#💬|general-chat
1 messages · Page 117 of 1
my background is about a month of obsessive self teaching. so, the small details are important for me lol
well i have the processor create the pose with the open pose
but then it still ignores it and does its own thing
preprocessor
i will @ you on general with images, and show you the process. i suck at explaining , so ill just put screenshots on there
kay
have you downloaded any controlnet models?
i have the 1.4 or idk
you need to put controlnet models into the folder the extension looks for them in
oh my control net generally works i think and i did that in the past when i first set it up
actually starting to think its a model issue
😦
you have to use the contorlnet that matches the model architecture. if you have sdxl loaded, you need sdxl controlnet models
^
Dang… now Emad is stepping down too? What's to become of Stable Diffusion?
blockchain
i think all internal people are told to stay quiet for now
Stability is going to float around in economic space until a fish big enough and bad enough swims by to swallow it whole.
Meanwhile, on land, bears and dragons roam
whats the best tool to change clothes on a photo I have? and put a "logo" on it? can stable diffusion do that with inpainting or is there something faster/easier to use?
they've partnered with blockchain companies so i mean, they've sunk. blockchains don't float. only held up by massive institutional fraud
in an ideal world, NANO would’ve taken off more with its mining-free approach
that’s another story tho lol
lol wtf dude Nano is 2017 scam
blockchhains are a whole lot of idealism
exactly my point
full time crypto trader/investor here
basically a gambling addict. hope you''re okay
thats how it always is with revolutionary tech
yeah im okay enjoying my freedom from hard work
the only thihng crypto has managed to revolutionize is crime markets
Well, it comes down to human nature. Humans don’t want public ledgers.
I have nothing to do with the availabilty of the bots 
anyone have answer for this?
its not even decentralized. crypto verification nodes are owned by a few and will only continue to consolidate
i’d recommend using Stable Diffusion along with some control nets, like canny maybe
why control net for what purpose?
mostly for isolating the features you’d like to change.
otherwise, you can try photoshop’s inpainting feature
didn;t knbow they had inpainting feature. would that be for the logo?
the shirt, mostly. the logo depends on what you’re wanting to use, whether you want it to be in-painted along with the shirt or if you want to bring it in from another process / workflow
agged u in #🏞|general-with-images if u have the time could u plz advise 🙏
so I’m not aware of any custom projects that’ll take logos and automatically work them into an image (there probably is but I haven’t done my homework on that particular application) but I would imagine this is a matter of using a controlnet along with a mask that has the contour of the logo as well as the shirt.
so ideally you’d want to do a mockup in PS, where you’re just doing a basic superimpose on a shirt. Then you could try doing an img2img using that as your base input.
crypto being decentralized is a big ol lie. It's in control of those with the most wealth, as usual
it’s reality preying on the ideal
Anyomw who does any crypto stuff is scamming or money laundering
There are individuals with wealth who are trying harder than others, i.e. Mark Cuban
its the kind of shit that gets everywhere and people fetish themselves over
with his pharmaceutical ventures
Crypto wastes too much power
Bill Gates with the B&M Gates Foundation
Mining etc graphics cards
thank u kind sir
yep, say what you will about him as a human but relatively speaking, he and melinda have both done a lot of philanthropy
they've got rid of that con i think. a few networks do the value through wasted energy thing still. the big one just cuts that middle man right out and gives control to those who are rich , straight up
He didn't do it for charity he did it for brand image
And appearances
Fuck him he hates open source
well, he could’ve gone steve jobs and done zero philanthropy
Bill Gates can jump over a chair. Show me another Billionaire that can do that
you think zucks is going to let people visit his bunker?
go?
how do i tell stable diffusion to give me photo of woman from above torso
💀
"cowboy shot" but then sometimes they're just a cowboy
Put naked and nude
In negative
There is a lot of bias
Especially if you only put "women"
also try half body portrait
in your prompt
I think it’s time to make a Mortal Kombat parody but with all of the big AI companies as the characters
beautiful
lol
just reset the whole system. works now
lol
well... I couldnt figure out what the missing thing was
thankfully its not a huge system to re-build
So now Im trying to figure out the better workflow for video if its SVD or animate?
ok i m blasted by webui forge .. for those without rtx ..it speed up my process by 60% ..ok gtx user it work ... a special thanks to @mojo... 😉
That depends lol. SVD is good for animate still images, you can use a workflow to use prompts to generate an image to use as the int for SVD, and etc, but, it's still Animating a still image, but it does it about as well as Pika Labs and RunwayML Gen 2. Better sometimes but you gotta play with it and tweak it. I use loras with it to speed it up, but I prefer using it in forge.
AnimateDiff on the other hand. You can work it into vid to vid workflows. Use it for text to Image, or set it up for img2vid. I use it to convert 3D rendered vids into realistic vids or comic book styles and etc.
Both are good, but different applications. I'll often extract frames from my svd vids, use a frame as an int for the next svd run, discard the ones after, then stitch them together to make a cohesive video. But that's usually when I'm going for something more artistic representation and less story centric.
full transparency, Im working on stage thing that doesnt exist yet 🙂
I dont need a full on txt->img cuz I have that concept down
its animating the still images
ive been working with the core basics of SVD most of the day now that I have it working again. trying to find the limits
I've worked on many things that didn't exist yet at the time that now do 😜 😅.
Animating still images SVD is best, it's pretty simple.
animatediff has sparsectrl now but i haven't tested enough to decide if its better or not
Plug in image, insert prompt, video comes out lol
AnimateDiff doesn't really shine unless you dive into crazy workflows.
Xl
automatic1111 with the cn and ad extension shine. keepin all possibilities under one umbrella. there are great workflows for comfyui in the realm, but so much is possible easily with the extensions
stable video diffusion is based on sd2.1
What I will end up spending my life learning... is how to make the video as smooth as possible, then taking those last frames and blending them into a new thing.... and compiling to a final output
animatediff works on sd15. beta motion module for sdxl but its beta
This is true, the xl is beta still. 1.5 is more predictable and stable
Svd xl is Different, the XL version works better than the 1.5 for svd.
Ive been looking at a workflow on civit?
