeed!
Correct so I never use auto
I use the same VAE for everything
Can I rename the VAEs to match the model?
I find it just looks good across everything
yeah!
yep
I use VAE-ft-mse-840000-ema-pruned, which is from open AI
What about LoRas? Can I rename those as well?
yessir
their trigger words will stay the same
you can rename models as well
all of my models are renamed for organizational purposes
Hey thanks! I was afraid of breaking things. LOL
no worries haha
Yes, they put version numbers that make it hard to keep track when typing
I like to sort them based on V1.x or 2.x
just make sure to keep the file extension (.ckpt, safetensors)
Of course. 🙂
you just have to press the little reload button in whatever area you rename in
those little reloads. There is one with LoRA's as well
I didn't know Loras needed to be reloaded.
if you add new ones or rename them. It just refreshes what it sees in the folder is all
I just add them to the folder, and Auto1111 just uses them without error.
that will work if you add them before it launches
I has worked for me even after launched. I was surprised too.
As long as I get the name right, Auto1111 just finds them
ohhh, do you not trigger by clicking?
No
the reason we have the layout is cause if you click them, it adds in their trigger word
I only click the reload button when I add new checkpoints, not loras
if you click them, it adds them in, or removes them
if you go into the LoRA UI in SD, you can just click on them to trigger them
I didn't know that. LOL
click them again and it will remove the LoRA
I probably don't even know where. I am sick and tired of searching for the files to recall the proper name. I am sure there must be an easier way
LOL it was the WHOLE TIME, and I didn't know! LOL
yup haha
its way easier to use like that haha
it keeps all TI's Hypernetworks, ckpts, and LoRAs
Thanks goodness, that was driving me CRAAAAZYYYYYYYYYYYY
That is the sort of thing I want in comfy to make it actually comfy
glad I could help haha
Very much, thank you!
The washed out images were driving me nuts as well
did the VAE fix help?
These images I am lovin' it.
I don't know yet. I have to see where in the settings, using your screenshot and see.
for reference, if you ever use Anything Furry V1 (best generat anthro model), it will crash with no VAE
I was lucky so far - no crashes for lack of VAE. In my innoncence, I assumed I could just place the VAE on the same folder, and that would do it, but noooooooooooooooooo
They have to be ACTIVATED. No one says anything about that at CivitAI.
They just tell us to put it in the same folder.
Thanks again for making life easier again. Auto1111 is good, but only if we know these details. 🙂
I've used around 45 models, and I've only found one that crashes like that. And that's the anything furry model
But I have heard of it causing issues with other models
I am trying to keep models under control - SSD can't afford such sizes.
They are crazy large for SSD standards.
Can we change the location of models to another drive, where I have more space?
symlink
Here I am at 47 models installed

And yet I only use three of them
But I always have the other ones in case
But really though, I really need to delete them to save the space. I have outrageously fast internet, I can download them in like 10 seconds
I have all of my models on the HDD symlinked to the SSD
A "just in case" that kills our disks like they were nothing.
I am assuming symlink is yet another A1111 extension?
I have 4.5 terabytes of high speed NVMe storage, I'm not exactly hurting for capacity at the moment
Like a shortcut?
It links multiple drives together
If I remember properly, it might even cache files on the SSD temporarily that are commonly accessed from the hard drive
My system drive is only 1TB, so I don't have that luxury.
I also should be doing that, since KNOW which ones I use the most. 🙂
The problem is the "just in case" LOL
link to the upscaler?
Hey Sytan you have a link to ultimate SD upscaler? You ever tried it on portraits? @stark vine and I were talking about the 1.5 and 2.1 portraits and looking for an upscaler to test with
is it this one? https://github.com/Coyote-A/ultimate-upscale-for-automatic1111
if that's the one that's pretty interesting
although doesn't look like something Topaz can already do?
No, gigapixel/topaz hasn't updated their models in a long long time now
Not sure wtf is wrong with them
I'll have to check it out
anyone knows if it works better for inpainting than lanczos?
Oh, my
Is there a way to find out to what model a VAE belongs to when they come with random names? X___x;;
Nope
Makes it hard to track them down
Indeed, pretty much
Can a VAE be used between different models? Is there a dependency?
I guess I will try and try and find out. 🙂
I have 45 models, but only 7 have VAEs.
no dependancy
Ohh, I see some chance of improving the looks on some models that looked washed out then. ^_____^
Well, post work can help it even more
even a photoshop auto tone/colour/contrast for the lazy can
I don't trust anything "auto" in my postwork.
