💬｜general-chat

[sullen fossil]

every time i reopen the thing it takes 20 minutes to redownload the same thing

[nova zodiac]

what are you opening - are you starting a runpod or something?

and are you using comfyui or a1111/forge?

[sullen fossil]

a1111

i figured it tho

thanks

[solemn marsh]

is bot down

[mortal delta]

Am i allowed to talk about ai animation with Stable Diffusion here? or is there a better channel for that?
I always double check when it comes to channels, by asking.

[slate dust]

Sure! You can also try #🎥｜animation

[mortal delta]

I wasent sure but i guess ill start here.

[solemn marsh]

is the bot not active?

[mortal delta]

is animatediff the only animation workflow for comfyui or are there better ones to use for comfyui?

[mortal delta]

I would assume so?

[nova zodiac]

#1047610792226340935

[nova zodiac]

there is svd, hxsvd, and deforum

[mortal delta]

which one would you suggest for a beginner if i might ask? im thinking of trying to make a super short film if i can.

[nova zodiac]

try em all tbh

you'll find one you like and understand

[mortal delta]

Ok i just googled them and im debating between hxsvd and deforum but i guess ill try as many as i can first like you said.

one more thing where should i share the ai videos at?

[fervent thunder]

I would be interested in looking at the videos. I have only seen examples on SD reddit

[vestal laurel]

haha that commerical was just making fun of AI art

[potent spire]

Fooocus is pissing me off lmao

or is the model behind it to be blamed

[cedar salmon]

i get mostly ID-10-T user errors

[spring eagle]

where can I find artists to hire that are not from civitai? I've been through a lot of them already

[fervent thunder]

What do artists currently charge?

[spring eagle]

It ranges based on their skill and their ego

Some charge $30 USD, some $120, and some over $500

[fervent thunder]

for how many images?

[spring eagle]

For images? I'm talking about lora creation

[fervent thunder]

oh ok I have no idea what I am doing with SD really. I can't get anything even close to midjourney

[slate lintel]

anyone know how to install open_clip dependency?

[nova zodiac]

pip install open-clip-torch

[slate lintel]

thanks man, figured that out a few minutes ago. Now I’m just getting the configuration set up correctly. For text to video do you know off hand what the YAML is for that?

[nova zodiac]

no sorry

[drifting flame]

can we put link here? cant find rule against it. I have something to ask

[oblique willow]

ye you can paste it

[drifting flame]

https://www.youtube.com/shorts/1mwc6gk71Mw

im asked to imitate this photobooth installation. I only given 2 months time to research or I will lost my job

im a graphic designer with 0 programming skills, do you think this project is possible to reach? 💀

And I wonder if Stable Diffusion / ComfyUI is the right tool to reach that goal

[nova zodiac]

Yeah that doesn't look crazy - it looks like it takes a photo input, uses something like rembg or segment anything to remove the background, and then it is either photoshopped onto a movie poster, or onto an AI generated background (the text in the movie posters looks too good to be ai)

If you have a stupidly good setup (like a desktop pc with a 4090, and a good digital camera), you should be able to do each of the steps manually in a week, and then you've got a few weeks to figure out how to put it together in a sleek and automated fashion

[hallow sable]

still dont know how people dont throw up when you buy a 2500 dollar gpu

[drifting flame]

Hardware is not a problem, i got two A5000

[nova zodiac]

My car isn't even worth that much!!

[nova zodiac]

Sweet..

[hallow sable]

why buy the a5000 when you can just get a non server grade gpu that has the same vram

[drifting flame]

Im not really sure if its manually photoshop or not, but my Boss expect the generated image result in less than 3 minutes after the photo was taken

[drifting flame]

the PC often used for rendering for days, I think server grade GPU is best for constant work

[trim magnet]

you can do it in like 2 mins if u have the backgrounds and the text already generated

[hallow sable]

amd has a server grade 7700xt called the W7700 and its basically the same thing

but the W7700 is 1000 while the 7700xt is like 500

[nova zodiac]

It should only take a few seconds to save the file and then strip the background from it, then about a minute to generate a set of potential images

[trim magnet]

also u dont need rembg or segment if you take the picture using a solid background like the one used in the video,its easier if u have a green screen in your photo studio

[drifting flame]

we also have a wx9100

[nova zodiac]

good point

If you have the backgrounds and text already generated, yeah l can see 3 minutes being more than reasonable

[pale latch]

if you do the soliid colr background, always a good idea to set up a back flood light so that color splash doesn't happen

[hallow sable]

nearly every program I know slacks when it comes with AMD

[pale latch]

fill light? i'm not entirely sure of the technical terms. don't want green reflecting back is the goal

[nova zodiac]

I feel like there would be people in here having that going in a couple of hours if they had the hardware and the boss was paying them to do it

[drifting flame]

I dont really understand about hardware, my office is willing to pay stupidly amount money for hardware but no for human resources. Thats why they asked a graphic designer to do this instead hiring suitable experts

[nova zodiac]

so yeah, 2 months is more than feasible

[trim magnet]

i remember when i tried daz studio with an amd gpu and it was so slow rendering with nvidia iray

[hallow sable]

I hear Devs talking shit about AMD and then I hear AMD talking shit about dev studios

[drifting flame]

thanks for the input

seems i came to the right place

[hallow sable]

like with blender... I think its the studio at this point because they keep adding junk instead of improving stuff

[drifting flame]

Can you help me by giving me more insight, what are the important points I need to learn im really new into AI stuff

[nova zodiac]

look for a tutorial on inpainting backgrounds

[drifting flame]

thanks

[mortal anvil]

I have few pages on facebook, all 500k-1m followers, I want someone to JV with me. As I don’t create my own content? Anyone up for it? #content #help

[nova zodiac]

jv?

[gentle bloom]

where do people go to ask question about stable diffusion, seems asking here it's hard to archive and find it back in the future

[nova zodiac]

depends what the question is..

[pale latch]

how do i use stable diffusion to steal the declaration of independence? the real secret one kept under lincoln

[gentle bloom]

I have over 200 2D design, in both 2D sketch and it's final 2D polished if you will. how do I go about it to train a model for it to learn the transformation from first pass 2D sketch to final polish drawing, so that I can pass the model new 2D sketch and it spit out the final polished image that're both in quality and style match the previously 200 final polished design

[gentle bloom]

getting my question ready

[pale latch]

i would train a lora on your finished work and use existing sketch controlnets

[gentle bloom]

is there a "beginner's guide" video that you can point me to? I'm total beginner

[pale latch]

https://www.nextdiffusion.ai/tutorials/how-to-install-controlnet-extension-in-stable-diffusion-a1111 good site with tutorials here. there do videos and text guides.

learning how to train a style lora is going to be a little more involved. easy enough but you will have to experiment a bunch

https://www.youtube.com/watch?v=RT2jj-5t8x8 the scripts kind of update over time, so this is a little old. like most videos. but yeah. good primer for it still

[drifting flame]

hey did anyone familiar with integration between Stable Diffusion and TouchDesigner?

[gentle bloom]

watched all of it, so much concept are unfamiliar, lora, control net... don't even know how to dissect it, for image to image and with my use case in mind, what kind of project that are be suitable for me to start experiment on, or what keyword I should search video for

[pale latch]

You've bit off a big piece of cake to chew for someone who has no experience. Try starting with a glossary of terms

[subtle steppe]

Huh

[nova zodiac]

Uh

[wary wren]

hello

[fervent thunder]

hey wassup guys am new here

[halcyon willow]

What is the next logical step for Stable Diffusion? More SVD development? Improving SDXL turbo?

[copper orchid]

What is best model/method to make dating app pictures atm in 2024?

[sudden ruin]

With a camera

[copper orchid]

Hahaa might be true. But i would love to know latest thing right now and test it. Im so bad at taking pictures of myself 😦

[sudden ruin]

I dont think people appreciate if youd use AI, just be yourself and youll be fine

[lofty basalt]

hi

where are the bot channels? 🤔

[copper orchid]

You are propably right. But i would like to test and see What are the results. At somepoint dreambooth was way to go, But not sure What is bleeding edge technology right now

[sudden ruin]

#1047610792226340935

[sudden ruin]

IPAdapters probably right now

Or using a face swap tool

[copper orchid]

Okey! Thanks alot!

[split tapir]

what version of automatic1111 is best for amd gpu on windows?

[warm junco]

Depends on what you want, Pure Speed or (Usability/Compatibility)

[split tapir]

pure speed i trried comfyui but i only get 1.4its

and it doesnt use all my gpu ram so it crashes trying to generate anything above 512

[warm junco]

whats your gpu?

[split tapir]

6700xt

ik i shoul have got nvidia but i didnt know about this AI stuff and upgraded from a 1060 😔

[warm junco]

then i would recommend using the amd version of auto1111
you would get the most out of it

[split tapir]

do you know what frok that is ?

