#💬|general-chat
1 messages · Page 176 of 1
i use bmaltais kohya ui, and taggui projects to train flux loras
does anyone know a comfyui node that takes a 1-d tensor as input (or a list of numbers) and plots them as a graph?
MTB nodes
I don't mean to make a plot as an output image (I would just use matplotlib for that). I mean having a plot for visualization in my workflow.
The checkpoint is the model
So you can’t run two models at once
No but if its a merged / you merge them its basically the same thing
can you send me a link? Flux gym is not working
https://github.com/bmaltais/kohya_ss
https://github.com/jhc13/taggui
two separate projects
Good Morning Everyone
ok guys i downloaded flux and got everything working and gen'd pics. and theyre really good. buuuuuuuuuuuut theyre not that much better than dreamshaperXL or epicRealismPureEvolution. the quality is better but it becomes a wash basically as i put it into kling and degrades the quality and including the time it took to gen im thinking of switching back. what do you think? am i making a noob move?
how many steps are people using for renders? i see 20-40ish being normal but some of the high renders on Civitai being upwards of 1200steps (on my comp i think that is like turning it off for 5 days)
just want to ask, is it possible to install hallouAI or other video AI for forge/automatic1111?
it's much better. You see the difference if you use more complex prompts. SDXL has poor prompt understanding and makes a lot of anatomy and consistency errors in comparison to Flux.
Guys, I have a question: what image to video method would you use to animate a boy's mouth for a few lines?
Flux is too good
I also disliked it at first, you can see old comments of me saying I was gonna stay on SDXL lol
but Flux is just such an aesthetic model with the way it does colours, shadows, lighting, blur and particle effects
and various people have actually managed to pull off putting CFG back into the model, which I thought would not be possible
CFG is not that important. I find that good prompting, adjusting weights and using loras and other techniques gives you much better control over the output. Using CFG or even guidance is just trying your luck. I keep those values at a sweet spot and focus on other parameters.
if you are using a technique that works via the negative then CFG is required though
its not always optional
Yea negative prompting is the exception
at least with the current models I do also think you get a bit more quality out of them by running them with CFG rather than without
Has anyone used segment anything video version
Hello Please Read.
I would like you to imagine getting home from school and only seeing rubble not knowing where the street is that leads to your house is it's all just rubble. I would like you to imagine dead kids all over the streets from starvation and diseases from not drinking pure water. I would like you to imagine sleeping at night hearing explosions at the distance getting woken up. I would like you to imagine a finding out that your home got Obliterated by a rocket. I would like you to imagine how that feels. Because that's not imagination. That is a reality there are mass genocides happening right now in so many places. It is really tragic, but we can do something to help these people. We can:
Donate to charities (From a trusted source) that give food and clean water to people in need in these places.
Spread information about what's happening in these places.
Sign petitions or take part in protests to cease fire in these places.
Stop buying products from producers that support this genocide.
If you can't do any of these things you can tell someone who can do these things.
But I hope I could've made a difference.
Free Palestine.
Cease fire.
yep, its very fast and good but I think an improved version called samurai came out too.
hallou is not open source, hunyuan is probably what you are looking for but no image-to-video yet. Also, for most devices its going to be very slow.
Ltxv is decent with image-to-video but considerably behind minimax, however its really fast and still decent. On a 4090, it can take as little as 6sec to generate a 5sec video and on h100 it only takes 2sec.
what's the best way to create video on auto1111? I only found animatediff
yeah hunyuan is very slow so be warned(25mins on 4090), but it is best quality for sure, comparable with stuff like sora now.
ltxv is lower quality considerably but its really fast, can be as low as 6sec for a 5sec video on 4090.
I believe swarmui and comfyui support them.
-
charities are often fraudulent endeavors and the money isn't actually going where you intend it too.
-
Information can commonly be skewed by biases.
-
petitions are just nonsense tbh. Often are just bait for fraudulent charities
-
boycotts dont work and you're likely buying many still
-
telling people to care about other people's problems that have no correlation to their own personal issues in their region is a pretentious thing to do.
-
palestine needs to be freed from hamas first. because you know, terrorism.
That is all.
I have nothing against muslims like i have nothing against other religious people. Christian sects like Westboro baptist i think are not genuine and greasy people who thrive on hate. I feel even more strongly about Sunni sect of muslims. They're not true muslims like WB are not true baptists.
sir this is a stable diffusion discord politics shouldnt even be discussed 🙄
didn't see you speaking up to the other one. someone had to do a reality smash on it ¯_(ツ)_/¯
those mfs killed our aid workers and no one has ever been arrested because of it, netanyahu is a war criminal and if he ever steps foot in our country hes gettin arrested
Lol it's either a bot by some activist group or the guy is so delusional that he thinks shitposting on discord is gonna change the world. And yes charities are a scam. Every business will always converge towards the most profitable business practices and dishonesty always makes more money than honesty. And it's ironically people like him that make these fraudulent businesses thrive.
Same bs that got internet archive targetted. 🤦♂️
The thing with publishing copyrighted books? Yea that was weird. As if people don't know how to pirate books themselves...
No they got hacked as a protest move. Youre referring to the pile dataset used for training LLMs. Pretty sure. But few months ago Internet Archive got targeted & had all users credentials exposed. Like in civil protest.
Yeah the Pile case was over GPTs & stuff being trained on public data, & some non public data. I have the stats somewhere but id have to dig. A lot of it was legal public content. The main thing was it was scraped (illegal aspect of it) from a premium 3rd party hosting service of books & stuff like that. Some user, scraped that entire set of data. & thats how The Pile got it. Think all LLMs used some version of it. Atleast GPT text based ones afaik. Harder to get books. Which lead to the lawsuit. Pretty sure even nvidia used the datasets. 🤷♂️
Think the service closed doors after
Internet Archive went down from a JS library exploit but it's all been improved upon since then afaik. But the users are all in HaveIBeenPwned now.
80-800GB of data (im guessing from memory) was used in training the pile. That was suspect or looked at in illicit manor. Rest of it was legitimate.
HaveIBeenPwned has more users than Facebook
dose anyone use amd gpu to run stable diffusion?
Think there's an AMD guide some where in here
I cant remember the number of users whos emails n pass hashes were leaked from IA. Pretty sad tho. Neutral party support peace & knowledge for all, hit by that stuff.
Does anyone know any nodes to mess with the latent layers (after inference)? Has anyone done any experiments like changing the weight of specific layers?
latents only have one layer
Maybe I'm using the wrong term here. The Flux latent has 16 "layers" for lack of a better word.
oh yeah the term you need is channels
on SDXL and SD 1.5 I used to manipulate channels a lot
but for Flux its hard to know what each channel does
pretty much SDXL has two colour channels, one for intensity and one for pattern/structure
is there a flux paper yet?
Interesting. One of the Flux channels looks like a lighting or highlight map. Would be interesting to see how much they contribute to the final image. So do you know any nodes that let me change their strength?