- AnimateLCM Lora & Motion module
- Dreamshaper SDXL Lightning (or any SDXL checkpoint combined with Lightning Lora)
- Photon (or any SD1.5 checkpoint)
- SparseCtrl rgb (available in comfy manager)
- IP-Adapter (available in comfy manager)
- Clip vision for IP-Adapter (available in comfy manager - CLIP-ViT-H-14-laion2B-s32B-b79K)
but this is a leap toward animate?
would someone be so kind as to give me suggestions on how to make a hologram-like picture in stable diffusion? i want it to look like cortana from halo.
Photon LCM is best, same model with the LCM lora baked in. Saves you some vram from loading the lora, the combination with the dreamshaper lighting model let's you drop cfg down to 1 or 2 and my steps are usually 2 to 4. This significantly increases generating speed.
is it kinda the same end result? only more control over the model ?
Same end result, the lora is just baked in so you don't have to load it separately. Or if you meant cfg and steps, steps significantly more control. 1 through 9 is one possible image with varying quality, 2 to 4s usually good for me, but 9 can look way better or worse lol. Then 10 through like 19, and etc. Once you cross that threshold there's a big shift usually. So more control in that sense.
Cfg pretty much has to be in the 1 to 2 range, maybe 3. But low low cfg is usually required. Steps can vary depending on what you're going for.
ok, I havnt taken a dive int othat other workflow yet. I just love the end results i have seen
Otherwise you won't benefit from the lighting model or the LCM models.
The workflow you posted is good.
it lo0oks amazing.
ComfyUI became super easy awhile back with the manager. Just click on install missing nodes lol
Bam, workflow
I wanted to explore the SVD workflow to see what was cool
thi 25 frame stuff sucks lol
I dont think I understand the condition as much as I thought I did... im trying to translate film into this thing
video frames = the length... but falls apart after 25 I think if I read this right
motion bucket is how much action to push in that short time
then just take notes duringa 20min video walkthrough on how to use each and every workflow
also, you ought to be vetting every node you install instead of just trusting everyone else did
but fps..... is not frames... so that one confuses me
nothing significant. just make sure that it's not something causing electron mining or anything
that swhy I killed my files and rebuilt
the "fps" setting in automatic1111 extension only determins the rendered file's fps. nothing to do with the generation process. this may be the same on the node side of things
i often do 60fps files , albiet with the frame interpolation as provided by the deforum extension
first time ive seen staff being in here posting sd3 images, what a time to be alive
i hope so
whats going on
nothing is going on, haha, idk how often they are here posting images
I usually peep the code and see how it works 😅.
#🎥|animation message 80fps render. 36 keyframes
took 3 and a half minutes
people keep telling me that comfyui is better for animatediff, but i dont get it.
i keep hearing it about animatediff
i'm throwing a dancing openpose video at it now. 250 frames. gonna go have some dinner and will upload it when i'm back.
the extension on a1 loads the motion modules
(im gonna ruin terms till I better understand them.. so sorry)
your spy name is the term ruiner
makes me happy, first time ive given the grpahics card this much of a workout since mining coin
controlnet + animatediff for 250 frames is only taking my gpu to 13.5gb. fp8 support babyyy. harder to set up in comfyui. checkbox in a1
whats your gpu?
4080
3090ti here
well, if anyone works inside comfy and has any video workflows they dont mind sharing 🙂
Ah, it's because you can chain nodes to lora and controlnet stack, on top of adding in reactor, clip, interpolation, face swapping and fixing, and etc with significantly less vram 😅.
I also use forge and automatic1111
They all have their strengths.
i hear that too but i never have vram problems in a1111. i don't do face swapping anymore. ip-adapter is better since it noises the swapped face into the diffusion steps.
I'm running an rtx 4090, it depends on what you're attempting to do.
whenever i turn fp8 on in comfyui, most of the custom nodes flop. workflows aren't set up for them. so i guess i go learn entirely new workflows? meh
But the vram definitely has been more of an issue with automatic than comfy with animateDiff. I can do much higher resolutions and much longer videos in comfy.
i have a 4080
i only get memory problems at higher resolutions
if i switch on comfyui's fp8 mode, i get the same max resolutions in it as i do with automatic. 2048x2048 ish. not all prompts or settings are the same
and at those resolutions, more tokens can make all the difference
My videos are usually 2 to 4 minutes, so squeezing out the extra vram is a must for me lol.
i think that once you know how to use both systems, they're really all the same underlying libraries. they have the same peak efficiency
video length doesn't matter. just context batch size.
and animate diff can't go over a certain batch size anyways
added things like controlnets are the biggest memory hogs. or lora's
Yeah, after working with and on extensions, nodes, and etc for all 3 I can say. ComfyUI is the most optimized, making it faster and allows for higher resolutions, offers the most customization. Automatic1111 is the least optimized but the most robust, it has more extensions than any other and it's well maintained. Large community. Forge is a bit in the middle, most of Automatic1111 extensions work in it. Some that use to work in Automatic1111 still work in Forge. It's more optimized than auto with more extensions that comfy putting it somewhere in the middle. SVD XL works pretty good on Forge as well.
nah. confirmation biases go hard. i'm not sure what you're doing that makes a1 1.8 eat the most memory, but it doesn't have to.
Nah, just actually dove into the code for all 3 and use all 3 regularly 😅.
forge uses the memory management code from comfyui and consequently has less efficient fp8 memory usage than auto1111 1.8 does.
the fp8 support is just bonkers in comfy ecosystem. it uses half the memory footprint so i mean.. okay.. squeeze your ada card
I mean, you can literally monitor things and see 😅. It's not an option, that's just the objective strengths of each 🤷♂️. To eliminate biases tools to measure resource usage are used and the code is analyzed 😅.
How do I switch to fp8
i code surf all the time but it doesn't mean i'm an expert on this stuff. i just hope that if i stare long enough it becomes apparant to me whats going on.
usually i rely on metrics
in automatic1111 its in settings optimizations. checkbox
I'm on comfy
comfy its like 1 of 3 command line launch switches but you have to use custom nodes that work with it
Ah, screw that then
But to each their own. I'm mainly here to see what's going on with SD lol.
forge has the checkbox too but it uses the udnerlying comfyui memory code
its literally half the memory footprint and people are so often "nah" cause it has shit support
and on Ada cards, its got hardware support for it so theres no speed hit
but people are like "nah. comfy is better"
comfy is better if you have more vram I guess
A1111 and forge are nice but they lack the same amount of fine control
even then... that more vram is effectively doubled in fp8
i don't find benefit in most node graphs. there's few i seen in comfyui that i think would be difficult to replicate in a1. animated regional prompting for example.