Sometimes auto-contrast works, sometimes it destroys the image. Better do it by hand. It's faster. .
Auto settings are like barcodes - they are great when they work. LOL
I like the look of the too much post over the blah of #2
For me I am not going to spend hours refining something that I am not getting paid to do. Just not my thing as I can spend the time better elsewhere.
that is the master one
pretty much
Thank you again for the help. 🙂
You're welcome
Is a LoRa version of a full model comparable?
I mean, 5GB to 200MB. What can we expect from that?
you can expect quite a lot actually.
a lora isn't really a full head on model though
DB>Lora>HN>TI
I squarely blame SAI for the death of hypernetworks for never having refined them.
they were so good but took so long to train as they had no optimizations
I still, to this day, use my HN I trained as a general all around. If I could find the data I would see to train them as lora
So, are you saying the ARE comparable, even if LORAs are not full models? What gets lost in the way? I guess that's what I was wondering.
Sounds like magic 5G -> 200M
LOL, no
A prime example is you ask for Spuds McKenzie. Does the model give you something that looks like him but not quite? A body double? Cool, train in Lora and now it knows it precisely. No? Ahhhh, well DreamBooth full train a new model.
I have no idea about all the shit out there (and there is a ton of it) for 1.x as I find much better out there for 2.x
The day 2.0 dropped I tried it and dropped everything from 1.x
Some LORAs are made with 4 images only. Some with 800. I wonder what is the sweet spot?
that which works. I am being serious.
So it's not about quantity, but quality?
BINGO
I do styles though and it takes a lot more than on a subject
Styles sweet spot is 50-300
How long on a RTX 3090?
10-15m
Let's say, 300 images
10-15m
Impressive.
Friend trains on his all the time
What's the image size for that?
768 sq
So you have to crop them all manually?
We only do 2.1
That's high res, so it makes sense it has to be 768
We do as we both do not trust the invisible machine to do it called buckets but others use buckets
I only know render buckets, don't know what it means for AI training
buckets are nice if you have a lot of weird shaped images and I mean a lot of them
Like fractal images?
each bucket is a resized image it did for you to match your dimensions in latent space
Hi, I just joined the server
Someone can explain me where can I generate images?
Not on this server
We all do it locally
One more victim of this timesink. Say goodbye to your friends and family. LOL
HAHAHAHAHA
And buy a bigger SSD!
See, I honestly have little enjoyment from generating (inference) and a whole lot in training but colab dang near killed me. I was so burnt out
My 1060 can't do it
Wow, I didn't know those even had Tensor Cores
LOL, they don't on a T4
10m 3090 same exact data on T4 is 25-30
plus 10 more minutes just to fire it up
I can only imagine. Never trained my own model... yet.
I did DB in 1.5 and tons of TI embeddings on 2.x that I released and having a lot of trouble with Lora due to 2.1
2.1 fights ya
I have a gallery with 17 years of my 3D renders. I was thinking to making a model of my own stuff.
See, just to show you how I am, as opposed to @smoky oak who does models I saw 17 years of 3d renders and drooled to get them and make them a style.
That was exactly what I was thinking of - making a style of my own work.
No idea how, though.
The issue with styling is finding a common theme among them.
You need an eye
I have 2!
You can't throw picaso and Banksy at it as it would come up with shit or something inbetween
most times shit
no, an eye for see the common
I did mostly pinups, so that sounds like a common theme.
like how critics can see a painting and tell you who made it
People tell me they recognize my 3D renders, so I likely have a style.
Well, show me three random ones in private if you wish and I will see what lora might do with them
need at least 20 though
I have 2000
if you have 100 all the same theme/style you are golden
I think so.
you probably do
Even more in the same style.
No more it makes Lora vomit
chill
Thanks, I am keeping this link for later research. ^^
Maybe I can automate the tagging with Python?
No, it is never as good as you are
One day I hope
It loves say 1girl yet there is 1boy
or vice versa
There are virtually no men in my renders.
I guess I don't like men as much as I like WOMEN. LOL
Oh, so SD interrogates the images, tags them with what it THINKS is there, and we have to fix them by hand?
yeah, via an extension
Now I understand why all these LORAS made with 4-10 images. LOL
some people have no loved ones so they sit in their basements for months on end and just do it by hand. They have become the machine.
One day archeologists will find these mysterious mummies on basements, and wonder what they were doing there.
And I thought cropping was the hard part!
full finetuning they spend months at a time on doing 300k images and tagging
Not for humans....
not for me at least but they do it
One day, AI should do that part - but then we will be useless anyway.