time to see how intuitive comfyui is
i've been using a1111 and its derivatives for years, i'm going to see how long it takes me to produce an image on comfyui that would be the result of my normal workflow (generate image, upscale image, use adetailer on the face) without any guides
I do recommend looking at swarm ui aswell
It has a comfy ui backend AND a nice front end you can use
And support is pretty good
i've been using the old version of forge for so long 😭
because it's still faster than the latest version of a1111
I switched from a1111/forge to swarm really early and its simply the best choice ive made
I do got a seperate comfy ui setup but thats because swarm doesnt like comfymanager
We will publish a more detailed tech report in the near future. their original flux blog post said this
i would just switch to the latest version of forge but it breaks this extension that i get a crap ton of use out of, where it's a few buttons where you have resolution presets
lol comfyui is not intuitive. every new workflow has the author's own peculiar way of using it. theres no standardization. it's just spaghetti and nodeballs
that's what i assumed, which is why i've stayed away from it for so long
i dont know how i set it up wrong but swarmui doesn't load flux for me. it loads the nf4 version after it prompted to install nf4 support. and that works, but is slower than forge ui. i bet i'm just doing it wrong though
apparantly there's comfy desktop now. a more robust app or something
which quant of flux do you think i could run on a 4070 Super? i was thinking probably fp8
i just got this GPU a few days ago and it's a monster for SDXL compared to what i was using before
4070 has the hopper transformer engine so it handles fp8 loading a lot faster. it should be good to go for that. i have the 4080 and prefer the nf2 or fp8 version
there's not a lot of difference in the inferrence for it
i keep feeling like i should have spent another $100 for a 7900xt
i think it would be slower but it could handle larger image and text models
since it has 20gb of VRAM
llm quants work well. and you can always rent cloud gpu to see what you were missing out on. $100 would go a long way towards cloud services
the 7900xt would also be better for gaming i think
well, a bit of a way
debatable but probably. if you're just doing 1080p or even 2k gaming, i don't think there's much benefit of one over the other
back when i was still on my GTX 1070, i used paperspace to try out SDXL
i had a vega64 when i began my sd14 journey. i was using collab and then eventually figured out how to run auto1111 on my arch linux setup
v64 was slightly better than a 1070 and slightly worse than a 1070ti, but had no cuda. so you were better off than I
paperspace gradient is like $8 a month and you can sometimes get access to an A4000 for 6 hours at a time
it runs jyupiter notebooks though
there's always tons of people converting every new LLM model to GGUF quants. its a good scene
I used this for SDXL but its not been set up for Flux https://github.com/Extraltodeus/pre_cfg_comfy_nodes_for_ComfyUI
i can experiment with them locally really well to learn all the particular settings. but if i really wanted to do one seriously, i'd rent cloud space for it. then use the knowledge and tricks i've picked up locally on that
flux probably won't run on the old pre-gradio 4 build of forge
so i'm installing swarmUI if that can do it
i also have the latest automatic1111 webui as a last resort
i doubt that the actual forge project is going to get a lot of updates going forward too. i think it's kinda dead in the water and the author has lost his enthusiasm since its gone unappreciated
there's the reforge project which i think is more updated than auto1111, and has the comfyui backend so it's faster.
but it don't do flux
Hey there
Hope you are doing well
I am a senior full stack developer who has full experience in web development | design and AI development
So if you have some projects, please let me know
thank you
What website can I use to train my own checkpoint?
Civit
I can only see train Lora
oh you want an actual checkpoint trainer? thats local
why do you want a full checkpoint? lora could possibly achieve what you're looking for
Yeah, I want to try to find a way to convert my current SD 1.5 Checkpoint into a Pony one with own training
I am using my current SD1.5 Checkpoint that does my style atm. Like how the images turn out.
Lora is more just to make it look like another character right?
I am using this one "https://civitai.com/models/25132/tfm-cutesy-anime-model"
and I wanna update that oen with those image I have done but to a Pony checkpoint
lora is like a tiny checkpoint addon that helps reinforce what its trained ontop of the model's knowledge
Or should I just do a Lora instead?
i would yea
no sites that i know of (considering i didnt look hard cause loras is enough for me) that lets you train full checkpoints, i know theres local methods but you need allot of vram
you can loras on cvivit for cheap if you dont have the hardware to train locally
What settings do I put so that I train a Pony Lora?
Pony is just a variant of a sd model
If you train a lora using autismmix ponyv6 for example for a certain style/concept with the proper pony tags its a pony lora
But i would check on civit ai articles/guides they are pretty good
so only comfyUi for those? Guess i'll have to try and install it too
Iirc swarm has it supported too
is there an extension or prompt method to apply a prompt only to one character?
like only one of the character having glasses
Depending on your ui. Base concept is called regional prompting. Extension is called “couples” but comfyUI has nodes for regional prompts.
Damn Hunyuan video is good. Wish it has image to video.
cant believe it runs on 12 gb vram now
hello
doesn't run on 10gb i guess?
Can I DM you about training Lora's?
idk, you can try....
i heard rumors of ti working on 8 and even 6
created a game, simple as idea but hard to play
https://guessthepainter.vercel.app/
Give it a try 🙂
I'm a newb and i may crash and burn but can any point me in the direction of a site/pdf/doc that I can learn about training AI's to generate images of a specific character. I want to take an existing character and create images or inpaint it onto someone. Wish I could ask the question better but no idea what doing
i think 'training' is the write word but not sure
Very frustrated with blowing through tons of buzz credits on Civit.ai trying to get concepts and image styles to work. I'm mostly using pony (and variations therof), but it's woefully garbage at understanding some pretty basic concepts. I've wasted HOURS on concepts like "woman riding a horse" (which utterly baffled it, with some really nightmarish results), "woman on a hammock", etc. What is your approach to iterating concepts without blowing through credits? Do you start with a simple prompt like "1 woman, laying in a hammock", generate that, and then go from there with more and more details? Or do you type in the most accurate detailed prompt you can on the first try, and if it does seem to undestand, change base models?
Then I find a draft image I like, and I go to remix with seed and up the sampling rate, and the style ends up being really not what I expected, even based on the LoRAs I'm using.
i've been playing with generating images to 'find something i liked' and that can severly expensive
im new at this so get a vets opinion but you talking alot of do
I've only been at this about a month
have you thought about doing comfyui or some other ai generation app that you can do for free until you see something you like?
pick the mods, checkpoints and other filters you want and keep going till you get osmething close. im using one ethat i can generate 4 images in about 15 secs...the aren't great but give idea on how character and background will look
can i post a pic here to show you?
Aero: sent you some messages. Have to afk a bit. GL 🙂
Any1 to tell what AI program this account uses? https://www.instagram.com/pia.nea8/
I tried comfyui, drawthings (I'm on mac), and regular SD webUI. Drawthings, i just get mostly terrible results overall, maybe it's my settings. ComfyUI works well, and so does regular webUI, but I also run into memory issues when using those, even if I'm not using a larger model like flux-s. I want to revisit Comfy when I have a bit of a better handle things, and also once the dopamin drug-effect of churning things through Civit.ai dies down.
Anyone using Kohya_ss and could help me with how to setup stuff in there or to give me guides? 🙂
Why am I unable to use certain models after I downloaded them and put them in the model files?
They are available but I can't apply them
First of all try to lookup some pony prompt as most start with: score_9, score_8_up, score_7_up, 1girl, solo,
To avoid low score images. Also the tagging is different from the sd1.5 keywords….
Second you could try using diffusion bee for Mac (download current version from GitHub) it uses low quant version of flux and runs on apple silicon.
Yeah I have all the quality (and negative) parts of my prompts. I've made some very good looking stuff that I'm happy with, but sometimes it's really frustrating blowing through credits trying to get it to understand just a composition concept. I'd share some examples, but the stuff I'm working on right now is NSFW, and I know the server rules. 😛
is there an online service that has good image editing capabilities with stable diffusion or something like that?
So on comfyUI, is there a way to move multiple LoRAs in and out, rather than plopping a bunch of LoRA loader components, connecting them all, etc?
I've tried a few dragon ball/toriyama lora, but i can't get the look at all unlike other lora which capture the style quite well... what am i doing wrong? there's no trigger word for them, aside one, and i'm using it
rgthree node pack has a nice lora loader can add many and switch ones on and off easily in one node
Morning
I'm trying the quantized Flux Schnell in Invoke, and it needs the bitsandbytes module. Anyone know where I can grab the right thing? Im searching Huggingface but not entirely sure
Just do pip install bitsandbytes
Yeah I did. Do I need to restart invoke?
Edit: think I got it, looks like it wasn't actually installed the first time. Then had to reset InvokeUI.
some people on civitai have a full on checkpoint, and the same style but as a lora, and the pictures i've seen on both looks super close to each other, so why would i use the checkpoint instead of the lora which is way smaller?
looking for pros and cons here
some loras can conflict with eachother and some can influence other in ways not intended
checkpoints are generally more braod
broad
Ah it's weird, it's still saying the module isn't available.
that is strange, you're installing it after activating the venv masnually?
also are you using stability matrix by chance?
.. \ .. \ venv \ scripts \ activate.bat
(without spaces) then pip install bitsandbytes
try that
Yeah says. "already satisfied"
Me: "What is DiT?"
ChatGPT: "It seems like "DiT" isn't a widely recognized term, and I couldn't find any specific information about it"
Me: "You know!"
ChatGPT: "Indeed, I do! DiT stands for Diffusion Transformers."
Hey when using flux why does inpainting take so much longer than img2img? Shouldn’t changing the entire pic take longer than sections?
because when we look at model size in terms of parameters or gigabytes, that's only an estimation of how heavy they will be to run
different pytorch operations and different blocktypes have different running durations
Anyone here buying a 5090?! Let's get some hype going!