Used 3090 24g vs 4070ti super 16 g ?
3090 every time
really have to wonder when forge will get more updates, it's been two weeks, and they just happened to have stopped taking merges from upstream when my refiner fixes got merged
and the other thing I got PR'd into A1111 is also coincidentally broken 
comfy is really good with vram
oddly enough i'm finding i'm getting the fastest inference with forge now, faster than comfy, idk why
a1111, fully updated etc, still 2x as slow (on a 4090 5950x 64gb ram setup)
I believe the owner of the repo is focused on course work at the moment.
Though I may be mistaking it for another lol
one of the updates busted regional prompter and that's been known for over 2 weeks
i had to roll back
Ask away, I'll probably go afterwards 🙂
I have pic already generated by stable diffusion…… i want to replace the character
How i can do it
You can do it using inpainting most likely 🤔 It depends
More involved methods would involve using controlnets
у кого стоит нейронка на амд? у меня rx 7600, и уже много что попробовал не запускается
This might be a bit of a smoothbrain question, but, is it better to always generate stuff without HiRes fix and then, when you want the image Bigger+Better, run the exact same seed with HiRes Fix?
I'm on Euler A, so I think seed reproducibility SHOULD be on the high side, but I'm also using Dynamic Thresholding which might skew the overall behavior
I don't think the seed actually matters at all tbh
It only matters if all parameters are the same including resolution
the seed is just how the initial noise is calculated
anyone experienced with the Krita generator?
not really what you were asking, but generally ive found results are better if using sd ultimate upscale instead of hiresfix. usually run 4x ultrasharp on it
Remind me how I use that?
I've had some trouble figuring out where that's actually kept in the UI
Extras?
anyone have preferred browser to run their ui on? i normally use firefox for everything, but it does not play nicely with the webui's , it seems. chrome, it ran a little better but i just really dont like the browser. tbh edge seems the best so far, just wondering if anyone has a better alternative
Firefox works fine for my use cases.
scripts,. although, the caveat is that its only usable in img2img, not text to image. but, if you copy over seed , prompt, all the settings etc from a text2image generation then it shouldnt be an issue. wont change the image at all aside from upscaling it and usually smooths out some minor imperfections depending on how you use it
Ah, okay, that's why it wasn't appearing
yep 😄 i was scratching my head looking for it the other day lol
There's a button to send image and generation parameters to img2img, so
For clarity this is on an XL model, but, there's a lot of options here besides the upscaler
do I need to change Type, the tile settings, padding etc or can I leave those alone
my biggest complaint with firefox, and this happens in both forge and a1111, even with a clean install and nothing but default extensions, is that refreshing anything causes everything listed to be greyed out. not a huge deal, just mildly annoying. also have to change the canvas zoom hotkeys from default or change hotkey settings in firefox to make it work properly
i usually leave everything alone except tile sizes
My base image was 832 x 1216, and I was thinking of just 2x sizing it, so
oh, and i set it to scale from image size
how should i configure the tiles?
yeah, just do uhhh scale from image size, 2x, and bring your tile width up to as high as you think your gpu can handle
what about height, and, should tile width not exceed a certain amount?
like i guess i'll try 1024
tbh, i havent tried setting it over image size. not sure what that would do, try it and let me know 😄
also if you fuck around with the settings just right, you can abuse the tile settings to create mosaics !
it didn't copy the seed over
rip
should still be in txt2im g tab. or, worst case scenario, just drop the original generation into png info tab. thatll have seed
yeah no worries i'll just copy it over next time
this so far seems
SUBSTANTIALLY slower though, although usually mine are 1.5 on hires, not 2x
so that may be part of it, and the tile width is just 1024
it is on 4xultrasharp_4xultrasharpv10 (why are there two of these in the settings on forge? idk)
oh that was pure chaos
that did Not work properly
i'm gonna leave ultimate upscale alone for right now
i need a clearer correlation between img size and tile width or else it'll be a lot of spent time for weird results i feel like
so 430ish. should run much faster that way
man i still dont grasp it. but i know its fun as hell to play with 😄
best i can tell, setting tile width to half image or full image width is best as it minimizes seams, and dont have to run seam fix that way.
a future showroom of environment & desert
bots offline #1047610792226340935
#1047610792226340935 a future showroom of environment & desert
a future showroom of environment & desert
#🏞|general-with-images a future showroom of environment & desert
#1047610792226340935 bots currently offline
Dont bother trying tbh, you post that bots are offline and one message later someone else tries to prompt or asks where the bots are 
I know, but those messages makes my head hurt 🫠
Hi all. I'm new here and was just wondering where to post questions. I see there is a prompting-help channel but my question is not realated to prompting. Btw, I totally see the irony in posting the question: where to post questions 😄
It depends on the question!
What is it about?
Hey! It's a question on lora training. Should I just ask it here?
Hmm, you can go ahead, I can't think of a more appropriate channel 🤔
Ok thanks, here goes:
Let's say I want to train a lora that blurs faces. I would think that the best approach would be to take some images with faces and blur them in a photo editor and use in my dataset.
What I'm wondering is: how can I use the images without blurred faces to "show" the lora specifically what to add?
I'm hoping this could make the lora very "pure/clean" so It only blurs the face and does not alter any other part of the latent image.
The reasons why I want to try this are:
- it would produce a "clean" lora that does not affect anything except from the face (no unwanted style etc. added).
- I would skip the whole tagging part of dataset preparation which is a bit of a pain.
regularization images would be your unblurred faces
Hi, thanks for the reply. That's what I thought as well but after a quick test of blurred images in the training data and un-blurred as regularization my model is heavily influenced by the subjects in the training data. So ex. the faces might be slightly blurred but the person looks like the person in the dataset. I might have made an obvious error as I'm pretty new to lora training.
are you training for 1.5 or xl? xl is kind of hard to train the text encoders for. also, there are a million ways to screw up training by using the wrong settings.
for this I used SDXL
which optimizer are you using and what settings for it?
btw, by "text encoder" do you mean altering clip weights with tags?