A micro finetune is consider 1k, or less, images and captions
That's crazy
yep
I would want to automate that with Python.
Yes, I can see that now
the tech is getting better but by hand it is just far better
Based on my experience with image interrogation, I can see it's far from good enough
When the day comes we slide an image to the ai and it says Jennifer Anniston in blah blah standing next to blah blah we will have finally arrived
right now it gets a lot right then so much it doesn't
That will probably be "Jennifer Anniston mummy" by then
six months ago it was all rubbish
I am impressed how fast AI progresses.
What used to take years to advance now takes months
Sometimes weeks
But I am happy enough that AI still won't replace me.
Evetually nothing is safewhen AI/robots can already replace some surgeons.
You can find it in the extensions tab. You just scroll down until you find ultimate SD upscale
It's an extension of which adds a script that you run an image to image mode
Ai can already replace lawyers - well deserved!
I do not trust AI until it can learn on its own and purge any bias it was programmed with from its devs.
Sorry, I am really confused about this
I've done 3D renders/art for around 5 years now as well
I am the one who has done it for 17 years. 🙂
Why in the fuck would you be confused? You train models I train styles. I see it I don't care about the models I want those styles you probably want the models.
One question - can I put VAEs into a subfolder?
There must be a memleak as the more I gen I will eventually run out of vram and have to close it down and start over
Would A1111 still find them?
they are already there
'
don't you have to download extensions? is this the specific extension? https://github.com/Coyote-A/ultimate-upscale-for-automatic1111
I know we can put LORAs in subfolders, but what about VAEs?
Did you see?
I don't know who's talking to whom anymore. LOL
Those images were to you
that is how mine is set up
those are all on E a HDD and Auto is on D the SSD
I don't get it - what does the upscale extension has to do with this?
It doesn't that is another convo
I don't know, I know that it's built into the UI already
LOL
You install it from the extensions tab
So... we can place the VAEs on a subfolder?
I will restart A1111 and find out. 🙂
If I put VAEs on a subfolder, A1111 can only "see" half of them for some reason
I wonder why?
it is not on the collabs i am using. I guess I have to install that one and try it out. I can install extensions when i have the github link. But as far as I can check, it did not come with my collab as standard.
Ah, I see - it only sees the ones that end with "xxxxx.vae.pt". It won't see the ones named "xxxxxxxxvae.safetensors"
It only sees them if they are in the same folder for some reason
Oh, I got it now. I have to rename them to "xxxxxxxxxxxxxxxxx.vae.safetensors" The dot is important.
Ok, now it sees them all. Strange that it could see them as they were when placed in the same folder.
But require renaming if placed in a subfolder.
Great, now I can separate my VAEs from the actual models.
People told me the generic VAE for realistic style is the 84000, while the Kenshi VAE is generic for Anime style.
I thought VAE was color correction. I didn't think the styles mattered
You are massively overthinking it. Anime has their own colour scheme and flatter
But I assume it does matter because Anime style has a brighter, more vibrant color scheme
that as well, but not always
So now I know what these VAEs will do. The Kenshi one will give more vibrant and possibly more saturated colors
Could be the other way too
I am new to VAEs, so I will have to play with it to see. 😉
We don't use vae in 2.1 as they are baked in. Sad about that really.
Until now I was fixing colors in postwork, since I have to fix hands anyway.
SDXL I have no hopes for
I don't even care for the name as it sounds like a used car dealer trying to unload a 1977 Ford Pinto for 20 grand on me.
I prefer 2.1 for inference and for training 1.5 due to the CLIP. CLIP is fantastic, and can actually count, but fights you for training
CLIP is the image interrogation part? Or the text to image?
text to image?
I used to know what CLIP means, but my brain decided there are more important things to place there.
Yeah, OOM error so dump this to make room
no upgrades yet but Elon is working on that
LoRA name?
If I wanted to make a model hash search Python app, is there a repository to access models by hash code?
I know CivitAI can do it, but I wonder where they are getting the info from?
Not sure I like the idea of a central repository for said hashes
Sometimes prompts only provide the hash code, and I want to know what model that is.
I usually go to CivitAI and search there
Would be nice if I could search directly from a Python script.
Please help me
Is there a way to identify what LoRa created a specific image?
Can anyone tell me how to embed hugging face embed onto stable diffusion webui google colab?
Have you tried PNG Info?