I don’t have enough in my piggy bank 😭😭😭
I think NZXT said you could win a fortnite tournament or something
If ur buying it not for gaming and just ai gen I would maybe wait for an ai targeted card to come out within this year
Whats the price point on 5080 & 5090?
expensive and more expensive
It’s just rumors they say they don’t decide the price until Jensen basically gets on stage, but they’re saying $2.6k for the 5090 and $1.4k for the 5080
😭😭
$2.6k seems ridiculous........
With tax that's easily close to $3k
Before gougers too lol
Can't wait to pick up one of these for $10k on ebay
True
I wonder if Nvidia will have a trade in program.....where you can trade in one kidney for a video card.
I’m not good enough at fortnite 😭😭😭
is there any shortcut do activate/deactive adetailer?
T-T i was gonna buy a 5090 but i guess not
Never give up hope!
https://www.theverge.com/2025/1/6/24337530/nvidia-ces-digits-super-computer-ai new $3000 mini pc with 128gb of unified ddr5 mem. alright then
it's only $3000?? that could be pry win... if I actually had that much extra money
need some 3rd party benchmarks for all this stuff.. but I bet that min comp would be epic for image generating
starting price. but it's a new form factor that i expect ot make a huge disruption
the digits pc is not the 5090
oh you were replying to the rumors blu talked about
i wouldn't expect retailers to sell non founders cards at MSRP. they'll all be jacked up with rgb bells an whistles, and traditional price gouging.
MSRP is just recommended
also, incoming tarriffs. these cards aren't made in america
https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/
the fp8 support on 40 series has been good. now we gon see fp4 support at the hardware level. such a boon.
Hmm im just unsure if to grab a founder card or to wait a little
I dont like the rgb stuff, just want a slick look
just budget a little higher. all the founders cards will be scalped in a quick minute and then marked up to what the market value is
preorder from Nvidia if they do that again
i dont think they allow preorders. and if they do, scalpers will buy up all the preorders in a quick minute
ya i dont know anymore what they do, i did that with the 2080ti
market value is market value. theres no sense fighting it. it was never a mining problem. it was always people capitalizing on the value of efficiency
Does someone else totally geek out about the technology? Like the techniques you can use and stuff like that have become a special interest of mine and I sometimes just unload it unto friends :D
this is a server dedicated to that so , probably.
friend's dont like info dumps so much. i've been told i do it a lot
Pfff basically all my friends have autism as well so they at least understand :D
Its usually a friend and me dumping info at each other
vibe
But I dont have the balls of posting that stuff on BlueSky because I follow a loooot of furry artists. And furry artists hate AI with a passion
So I feel the itch to share
hello
there was a LOT of money to be made in that realm, and now people are commissioning a lot less
I know :D
I am in the community myself. And I have seen the ripples
guy i worked with was dating a furry artist and that dude was banking cash. especially during the pandemic. just hauling it in fist over fist
And now you can just kind of get that yourself
on demand furry art yup. lol. not my style.. i'm not really into it at all, but i respect the fun people have
its not just people who want stuff doing it either. other people are answering commissions and now the market is flooded with other artists. but they're all using ai
I can recommend at least playing a bit with it. Since furries are unique with textures n stuff I think
so is a lot of art. the anthromorphs i like are dogs playing poker. cats in hats. the classics
animorphs were popular books when i was growing up but i was always a goosebumps lover
Hey guys! I am an aspiring content creator and was just getting into the realm of ai. Would someone be able to help guide me into what sort of computer I would need in order to run stable siffusion? Would I need a desktop or could I get a mini pc? Any suggestions would be greatly appreciated bc I am not the best with computers. Thank you 🙌
Afaik you can run it on a potato :D
The better your tech the quicker the generation basically
people can manage it with laptops, macbooks, amd systems, but really... i just say go full system. Get a desktop, get a proper nvidia dgpu, lots of ram, and a fast SSD
you can also rent hardware on the cloud if you want to dip your toes
keep in mind that linux is king in the realm of ai too. Python is the language. you're going to want to learn system terminals.
what does this mean exactly?
gotcha, any suggestions, preferably budget friendly?
https://blog.runpod.io/runpod-slashes-gpu-prices-powering-your-ai-applications-for-less/ runpod is doing this big pricing slash thing right now. its a pretty competitive field
okay so essentially you just pay a certain price for a temporary boost
suggestion is to break the budget and buy the new upcoming 5090. or the cool new nvidia digits pc
get a bank loan. max your cards. its fun
you only pay for what you use, and it's often comparable to what electricity would cost you to run at home
a lot of content creators run their generation stuff on cloud services. depending on how much power costs its the easy choice
interesting, so for example a mini pc wouldn't run it very well?? like a Mac mini or beelink?
not very well at all. you need a powerful ML specific chip with lots of memory
macbooks have the neural engine, but there's not a ton of software support
okay so when you say buy the new 5090, you mean to purchase a desktop and the GPU in addition?? sorry, I'm a bit slow when it comes to computers 🤦♂️
https://www.nvidia.com/en-us/project-digits/ nvidia just announced this thing too. its a mini pc specific for AI that costs min $3000
I've personally run it on windows with like older hardware (10something nvidia card). Its slower but its very possible. And the key to a good image is not the speed of your PC
oh yeah. go hard. biggest desktop box you can manage. build it youreslf. get in over your head
I personally think you should try to get into it as soon as possible as cheap as possible and scale up from there
yeah windows environments are not unmanageable
i'm on windows myself but sometimes i run WSL2 (linux VM on windows basically)
Why xD
Ah
okay, are there any resources you know about as far as how to buy pc parts and build it yourself?? I'm a complete noob
security
damn 3,000 for a mini pc
Oh that has me even more curious
yeah but it's 128gb of memory on a blackwell gpu
we're installing a lot of random scripts and custom nodes. i'm trying to be better about it
Try looking into stuff like target pcs or something. At least in my country you can get solid starter pcs for cheap. And then an older (used?) GPU
Oh wow. But why?
attackers want to get viruses on AI capable machines so that they can hijack the compute tehy're capable of . there's been a few situations where keyloggers or miners are installed
No I mean why the scripts
when i was first getting into video editing i bought a prebuilt lenovo ideacenter from bestbuy on black friday clearance. I know how to build pcs but that was cheap af
every model you run needs inference code. those are scripts
Yeah exactly. And you only really need to change stuff like maybe a bigger hard drive (for more images saved) and ofc the GPU
Huh. But like is the A1111 stuff not good enough?
a lot of hobbies or specialized tools generally won't be cheap
Buuut as with any hobby you can slowly scale up
a111 extensions have been compromised in the past iirc. and it's not updated often so i'm trying ot get back into comfyui and deal with all it's quirks. A1111/forge is very limited in it's scope
What extensions T-T
i can't remember. it was dealt with overnight when it showed up. i might be misremembering. there's been a couple comfyui custom nodes that were viruses too
but anyday one could show up. they're scripts. anything you run on your machine could own it in theory
I am personally using default A1111 mostly :D
Theres some stuff you can practically only do in Comfy.... but I use StableSwarm for that
i use a mix of forge ui, swarmui, and fooocus
sometimes i play in 0most for funzies
True just wanted something as basic as possible for starting out
buy a second hand 3090 when people upgrade to the 5090
And thats great I think. Dont pick up tools you might not need. What if you don't like this at all :O
still recommending a full on desktop even for a basic start. since desktops scale up better than laptops or minipcs
its like lego its fun
And cheaper
Yep exactly
here is what grok suggested:
cpu: Nvidia RTX 3060 12 gb
cpu: AMD Ryzen 5600 or Intel i5-12400
RAM: 16 or 32gb
Storage: Samaung 860 QVO
Motherboard: B450/B550
PSU: 550W-650W 80+Bronze or higher
Case and cooling: NZXT H510 or Cooler Master NR400
I dont know much about the specifics. But focus on affordability :D
You could basically ignore anything but RAM (which is cheap), GPU and Hard Drive Space
12gb of vram is okay but 16 is better
24 is the best you can get at home without going overboard and getting datacenter grade hardware
32 soon though 
What do you even do with fast generation like that xD
When is this coming out? Do we know the specs how long with img2img take on flux?