I was using adafactor with constant scheduler and 0.0001 learning rate.
yes, it's a toggle on/off in both koyha and onetrainer. don't use them for xl, it's not really worth it. and adafactor is a dynamic optimizer where you set all learning rates to 1. it manages them on it's own. personally, i'd recommend using prodigy+constant with bias correction and a weight decay(make sure it's the optimizer one) between 0.05 and 0.1
are you using koyha?
if you are, load the preset called something something "now prodigy" and change the batch size to match your gpu limitations. it will have the "weight_decay=0.01 decouple=True d0=0.0001 use_bias_correction=True" tags already with it. you can adjust other settings as needed. i would add --network_train_unet_only to the additional parameters(under advanced) line though, this is what makes it not train the text encoder
yep, using koyha. "yes, it's a toggle on/off in both koyha and onetrainer." - which toggle is this?
I've heard of prodigy but does it now requre extra arguments? I've not dared to get into that just yet.
"bias correction and a weight decay" - this is over my head 🙂
yeah then read what i just said
i'd also adjust the epochs depending on your dataset size as well, 160 would be overkill if your dataset is like 100 images or something
but other than that, you're going to have to do a lot of experimenting and a lot of reading up. most of the guides you'll find are centered around 1.5 and not sdxl and what works for one doesn't work for the other really; in terms of settings
oh and you'll also find a ton of terribly conflicting information as well lol...
thanks! I find that the guides are usually targeted for training celebs or anime characters so specific things like this are not well documented and I fully agree with the conflicting information. one more thing: when using regularization images, should the repeats match the repeats of my dataset?
Hey all! I'm looking to migrate from Midjourney to Stable Diffusion. Is "Draw Things" a good way to start?
not completely sure, i dont use them much. i think it's more about the training data:regularization data ratio. like 2:1 or 4:1 or something. i could be completely wrong though, so do some looking around it on. i'm not educated enough on the matter to really have any solid word about it lol
ok, thanks anyway!
np, gl
Really good if you're on a Mac/iPad/iPhone
hello
@pearl ocean
https://m.soundcloud.com/4dreamsy/love-you-like-i-do
good day
damn people are freaking out about SD3 possibly being closed source
Personally I suggest going balls deep with comfyui
After that everything is easy and you learn a lot about diffusion model's inner workings as well
hi all
Does the controlnet tile method of upscaling not work anymore?
Seems controlnet changed its layout and I'm not seeing the needed options.
It’s a big deal—SD3’s multi-diffusion capabilities mean one-shot professional marketing. We’re seeing the power of corporate influence right now.
yeah its unfortunate
It truly is.
Honestly, it needs to be. They need money to keep progressing. Weirdos obsessively sitting at home making waifus aren't footing the research bill. They might release a smaller stripped down version of it though like a 2B variant
What’ll happen next is that laws will set in, and folks will go “…I understand why there are laws, but what if I really did want to just generate a picture of X reading Y, etc.?” And the official response will more or less be “if you want to crowdsource your own data and build your own model from scratch, go for it, etc.” which will give companies excuses to prey after folks’ personal data troves
So it’s just a mess
SD3 Turbo looks good only around 4 steps
and making it a service would just cost them more than SDXL Turbo
Can't run on home PCs, they already explained that
well 24GB is possible
but I get it
most people have 12GB and lower
like 90% of the community
💀
It’s about control…an abstract dial tuned to the 1%
No, it would be a reduced quant that runs on 24gb
Afaik, sd3 turbo likely has to stay in all f32 to be worth it as an upgrade to lightning/turbo
Any released date for SD3?
No
😦
We're mostly just currently hoping it gets released at all
I'm guessing it still will be but I'm admittedly guessing
why
It’s currently “shaking up the industry” behind closed doors.
Emad was pushed to resign
Powerful people don’t want SD3 to be released because of its ability to mimic reality, essentially.
And said some weird stuff about how he's leaving to decentralize AI even though he's the majority shareowner
So the big q I think is whether investors have some kind of agreement in exchange for VC that allows them to overrule decisions by emad
I think they’re all getting equally “pressured”
And force a direction that's more closed up that is more immediately focused on profit
I don’t think we’re seeing this bad boy the way it’s been presented to us
Last thing we need is another (overly) curated AI model, I understand the need to be sensible with it but the way openai limit the things you can do is just annoying. for example "a man on a boat with the women wearing a bikini" > flag! beware! ... stupid.
Yeah I have zero interest in openai bs
It's open or it's not and if it's not I'm not bothering
I’ll be honest, I didn’t think this moment would come so soon—but we are living in a desperate world, and even powerful people can be desperate sometimes.
I want to use SD3 so bad xd
IDC if it means I need to drop 5k on a massive PC upgrade
Same
It'll be pretty interesting once ppl realize that most image and audio evidence should not be admissible in court anymore in the way it has been
Yeah, but when open-source becomes unstable or banned or anything of that nature, the lawsuits against the for profit organizations begin 😅. Adobe, Google, Midjourney, OpenAI, Microsoft, Reddit, Facebook, like they are all going to see people's who's data got harvested coming after them, open-source devs with class action suits for using code, concepts, and etc.
The issue with getting Billions of people involved in the creation of a tool, once you attempt to cut them out, they're coming for their share 😅.
This is what I mean by open-source being a buffer against the anti AI crowd. Once open-source is no longer considered a thing, it's all going to fall. All the good will and buffers become additional issues to deal with lol.
It all needs expert analysis by ppl like us that know exactly what to look for to spot evidence it was generated
And that would only catch the sloppier work
Eh, after awhile it will be a matter of context.
And then it will be more subjective than ever, relying on actual AI models to detect
Anyone that has some decent Photoshop skills can patch up something so well it won't be spotted
Yeah ai won't be able to do it reliably
So it won't work for court
You must understand that these models generate what are essentially syntactic collages
And they can be dismantled as such.