[warm junco]

yea its the one from lshqtiger,
youll find my full install guide of it in the #🤝｜tech-support channel, in the pinned messages

used it on a 6700xt, 6800xt, Vega7 and 7900xt, and the usage is good, speed is okay, not the fastest

if you want max speed you would need to use the Shark webui from Nod.ai, but that lacks support of the most features, also no extension support

[last socket]

how to use the ai?

[warm junco]

the bots here to generate images are currently under maintenance.
you would need to use an other service, or install Stable Diffusion localy on your PC

[last socket]

oh thanks

[split tapir]

I followed the tutorial and still get 1.4 its 😦

[warm junco]

yea it uses directml, the same as comfyui uses for AMD, but you can generate at higher res than 512

and you can even use Hires fix to upscale images when done right

[broken cave]

everything you are going to try will perform more or less the same under comfyui & directml on windows.

[broken cave]

improved comprehension of the prompt

[broken cave]

you can fine tune sdxl to learn your style. you can use controlnet to turn sketches into drawings.

[bold umbra]

hello everyone!

[robust sigil]

hi, im back after quite a while and am wondering, where i can type in my prompts, because the bot1-17 etc. system seems to be offline...

[sterile torrent]

guyw what techniq is good for upscaling? upscaling +ESRGAN+ 4x, hires fix, refiner? its a lot of nods and options and i dont know what connect with what 😄 SDXL and SD, or all connected in one? can someone recommended

For now im using two workflow, simple workflow with hires fix and another simple workflow with ESRGAN 4x upscale, but i dont connect refiner and stuff

[potent spire]

Fooocus seems simpler to work with but it has the same issue as the other ones: Performance on my laptop lol

[sterile torrent]

foooocus is to simple for me and lack of things

[potent spire]

depends on what someone wants, i dont need SD for more than some generations of things which i cant achieve with Firefly

but it hits my laptop hard so i guess i will have to go for some web UI

no local

[spark solar]

Is there an AI tool that is good at including text in images? I'd like to find an SVG generator that can input things like a person's name, and add decorations, embellishments etc.

[lavish current]

It's not SD but there's || Microsoft Designer ||

Iirc some SDXL finetunes are also good at this

As well as StarryAI and LeonardoAI

I usually use Bing Image Creator/Copilot nowadays

AI art rocks though, so many cool ideas I would've never come up with on my own

I just doodled something for like the first time in months, drawing is super slow but the AI helps a lot

[timid fern]

Is there any sdxl normalmap controlnet/t2i-adapter/anything? I'm not sure if there actually isn't one, or I'm only looking in the most obvious places...

[wind island]

Anyone here good with unerwater/landscape/simplistic cartoon like styles look at my post in community projects
looking to colllabbb!!👀

[potent spire]

i like MS Designer/Image Creator but i wish it would have less censorship since i already use another one that is heavily censored

[rich sapphire]

I have a newb question. I'm trying out a new (to me) SDXL checkpoint, and the images look great all the way til the final step, and then it switches too oversaturated, bad coloring. I'm seeing comments that say this is a VAE issue with unpacking? but I'm not really sure what to do to fix it. Any advice?

[nova zodiac]

a1111 or comfy?

[rich sapphire]

a1111

[nova zodiac]

if its a1111 and the vae is set to automatic or none - download this vae and set it to use this https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/resolve/main/sdxl.vae.safetensors?download=true

[rich sapphire]

Boom, fixed. You're awsome. Should I keep that as the default for all checkpoints, or only use it when I run into this saturation issue?

[nova zodiac]

default for sdxl checkpoints

[dusky stirrup]

Saw news about CUDA Library replacement for AMD GPU ZLUDA. Looks really promising for AMD folks

[stable pawn]

I do not like how AI use poor language, like you do not get a cat when asking for a pu**y, and what shall we write if we want a cherub?
No no no no no, that is not a small child, a cherub was horrible monsters with a lions body, eagle wings and a human head with a beard and also goat hoofs. Will it ever come a grammatical correct AI.... did the dyslexic person asked.

Fgs, even this server blocked a correct word for cat.

[fervent thunder]

Hey guys i make 30k with dropshipping thank you lord

[iron imp]

It uses whatever it was trained to use.

[mild geyser]

Hey G's how y'all doin, i need help with stable diffusion set up when I open the webui-user.bat file and try to run it to gain access to stable diffusion interface it gives me this error, Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check, any ideas on how to fix this

[stable pawn]

Yes, the downfall of AI is that it is trained on the open Internet when all good info is behind pay walls 😄

[iron imp]

Just call Discord by it's name 😂

[stable pawn]

On the other hand, the Discord AI do kind of funny images as long as you do not use educated language, you never know what you get when ask for a pu*** in Boots.

[mystic star]

Someone knows about the amd gpu issue?, i have a rx580 and it dosen't use my gpu,it uses my cpu

[warm junco]

there are various models for controlnet, like here some:
https://civitai.com/models/38784/controlnet-11-models

[warm junco]

can you share some info in #🤝｜tech-support , mostly you run the wrong webui

[lavish current]

It depends on what you are trying to generate

[timid fern]

Unfortunately not for SDXL though 😦

[pale latch]

lots for sdxl too

[pale latch]

https://huggingface.co/lllyasviel/sd_control_collection/tree/main

[timid fern]

thanks, yeah seems like there's no normal map unfortunately. will try a bit more with midas or zoe depth estimation

[pale latch]

https://huggingface.co/bdsqlsz/qinglong_controlnet-lllite this guys stuff too

[timid fern]

wtf wow thank you for finding this I would have never seen it. Seems like it was only trained on anime, but idk, maybe it'll work decently for photoreal as well. I hope it works

[pale latch]

yeah theres lots of obscure efforts

[timid fern]

curious how you found it? What search terms did you use? B/c even when I use huggingface search I didn't see it. If there are other obscure efforts like this I want to be able to find them

[mortal delta]

Do any of you know if there is something for stable diffusion that turns an image or text and turns it into a 3d model, or is that not a thing yet?

[potent spire]

Stuff like Pokemon and so on works...but when it gets more controversial it doesnt

I cant make parody of Jesus or prophet Mohammed for example

Jesus works with some things

[potent spire]

There was something but a) if i remember correctly last updated like 6 months ago and b) bad quality (like with all of them tbh)

[mortal delta]

Ill be wiling to check it out do you remember that name of the tool/extension? or should i just learn 3d modeling...

[nova zodiac]

Zero 123

[potent spire]

Stable Dreamfusion.

I encourage people to give 3D modelling a try but the question arises what do you want?

I mean the same applies to 2D as well

[potent spire]

Yeah this one as well

For me what im excited for are actually those 3D gen AI things that generate point clouds instead of meshes

Meshes are "dirty", have bad textures, topology of doom, mesh anomalies that have to be cleaned up as well and so on

[burnt jasper]

Hey everyone i have a question about loras something I really don't get
Is it useful to use multiple lora's with 1.0+ each or adding to over >1.0
i feel like i've already seen many prompts that do but i am not sure because at first read the total needs to be <=1.0

[plain raptor]

i never heard of going over 1.0

usuaully u go, 1.o or less, 1.0 on the lora u want more of on the art, and the other loras under that, so that they appear, but not as pronounced

and from there u can kinda tweak the values

i dont think going over 1.0 is even supported and of it is, it prolly has nu effect on how much its used over the other loras is such value is assigned to that lora

bc nu matter wut, the loras under that one, set to 1.0 will always be less than the primary

its kinda like, think of 1.0 as 100%, and anything under like, for example, 0.9 or 0.6 like 90% 60%, etc, i think anything above 1.0 is still techically 1.0

:3

[peak wolf]

https://ja.stability.ai/blog/stable-cascade 🦗

[nova zodiac]

Some loras will go from -5 to 5... just depends on the training

[plain raptor]

whaa

<v> but then that begs the question, where is middle

if a lora can go to -5, a negative integer cant possably be valid

[inland terrace]

Hello everyone!

I have a question about SD1.5
I'm trying to generate a human character with cat tail.
But whenever i try to generate the image, it always has multiple tails or misplaced tails.
I tried positive, negative prompts several times but it hasn't worked so far.