Many images, or just video gen at nice speeds
Ok video gen makes sense. But I can not see the need for that many images :D
I make like 1-2k images a day so for me it would save time lol
And using it to generate images for new study material in edu (hand edited after)
Huh why though?
Except for like video, which I can understand somewhat
Comissions, backlog for a eventual patreon, and certain assets for "games"
Even like 20 images a day would be way too much for that though :D
Yeah but how many of those 20 are great
But then I would recommend training your prompting :D
My prompts are fine lol, i get really good results but sometimes not exactly what i want
Then your prompts are not fine xD
Certain concepts etc but its also useful to train eventual concept loras
The model doesnt know everything yet
Like I can get from idea to image in like... 5 or something in-between images so to say
Not literal thousands
Certain background layouts, colorschemes, skin colours, and wildcards
And the bane of my existence, proper multi characters in one image
I mean if you really do want something very specific AI is not the right tool I think
And not to forget its also fun
Ai is better at spontaneous creation
Yeah i know, been using it since 1.5 release
That part I can get. But not like 1k images a day get :D
God forbid a man has hobbies lmao
I mean you do you. But to me it seems like theres flaws in your process that lead to waste :O
Huge huge waste
Mostly experimental gets waste tbh
So much waste that its more in a day than some users have created put together
But its all digital so its not exactly hurting anyone
But my autism kicks in :D
Trying to get mostly ai feel out of them. 40 variations of a after detailer so i dont have to do it manually etc
Inpainting*
Some ai images get this soulless look and i really dont like that so i try not to have that
Or the blank eyes, the way certain elements interact etc
I get that :D
But like... comparing 40 variations of an image before changing the prompt?
No
Lets say 10 variations of a single prompt
I like 2 out of the 10
I impaint the things i want to change, and let it run for 30-40 variations of the impainting process
Because impainting still messes up hands etc
Ah I see
So i dont have to babysit it
Inpainting is the issue here imho
And if you do that a few times you hit 1-2k easyly
I just... go with the flow so to say
Yeah but its no biggie, i just dont want to babysit everything
Wouldent call it a issue
Because of stuff you mentioned with like hands n stuff I personally think its too much to really work with. For me
Its much better to try another variation or other weights/ keywords
Its not that big of a problem, sometimes the hands come out fine without any problem
Yeah exactly :O
From the getgo
But if i do game cg/assets well you need a LOT
Rpgmaker/renpy games comissions etc
Depending on what you do :D
But depending on what you do AI is also a bad fit
Even had a few html
Eh if they request it knowing it its not my problem
Also using it for new study books at my work as theres been an budget cut again
For like stock images?
nvidias new 3000$ pc seems a banger, able to run 200B parameter models, so im assuming it will also be able to generate 15 flux images at the same time
That, and just generic illustrations to liven some stuff up
Same time? Nah lol. It can still only do 1 at a time. But it's the question of how fast each image will generate at. And the biggest question, can it effectively convert a flux to TRT this time lol
hey guys im using flux in forge with a 3080 and my generations went from like 3-5 minutes to now 20 minutes. i dont know what setting i changed if anything. i have 64gb ram and its utilzing almost 100% when im generating something
am i doing like something obv stupid that im not seeing
Hey, you need to use the smaller flux models (fp8 or Q8 gguf). Or Q6 as a 3080 only has 10gb vram.
Also dont use the t5 fp16 encoder. Better use the t5 fp8 or gguf q8
appreciate the reply bro. ya i saw something similar on threads online. but im curious on how my build was able to gen pics before reasonably and now its taking forever? it actually just went back to normal too after i closed stremio and some tabs. could that really have made the difference?
You need to close everything that uses GPU.
So streaming stuff, games, Wallpaper engine and browsers tabs should be closed
will see, but the way i understand it is that the 3000$ pc is equivalent of 20 nvidia 4090s
the caveat is that its only for ai models, not gaming
hence why i think that if you can generate 20 images at the same time in 20 4090s you should do something similar in that 3000$ beast
Aye. I know it's not for gaming :P Link to it actually, wanna keep an eye on it. As if it will release in a "formfactor" where i can install regular server OS on it, i might consider getting it as my unraid server, and host LLM's/SD on it 
I always forget how much denoise to use- if I'm taking a illustration-style image and I want to upscale it with a photorealistic model, should I use less denoise, more denoise, or in the middle?
BFL is releasing a specific fp4 version of flux optimized for blackwell hardware. Not like that'll matter since there is 128gb of ddr5 memory the gpu has access to. It'll load the full weights just fine.
Hard to say what the specs are. The blackwell architecture is going to be a leap. This should be faster than a 4090 if i am seeing it correctly.
There is a good Int4 version of Flux here too, its been out for about 2 months. They have FP4 version as well, but the Int4 version is useable already without new GPU. https://huggingface.co/mit-han-lab/svdquant-models
There is no rule. That totally depends on the illustration and what type of photorealism you're aiming for. Your best option is to make multiple images with different denoise values and pick the one you like best.
denoise numbers change based on your schedule, resolution and sampler
so we can't really help
if you are in Comfy you can use the "split sigmas at sigma" node from Comfyui-predictions
whichever sigma number you choose for this node will work the same across different schedules, resolutions and samplers
timesteps change but sigmas don't
even then, big enough resolution differences will change it
hey fam
No it's not 20 4090s. It's more like two 4090s with 5x the VRAM. I did not look at the specs, but there's no way you get 20 4090s inference performance for $3000.
NVIDIA Project DIGITS is slower than a 4090
it just has a lot more VRAM
DIGITS is "up to" 1 PFLOPS FP4
4090 has 1.3 PTOPS Int4
5090 has over 3 PFLOPS FP4
note I gave the Int4 number for 4090 cos that's the 4 bit number format it supports
its a fair comparison though
seems like marketing for computer labs in schools, just buy this and you can plug it into your Dell
lol
It'll be predominantly used by call centers for customer "service" lol. Or by Indian scammers.
probably haha
No? If you mean ai filtering copyrighted images
Its free below 1m in revenue?
Yeah if i was i wouldent be talking about it lol
No worries lol, i have done some reading up before doing it so its fine
its still pretty convoluted, its 1 mil from any source, like you made 1.1 mill plumbing and you want to use SD for a flyer, you need a license
Granted i earn 40k ish a year i got a long way to go before i hit 1m so no worries!
What no my real job is 40k
This doesnt hit anything close
Its pocket change lol
I prefer xl still
No not turbo
Just sdxl generally
Mostly due vram constraints:^)
Gm
The speed is similar to Flux, but the image quality is worse. However I prefer the composition and overall appearance over Flux. It could be useful to create an initial prototype with 3.5 and then use Flux for refinement.
How many tokens can I pass in a prompt. I have read 77, but also 512 with T5. I'm still new to all of this and continuing to research.
https://github.com/NVIDIA/Cosmos is this a big deal? i'm not in the right frame of mind to tell yet or not
https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/ they use this license which seems permissive
Yes, it’s hunyuan level in quality(14b one is similar, 7b is close, didn’t test autoregressive ones) and all supports i2v too. There are two flavors
Diffusion based - 7b, 14b
Autoregressive -4b, 5b, 12b, 13b
Autoregressive ones are somewhat fast too, 4b can as high as be 10fps. Diffusion is ok speed, 14b one is around hunyuan.
neato!
On flux why do gen times like basically quadruple if you increase the CFG scale at all or put any word at all into the prompt? Is flux made to not use the prompt and swing different denoising gens?
shouldn't quadruple. only double. because CFG is 2 pass per step
anything over 1
I have found myself going back to epicRealismPureEvolution for editing. Like I create batches of pics I want, then I change it to flux and run it through im2im at a low denoise and basically use it as a glorified upscaler. Thoughts? Dumb/smart/neither?
would be more efficient to fine tune flux to get the look that you want
without needing to go through lots of seeds
loras are super useful for flux to cut through the guidance
hello
Hello
was looking for a community for SD like this. was wondering how people were getting hyper realistic results with their avatars / characters.
who was enough knowledge over stability matrix (loras, etc) so i can ask for help for something
What do you wanna ask? On how to train a lora?