If it's something that came straight out of A1111, PNG Info will extract the prompt for you.
im so confused I ran the command and its not in the embed list
I did this long ago when I first installed A1111, and I can't recall how it works. But I do recall they give instructions about this.
you mean for the hugging face embed?
You mean the code they give you to use their model?
It works like an activation key that recognizes you as the user.
LOL I tried making characters ride bikes before, with hilarious results.
Im trying to embed one of their models from hugging face
Embed? Isn't it just placing it into a specific folder?
do I download the .pt and put it into the folder then run the command?
If A1111 is already running, you first need to press the refresh button to make the program see the new file.
how could i
20 steps it was missing the handle bars so I had to do 30
use the stable thigs
so all I gotta do is put it in the folder?
Sometimes adding steps can also add more arms and legs. LOL
Yeah, which is maddening
Yes, put it in the folder, and then press the refresh button in A1111 if it's already running. If it's not running, just start it, and the model will be available from the list
Maddening is what SD does to hands.... <___<
One day.... One day....
More than just hands
Well, controlnet allows good hands
I don't use controlnet though
Welcome to the world of mutants!
Funny thing is I make a mutant it has human hands I make a human it has mutant hands
For those lunatics who claim AI images are "easy" and take "no effort", they obviously haven't tried it yet.
Hands are getting better with some helper embedding now. Better than nothing
I refuse going to Blender to pose a hand, extract the skeleton pose, and bring it back to A1111 hoping the perspective was right.
Not for everyone
I hate blender for decades now. No thanks to that.
I started with 3DSMAX, so Blender feels like torture to me. Can't help it.
What I can do in MAX with 10 clicks, it's 100 in Blender.
Blender was never really made for what people use it for though.
there is a reason its UI sucks balls
People seem to love it though
The UI got better, but the workflow sucks if you have ever used other 3D programs. Very inefficient to me.
People who love Blender have probable never used anything else. You can't miss what you have never had,
C4D, or 3ds or pre AD Maya was so simple to get around in
Comparing to Blender! LOL
Why this driving need to make shit overly complicated?
Same thing I asked myself. Good to know I was not alone on this.
Ken in Blender: I want to edit this vertex. I can't!!! Oh, I have to switch to EDIT MODE....
Hell, I see it in Windows where something you used to just right click to get to the next version is two then the one after is so damn deep you can speend a lunch break getting to it
The art of Windows "improvement".
It's like arranged marriages - you will love her over time.
some small changes but then he came in and it has never gotten better
11 is so deep on shit as if they are trying to hide stuff on you
I can't use it. Haswell-E here.
Sorry to bother you again. But its currently not in the embed lsit I think I have to run this but I dont know how to get the links. because these didn't work
Nor I but I see others and vids and just wow
If it's an embedding file, you place it in the Embeddings folder.
The PT extension means nothing.
is this how you do it?
yea it ist
isnt
Then you should be fine
BTW, you don't have to restart A1111 for this
If the file is there, and you use it on a prompt, A1111 will find it.
Same for LoRAs.
Just models require pressing the refresh button
Was looking at some prompts from CivitAI, and some images have sizes like 600x512. HOW????
That size is not allowed....
image
I can't reproduce these images because A1111 won't allow 600 as width
It has to be multiples of power of 2
The VAE files make a world of difference. This looks edible.
No mutant fries! LOL
Had to try several different VAEs, since the 84000 and Kenshi looked both too saturated to be real.
just click the number and you can type whatever value you want
I tried, but A1111 auto changes them to the closest power of 2
Ah, I have to hit ENTER! Doh!
just click out of the box after typing it
Yeah, I didn't do that.
Some prompts include the word "BREAK" instead of periods. Does that make it any different?
does matter
breaks being UK and periods American
this was trained in America so a lot of those words it has no idea
man, is there no batch interrogator yet in auto?
have a nice one for tagging
Funny that I can make the AI make me a burger with fries, but if I ask it to put ketchup on the fries, it sticks the fries inside the burger and throws it on me. Sorry for asking.... <___<;;;;
On my 10th try, it put the ketchup in the soda.
I see one so let's give this new extension a try
the thought of that makes me sick
milk in pepsi was a think but ketchup in coke
Point taken, the models have no clue on how to put ketchup on fries. But it can put it on the soda cup instead.
If the AI has never seen it, it probably can't do it.
First time I put ketchup on fries in Brazil, people almost threw up when they saw it. They told me I ruined it. Well, to each, their own.
Is "bad-artist" an embedding? It seems used a lot.
Oh well, it IS an embedding. Better get it
I don't use 1.5 so unknown to me
Pretty cool I see an extension to change my ckpts to safetensors which I was wanting.