Do you have any kind of ideas of solving this issue??
Thanks in advance 🙂

[trail lion]

kind of just depends, many times if you have more than one at >=1.0 then you will start to have lots of image artifacts, or lose the ability to produce things the model knows how to draw well. if I'm using multiple loras at once in txt2img, I often will set the ones high that will get the basic shapes in the image (think of like what you would want to see in a controlnet depth map), then using controlnet to maintain the integrity of that, you can play around more with the various loras without using really high values

[trail lion]

in general 1.0 only means, 100% of the strength of the training, but that could be overtrained or undertrained. that in a nutshell is why "it depends"

my personal anecdotal experience with loras posted in public domain, many are overtrained, and that's not a bad thing really, there are times you want to nudge the interpretation a bit harder, and you can always reduce the weight

[stuck sun]

https://ja.stability.ai/blog/stable-cascade

New SD version coming out in a couple of days

[bold python]

I'm Taylor David, a skilled AI developer with a focus on developing cutting-edge conversational AI systems. Creating AI chatbots, like ChatGPT, with a focus on voice AI, SMS, cold calling, and Twilio connections is one area of expertise. competent in creating Broadcast Chatbots, Auto-responder Bots, Telemarketing Bots, and Appointment Bots. competent in deploying frontline AI, VoIP sales, IVR, live chat, and conversational AI for Plivo SMS, guaranteeing smooth and efficient interactions.

[burnt jasper]

does it matter how many lora's you use together--even if they are all 1.0 or lower?

[mighty fable]

where can i create images guys ??

tag that chennal

[rocky lava]

Where do i make pics? Cant find bot1 etc on list

[rocky lava]

Found out...

Sad news..

#1047610792226340935

Read last message

[outer crescent]

Cold calling voice AI. Classy.

[vast berry]

and i'm a unicorn

[plain raptor]

increase cfg scale

by liike, default is 7, so try 9

[plain raptor]

also try my negative prompt,

"boring_e621_v4, aid291, wings, bad_prompt_version2, ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, bad anatomy, watermark, signature, cut off, low contrast, underexposed, overexposed, bad art, beginner, amateur, distorted face"

[stuck sun]

have you searched civitAI for loras that do this

why does it take more than zero seconds to refresh the lora folder in A1111, losing my mind

[plain raptor]

wym

[nova zodiac]

Because it looks into each lora to determine if it is the right architecture, and then only displays compatible loras

[stuck sun]

does it dial a modem first

[fervent condor]

well looks like SAI Japan was too fast with their fingers and published early major SAI announcement

https://ja.stability.ai/blog/stable-cascade

I wonder if they will release it officially today

or in few days

Looks like huge improvement over SDXL again. Increase in VRAM for default models but aesthetic score improvement means probably that it should be waaaay better at prompt following reducing mutant generations and so on.

[stiff quail]

hopefully competes with the big 2 other image gens rn. i need something to be able to work locally again with high qual for anime-esque pics.

[stuck sun]

aesthetic score compared to what lol

scored by whom

[fervent condor]

to SDXL

by same people i assume

They wouldn't be releasing worse model than SDXL because there would be no point

[stuck sun]

lol most ppl still use SD 1.5 instead of SD 2/2.1 and XL

but hopefully it'll be faster and easier to train than XL

[trim magnet]

isnt that what happened with 2.1?

[stuck sun]

yes

[fervent condor]

SD1.5 was not released by StabilityAI it was released by researchers who were involved in creation of SD and who decided against SDAI wishes to released model uncensored.

2.1 is better than 1.4 to which should be compared

[inland terrace]

Yes i did. But none of them give me exact 1 tail

[nova zodiac]

there is a standardised aesthetic scoring model out there..

[hushed quarry]

You mean all the ones who work at stability?

[fervent condor]

I see you were not there at release of 1.5 and following controversy spat between Emad/Stability and researchers

[hushed quarry]

I was definitely there haha. Since that the 1 researcher who was not at stability joined stability

Whole team is here now

[fervent condor]

Then why do you not remember that spat ? Should i post logs from hugging face or something ? They are still there. Clearly 1.5 was not meant to be released as such.

[hushed quarry]

I do remember it well. Tbh was a little silly at the end of the day. 1.4/1.5 were not all too dif from one another, given it was only 550k steps of retrain in a laion sub portion. Either case its long past now

[fervent condor]

yeah

[hushed quarry]

I guess I just like people to know on the research end we are all buddies at least haha

[fervent condor]

so was japanese SAI hot finger to fast or this is japan release ?

I mean researchers sure are buddies no question about it. management and corporation when money is involved are entirely different thing, especially laws to follow

Emad on twitter was raving about something to be released soon and i assume it was it

[hushed quarry]

Yea definitely, navigating the gen ai landscape can be a whole thing to be thoughtful in

[fervent condor]

i remember japanese courts outright said AI companies can use copyrighted data unlike US or elsewhere where it is still gray area

so maybe japan release has some meaning

[hushed quarry]

Huh wonder which thing it was, tbh I am not the best at following his Twitter 😅

[fervent condor]

truth be told SAI is cooking multiple things at the same time so it might be about something else

[hushed quarry]

Not japan specific, just think they are up early

[fervent condor]

nice, then so today release, hope whoever sleeps with button on hugging face repo wakes up early haha

Speaking of which. I would love to see proper image to text to image model released. So for example you talk with model :paint me a tree and then you say ok now make tree taller and have bananas on it.

in theory you can do it with vision model and then llm as fetch for SDXL creating prompt but live editing would require whole end to end model

I think creatively this is where goldmine is, more so that prompt style.

[plain raptor]

potato

why dunt u just work for erm

Microsoft bing team

[hushed quarry]

Definitely, iterative models are an area to be majorly improved on. We saw some early variants of this with say ip2p and emu edit, you can honestly get away with a lot just doing plain old noise inversion. There are some potential model changes that could be fun to help with this task but really I think the biggest barrier is generalizing ever form of edit someone might want both in the language side which likely will incorporate a vlm and the image model side to enact that change is on the data for training one. Open source VLMs still need some love and improvement and a lot of this will become easier

[fervent condor]

Shoudln't dataset be a problem ? I mean creating image from image text pair is workable, but teaching how to take language and modify image is entirely different problem.

I mean such dataset would need to be probably creating from scratch

and you have chicken and egg problem essentially

What gives me hope is that with huge LLMS you have emerging properties. Like spatial knowledge where model can infer gravity or other forces without having experiencing gravity

[hushed quarry]

Yea data is pretty much the biggest barrier to most problems I still see in the entire generative space. Making a set that encompasses enough radical and tiny modifications to generalize a model of that form to most things someone would want is not an easy task.

[fervent condor]

I still use 3 wooden box problem to test model inteligence. Puzzle goes like this. You have 3 wooden boxes in exact same shape and size you put second on top of first and third at the side of second. Puzzle does not contain any information about friction, gravity and other things you should take into account and most of LLM models fail at it aside from really big ones (or best ones) like GPT4 or miqu70b

Effectively puzzle of second order when answer relies on things outside of puzzle

[hushed quarry]

Yep, just need enough strong data points and a model capable of learning from the distribution and you get some magic haha. A very strong accessible VLM would be helpful in making a synthetic data set for a task like this or at least filling knowledge gaps in training

[fervent condor]

So if model is just next word generator it will fail, but if it had emerging property of reasoning based on experience then it will not

another thing to consider is probably compression artifacts.

the ability of 7B model to call up data from a terabytes/petabytes of data is insane

it leaves in dust even the best compression methods, but it has its price in form of compression artifacts aka wrong data, hallucinations and so on.

[outer crescent]

dude is literally spamming his resume as a spammer, lol

[red wren]

Do you guys feel any difference on using Comfyui or automatic 1111 on pinokio or external browser? for me automatic 1111 is way slower...

[nova zodiac]

try sd-webui-forge

a1111 is slower, but forge is as fast if not faster than comfy, but with the same webui

[timber axle]

Did you find?

[stuck sun]

hey question, i use remote desktop to connect to a couple of machines on my home network to drive them from my office but more recently one of the machines (4gpu quadro nvlink 32 core threadripper machine) machine has started kicking me off the connection when i'm generating stuff with A1111. I can reconnect again shortly afterwards but if keeps happening. Anyone ever experience anything like this?

network speed is 10g

[warm junco]

Maybe the vram is so maxed out that it can't keep the stream on

Because the remote software uses vram too

[stuck sun]

yeah thats what i was wondering so i plugged the video outputs into a different card and its still happening

[fervent thunder]

20GB VRAM, kek.

[nova zodiac]

Try using —listen and then use the web browser on the connecting system instead of rdping in

[vagrant gazelle]

Hi

[stuck sun]

yah i've used that before, just annoying because i still need to be able to drive the machine, not just for SD

[warm junco]

Im using Rustdesk for Remote control, maybe give that a try

[stuck sun]

thx will look into it

[undone garden]

https://huggingface.co/stabilityai/stable-cascade/tree/main

[oblique willow]

Do yall know any extension which can help me improve my prompts?

Like suggest me what should I add.

Or any other cool extension you'd suggest using

[sterile torrent]

what version download this model from files list?

[plush knot]

Hi

[sterile torrent]

Is here someone from Poland?