Good prompting and good models.
have a mix of both... but i feel maybe one or the other needs tweaking. i use a realism lora and chatgpt with an assigned role and context training to help tweak my prompts. still, sometimes the output can look a little airbrushed here and there.
what UI are you using?
what do you mean by "hyper realism" because that's typically an art style that sculpters use
realistic
and you're having difficulty getting flux to be realistic? forge has the sampler for flux that drives the realism home really well
im curious what combinations of lora other ppl use
Whats the question bruv
I like to combine a pixelart lora with other style loras to get a whole different feel
guess it depends what you're into also. some like creating anime, others landscapes, and models
im using ae, imageupgrader, flux realism, and a hand one so they don't get all distorted...
DAMN, that means if 4090 takes 30s for flux image, 5090 will take 13s O_O
meanwhile DIGITS probably takes around 39s.
for realism, i generally don't use anything for flux. It's aesthetics guidance by stock looks very aesthetically photorealistic already. I use the special sampler that forge ui has if i'm using forge. not much else.
if i want specific art styles or characters, i'll use a lora.
trying one generation without lora and the prompt i last used with a few to see what i get
about the same
well.. im willing to bet that it's people who can afford to use pro that get the next level of realism up lol
i dont know what you mean by realism still. Photo realistic is easy to achieve
the influencers you're looking at are likely doing post processing or upscaling with topaz or using multiple passes in comfyui. There are so many approaches to get a higher quality looking image, but "realism" is very vague. it doesn't describe your goal at all
dms?
Depends on the question lol, if i dont know others might
But i think if people have to dig to get the question out of you they arent gonna be stoked about helping lol
Ah based on that question, i would recommend you use either civitAI or koyha_ss if your pc is powerful enough
But if you want a more detailed response, https://m.youtube.com/watch?v=WPKPO-2WFK8
(mobile link)
But if you were getting stuck with training a lora i recommend #🤝|tech-support
there's also flux_gym for low ram local lora training on pinokio. koyha_ss is on there too though some use google kolab
It really depends on what model he wants to make a lora for too
true
If its for hunyan video, you can make a concept or style lora pretty easyly but a motion lora i have no idea
You train with videos but thats a whole different story
could not get hunyan to work.. or pyramid flow
i really want a local open-sourced image to video generator on my pc lol
if i can log in, then ill burn up my kling credits but ive been focused on generating images for like 40+ genres / environments
so will be doing videos after i edit and audit some more
i think my error was that the pc couldn't detect the safetensors or something in the directory. tried putting it in there manually but no luck lol
i have 12VRAM at least
Hmm what ui were you using?
i tried with pinokio then via github method
also after some time it would time out on trying to download the safetensors
Hmm strange
I never heard of pinokio so i wouldent know
Still strange it would time out
it's like stability matrix or the other one ppl use that has a GUI for a bunch of programs all packaged nicely so that they dont have to download one by one
i would use comfy_ui but the manager doesnt work for me 😦
i'm looking at that digit pc. a blackwell 5070 equivalent chip with 128gb of ddr5 memory to access
anyone know how people are making videos of models walking around and doing stuff? the quality has been excellent...not noticing hallucinations as much like months past with some generators... i would guess that it's kling or minimax plus some pose control / openpose dance.
havent looked it up on reddit or anything yet
they pay for one of the services or they use a local video model. some of which aren't as good
https://blog.comfy.org/p/hunyuanvideo-native-support-in-comfyui here's a strong post about one
anybody an expert in style making using illustrious and pony?
i need help
i can train characters but not styles for some reason
Hey there
Hope you are doing well
I am a senior full stack developer who has full experience in web development | design and AI development
So if you have some projects, please let me know
thank you
guys
i cant
why it keeps putting 2 girls on the image if i literally have it in the negative prompt
usual prompt is also fine
theres nothing that can conflict
What is the prompt?
Are there keywords that could hint at more then 1 person
Any loras?
1:1 image or is it like 16:9
1girl, anime, source_anime, line art, symmetrical_face, small_nose, simple eyes, large_eyes, almond-shaped_eyes, brown eyes, thick_eyelashes, well-defined_eyebrows, smooth_skin, round_face, delicate_features, short_hair, brown_hair, layered_hair, bangs, hair between eyes, (slim body:1.3), small ass, (big breasts:1.3), bookstore, excited, wide opened eyes, pointing to the side
line art style lora
1:1 image
i dont think theres something that could provide more than 1 person
what do we all think about the new gpu series? Is it not worth not buy a 4090 right now?
Try removing the ( )
( ) adds weight
And
1.3 also does
hey guys, whats the best ai tool for turning photos of clothes into photos of models wearing the clothes?
What's the resolution and model?
Hello guys!
I wanted to ask if anybody here uses Stability Matrix on some regular basis? I wanted to Install forge and I did some digging in the old reddit threads where people suggested using Stability Matrix for better organization and later moving the folder if you ever need to change drives or PCs. But others also warned that there are a lot of bugs and it doesn't work properly.
But all of those are quite old reddit threads and i've yet to find out what is the deal with Stability Matrix right now. If you are using it for running multiple webuis, please share your experience, I would much appreciate it
🙏🙏🙏
Hi
hello from story !
Would it be possible to train a lora on Illustrious, with only 8Gb of VRAM?
Yes, but check the comments, dont think your gonna be happy with the time it takes
by "comments" you mean pinned messages?
those are +1.5 year old, I guess it's outdated
I would imagine it'd take days
surely it can't take that long, unless I want to train a lora using +100 pics
yes but i think it takes a few hours with 8gb
That's fine I guess
yeah guess it depends on how many images you wanna do too. epochs = 10 steps per image iirc
I don't know yet, but loras can work fine with way less than 100 images, right?
if you have 30 quality ones from different angles and environments i think it'd be fine.
Does it matter if it's for a character?
that i can't tell you as I've only done characters
specifically, it's for the eyes of a character
because they're often messed up with no lora
Hey there, future dropshipping& reselling rockstar ! Let me ask you, are you tired of feeling stuck and frustrated with your online business? Are you struggling to find the right products, suppliers, and marketing strategies to drive sales and growth?
I've Been in Your Shoes
I know what it's like to feel overwhelmed and unsure of how to succeed in dropshipping &reselling . But I also know what it's like to break through those barriers and achieve success. And I'm here to guide you every step of the way.
drop shipping stopped being a thing in 2010 when Aliexpress opened and factories in China were able to sell to the US directly
Either a bot or a desperate guy
there might be some eye loras out there on civitai.
it's eyes with specific shaped pupils
ah
yo what up
hello, hoping someone can help me out here. I installed automatic1111 on my pc last night. I'm wondering what's some good models for making monsters and such? kinda like bing ai before it got censored into not being able to make cool stuff
hmm i know a few but are you aiming for realistic or cartoony
realistic, thanks for the reply
hmm what is your GPU?
2070
this model is capable of monsters, since you got 8gb vram it should be able to run it but it might take a few seconds https://civitai.green/images/48983430
check resources used for the model
thanks i'll try that out
of if you want evil themed images: https://civitai.green/models/626819/beauty-in-evil-by-hailoknight?modelVersionId=700704
its a lora
works on both anime ish and realistic i see
"There is a lot of horns in the dataset, if you dont want that add horns to negative prompt." thats all you'll ever find on civit is horn models. HEYOHHH
nice lora though thanks for linking
Hello! Help, tell me which LoRA, or what extension can be used to get a square empty frame for a game avatar or card game. I use SD 1.5
You need a mask. There are video tutorials out there that show you how to do it.
Can you share an example? I don't know English very well and I don't know how to properly request such a video
I don't have a link. But I have seen such videos before. You could also just do image to image and start with a black image. then you inpaint the frame and in the prompt you describe your frame.
thanks, I'll try that
is civitai down
ye
hey guys, does anyone use FORGE UI ?? I read it was good so I use it... but now it seems like people only use comfy
forge and reforge are good alternatives to a1111 since theyre essentially forks of it w better optimization and extension support, if you want visual gui like a1111 but power of comfy use Swarm but comfy is what most extensive users utilize cos of the sheer power of customization it offers
@slender vault does comfyui create images with simple prompts like Fooocus? I’ve been trying AUTOMATIC1111 but I can’t get consistent results or even anatomically correct images. I just started using Fooocus but it seems a lot slower to me although it’s very easy to create good and consistent images. Also, the stop or skip buttons are unresponsive for me.