Takes 2 seconds to make deep fakes with photoshop using contracts 😅. Like good ones lol.
Using what? Lol
Ive primarily used affinity photo tbh
Does have some features Photoshop lacks that I really like
Example of what you can do with swapping faces alone with comfyui and a thorough understanding of frequency separation/image structure/chaining tonal information
FaceFusion is nice
Ignore the fingers the only point is the face, you can throw in any image I just gen'd one within the workflow out of laziness
This is better than anything I've seen
Though the inswapper devs were equally pressured not to release any of their models about 128
above*
Uses every trick I know of that doest involve manual editing
Yeah
It's not as good as this either anyway
Look at how the tonal information, the reflections, shadows, positioning and angles of eyes, nose, mouth are all conserved
That's the biggie and this is just me, an amateur, doing it
Eh
Imagine if there's some real financial incentive, and a pro team is behind it
Various people are reaching uncanny valley at their own rates.
That’s happening everywhere already man.
I'm sure it is
Big Fish don’t like Little Fish.
You load up the original image, the face you want, use the context aware selection tool (AI tool thats a pre stable diffusion release tool) or lasso tool, then it's just about blending. It's super easy these days even pre widespread AI. There's a super simple way to do the blending to make the New face display proper color and lighting for the subject and the scene. If you good easy photoshop faceswap it will show you how.
It's actually easier to do in photoshop than with AI 😅. Quicker to. Lower end system requirements. Like in every way it's the better way, but, that's not the purpose of these types of groups lol. Deep fake is it's own community.
Do we have anymore news on what is going to happen with SAI?
I'm personally a little bit concerned..
Oh, yeah. I've heard about that, haven't tried it myself
And it’s also a subjective process, made subjective by the various directions you can take these workflows.
🙂
The workflow I linked to was the first time I got it to the point where I couldn't find obvious issues
Cool.
Usually even if it gets the face structure right. The angle will shift, and even if it gets that right, and the tonal info doesn't drift... Color temp etc
The lighting will change. A reflection on the corner of a nose will be gone, or diffused etc
This preserves all of that
I think the theme for today is, “whoever is winning the game is actually losing the game”
#🏞|general-with-images message an example with the lighting... There's a lot more that can be tweaked to improve each set, this was just from me bashing out a big list of prompts with generic settings
Yeah
If you're really patient I could prolly get it to pull out all the mid frequency data too to really nail the reflections, dirt on the face, etc
I think it’s super impressive that we can find these techniques and use them so effectively. I worry about this week in particular and all of the news developing, because it’s a sign of precedent (i.e., a sign of what will continue to happen with every other open-source startup like Stability from this point on, until actual laws take hold)
What we REALLY need is for someone to break nvidias monopoly
That will take politicking of a different order
If vram cost what ram cost, we'd be able to crowd source funding for sd3
Crowdsourcing hardware could very well be a grassroots cyberpunk vision of the future, tbh.
We'd have cards with 96gb of vram for under 2k
The direction apple is going is pretty interesting imo
Just using ram and cranking up the bus speed
We don't need to have the fastest cards, we just need vram-liku speed not the 30x slowdown or whatever it is we get shuttling from Nvidia vram to ourmemory sticks
Yeah a100 is slower
no it isnt
But 80gb vram so it's univeraslly used in place of 4090 despite being 10-15x the price
You know, what could happen with gatekeeping models is that people generate leagues and leagues of images with those models, and then they take those images and make their own spinoff model from it using the same multi-diffusion principles, etc.
so, even if these models went closed source, after awhile folks could theoretically farm enough fidelity to make their own
https://forums.developer.nvidia.com/t/which-one-is-more-suitable-for-my-needs-a100-or-4090/252853/6 here ppl say that the 4090 and 4080 are faster. I wouldn't know cuz I've never sniffed an a100
(just musing)
Yeah ppl are doing that now with MJ and dalle3
exactly
Shit I've used that method with LORAs
an a100 gives u like 50 it/s on 1.5 vs the 30 ish it/s from a 4090
What a time to be alive. Heh
Yeah, I made a base model entirely out of the oldest color photography set back at the end of 2022.
Huh. What about sdxl, or espec training? The training is the key part here id think
I know there's weird shit with how Nvidia drivers or the 4090 idk which handle stuff too
I bet that stuff only gets weirder lol
Grotesque example is the 4090 is only about half utilized by training SDXL
because ADA cards have better AI support on their drivers
It's not the CPU, I thought it was but I talked with someone on the OneTrainer discord with a lot of exp who has about the fastest Intel chip there is
Significantly faster core by core performance vs my 5950x, whatever that chip was
And he had the exact same it/s and exact same issue
Drives me nuts
It will get very weird
Tbh if I'm one of the "powers that be" so to speak
One of my top concerns is AI malware
Yes, that is absolutely a concern.
Espec polymorphic code
Opportunity for the Terminator 3 scenario
Unlikely to be as dramatic
But could still be very destructive
Baking polymorphs into datasets such that a specific bias can be flipped to dramatic consequence is a total risk
Just like reality’s ecosystems, as the artificial ecosystem grows so do all of the variables that inform / structure it
Models will catch colds, the flu….COVID….the Black Plague, etc.
lol
I'm thinking in terms of bonnet malware that uses its distributed computing resources and an AI LLM to evolve its own code
Spam it into the wild, the hallucinated and buggy shit just dies the rest catches fire
Well, that shit is going to devastate by principle.
That’s the stuff that will be the most aggressively pushed back against first.
So, checkpoints actually got phased out for this reason lol. Safetensors became the new standard to avoid that type of scenario.
Can still be "tainted" while training, but only the nee model created. Can't really burry code in them anymore.
Big issue there is it doesn't matter what our laws are
But you used to be able to
But, much like how speedrunners figured out how to do code injection techniques, the same techniques could be improvised in this fashion
I'm thinking of states like NK in particular
Our reliance on global information infrastructure has never been more transparent than it is today. Having different countries means having different degrees of enforcement, so…yeah, how you gonna shut something like this down planet-wide?
You don’t
It just happens and happens and happens
So you can take the stance of “I dOn’T lIkE AI, etc.” but that is now equivalent to turning your back on a big boogie monster.