[outer crescent]

yes, two things: dynamic-prompts has a 'prompt magic' feature that runs your existing prompt through one of several language models and ads cooler stuff. you might also want to look up some styles files, which are snippets of good prompts for certain styles, like vector, photorealism, retro, cute, 3d render that can be appended to your prompt and beef them up, kind of like an autotune for your existing prompt. 😄

[frank mirage]

Hallo

[steel garden]

which text2image model (local) has the best results for a A100 GPU?
maybe on dalle3 level
i want this art style: https://24posters.co/products/crypto-kids-6
maybe someone know a specific model/checkpoint

[oblique willow]

You are a lifesaver. Was looking for something exactly like that. Tysm

Any chance you can paste the link of that dynamic prompts extension ?

I struggle finding the right one.

[outer crescent]

https://github.com/adieyal/sd-dynamic-prompts/ sadly im not sure if they have a comfyui version or not. if you use A1111 install this via the UI rather than outside of it. (extensions > from url)

[oblique willow]

Thabks soo much i use auto11

You are awesome

[outer crescent]

https://github.com/adieyal/comfyui-dynamicprompts oh the dev DOES!

[oblique willow]

: )

[outer crescent]

that's cool, no probs. One thing i will say about using magic prompts. never use any of the models tagged unsafe, they contain some really horrible creepy stuff that would make a 4channer blush and you really don't want something randomly generating images on that.

its actually worth looking the extension up on youtube too, it does quite a lot

[oblique willow]

Oh : / . I will not download random models then

Do you the best one?

Just 1 model for general work. (No nsfw).

Which was trained on most prompta

[outer crescent]

yeah, the gustav model is great.

[oblique willow]

Ty i will get that one then

[outer crescent]

a1111 will download the model for you first time you select it

[oblique willow]

Oh wow cool.

[outer crescent]

i forgot, it also has "im feeling lucky" which will take random prompts from lexica.art and play with them too. if you cant think of anything at all to prompt.

[oblique willow]

Thats handy too. Tysmm. Is thwre any other cool extension you suggest ?

I am noobie.

Or Tool

[outer crescent]

controlnet is very very good too. i would suggest playing with dynamic prompts first and then getting controlnet. or controlnet first. both are kind of a whole thing to learn so you dont want to have 2 extensions and not be familiar with either

[oblique willow]

Oh I got that one. Controlnet.

[outer crescent]

yeah i dont think i know anyone that doesn't by now

[oblique willow]

XD

I'd suggst, Lama cleaner it helps a lot. In case u havent tried yet maybe..

[outer crescent]

dont think i have. i just between comfy and a1111 a lot

[oblique willow]

You should definatly get it

It removes things from picture with your brush.

Just paint on what y dont want and it will fill

[outer crescent]

https://github.com/huchenlei/sd-webui-openpose-editor this is good too, it works like an extension for controlnet to let you edit the open pose models and move little things around, or attach a hand if one wasn't detected.

From ControlNet extension v1.1.411, users no longer need to install this extension locally, as ControlNet extension now uses the remote endpoint at https://huchenlei.github.io/sd-webui-openpose-editor/ if no local editor installation is detected. nevermind, its actually integrated into controlnet now, im behind the times 😄

i will look into lama cleaner though, i tried something similar last year, but it was a bit underwhelming

[oblique willow]

Oh. So it lets me edit pose image?like i can move hand points around with mouse?

I didnt see this feature yet

I thought it will need a 3d program or somewebsite or photo editpr

Bcz it generates a 2d pose flat image ig

[oblique willow]

Mikuu

[outer crescent]

no its tucked away. i cant share a screenshot with you but if you press the little explosion icon beside preprocessor (using openpose) it will generate the pose model for you. on that image it will have small buttons. edit, JSON and close. click edit and it will open up a finetune editor. its a bit janky but still good.

[oblique willow]

So its already added to controlnet?

I will definately look for it next time I use

Thanks so much.

^_^

[outer crescent]

no problem, learning and sharing is pretty much the whole AI community. we're all running blind figuring things out from each other since theres no definite academic courses invented yet 😄

[oblique willow]

: )

[fervent condor]

any front end using stable cascade ? Comfy UI ? Automatic ? etc ?

[agile tusk]

Not yet

[fervent condor]

damn it, from online demo it works amazing

it is really great imrovement in following prompts

[agile tusk]

Nice! I've not had chance to try it yet.

[fervent condor]

there is online demo but it is currently hugged to death lol

https://huggingface.co/spaces/multimodalart/stable-cascade demo

Example: Red squirrel sitting on motorcycle in middle of shopping mall
output on first generation:

https://multimodalart-stable-cascade.hf.space/file=/tmp/gradio/ef8601e3c1744bfbf161140673733dfe8fdead76/image.png

damn it online demo deletes fast generations

but it followed really really really good prompt

Kind of like DALL-E3

[agile tusk]

Looks promising. Can you try some with hand in and see how that copes?

[fervent condor]

from initial reviews hands are still problem, dalle3 still has those issues.

still to early to tell. If it can improve in that aspect by 50% that is still a win

[agile tusk]

Not even a little better hit rate?

[fervent condor]

online demo is hugged to death by public rn

generating anything is almost impossible

[runic temple]

hi

Stable Diffusion WebUI Forge needs xformers?

[dim surge]

is there something similar to photoshop's "puppet warp" here in SD?

[tropic frost]

say theres something im curious about, saw it a couple of times but dont know what its called. Basically before you use a image as reference for a generation, you grey it all out so only a sihluette of the original picture remains.
Any ideas whats that called?

[fervent condor]

looks like online demo got some improvements and now there is proper queue and you can see generation in real time

woman buying dragon egg in fantasy mall

https://i.imgur.com/Devq201.png

war scene from movie, lizards attack human knights

https://i.imgur.com/0Sy9YWc.png

[warm junco]

Maybe you mean Controlnet with Canny mode

Or Depth

[tropic frost]

deph, thats the one

[fervent condor]

scene from sci-fi movie, spaceship battle over planet

https://i.imgur.com/gIYMcOq.png

those all first generations no redos

[plush knot]

https://huggingface.co/spaces/enzostvs/lora-studio

Free Lora gen

[grave depot]

Stable Cascade looks pretty promising.

[fallen gale]

does the bot use stable cascade now

[agile tusk]

There is no bot #1047610792226340935

[fallen gale]

ok

[outer crescent]

depth can be very good at getting poses correct. a lot of people dont really use it, but if you combine it with openpose as the second controlnet it works very well for fidgety poses like dance or crossed arms or legs that pose just doesnt quite work 100% with, like understanding which leg is in front of the other.

i mean openpose as the second controlnet and depth the first, my grammar is weird 🙂

[slender field]

training needs 20-24 gb ??? but waht about running a model ?

[nova zodiac]

Depends what batch size they recommend for training but given the numbers of parameters involved it should be similar to sdxl

Assuming stage models are swapped out

[slender field]

i am not an model trainer so my question was can i run it offline with 16gbvram

because it is very near dalle3

and to my wonder its based on wuerstchen

@nova zodiac

[nova zodiac]

Im guessing yes but that may be premature

[rotund forge]

Guy in the DB training server said it was ~9gb if you use gpu offload - and there are multiple models so will just have to see....

[spare plume]

Oh Nice, stable cascade. Sounds like a nice excuse for me to ditch my RTX2060, or will it run?

[long wigeon]

I admit I haven't researched this at all, but what exactly is the difference between SD1.5, SD2 and SDXL?

I assumed SDXL would simply be the newest and best version, but for example on Civitai there's far less LoRas for it

[iron imp]

There are other deeper differences but SDXL is 1024x1024 it requires more time and vram to train. That's why there are less on Civi.

[long wigeon]

Ah interesting! What are the other differences then?

I am currently using SD1.5 and generating my pictures around 400-600 pixels depending on the aspect ratio I need.
If I used SDXL for that, would that also take longer?

[iron imp]

https://sandner.art/sdxl-vs-sd-15-a-deep-dive-into-image-generation-ai-performance/#:~:text=SD15 base resolution is 512x512,and full fine-tuning support.

[long wigeon]

Thanks!

[stiff quail]

stable cascade is the newest and best version, just came out today

[long wigeon]

Now if that isn't an amazing coincidence

[stiff quail]

yep; https://huggingface.co/stabilityai/stable-cascade

but i'm not sure if anything has been updated to be able to use it yet

there's a demo for it on hugging face @ https://huggingface.co/spaces/multimodalart/stable-cascade but it's being... hugged to death if you will xD

when it works the image quality is great.

[pale latch]

there's standalone scripts released for it too. no ui. its fun to use generative ai that way sometimes

[stiff quail]

where's the standalone scripts? i don't mind it being clunky, haven't found anything yet tho.