Thanks.
you can with use of custom nodes. promptgen and whatnot. I imagine foocus uses something similar under the hood, personally i run ollama workflow that lets me create diverse prompts out of a simple subject input
Pony has catastrophically destroyed knowledge of the base model already, and disaligned the network from the text encoder in a big way. Don't try to train styles for pony xl. just use a better model. Train the styles on the base SDXL model, then use it on any SDXL model. If you do it with smart enough settings, then it should even work on every pony model, since pony is'nt a base model. it uses SDXL as a base
what about illustrious ?
i would still train your dataset on the base model and then use it on illustrious
people say its better
is it compatible ?
It's an sdxl refine. that's how base models work
i see
but still how do i make a good lora
all my attempts end up as blurs or blobs of color
makes sense that the encoder got destroyed for pony
i would explode myself if i was forced to observe horse genitalia 616 million times
trial and error. learn the tools and practice until they start coming out good
i made 50 style loras and i still cant get it right
did you try 50 different approaches?
autotagger
Auto tagger but did you use a unique identifier for the style?
do i need unique identifyer for the style ?
like a trigger?
I see a lot of styles do it
your tagging might be the culprit. the dataset is the biggest key. settings while training only drive it home
i can show u my dataset
Civit aj auto tagger is wunk, i had images of perfectly pg13 labeled porn and actual nsfw as pg13
i dont use the civit ai tagger
describe which schedulers you've tried, because i feell this is suspect claim. "how do i make a good lora" is not a very well informed question, to be blunt
how would u describe an image li
liek this
ill train for both pony and illust
but illust is the priority for me
cnh cartoon, stick figure smiling, green shirt, inside, line art
1man, clapping, bald, dot_pupils, green shirt, line_art
Depending on the action you want it to do ofc
A still image with hands up can very well "raised arms"
so you don't know anything about different schedulers?
just do characters
With image generation if you forget stuff it's fine. With loras it gets wunk
i have 2 brain cells
left one says dataset amount
right one says repeats
I mean if the character has raised arms
You gotta tag it
and thats about it when it comes to lora knowledeg
so you haven't really tried 50 different approaches then. Why lie when you want help?
most of the tags have that
but like the stick figure, black dot eye tags
Did you follow a tutorial for that specific lora trainer or did you dive in blind?
i dive in blind but that's just my style
i did try 50 different approaches
unless approaches is a word when it comes to lora training like encoder and repeats and steps
for a style i find less tags are better. otherwise you end up with the lora requireing "dot eyes" to do it
i use collab since its free
Same but if i get stuck i lookup a tutorial, i rather grind my gears for 5 hours then read a 5-10minute tutorial
So no tutorial?
nope
I recommend following a tutorial <3
there arnt any good onesz
i made my own tutorial for characyers
https://civitai.com/models/134056/explosm-cyanide-and-happiness-style can just use this too. why reinvent the horse?
sadly my pc cant run it
a competitive market is a healthy market

you aren't competing . you don't even know the basics
i'm only being honest to keep you in scope. you don't seem to really want to learn even
like extracredits/extrahistory
you want it to be one button easy. well, i gave you the easy way
i made successful styles for sd 1.5
find someone else who done it
A quick google gave me a few results but it also seems like:
That sampling doesn't always work as colab stops it in the free ver
From lastyear
thats why im here looking for someone
in my tests aslong as its less than 2900 steps
i'd recommend schedule free training but you don't even know what you've used as a sampler/scheduler already
Doesnt style use like over 3000?
me neither
damn
I get that this ai stuff is exciting to mess around with but i do recommend a little more research before diving in
That and articulate the problems you with the steps youve already done. (Not just you tho its a common thing in this discord lmao)
i tried to do some research but it always comes out weird
i know
i help those that i can assist
Good Googling is a skill too nowadays
lmao
But
https://civitai.com/articles/6182/how-to-make-a-lora-on-colab try following this
https://civitai.com/articles/7993/lazy-tutorial-or-how-to-use-trainer-lora-on-colab-or-sd-15-and-xl-by-mikus-silly-and-easy this is my favorite appraoch for style lora's in my experience. smallest captions possible. its good for world morphs.
i follow hollowstrawbweey
i followed tha ttoo
I meant the tutorial choom
my bad 
i doubt that. since you coudln't mention it earlier when i asked how do you caption.
i honestly doubt that even 5 attempts have been made. i feel strung along by a dishonest person. and because of that, i'm abandoning this effort. it's like trying to help someone up while they're dead weighting their body
i was the guy that asked him to make the article so that i can understand how he switched from 1.5 to xl
doubt again
You'd face the same problems with sd 1.5 style training
even professionals coulnt help me
my sd 1.5 styles were succesful
don't blame other people for your lack of honesty
lying about what
im not a proffesional at all
u can see the settings in my loras
all of them are the exact same settings
only difference is the repeats sand image count

i do that too
character loras come out perfect
style not as much
i even have a bounty on civit ai for a person who could teach me how to make a style lora
is @signal palm adding anyone else here ? they've never spoken in the server and requested a friend add to me. shady behavior for discord
Seems like @near silo 's alt imo
What?
I don't have an alt in this server lol
I have no idea who that is, and I am not associated with them. Not sure if they are causing issues or something
@copper crystalPinging real quick so you know I responded
they're pretending to be a dev. your M.O. convenient that you could answer so fast
Modus Operandi
it's latin
Anyways, I am not associated with them, whoever they are
I have no reason to have an alt in this server, I am not even active in it. I just saw a discord notif and responded
Also, I have never once pretended to be a dev, so I am not sure where you are getting that from
Ok. You're dismissed
Dismissed? lol
Ok then, I guess lol
Bit of an interesting move to ping me about being associated with some rando, claim something completely false about me, and then "dimiss me" but go off I guess
K
Damn, you are next level insufferable. Thanks for the slander, have a nice one lol
yeah that dude is weird
lol simp for him all you want. he loves that.
Glad its not just me that feels that way lol
Bro, what is your problem with me? 😭
Its been months since I have messaged in here lmao
remember when you told all the girls you were a mexican guitar player? that was fun
What the hell are you talking about?
that time you were like "i play the guitar all the time" but then i found the source video from 2004.. remember that?
Dude, I seriously think you are remembering a different person lmfao
Whattttt are you talking about? I will pay you $1000 to point me to where I ever said that lmao
just block and move on lol
I was born in 2003. I AM Mexican, but I am absolutely not a guitar player? I genuinely have no idea what you are talking about 
#🆕|sd3 message guess that aint you ok then
Its... not?
I have no idea who that is 😭
I am being 100% serious, I have no idea why you think thats me. I am not that account?
K
I have been in this server since a week after SD1.0 released. I've never done anything like what you're talking about lmao
I don't play guitar, I have never claimed that I do. I would never show off to "girls", I am a gay man in a happy committed relationship. I AM Mexican, so I guess thats right? But I am more amused than anything about this
This is at least very amusing lmao
I am curious who you are getting me confused for 😅
@copper crystalSeriously tho, if somebody is saying things or acting out using my name, please let me know and link to it. I genuinely have no idea what you're talking about, but if somebody IS doing that with my name, I would appreciate knowing
i'm not interesting in helping a troll out. thanks but no thanks.
Alright, I am just gonna have to assume you are mentally unwell or trolling or something. I'd appreciate you either being a man and pointing to what you are talking about, or stopping, cause I don't appreciate being slandered like this
lol dice by all accounts. comments like that shine.
You think I am a different account?