If you haven't heard about that search for the wired mag article for a good summary
That's ancient technology now
I’m 40
Ahh k so you've seen it too then prolly
🙂
But yeah espec with AI being around now this is way way way old tech
And that shit was crazy
Blew my mind when I saw it and took a look myself with ida pro
It was like finding a crashed UFO
honestly I am not even that hyped about SD3 anymore knowing that only the 2B model might be viable
Yeah, and the principles behind it are also old. I think people are looking at past precedent a lot right now as a means to understand the phenomenon of AI, but it isn’t working very well 🙂
hahaha
And that was done just by ppl with no AI shit and probably a small team on behalf of a government
and that model might look like what 8B looked like in it's half-baked state around very early march and end of february
Y’all familiar with the Gutenberg Press?
Yeah I just bring that up as an example of how sophisticated and sneaky malware can really get, and what state actors can produce with dedicated teams and the resources of a government... Access to treasure troves of 0day sploits etc
I am
Before 1440, you couldn’t copy anything. Writing was effectively holy. Closed source.
🙂
We are living in a Renaissance-like time, only it might be the end of the first one.
Yup
Sigh
Its the most exciting, promising, threatening, dangerous period in history
You know it
I don't think the future has ever been more unclear
Mhm
I’ve got a radio show starting tomorrow and I’m putting up a lot of music that has influenced me through this stuff.
First track is “This Is Now” by The Knife… good stuff 🙂
you a radio host?
90.3 / 98.3 FM Freeform Portland
awesome 🙂
Streaming worldwide at http://www.freeformportland.org
(volunteer-run)
I have a two hour block tomorrow, 10 AM - 12 PM PST
😄 AI voices throughout.
lots of good stuff
I’m making it like a Dr. Demento sort of old school radio show
With sound effects, etc.
i've def heard that track btw
I’m aiming for a lot of unique dance tracks and some indie rock, some hip hop too
sonuds like a good mix
i had some friends that had their own radio station in buffalo a number of years ago
kinda, uh, sorta, well, "shared" a frequency with a local christian music station
i remember once coming over a hill on the freeway and they had that station on and i was like uh why is this on? then it suddenly cut from happy hymns etc to some stoner rambling about how to remove piss stains from a mattress lol
(it was pirate radio)
lol yeah
have you done any motion projects?
only one tiny silly thing
Well that’s something 🙂
There’s the next big action set piece for Action Hero Movie #28372735852: “A bunch of masked criminals wearing extra-digited fingers, with some of the higher up criminals wearing cloned heads, attacking a…BANK!”
yuuup lol
that image is hilarious, and a joke, yet.... honestly, a good idea lol
and def where we're headed
get some ingrown hairs going on your forehead and it'll look like your hairline was edited by AI and maybe your face was swapped
it’s funny how Minority Report harnessed so much truth. I mean, Philip K. Dick was good like that
I think A Scanner Darkly is also super relevant right now
more than when it even came out
definitely
more philip dick... https://www.amazon.com/Philip-K-Dicks-Electric-Dreams/dp/B089VKJKJR
first episode blew my fn mind
and is going to be extreemly relevant before long
the rest are alright, there's another that's really good but that first one "Real Life" stood out
I really liked the episode with Janelle Monae
which one is that? tbh, i was deliriously tired and kept falling asleep in most of these
so my memory is hazy with a lot of em
it’s the metaphor for an AI-operated amazon warehouse that delivers stuff via drones
i can’t remember much either
oh, yeah, talk about prophetic
lol
Eventually you get numb 
MJ and DallE are ass
I personally like Midjourney the best
Is it safe to assume the bot channels will never come back again? I may have missed something, not sure. Will we ever be able to make content in the server again?
Who voted Midjourney? 
Just wait…the Noob Saibot of diffusion models is going to drop somewhere eventually
“A NEW CHALLENGER APPROACHES”
prolly the poll creator
how to use stable diffusion to generate image?
MJ is nice and all, but it's childs play compared to the world of possibilities we enjoy
if i had a free sub to MJ and Dalle3 but no local GPU, i'd get bored with image generation within a few weeks and be done with it
You can't, they are immune to damage in literally all games
once uve had a taste of the freedom of sd, its pretty hard to go back
Ooo a parody account
i cant believe sdxl has only been out for 8 months, i thoufh it was older
I did lol
Do you have stairs in your living space?
They are probably 11 years old
haha so it rly was the pollstarter who voted mid = )
I hope you wear a helmet when you’re changing elevation, being out in public, etc.
Why u assume that Midjourney is so bad and everyone that uses it is young
No I mean the other guy
you clearly have a few learning disabilities
maybe we have to dumb down our responses abit
I mean they are probably 11
you squeeze your step dads what?
If this guy was an AI model he’d be about 500m parameters
Which isn’t much
Is that some special year or something you’re all about?
Silly little troll
Go learn more words.
or she
This is for artists
Go tell your parents what you are doing and they will tell you it's wrong
what parents
"im gonna make you my girlfriend"
Promptists are now called artists?
imagination is key
Uh oh….here we gooooo
It was a question, I wasn't gonna say anything else
It’s subjective that’s all
prompting is an art!!! (1girl, big boob, intricate) 🔥 /s
so i see that ur a Ice Poseidon fan
technically this is for generative AI, it generates images that CAN be considered art, as well as artist consider a banana pasted to a wall art
lol
money launderers consider that art
Ok everyone. Lol the stable diffusion discord is funny
I remember just minding my business and someone just said slurs and posted the SS symbol and got banned
that was like a year ago lmao
It’s like a Twitter bot somehow found its way into a chat room
You left out some words up there
Nah twitter is better than x
It’s incredible what you can deduce from what is unspoken
Wait @native plover if that´s your nick...how you still alive?
Well they’re going to keep talking until they get bored
So might as well feed them
Right?
Feed the troll!
somsone on his older brother computer lol
LMAO
Someone is on the primary school computer
u know all about that dont you
The what? What’s that?
Anyone want to google that word he dropped?
Or is it just a misspell
he tried to spell a word and failed, prob couldnt reach the keyboard properly
lol
based
Who is your main character in Genshin?