[pale latch]

https://github.com/Stability-AI/StableCascade

[long wigeon]

Ah so it's not possible to load a stable cascade checkpoint into Auto1111 or Comfy?

[stiff quail]

not yet, it's the very first tri-model

sdxl is a dual model, cascade uses 3 models in its inference

[long wigeon]

Where can I read more about what that is?

Tri-model and inference

[stiff quail]

https://stability.ai/news/introducing-stable-cascade this is the official announcement

[pale latch]

i know they support the wurschein (sp?) architecture with the first model they released. shoudn't be long before cascade gets ui support

[long wigeon]

Great

[stiff quail]

ugh; the scripts are in notebook format, i'm not very good at setting up notebook stuff

[pale latch]

wurstchen! with an umlat but i don't have access to those characters

[keen ruin]

I got it running locally with this: https://github.com/EtienneDosSantos/stable-cascade-one-click-installer

[stiff quail]

nice, ty.

[long wigeon]

Würstchen !

[stiff quail]

wait that's wuerstchen

[long wigeon]

Yes ue can replace ü

[keen ruin]

That said, I think it loads all the biggest models, so YMMV

[stiff quail]

is wuerstchen v3 and stable cascade the same thing? i've been seeing people say that

[keen ruin]

They're built in similar ways, I believe

[stiff quail]

that script is specifically for wuerstchen v3 i think

but i'll try and get it setup & see

[keen ruin]

It's not pulling wuerstchen, it's using StableCascadeDecoderPipeline and StableCascadePriorPipeline? But I've never used wuerstchen, so I can't really compare

[stiff quail]

ah i was reading the code directly and it says "pip install git+https://github.com/kashif/diffusers.git@wuerstchen-v3"

[cedar salmon]

there are demoe on HF, quick give me a prompt before it gets hammered

[keen ruin]

I think that because they're built so similarly, they borrow some underlying base code. But the models it uses and the diffusers it uses are all Cascade

[stiff quail]

oh ok

also i never used diffusers before, is it still like 100% local?

i always used the manual setups/webuis/etc

[keen ruin]

What I just sent you is, yes 🙂

[stiff quail]

alright. ty. gonna run it now and see

[pale latch]

Umlauts are so metal : 🤘

[keen ruin]

It uses just around 20GB vram with whatever models it downloaded - I think that it's using the biggest models. I'm not quite certain where they got downloaded to, so I'm not sure how one would switch them out

[stiff quail]

np; i have 3090

i might have to close youtube tho

or nah i have more than enough leftover even with yt open xD

[honest spear]

will cascade finetunes be 3 models or only one will be the finetune and the rest is the same?

[pale latch]

YouTube using vram? Weird

[stiff quail]

sometimes discord/youtube open total to 1GB vram but not right now

seems i'm only using 0.5

[pale latch]

Vram?? Or system ram?

[stiff quail]

vram i think, when i used to use kobold AI and maxed out my vram i had to close everything out

[pale latch]

I've never encountered such things..nifty

[cedar salmon]

it generates at 1536 x 1536

[stiff quail]

that's the max u can go before quality loss?

[tall musk]

@keen ruin it doesn't use 20Gb it uses just under 15Gb (bfloat16) , it would work on Colab free tier if float16 was not overflowing, and thats the full fat version, the lite version should be significantly smller

[cedar salmon]

dont know, just on a huggingface space and it let me gen at that

#🏞｜general-with-images message

[keen ruin]

Well, the script I linked above is using 20gb - I can't exactly tell you why, but it do!

I've been switching between the Cascade model and a base SDXL model and this may be anecdotal, but **hands **generate a lot better! I'm having more trouble breaking hands

[stiff quail]

2.1.2+cu118 won't work, will it do fine if i do 2.2.0+cu121?

o same with torch vision, also fails install so need to bump up the ver

[stiff quail]

did you have to do anything special to set it up? the torchvision version and pytorch version it wants to dl don't work & it fails to build for me.

[keen ruin]

I did not - I can see the author is currently adding commits, though, so I have no idea if what you cloned and what I cloned is going to be the same 😅 I have the commit with Code_Example-v0.0.3.py

[stiff quail]

the install.bat is the problem for me, i can't get the specific version of pytorch/vision it wants apparently

[keen ruin]

That's odd, could be a Python version issue? Mine seemed to install fine without any errors. I'm not sure if it pulls any global dependencies or if it's all in the venv (I'm not too knowledgable with Python specifically)

[stiff quail]

using 3.11 or 3.12?

[keen ruin]

3.10.11 actually!

[stiff quail]

"Could not find a version that satisfies the requirement torch==2.1.2+cu118 (from versions: 2.2.0, 2.2.0+cpu, 2.2.0+cu118, 2.2.0+cu121)
ERROR: No matching distribution found for torch==2.1.2+cu118"

ah okay i'll downgrade

[outer crescent]

its silly, but there are so many dependancies on it they have to keep it there for now. like how big military bases all use old computers. it would take forever to upgrade everything and make everyone else do it without issues.

[stiff quail]

yea; i know newer versions sometimes don't work, used to a lot of the stuff being made for the new stuff.

[outer crescent]

also if you can, use pip to downgrade

[stiff quail]

yea it's downloading it fine now lol

[outer crescent]

its annoying when you have drivers and stuff screaming to you to upgrade and also being told to keep it low. i remember the little screaming fit moving up from torch caused 😄

[keen ruin]

I have been using 3.10.11 for like 8 months without issue (knock on wood), I'm scared for the day I need to upgrade

[outer crescent]

also if you ever need cuda different versions (you almost never ever should unless youre devving) you can log in to nvidia with an academic account and get builds.

[keen ruin]

Thanks for the tip!

[stiff quail]

getting no module named 'torch' error now ;_;

[outer crescent]

have you maybe missed a step in installation? like does your computer have the binaries?

[stiff quail]

well with the install.bat i'm getting some kind of error at the end

ERROR: Failed building wheel for insightface
Failed to build insightface
ERROR: Could not build wheels for insightface, which is required to install pyproject.toml-based projects

o this is hyper-manual, i need the visual studio c++ build tool things maybe?

[outer crescent]

yeah the build tools

and it might also cry that you installed them after instead of before. first time is a headache but youre moving through it quite well

[stiff quail]

i've done some of it before, just on a fresh windows install now >w<

[outer crescent]

oh thats a face swap thing. thats a whole iffy issue i wont give tech support to. just because. anything else, sure.

[stiff quail]

"one-click install" is def deceptive XD

but i'll do as much as i can to get it running, it shows promise

[cedar salmon]

install wheel

[stiff quail]

pip install wheel?

[cedar salmon]

ya

[outer crescent]

its worth the tears

[cedar salmon]

then try again

[stiff quail]

finishing vs build stuff first

lemme upgrade my powershell to terminal again too one sec

[outer crescent]

a good powershell script would be great for new users

instead of a bat.

dont worry though, you'll get there 😄

[stiff quail]

i already had pip wheel

checking how the install goes with vsmake stuff nowh

[cedar salmon]

im sad i dont have time today for this, maybe someone will implement the diffusors pipeline to comfy nodes 🙂

[stiff quail]

ehehe. i wish. cuz it also uses a better version of pytorch with a more efficient

thing that makes it run better

xD

[outer crescent]

you might scream "never again" but you will get it. getting AI working on your PC is honestly more than any academic course would ask you to do. if you dont know about python or wheels or orders or venvs its especially infuriating. the should be a small cottage industry of people installing it for people

[stiff quail]

i knooo i'm not a total novice, but i'm on a fresh windows install so i have to get every single thing it needs manually

it's progressing more than before

i was expecting one-click install to be one-click install xD python is actually really intuitive

cuz it gives you nearly all the info you need to resolve a problem luckily

[long wigeon]

I’ve read the stable cascade introduction. It mentioned that it’ll be more efficient and better for lower end hardware, but at the same time mentions 20GB VRAM

I’m currently using SD1.5 as my setup takes too long with SDXL. Can I expect to be able to run Cascade then?

[stiff quail]

the screenshot the creator of one-click script uploaed shwoed it using all their 16GB VRAM

[outer crescent]

no no, its not a trap for idiots, its just quite unfair if you've not got a pristine clean machine. 9/10 times for me its a venv issue and im accidentally in the wrong environment.

[stiff quail]

ah yea, i made sure to read the scripts

and saw that it's properly activating the right venv n stuff too

finallly it finished

[outer crescent]

before/after cooking videos 😉

[stiff quail]

xD

[waxen moss]

How is Stable Cascade compressing the images so heavily at the first step different than what a U-Net does? It sounds very similar to me.

[outer crescent]

christ im sure its a white paper i couldnt even real the first sentence of. everyone lost me at LCM 😄

[long wigeon]

Well that could depend on what they’re generating. But how does it compare to the SD models relatively?