I am using tiles and have an image with 4 different pictures, 2 on either side, 2 in the middle. I can't seem to get it to fit into the save/load box its supposed to show in. It is cutting off the outside edge.. This is from Pixaroma's Lesson 28 (if that helps)
tried redoing the image in gimp according to pixel settings and tried changing setting is workflow but im missing it
I have been in this server far longer than that account, so I have no idea why you would assume I am them. Whoever you are talking about is not me, and you have literally no evidence to think it is. I've been in this server for well over 2 years, and I have been a very active member (in the past) with literally nothing you are talking about
Just cause some previous idiot thats not connected to me got you upset doesn't mean I am them lmao
lol nah.
guys...skip all the hate...answer my question..plstxok
#🤝|tech-support Will get you better and faster responses. General chat isn't make for tech support
good point sytan..sorry..not enough coffee yet
Alright, I am just gonna conclude that you're a troll. Please keep my name out of your mouth if you are planning on spreading continued false statements about what I "supposedly did" when you have literally no proof. It makes you look crazy, not gonna lie. Especially when you are targeting one of the literal first members of this server 😭
Sure thing Will
you totally dont have a sea of obvious alts like dice ai dev. don't sound anything like that fella
helps that the links were scrubbed last week huh? now you can be like "blahblablablahblah deny deny deny"
@sudden ruin @bleak matrix I am so sorry to have pinged you guys, but is there any way we could have this guy stop spreading false information about me? I have employers in this server, and I don't need them thinking I am some "evil mastermind" or something
I have been a member in this server for well over 2 years with well over 30k messages, and this guy is projecting some weird obsession with somebody else on me
The link is loose but tangible. https://www.youtube.com/@Dice-Dream-Diffusion it has since been scrubbed.
that profile tied sytan's discord to their discord
OK? I am looking at what you are sending. This looks like some crappy slop channel
Wait, what the hell?
now alts that share the same ties are messaging me personally. it's all very sus
Ok, if what you are saying is true, I need to figured this out. I am not linked to whoever the hell this is
sure thing
If somebody is trying to hijack my identity and using my discord, I need to know so I can stop it, cause I don't mess with this shit
Says bro is in Australia, I am in California
yeah they had changed that. they're not in aus. its an alt
@finite cloak Hey man, can you please like... tell this guy I am not whoever he thinks I am? You have known me for a long time 😭
Alright, I just joined that dudes discord. I am getting to the bottom of this, cause I don't need some stupid ass AI hype beast slandering my name and my account. I am not whoever the hell this is
dafuq
@copper crystalAny information you have on this would be massively appreciated. I am sorry I thought you were trolling. If somebody legitimately has been linking to my discord account, that's not cool. I do business on this account and I have employers in here
Apparently this guy is saying some rando guy on youtube who was a menace in here was claiming my discord account as his own or something
I just joined his server and messaged him, cause I don't want to be a part of whatever the hell this drama is
its not my responsibility to sort your alts out. i pointed out to your dice account last week that you slipped, you scrubbed it all, made it look australiaan for some reason, then today you alting on me with a new account. now you're calling all the boss men acting like a victim. sort it out yourself.
Bro WHAT are you TALKING ABOUT
I have no evidence of if what you are saying is true, so I am taking it at face value. I have no idea who you are talking about or what you are talking about. I have been in this server since basically day one, and I have never claimed to be good with code, or be a guitar player, or anything like that. Whoever the hell this is is trying to steal my identity or something
I'm looking at his youtube right now, and it looks like a cesspool of AI slop
This dude's face has to be somewhere. Hes literally not me
I mean I just wanted to vent something else here, but you do sound very suspicious
'I am totally not that other guy. Heres my 20 posts proving that. Absolutely not that guy. Nah uh'
What am I supposed to do? Just sit here and let people think I am somebody I am not?
That dice dream youtube account is not Sytan and I have no idea what would lead nuuideas to think it is
the evidence got scrubbed. i should've showed other people instead of just laughing at dice about it. now this
the owner's face is in fact visible in its history https://www.youtube.com/watch?v=vWdMXTk4zRo
some random older dude
likely not him. he would link other guys playing guitar that weren't him here too. it's an alt with a crafted identity
I am talking with this dude right now and hes as confused as I am
#🆕|sd3 message remember this? lol
to what possible end would sytan have spent actual years building up a random side identity just to eventually use for an alt in relation to tech that didn't exist at the time, letting him be a ai slop youtuber separate from his usual activities as a fairly intelligent ai researcher online
I mean to someone just coming in to the discussion thats just how grifters/ scammers throw out their hooks isnt it? Have hundreds of alt identities?
Heck you can even sell/ buy them :D
Feel free to look in #🏞|general-with-images where you can see I am literally not whoever you guys think I am 
I mean I really dont care either way and dont want to come off as hostile. Its just that you really do sound suspicious the way that you vehemently defend yourself against a single accusation of a regular user?
Like its not like a mod wanted to ban you or anything
It sjust weird to me is all. I am one of the OG's in this server, and I have never had somebody think this about me before. I am more confused and entertained than bothered lmao
Try to think of it from Sytan's perspective here. Wild accusation thrown at him, he counters the accusation, "the fact that you're arguing just proves it more". Like, hwhat the witch trial hell
yeah it blew up. i just caught the vibe that synxaltz was a new alt that was trying to bait me into running a game server. no one else was catching messages from them, it was just me. soon as i showed "dice" link, it really went off
Again no offense meant or anything. Its just that to me you do sound suspicious :D
But I really dont care. Be whoever you are
Wheoever this dice guy is is saying he doesn't know whats going on either. His youtube links to his discord server. I didn't see whatever you're talking about with him linking to me. I am not saying its false, just that I have no idea
We are already a community of underground users of a cyber technology we modify ourselves 😎
yeah i stopped caring. maybe now the alts coming at me will stop since it's obvious
In what way are they coming at you anyways?
asking me to join their game server which presumeably requires running code
Idk just block them?
that sounds like just the usual malware spambots? they go around at random all the time
And with text-AI getting better I think we will see more Bots over time anyways
Anyways I came here to vent :D
i find people tend to follow patterns, and sytan and synxaltz is a pattern. "altz" is also like, throwing the obviousness at me lol
Theres a vocal anti-AI crowd. Nothing new so far. But it hurts me to my core how... misinformed they are. Its like they treat information about AI as this blasphemous thing
Oh AI takes enough energy to suck a lake dry? Why can I run it on a potato then?
It searches (?) the internet for images to steal? It works perfectly fine offline....
And like the whole debate over how it never 'actschually creates anything'. It does create unique output. And yes using AI as a tool for art is, well art. Otherwise paint splatterings on canvas would not be art. I thought we established that anything can be art
It really grinds my gears sometimes. Instead of a constructive discussion its all just BS thrown at each other
they'll putter out and be lost to the winds of progression sooner than later
No because they can do lots AI can't. Which they knew if they learned stuff :D
Extremists will be extremists will be wrong. There's extremists on both sides, they're both wrong. Non-extremists on both sides complain about the extremists on the other side. While you complain about the people that treat AI like the devil, people that are reasonable but dislike AI complain about loud twitter AI bros yelling that AI is going to replace all artists and make instant personalized movies and etc. etc. nonsense.
imo just ignore the extremists and focus on conversations with the reasonable people in between.
I try to. But its hard since I am in the furry community :D
Lots of butthurt artists. Even though their nieche is less in danger than they think
Even just posting an AI image on bluesky could possibly see me blocked from like half the artists I follow (if it pops up on their radar ofc)
guy i used to work with, his boyfriend is a furry artist. he has lost substantial income because the prices have severely dropped for commissions. AI artists are bidding a lot lower where before, he coudl sell his art for a good price because only him and a few others could do it
There is a reality to the lost income
i don't talk to them about AI at all. it's a touchy subject for them
got a lot of respect for my comrade in kitchen arms
not really. Their skillset's worth diminished a ton over night. I get why he's a mad guy about it
for a freelance artist, that's a huge deal
i think you're underplaying it
I dont know where you get that from
But still misinformation does hurt both sides. Social media being what it is, it would be hard to counter though
hi
Moin ^-^
how can I draw an image
this downplays the severety of a freelance artist overnight finding that his practiced skill is now 100-1000x more competitive
Pen and Paper ;)
I have a simple guide for installing A1111. But websites exist to get your toes wet
brainless
they don't have a job to begin with. they're their own boss. and when the market they sell on starts paying a lot less per image, while also having to compete against so many new people offering services in that market, it's a big deal. it's like they lost their job and can't even get a new one in the same field now
Idk. Is it so much different than me getting fired over night because of like idk lost revenue or something
Losing your job is a reality in this day and age. There are no guranteed/ stable jobs anymore
i got out of web development because outsourcing happened. local hires were offered minimum wage and i could make more money doing service jobs. so i was edged out of that field. It's not quite the same situation, but it's very similar. I do have sympathy.
They'll have to come to grips with it. the money train for furry art is over.