Wario
thats for u
we know u are going thru stuff, but we still love you very much, its okey kid
i dont love u
What is that is that an inter net web sight?
Does that go to a web page?
With web page things on it?
i mean it says gore in the url
3k hangs there every day
its alright to just block, alex
what plugins do you guys reccomend i use if i want the AI to generate text perfectly
for logo's
i think he meant my mom
That’s true. I suppose I’ve had my fill for today. Thanks troll ✌️
Slava
Whew
Anyway, moving on.
I have never been as excited for something as SD3, so I'm really hoping Stability gets well soon
x2
I am excited for it, but…
same
I don’t expect much these days
haven't heard of it
have anyone heard anything from emad today? any update from him?
but now that I expect only to run the 2B model I might not even care as much
planned to be released like next year or what
It'll probably be so much inferior to 8B and we'll have to wait for Finetunes
No updates yet on Twitter
SD3 triggered the whales
did you assholes remove my post?
which post
no
?
it's there
what plugins do you guys reccomend i use if i want the AI to generate text perfectly
for logo's
ohh
i didnt see it
Rotfl
no idea
No idea man.
sad times
Stable Diffusion 3 can!
coming in here calling people assholes immidiatly lol
SD3 is just better than anything we have. The possibilities are endless for an open source model as powerful as that, it's not just hype
i said make a logo that said "logo" and it wrote "lioga"
Sounds like convolution alright
mods deserve it
is SD3 released?
like when mods randomly remove ur msg and u start gaslighting urself like
tf
didn't i type this
where is it
Because it promises too much for the “average” person.
It does something that a professional marketer already works their ass off for 100k a year to do
Those poor marketers!
I think it will be a massive loss for the world if SD3 isn't released in the long run. Just look at what a big open source revolution was led just by the release of 1.5, and it might have a massive effect long term compared to if 1.5 was never released
I believe it.
I do not.
no way man
I expect resistance at every turn
You camt hold back progress cuz some people will lose their jobs
nuera link just got implaneted in a human a
The world moves on. With or without you
i still dunno what the release date is yet. there is none? then why even hype it up
when we get new powerful gpus, sd3 will take over and become norm, like sd 1.5
and were talking about somoneone gatekeeping a llm
I think we’re members of the same choir, sir
lmfaoo
if it gets release open source
I am a baritone
my point
i have a baritone guitar
Nice!
what does this mean
Nothing lol
these idiots are going to release new tech even if it kills them.
oh okay lol
Well, Emad was trying to release SD3. He went from “it’s almost here!” to “I quit”
So…
it might still happen
The brick wall has been encountered.
The problem is that everyone boicots ONLY stable diffusion and not midjourney or dalle, or at least the majority of dumb artists...that probably get paid by openAI and Microsoft
Open source should never be boycotted, only closed source
exclusively not
AI cannot generate the connection of love between art.
Nonsense.
since some people have gotten hold of sd3 alrdy, doesent that improve chances that its going open source?
The same artists that boycott open source are using photoshop and didn´t say a word against adobe generative fill
therfore the real movement in the world is artist
i boycot openAi only
I can only speak for myself, at the end of the day.
No, it's not dumb, it's only dumb to be a hypocrite
oh okay. i see some AI guys being a bit harsh on artists in general, but it appears that's not what you meant
Humans enrich the data; synthetic data is still data grown from an exact pool of variables.
It's an adult male singing voice between tenor and bass (I am a tenor for example!)
I assume not.. 🥺
oh yeah, my guitar is downtuned a whole step
The problem are not artists @fervent thunder the problem is that some artist "steal" from other artists and nobody say a thing, but when an AI does the same thing (it its not the same thing), many artists lose their minds
yeah that's hypocrisy
It challenges the notion of stealing by showing a neurobiological representation of creativity within a synthetic environment
yeah it really isn't even close to anything new
well, there's inspiration, there's stealing, and then there's everything in between, which is kinda everything tbh
straight up copy paste is a bit diff tho
should of never releases colored pencils, its been a downward slope for artists since then
it's usually obv as f though if someone is traight up ripping someone off
yea
As human creatives, we think symbolically and diffuse our ideas with similar symbolic weight and bias.
That’s what creativity is. Remixing. lol
Similar technology isn't as dismissed, likely because it's labeled as "AI". Snapchat filters are fine (they are) but throw in AI and you might upset a bunch of folks. I went on a tangent about this the other day, and here's one of my paragraphs:
If the tools that had been created up until now didn't cause such a huge uproar, such as digital audio workstations, image manipulation software, video manipulation software, grammar checkers, and so on - why does AI do that? I think many people forget that AI isn't exactly "AI" - it can't think or reason. And it's human-made just like the rest of the software we use every single day.
Another problem is that this AI generates IMAGES , not art, the ones that say that generates art are the youtubers, and news sites that use clickbait so artist can fight with people to use AI and they get a lot of views, comments and things like that
“Improving” the model? It’s at least as good as Cascade from the look of all the samples.
invites to the test bot were being rolled out. we're still 2 steps behind where we were in the roll out
thats still the guy taking over after emad saying they are going to release sd3 open source, is it not?
Sometimes it feels like many people have no idea what they are talking about either...
unless my eyes are broken
They are probably doing what has been suggested earlier, a lighter weight variant
i dont believe wer'e at a level where we can compare image quality anymore. or human preference of each images. Prompt comprehension and how easy the transformers are to further refine, is the big concern
cascade is still a unet model. can't compare that as easily
I just have this instinct that there is some homogenization happening on the dataset level
True.
consider that it has 8b parameters. a dozen images aren't going to reveal many truths about the capabilities
I’m basing this off of all the drama, as well.
no doubt i'm worried too
I wonder what kind of dataset used midjourney for their v6, if it was just quality or a lot of images
can someone who has twitter link that tweet in here?, so we can confirm it?
I guess time will tell.
photo shop programs are full of ai.
fillers, ai to stright lines, ai to correct smudges, ai that predict the intentions of the artist and assist in it
AI is 100 years old
my guess is their partnership with render network has them holding back any information until render gets their shit together and reveals something
yeah cause it was developed 100 years in the future and they sent agents back in time to recreate itself. we know. we all saw the documentaries
AI was invented by a man named Gustaf Karllson in 1927, out of a potato control net!
wrong
why shouldnt i boycott open source ever?