[stiff quail]

Will see in a bit, i was able to gen a couple things with it and it looked like early dall-e 3

couple things with the huggingface one

[cedar salmon]

i posed 3 cats in gen w images

[tall musk]

@keen ruin Are you doing a lot of images per prompt, they've gone for the way of doing thats uses more memory than necessary in the script, assmuing your interested in using less memory

[long wigeon]

Are you talking to me?

[tall musk]

@long wigeon Sorry no didn't notice the @ thingy didn't pick up a name

[stiff quail]

it's definitely improved over sdxl but it's not as big of a leap was i was hoping for; still will be worth experimenting with a lot.

[keen ruin]

No - I was only using one image per prompt.

(I am fine with using 20gb on a 4090, my note was more for anyone else)

[cedar salmon]

i think its phenomenal, huge detail first pass

[stiff quail]

It's much improved over SDXL, not as much as I'd like for anime style images.

[cedar salmon]

ahh havent tried

[stiff quail]

i modified the script a bit already to crank out gens faster since i'm not changing height/width/num of images

seems realism/regular drawings improved a lot too

it follows prompts well but i think it needs more than 30 steps

[long wigeon]

That’s great

[stiff quail]

oh maybe it just needs less guidance scale

the ones made with like 6.5-7+ didn't come out well but 4 came out really well

[lusty beacon]

Stability just dropped a brand new model they've been teasing for months, called Stable Cascade. It's really good at handling text (Including text into images)

[tall musk]

@stiff quail Yep the sample code use 4 for the prior and 0 for the decoder call

[nocturne pewter]

Was just reading about it. Looks promising. Did not fully understood but you can’t use it with A1111 or Comfy, right?

[lusty beacon]

im not sure

I dont think so

[stiff quail]

not yet. there's a script for it but you'll need vsbuild , python 3.10 & git all manually setup

[lusty beacon]

you can use pinokio

[nocturne pewter]

Yea, thought it would be something like tha. I found the three models but thought it’s not possible to yet use them like that. Better to wait then. As far I understood lora training too is improved greatly.

[tropic frost]

heres a funny one

if i generated a image with a character

is there a way to remove the character, so that i can showcase the background behind them?

[nocturne pewter]

You could try experimenting with inpainting, mask the character and generate the background on it

Lower cfg and medium denoising might be something I’d star with.

Then experiment with fill/object and whole image/masked. Object/masked could work

That would be A1111 but those translate to Comfy too I think

[lusty beacon]

https://huggingface.co/stabilityai/stable-cascade

[fervent condor]

downloading via pinokio let's see what it can do

[woeful maple]

does it take the a1111/comfy authors to do some kind of update for it to work or is there something you can do manually?

[nocturne pewter]

I’m guessing a1111/comfy needs updates or patching or something to add the functionality

[stuck sun]

quickest way is to use photoshop AI object select and erase tools

[lusty beacon]

its unusable

[fervent condor]

it works great for me

[lusty beacon]

its slow

idk why

[fervent condor]

uber fast for me

faster than sdxl

[lusty beacon]

are you in discord of pinokio

[wanton siren]

any models for a gpu with 6gb of vram?

[lusty beacon]

?

[lusty beacon]

check dm

[fervent condor]

it takes around 9 seconds for me at increase steps compared to default and 1024x1024

from my testing this is huuuuge step up

almost as big as from 1.4 to SDXL

[lusty beacon]

does it understand general text like dalle?

[fervent condor]

the biggest difference so far is lack of "wrong" parts. If you say to generate something it will probably generate it in just one generation correctly rather than you repeating generations until you get something that is "correct"

yes, understading of text is much much better

[lusty beacon]

awesome

Can I run this on Huggingface?

[fervent condor]

as for censorship it is slightly less censored than SDXL and can do better faces from get go which was SDXL problem

[full lark]

Do you guys think it would be possible to run that on a 12gb gpu? 🥺 , I tested it but only the online demo

[fervent condor]

there is online hugginface demo

[lusty beacon]

where

need it

[full lark]

https://huggingface.co/spaces/multimodalart/stable-cascade

[lusty beacon]

thx

[full lark]

Here

np

[outer crescent]

12gb gpu, yes absolutely.

[fervent condor]

original post from stability is that parts of stable cascade can be "replaced" with lower models or something like that

the main benefit of this model seems to be (according to SAI) much better and faster finetuning

[stuck sun]

its supposed to be easier to train right?

[fervent condor]

so far from my testing it is 90% DallE3

just without crazy censorship

whatever it is higher or lower than 90% dalle3 i need more time spent with it

hands wise it still has issues but dalle3 as well has those issues but generaly i see improvement

[outer crescent]

please stop calling a thing not doing boobs and things censorship. It isnt censorship when the guy at subway, or your tax guy or whatever doesnt flop himself out at you. he could, doesn't. hes not censoring he's just being normal.

[fervent condor]

models are also better spatially aware, so if you have sword characters seem to make less mistake with how swords are held

[full lark]

Sounds amazing if that´s the case, remember its the base model, the community will fine-tune it too like sdxl or 1.5

[fervent condor]

Yeah finetuning of both 1.5 and SDXL directly increased their capabilities by a lot

[hallow sable]

well it is censorship... anything that deliberate blocks information is censorship

[fervent condor]

What i am waiting for is proper frontend integration with control net

stuff should be crazy, or inpainting

[lusty beacon]

Check this out https://www.nvidia.com/en-us/ai-on-rtx/chat-with-rtx-generative-ai/

[stuck sun]

yeah literally, definitionally, it is.

[outer crescent]

thats really really silly. so anything that could and doesnt is censorship? that's not what censorship is. anyone could do anything. that they don't doesnt mean theyre censoring themselves.

[fervent condor]

If it can't generate anatomically correct human then it is censored period. The issue with faces in SDXL that needs heavy fixing via finetuning is also caused by censoring.

i don't see why you have issue with calling spade a spade

[outer crescent]

when you're saying censoring it sounds a lot like pouting to me. you wanted a thing and didnt get a thing. super mario bros was censored?

[stuck sun]

what do you think censorship is

[outer crescent]

ive just explained with examples twice. you literally know what i consider censorship and just not doing a thing.

[fervent condor]

i think no one understand what you consider censorship

[outer crescent]

you dont walk down the street censored do you?

[fervent condor]

because your definition of censorship is wrong

[stuck sun]

you just pointed to examples that arent censorship and said an example of censorship is that

[fervent condor]

When you want to curse around children and you don't do that is it censorship to your or not ?

[stuck sun]

https://www.aclu.org/documents/what-censorship

here is the ACLU definition of what is censorship

the suppression of words, images, or ideas that are “offensive,” happens whenever some people succeed in imposing their personal political or moral values on others. Censorship can be carried out by the government as well as private pressure groups.

[fervent condor]

there is also self censorship

[hallow sable]

gore and nudity is censored with SDXL. just because it effects the public image of SD

[outer crescent]

thats an intention you have. you're putting your intention on a model

[fervent condor]

either way offtopic and no need to continue debate

Stable Cascade is slightly less censored than SDXL from my brief experience

And about the same if you want to produce someone likeness

[cedar salmon]

what censored, i make boobs and nudes by mistake

the API?

[static badge]

hi how i restore comfyui to default again? loike when you delete the env folder in automatic1111?

[vast berry]

only on the discord because there are literally children in this server

i like when there are people in ai servers who are borderline nonsentient

[fervent condor]

another thing i noticed with it compared to SDXL is that it way better follows all kinds of signs

so you can type transparent saying XYZ and it will generate it

[long wigeon]

Is Image Variation a new thing with Stable Cascade?

[pale latch]

nope

[hallow sable]

if you are doing anything with words its better to just take it in to the GIMP and do the work there instead of trying to generate it in SD

[fervent condor]

Except when you try to come up with logo

vector art gives me really good outputs

[hallow sable]

I look at AI as something to storm new ideas not so much to do the work for me

[pale latch]

prompting SD to act on your masked controlnet words is best approach for words. HarrlogosXL is neat to play with too

keep in mind, as the end user, only you understand the key concepts of logo design and versatility. SD just knows how to denoise towards a trained goal.

[hollow mountain]

@fervent thunder your already here bozo

[pseudo jetty]

Are there any extension that add MP4 or Webp support to previews?

[mild ferry]

is using a mix of similar celebreties a viable way to train a face lora to get an "original" face?

[pale latch]

There's one but I forget the name. Would take each step as a frame.

[modest gorge]

Howdy all! Very early on in my SD adventure, but loving it...Q: trying to understand if there's a way to inpaint in a video?

[wheat wadi]

Where do we write prompts again, and how?