Oi. I dont like being sweared at
ya akho sharmoota. I asked you how to make an image ya 3ars
it's a greiving process
and you make fun,. kos ommak ya maniak
And I gave you like 3 answers. Draw. A1111. Websites
I do have sympathy :O
But I get jaded with the polarization and the misinformation
I mean I am not the one calling others names :D
I try to. Again I would never air out these issues in those communities anyways :D
Thats why I am here
avoid community that are welcomeing to that kind of discussion. there's lots of those on reddit. /aiwars/ would be a big one
Can't. As I said I am in one of those UwU
I get both sides btw. Artists and fans that hate AI. Friends of mine that are happy they dont have to pay hundreds of dollars for images anymore
just dont flaunt it in their faces then.
Truth will prevail. misinformation can spread like wildfires through LA (too soon?) but ultimatley. people know vaccines save lives still. in another 10 years, it'll be even more obvious. grief comes with pain.
think about the stages of grief and then when you see those people, just mark what they're doing on your grief bingo card
I dont :D
I do know where to say what. Being at the edge of society primed me for this :D
the outkasts are always the coolest.
alright alright alright alright alright alright alright alright alright alright alright alright alright
You ok there?
I think you should probably get some sleep :D
Isnt it like 3 am where you are
10 to 2
See
it was just a sick reference to outkast hey yeah https://www.youtube.com/watch?v=PWgvGjAhvIw
Ok?
the point is, we're at the edge because we're out casts. We got that cool perspective
You know, that was the number 1 song in the world on the day I was born lol
mine was olivia newton john's physical
Mine is nothing exciting (to me)
i was hoping for thriller but that came out way later
now i got physical stuck in my head
Ah its that song. I prefer the video of slugcats dancing to it :D
I am not sure I even know that song
OH
yeah, of course I know it lol
Had no idea who the singer was 😅
Same
And now I have Hey Yeah! stuck in my mind lmfao
she's the one that you want. (ooo ooo ooo)
OHHHH
Thats her?
I really like that son
wait-
I might be confused
I am a certified dumbass. I was thinking of Steevie Nicks Edge of Seventeen 😅
so many old songs
"Just like the white winged dove, sings a song, sounds like shes singing" (Ooh, ooh, ohh)" lmfao
I have no idea how I thought thats what you were talking about lmao
gonna bring some levity into this gong show with this fun song .. then i am off to bed https://www.youtube.com/watch?v=zGaqQiKYOF4
Oh wow, I haven't seen that clown in years
Does he even still do things?
Oh wow, he does. Looks like hes posting decently still
yeah very active chan
Hes got a very incredible voice, shame he doesn't get more activity
I haven't been this active in this server in months lol
I have just kinda had no reason to be here, cause I don't use SD3.x
Oh what do you use?
I use flux more, cause I am working on a paid project around it, but I don't do TOO much image gen anymore. Been way more focused on making ultra fast training configs and local run LLM stuff
Had some extremely promising stuff with said fast configs
but seriously i shouldn't joke about LA burning while it is. I really wish everyone there stays safe. I fear the city might be lost much like Jasper BC was though
Yeah... My best friend lives in West Hollywood, and they have been worried for a lot of their friends and family in the affected areas
Sorry, North, not West
i heard a few times over the years that if a wind hits a wild fire the right way, it could burn throuhg LA to the coast no problem. It's mostly a wooden city believe it or not.
I dont want to see that realized when i wake up
I live in the Desert region North of LA by about 100 miles, and we had the extremely bad Blue Cut fire back in like 2018, and it was terrifying
We also had the Line fire this year. Shit is scary
just a sec, I have pics from it actually. The sky was insane
i live in the woods. i'm so glad that this has not been a drought year. it's like living in a tinder box having drought for 5 years in a row. walking over my lawn like crricaack craciakckkiick
I live right smack in the dry ass desert, but we live in an area with really well trimmed wild brush most times for that very reason
neighbor lights off fireworks on some random ass day and i'm like "FUCCCKINNNNN"
my go plan includes rippin cables out the back of my pc and hauling ass
Or some gender reveal wildfires :D
bladerunner be like
Would you guys mind if we jump over to that chat? Just so its easier to send pics?
naw i'm out its 215. dont poke the bear. smokey was retired and it's ember the fox now
Ah, fair enough lol
This was one hell of an eventful night for me
oh wait, we are in the same time zone lol
Holy shit I hadn't realized it was so late
Wait you are a bear? UwU
Hold on there, scalie lmao
My day is just getting started (11:25 for me)
We got a Euro boy, I presume?
Yes ^-^
Most of my friends are in the EU lol
and my boyfriend is stuck in Australia, so far away 😭
food for debate, was smokey the bear a furry the whole time? #🏞|general-with-images message
I could never. But I sympathise
Yeah... Its rough, but hey
He has plans to come over around June. I wanna take him to his first pride in LA <3
I mean furries basically developed out of mascots in a ways
Nice. If the city is still standing by then
i thought it was sonic fandom
Furries are fun. And basically have been around since civilization ^-^
Uuuh a cheese
Then I rebranded when I turned like 15/16 to Sytan, cause I discovered I was a Satanist (no I do not worship the devil), and I wanted to keep my logo lmfao
Origin stories lol
Yeaahhh, I used to be lol
My sona back then was a sergal felkin hybrid
But I mean could Scyth not still work. Werent they like nordic (pagan to you) riders ^-^
Honestly, Scyth came from me being 11 at the time and not knowing Scythe had an e... So I wanted to change it up lmao
Also, Sytan just fits so well, cause its Satan with a y
Ah xD
You put the gen Y in Satan?
I'm gen Z lmao
Me too :D
Oldest gen z year
I just turned 21 last month, so not quite there lol
So in your teens when corona hit :(
Yeaahhh
I am sorry you had to go through that
Missed my Junior year
it was hell yeah
I DID at least get my prom, thankfully
But yeah, that was a hell of a time
How can I fix this issue (I'm still quite new to open source ai in general and don't know how to solve the issue): ComfyUI
GroundingDinoSAMSegment (segment anything)
Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument mat2 in method wrapper_CUDA_mm)
I'd recommend looking in #🤝|tech-support
That channel is better suited for faster responses with tehc support <3
I can also recommend A1111 if you are new. And in general :P
I know. It was for all of us
A1111 is prettyyy far out of support now days
But it still works very well :D
And there is no easier interface. Especially with hires fix
Yeah, lots of people still use it
With other UI I first have to jump through hoops just to restore two pass text2img
There are some that you can make custom. I made a few with comfy backend, but thats not exactly "easy" to do, so yeah, I would agree haha
A1111 is a good starting block, but I think most go to comfy after they wanna do some more advanced things
Idk about the use cases tho :D
Especially someone close to UX/ UI design
Like is it actually worth the effort?
... have you tried SwarmUI?
Yesh
extremely easy, great ui/ux, very easy to do highres fix
And I use it for one specific workflow I cant do in A1111 but dont bother otherwise :D
I mean its great UI compared to comfy. Because Comfy barely counts as a UI to me at all :D
But other than that you can do like 90% of the stuff I do in A1111
what's the 10% that you do in auto but can't in swarm?
(if that's what you mean by that?)
(if not that, then, just, generally: why not use Swarm?)
Theres a specific subset of two pass text2img that gets better with two different models :D
Because like fur gets fluffier and stuff
But mooost of the time I dont even use that
the Refine/Upscale param group in swarm can be used to do two pass with two different models pretty easily
Because as with all 2 pass thingies the original image gets weirder
Pff. I built my own workflow for that
But like. I dont need that often, and A1111 can do two pass easier. Like from a UX experience. Because its more of a finished UI
you don't need your own workflow. it's incredibly easy to do in swarmui
But I wanted to :D
??? you can't say "auto is easier" and also "I wanted to do it harder" at the same time
do you want it easier, or do you want it custom?
in both cases Swarm is the better answer
but you have to pick the question first
Yeah I can. Auto is easier in general. For whacky stuff I get out Swarm
In swarm it is very easy to do two pass with different models like you want
no need for custom workflows or anything. you just click refiner, click the model you want, and optionally fiddle other settings, and go
you can even save a preset of specific settings for this with again just a few clicks
But like in 90% of the cases I dont need that :D
I just need a UI I can use well and some basic hires fixin
swarm is easier for all the normal cases too
I mean... I have to disagree. At least with the version I last used
give me specific examples