? are you crazy?
boycotts don't work anyways. not in the mass connected age
What are you talking about
their problem is that, it bein open source, every idiot, even me, with half a brain and enough computational power can gen images
thats true
boycotts rarely work
That is not a problem
individually, we can boycott stuff, but to what end?
It’s about the prestige of the experience as well. That’s something people seem to want more control of, in the end.
i boycott gnome because the developer community around gnome decided FOSS principles don't matter anymore
what is gnome
thank you, appriciate that
when something became too diffuse, for who work with such stuff, is a problem
Aaah you mean to people who work with that, in that case, maybe
oh
i boycott emacs because vi is better
ohhh wait.. holy shit. emacs, emads, coincidence?!
or the lack of regulation
lowering the barrier to entry. often is a double edged sword
Not really, regulation is ALWAYS bad
Stability AI still has big issues to solve
Its used by all the government for censorship
No
or do you need extreme examples to showcase you that regulations are needed
Why do you think you cant generate putin or xi jinping in chinese AIs that run online? just to give a dumb example
no just with gen AI ofc
recorded audio, like music albums... so much better sounding back when bands needed a literal audio engineer on their team to produce an album that sounded good.
now it's just push button auto tune and only a small percentage of recording artists care about the nuances.
AI is very slippery. We evolved out of water, we have a tendency towards acting slippery
wdym with slippery? Flawful?
If one person tries to stabilize AI, another person will destabilize that
And so on and so on
It’s a slippery slope
intelligence in general is just very coy
i mean its far from being "too good to be true" as of now
Depends on the application, my guy.
in art (both 2D and 3D), in coding, in other areas
just corporations covering their ass
To not get affected by the government
its always the corrupt government fault (and corporations too, of course)
those that collude at the highest levels are trying to project this idea that corporations care about the health of society
Actions always speak louder than words
they may or may not but they have also obligations towards investors and shareholders
The older I get, the more I know this to be true
Nah, corporations want money, they would sell people if its legal
https://www.youtube.com/watch?v=dmZSGNW-QCU oh hey, speak of the devil. this documentary is now free on youtube
regulations are always and only in favor of the rich
and Stability AI is in the position it is because they couldnt drw more investors with more money to fund
draw
so would you
I wouldn´t
Nah, he wouldn’t and neither would I
"The Corporation" covers their pathology very well. the one i just linked is the recent sequel
Don’t project your own desperation on others
ofc you would
No, i would even starve to death instead of stealing food
doubt so
We all have instincts, Infidelis—but where do you live that you might feel this truth with so much conviction?
So you would do it? @potent spire
last i checked, stability was doubling down on block chain businesses. i think they'll become a shell corp and extract their value from block chain grifts
what i wrote is ofc partially what i mean and partially playing the devils advocate
sell people simply for profit? No. Stealing food if i was at the verge to starve? Yes i would in a certain situation
let alone if my family was starving to death
i would go well beyond that in that case
i've stolen food just because i was drunk and hungry
the video can't be seen in eu or usa. maybe in brics zone?
told mcdonalds they fucked up my order, but you know what? i never ordered
FBI open the door!
What video bro? 👀
that one
I for one am grateful we’re not at the level of this discussion being a reality
lol
It doesn´t work where i live
its a good documentary. will be available somewhere in your country. plex hosts a version
yeah, me too
Not even yt-dlp with a vpn can download it, so strange
you mean starving to death?
2 hours long. the original is good too. Analyzes how the modern corporation is a psychopathic entity
i'm jumping from a country to another, still no result
Yep, I’m not quite there yet.
well im glad im not there either lol
canada and us probably
Okay off topic but why won’t Pam from Cannot Be Tamed wear funny costumes now and then
usa doesn't work, netherland neither, maybe canada?
i'm in canada and it works. so if it's not then youtube is just detecting your geo trix
thats how it always starts
https://watch.plex.tv/movie/the-new-corporation-the-unfortunately-necessary-sequel this should work. plex is pretty borderless. maybe they've changed though
If only they released consumer grade GPUs with more vram 🥺
I’m sorry I said Willy Wonka, but I think I mean Aperture Science
with that attitude yeah
💀
lol
you just need to believe
but yeah mostly, existing stable diffusion models sort of use the text like guide rails in an open plane of latents. rather than like a specific track
prompt comprehension isn't great
theres some control
u tried it buddy
text with controlnet isn't that bad, at least judging by the 2 or 3 times I tried
There are ways to make text happen, but they involve more artistry
And those tools already exist in droves
But AI can get you 95% of the way there
people want that startrek holodeck. Where the computer never misses a beat and knows exactly what you want, even if you just give some vague description.
It’s as if they want something that does more than wait at a command line
we're not there yet. its more like an intelligent paint brush. the artist still has most of the agency
🙂
people are actively inducting reasons towards autonomous functionality as it relates to the user experience, etc.
and then powerful folks like musk are deducting from symbols of futurism like “the robot” and “the robot car” etc
like that one headline about how tesla is ready to shift to humanoid robot workers—like how is that not just clever marketing? how on earth is a humanoid robot more efficient than a specialized design when it comes to assembling an automobile?
The humanoid robot is just a show off , i dont think it would be useful for anything related to moving heavy things
Human body is pretty flexible ...
So if you need an allround robot that does human work ... ... ...
it’s his way of seeding the robot concept further and further but keeping it anchored to his brand identities
but, like the cybertruck, it’s his concept
🤔
Boston Dynamics said that for them it's too costly producing humanoids for single tasks, but also said there are other people that can make them and make it profitable
“profitable”
can we just take a moment to realize what are bodies are worth? i find it amazing how everything we created was inspired from the human body.
fuck AI we have souls
if a soul exists, it can be engineered
True 😁
self is the ego side of soul
self is the lie the evil sells us
theres no magic in reality. if it happens, we can machine it
in obtain for our soul
We could achieve a lot, thats true, but that doesn´t mean that we can´t do it more efficient thanks to technology