[hallow sable]

notepad ++

<:^)

[thin widget]

Yep, but it will require learning and effort. The Deforum Automatic1111 extension can achieve it using masks and ControlNet. AnimateDiff is easier but less powerful. There are other proprietary services that can do it even easier, like Runwayml. I'm sure others have more suggestions.

Does anyone know anything about why there's still no ControlNet Tile for XL? Is it difficult to produce?

[nova zodiac]

not sure tbh, but the same guy that did controlnet maintains foocus and now sd-webui-forge as well, so me thinks he might be quite busy!

[fervent thunder]

can anyone tell me if there is a tool that can allow me to upload a random face, and then that face can be used as a consistent character?

[nova zodiac]

FaceID, InstantID, PhotoMaker, Roop, ReActor....

InstantID for SDXL only, PhotoMaker probably best bets

[fervent thunder]

thanks for the response ill try them

[trim thunder]

Roop is dead i guess, you can try reactor and other options suggested above

[red hedge]

hey guys im looking for creative to help out in a project we are launching summer!

[nova zodiac]

#1092446741984444416

[hollow mountain]

bruh u forgot?

wait yeah

[hollow mountain]

how do u do the ai

[slate dust]

soft and tenderly

[nova zodiac]

first I take it out to dinner... 😛

[nova zodiac]

I mean are you asking how does one make an image?

[hallow sable]

ok xbox make AI

[sweet ore]

happy valentine's

https://wolffm.github.io/webcomicreact/

<3

[hallow sable]

valentines day sucks when your single

[noble plover]

我想要一个教室的图片，里面有中国女学生

[pale latch]

Stable Diffusion, i choo choo choose you to be my valentine! 🚂

[dim surge]

Would anybody explain what is the process to train a model with comfyUI

[pale latch]

you don't train with comfy

[nova zodiac]

you use kohya_ss or onetrainer

[dusky moon]

hi

does any one know how to generate AI images for Free in discord

[brittle bronze]

Can someone tell me how can I train lora for clothes such that it could pick up pictures on it and show same type of vlothing with picture in output

[nova zodiac]

in discord, not really anything free - you are better off going to the civitai website and using the generator there or going to aiscribbles.com

[nova zodiac]

https://www.youtube.com/watch?v=70H03cv57-o

you could also just use ipadapter instead

I believe Sebastian Kampf has a good tutorial on that

[lusty beacon]

Check this out https://boximator.github.io/

[pale latch]

stellar. 3months out. that'll be neat

[dim surge]

ty

[violet bane]

"Design a charming anime-style illustration featuring a lovable dog adventurer on a white background, set against a backdrop of serene landscapes or bustling cityscapes, capturing the excitement and wonder of exploration in every pawstep."

[nova zodiac]

#1047610792226340935

[static cape]

Dang, that looks very promising.

[lusty beacon]

ya

[long wigeon]

Is it possible to use Stable Diffusion Reimagine in Auto1111?

[steel garden]

how can i use this model https://civitai.com/models/271592/big-head-3dxl?modelVersionId=306137 with sdxl on Python?

i need a python code

so i can use my 2 gpus at the same time

[sterile torrent]

for beginners i recommended use stable matrix, u can intall there all one click, and u have there ready --xformers --vdram etc.... also python git etc

[bleak matrix]

Good morning, everyone! How are you all today?

[red hedge]

gm frens!

[trim magnet]

no pay ? then why?

[fervent thunder]

hey guys who is into dopshipping here

[red hedge]

its fun and meaninful and it could lead to a cool job

[trim magnet]

its not fun to waste time on a project just with the promise that you may get a job,thats why artists ask ppl to pay first when doing commisions

[red hedge]

well im not paying anyone

hey guys what the recommended way to use SD?i got something, i forgot the name, whats the version with the nodes?

[trim magnet]

u dont need to pay anyone,u have 2 hands to make it yourself

[red hedge]

i have this version: AUTOMATIC1111/stable-diffusion-webui

what version do you guys reccomend

i remember seeing a UI thats node based

[vale steeple]

it is not version, it is an application to run AI art stuff

Node-based UI is ComfyUI

made by ComfyAnonymous.

Most beginner-friendly UI is Illyasviel's Fooooocus, which give you some of basic settings like Guidance Scale and steps. He is also the creator of Forge, which is an optimized fork of A1111

[raven field]

bot is down forever ?

[fervent thunder]

Who wilp like to dtart dropshipping

[arctic coral]

Can't figure out why I can't reproduce an exact image that I have the PNG datas for, loaded into PNG Info, send to txt2img and selected the exact same checkpoint.

Also made sure other settings like sampler etc are matching aswell.

Even did a batch size of 10x2 images and none of the 20 were the same as the one I created yesterday

[zenith olive]

hey friends

[fervent thunder]

we fittin' to play RLCraft. we got 6 people so far but we need more since its hardcore mode. hit me up if interested (via dm)

[pulsar lintel]

What’s Dropshipping?

[onyx ether]

am i missing something where are the bot channels?

[fervent thunder]

Dropshipping is a business model where you sell products without having to keep inventory. Instead, when you receive an order from a customer, you purchase the item from a third party and have it shipped directly to the customer. It's a great way to start an ecommerce business without the hassle of managing inventory.

[tender cove]

Entirely off-topic for this server

It's a great way to sell stuff you've never seen and don't know the condition of

[red hedge]

hey guys do you know comfy ui?

any good template workflows you guys reccomend?

[iron valley]

Hey guys

Do you know if there’s any impact on performance or speed etc. if I have the checkpoints and loras in Automatic1111 directory while using comfy ui ?

[rich kestrel]

big if true

small if false

it looks way too good to be true, tbh

ya gotta remember the tech world runs on hype

[lusty beacon]

yeah hope its true

Check this out https://supir.xpixel.group/

[burnt jasper]

can SD models be used to produce wide format art--like 21:9 or more?
any suggestions

[nova zodiac]

Sdxl models can do 21:9, 1536x640

[burnt jasper]

thanks--i've heard the product of the HxW resolution has to equal an exact amount of pixels--if setting other then 512x512, is that correct?

well that 1536x640 has a much different amount tbh

[nova zodiac]

Both axis should be divisble by 64 to get the best results

[burnt jasper]

ohhh

[nova zodiac]

They do go best at a square image, but they do also train on other ratios and there is a guide for sdxl floating about

[last pagoda]

noob here but it looks like the bot is down...i am still seeing generations are these running locally and people are posting? ZDIdn't see much under announcements, please don't flame me

[nova zodiac]

#1047610792226340935 - its all good, we get this 3-4x a day so somethings not right with current comms methods

[last pagoda]

yah sorry about that

[burnt jasper]

what about if i wanted use it to produce super-ultrawide wallpapers---pushing it too far?
a 7680x2160 monitor is like 32:9 aspect ratio
i have two ultrawide monitors so it's probably almost 42:9

anything produced can always be upscaled after correct?

[nova zodiac]

Probably better off stitching two results together

[burnt jasper]

for SD 1.5 is there a limit to the pixel density?

like a number to stay under?

[nova zodiac]

Although 2048x512 should work for sdxl, def not for 1.5

[burnt jasper]

does cascade open more options though?

[nova zodiac]

I dont push it much past 768x768 unless i know the model is trained on 1024x1024

[nova zodiac]

Not sure yet, too new

[burnt jasper]

are most models suitable to use atleast 1536x640 though?

or will it cause rendering issues

[unborn nexus]

Is there a spot on the discord where i can post questions about DreamBooth?

[cerulean shoal]

Yo, is there a list of usefull prompts for pictures? And negative ones aswell?

[burnt jasper]

i've heard kohya-ss is preferrable to dreambooth for creating loras/checkpoints

[unborn nexus]

why?

[burnt jasper]

check civitai--in all the models posted there they usually give you the prompts if you click the screenshot previews

[burnt jasper]

apparently the usefulness varies by SD updates--less stable i assume

i think kohya is more specialized for it, but i haven't managed to get it working unfortunately

requires a small amount of command line tinkering

[rich kestrel]

false

[unborn nexus]

yea trying to get dream booth to work right now. Just tried doing a simple extension install but its missing tabs

[rich kestrel]

if you want info regarding dreambooth I recommend you check out the dreambooth subreddit

check out this guide I posted there as well

https://www.reddit.com/r/DreamBooth/comments/18o3dks/im_sharing_my_ultimate_dreambooth_workflow/

I've timetested this method

it's by far the best

[unborn nexus]

appreciate it @rich kestrel, would you mind if i ran a question by you?

[rich kestrel]

This only applies for dreambooth, for LoRAs you still want to use kohya

[burnt jasper]

I'm still rather intimidated by all the steps for either
I need to get my hands abit dirtier

[rich kestrel]

yeh

[unborn nexus]

yea, i figure it is like most tech software though, perseverence will get you results eventually

[burnt jasper]

what are you referring to then--what do you output with dreambooth

[unborn nexus]

I just installed the dreambooth extension onto stable diffusion but for some reason the input tab is missing from the gui. Have you encoutered this error before?