Darn, there is another one: "bad-image-v2-39000"
Where do I find that?
CivitAI and HuggingFace don't know it.
hey, what y'all think of my two girls? one norwegian and one malay
and how realistic are they compared to the most realistic ai photo you've ever seen?
They look good, but the eyes give the blonde away, and the fingers on the Asian.
interesting, thanks man!
Those are the usual suspects with AI. Hard to get it right, mostly luck sometimes. 🙂
Sometimes I get perfect hands, but the face is all wrong. Darn it! LOL
could try again with GFPGAN regarding the eyes, and i saw someone mentioning hands, what was it again
It's very hard to fool a human when it comes to human faces. It's engrained into the DNA how to detect the smallest flaw.
oh yeah openpose hand
ControlNet
You can also try "bad-hands-5" embedding to improve hands, but it's a hit or miss.
hmm
AI seems to like spaghetti a lot when it makes hands.
trying to do it in ruby with replicate and dreambooth: https://gist.github.com/basicfeatures/eaf7b414945c8cdc5f18441f51e1f731
Dreambooth has a lot of limitations
can't find no bad hands there. but did find https://replicate.com/collections/control-net
i see.. in what way?
I haven't seen it in a long time because I have A1111 installed locally. But last I've seen it, most options were hidden
we could only change very basic things
ah yes, i was planning on adding https://replicate.com/anotherjesse/controlnet-1.5-pose-template once i figure out how. then i can tell them to do stuff apparently
Control net is the most powerful extension I have ever seen for A1111.
Tons of vids on YouTube, with people drooling all over it. Very worth your time.
It allows things that were not possible in A1111 before.
Like this with a single prompt.
so i just add that stuff on top like gfpgan? and not integrate it with the dreambooth training somehow?
glad i could use ruby for this though. pretty isnt it? 😄
and i have a little secret too
i plan to use gimp in post production: http://elsamuko.github.io/gimp-elsamuko/plugins.html
man those are some nice pluggos imho (my girls above had double exposure to film grain)
or are there better choices available here in the AI world?
its like with my music, my plan is always to add so much vinyl emulation and analog distortion so that you can't tell it's digital and synthetic
They have tons of models to choose from at CivitAI. More than I can keep track of.
(whether the music itself is any good - https://soundcloud.com/haukeland/burzum-tribute - is ofcourse another matter 🤣)
cool ill check it out, cheers!
burzum, the famous black metal guy is who that scripts for. my godfather shared jailcell with him once...
Any hints on where I could find the "bad-image-v2-39000" embedding?
Thanks!
When I go to Hugging Face, and search for "bad-image", I get zero hits
I must be doing something wrong
interesting! only been there once before to get those ckpt thingies for my script:
ckpt_urls = [
"https://huggingface.co/BestJammer/HASDX/resolve/main/ckptSXDHAS.ckpt",
"https://huggingface.co/johnslegers/hasdx/resolve/main/hasdx_emaonly.ckpt",
"https://huggingface.co/johnslegers/hasdx/resolve/main/hasdx.ckpt"
]
anyone yall would switch out?
theres so much how to make old photos new but not the other way around afaik
There might be. Sometimes the challenge is finding things.
Good example right above - I have searched for this embedding at Hugging Face, but didn't find it, even though it was there.
aka. the art of browsing, very important concept in ecommerce psychology i learned the other day
music too, listeners are way more likely to share if they stumbled upon it themselves
what are you trying to achieve ken1177?
I am trying to replicate the sample prompts from the HELLMix model. But I am missing some embeddings.
cool
Finding missing embeddings has proven to be a challenge. I still don't know how to find that last one if I had to search it myself. I went to "models" and searched, but not found
theres a town here called hell actually: https://upload.wikimedia.org/wikipedia/commons/thumb/b/bb/Hell2.jpg/500px-Hell2.jpg
This one was "experimental", so it might be hidden somewhere
sorry thats not helpful
LOL
There was a street in Ohio called "Manko", which means "vagina" in Japanese. LOL
haha
I took a picture, or my friends wouldn't believe it
Now back to the embeddings from the negative prompt, they can change the final image completely.
without googling, what do you think this place is called? https://upload.wikimedia.org/wikipedia/commons/thumb/0/08/Craighouse_from_the_pier_-_geograph.org.uk_-_755742.jpg/510px-Craighouse_from_the_pier_-_geograph.org.uk_-_755742.jpg
vikings liked to give funny names to islands around scotland etc 😄