[burnt jasper]

isn't dreambooth for creating safetensor loras/checkpoints also?

or are you using the training folder thing

[rich kestrel]

yeh but dreambooth models are way different than loras

[rich kestrel]

got some sour news, but modern dreambooth extensions are pooh

if you want something half decent, you can use kohya

if you want the best possible output, you have to use an older A1111+DB commit

[burnt jasper]

does your dreambooth use a live folder for image training data

[rich kestrel]

live folder?

[burnt jasper]

like a folder you dump your images in and the tag uses them i guess

[rich kestrel]

yeh

[burnt jasper]

i have a very limited understanding how it operates

[rich kestrel]

yh thats how it works

[burnt jasper]

i am primarily interested in creating models atm

checkpoints or loras

also it seems to indicate dreambooth requires more ram/vram than i have (8gb/16gb ddr)

[unborn nexus]

i see. I guess i have a couple followup questions. 1) does this mean i will have to have a seperate dreambooth install(outside of stablediffusion altogether) or is there a way to install older versions of the plugin? 2) do you have links to where i can download or view older models?

[rich kestrel]

yeah ur gonna struggle with small vram

I would stick to loras in that scenario

[rich kestrel]

You would install a new folder with the old A1111, at least that's the simplest way I can think of

[burnt jasper]

have any of you tried running SD in a cloud server
i have been considering HF spaces or AWS for this

[rich kestrel]

I have before

[burnt jasper]

so you can use cloud GPUs/ram

which service?

[rich kestrel]

Too much pain in the ass

[burnt jasper]

seems like it yea

[rich kestrel]

I forgot the name

[burnt jasper]

but how does it perform

[rich kestrel]

runpod

it works well, but since its all linux and cloud, it's just soooo much extra work

[burnt jasper]

i have done the SDXL turbo demo on HF and it's quite fast

I doubt locally it would work well

[rich kestrel]

if you find a repo for training on it, then by all means

just dont expect it to be easy

[unborn nexus]

The way i have been doing it is running StableDiffusion in a virtual environment that i made with python. Is this what you mean?

[burnt jasper]

HF spaces has a python sand box--could you just dump a1111 files there

[rich kestrel]

all SD's come with venv

so the other one would run on its own venv

its just a matter of git pulling the right commit

[rich kestrel]

I cant say it works bc I havent tried it, but I would try it if you don't have any alternative

[red hedge]

hey guys do we know if the cpu version runs on multi threads?

for comfy ui

[unborn nexus]

@rich kestrel do you happen to have access to the dreambooth discord?

[graceful lodge]

Super newbie inquiry.

Anyone can offer tips and tricks keeping the imported image intact but have the QR code diffuses?
Meaning I want to keep the imported image as is.

[frozen helm]

tf

[iron imp]

@sudden ruin NSFW spam.

[thin widget]

So I suppose I should change the question to "Is Illya the only person who knows how to create ControlNet Tile or something?". Seems like there's demand for XL Tile, but no one taking it on.

[nova zodiac]

No idea sorry

[sudden ruin]

[frozen helm]

hammerbanned

[brave vigil]

Can I do keyword blending with more than two keywords?

[mild grove]

yep, concatenate contex node

[brave vigil]

XD I'm pretty new

Could I see an example?

[mild grove]

you use Automatic1111 or Comfy Ui

If it's automatic, I don't know how to help you, I think it only allows text input, but why do you need to repeat a command?

[brave vigil]

XD

I can't wait to get my new computer so I can run locally

I'm using colab right now, and its both buggy and expensive

Still, I havn't had this much fun in ages

[nova zodiac]

automatic is [token1:token2:ratio] in the prompt for concept blending iirc

[brave vigil]

Totally, I was just wondering if automatic had syntax for scheduling three or more prompts

[nova zodiac]

as in a queue system.. ie generate image 1, then image 2, then image 3? not that I've seen but theres probably an extension for it

[weak tusk]

Hey Guys - This is Ai music with sound engineering + Ai Art enjoy: https://www.youtube.com/watch?v=UgHmTPwkiY0

[manic grove]

With a good laptop, and stable local broadcast/other AI, how many seconds does it take to create 4 images (like on midjourney) ?

I'm increasingly considering buying a very good (expensive) laptop that could run all the big recent games, chat GPT/other local AI. But I don't know how to choose yet (1500-2500 euros in budget) because I'm still a noob on this subject.

[slate hinge]

When a model is asking for a image mask for inpainting, what format should that image be? I assumed white pixels where the inpainting should occur and transparent where nothing should happen.

I'm trying to interact with a model by api and the inpainting isn't working as I'd expect given my above assumptions. Curious if I have it wrong, thanks!

[brave vigil]

Quick check of my understanding

[A|B] toggles between prompts, first one on one step, then the other on the next, right?

As opposed to, say, trying to enforce both at once?

[still glacier]

yes https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alternating-words

[brave vigil]

If I wanted to enforce both at once then I'd just include both keywords, right?

[still glacier]

yup

[brave vigil]

My guys

I havn't had this much fun in years

Creative juices are pouring out of me

[nova zodiac]

I'm assuming that's spam

[brave vigil]

I'm doing the reading and experimentation of course

But could someone walk me through the process of establishing a lora for my own character such that I can start making all sorts of images of them?

Like, ELI5 XD

And, follow up. Lets say I sort this out. Now, I want to start building a big collection of works of this character. Would the next step be figuring out poses with control net and open pose?

[crystal edge]

hello

how to queue prompts to generate images automatically

multiple different prompts

[brave vigil]

Depends on the interface

[vapid tangle]

new model is so bad... 😦

air brushed af

[nova zodiac]

most base models are, as I'm guessing a lot of photos of people taken to try and look their best and so are either airbrushed, photoshopped, or the person is wearing makeup

and thats what gets thrown in the data sets

[hollow hound]

hello, ive been inactive for a past few months..and i foundout that bot channel to generate image is gone, something happened?

[hollow mountain]

yes

[nova zodiac]

#1047610792226340935

[nova zodiac]

try civitai's onsite generator, or aiscribbles.com, or happyaccidents.ai as good starting points

[ripe bane]

is it normal for a 600 frame (40 secs) deforum to take 6.5 hours to generate lol

gtx 1660ti and intel i59400f

[hollow mountain]

what

what happened to

stable ai

[nova zodiac]

dreamstudio is there, it isn't free; and as for the bot here #1047610792226340935

[nova zodiac]

what resolution?

[hollow mountain]

so

so what ai do i use

thats free

[nova zodiac]

the three above I suggested

[hollow mountain]

i need a picture of seoul destroyed in a post apoc world

alr

[nova zodiac]

when they said soul destroying they meant that it made someone depressed, not flattening the South Korean capital 😛

[ripe bane]

720 x 1280

[nova zodiac]

39s/it

[ripe bane]

6.5s/it lol

[nova zodiac]

390 minutes * 60 seconds/600 frames..

[ripe bane]

math checks out but idk my command prompt says around 6.5s/it and my max frames is set at 600

and of course the ETA was 6 hours

when i started

the one thing ive noted is only my gpu memory usage is maxxing out, everything else is below 50% usage

6.0 gb of gpu memory

[austere olive]

can someone help me with this ? https://www.reddit.com/r/StableDiffusion/comments/1ar454s/modelsafetensors_download_always_very_slow_to/

[nova zodiac]

In which case your going out of vram and swapping into ram - theres a setting in the nvidia control panel to stop that from happening

[nova zodiac]

Dont download through python - grab the file from huggingface or civitai and put it in the right location

[austere olive]

ok , but how cand I discover the file using python I'm using pinokio

[nova zodiac]

For a1111 you are putting the model into stable-diffusion-webui/models/StableDiffusion

If there are files there the system should try and load them first before resorting to downloading them

[austere olive]

I know, but how can I find the link ? in this case is just model.safetensors according to image I've posted

theres no information

[nova zodiac]

If not add —no-download-sd-model to the commandline argument

[nova zodiac]

Its usually downloading stable diffusion 1.5 model; i would recommend grabbing a different model though as that one is old now

[austere olive]

its automatic download I'm using comfuy through pinokio.

[nova zodiac]

In that case download a model from civitai (try ICBINP to start with) and put it in the models/checkpoints folder in your comfy install then fire it up

[austere olive]

ok but how can I discover what model is download only looking in the image I've posted ?

[nova zodiac]

You are just guessing (or heading into the code to find the url, but im 99% sure its the base stable diffusion 1.5 model)

[austere olive]

thats the point I need to know how to discover what model is being downloaded , what is the link and where to insert..

[nova zodiac]

@bleak matrix you around??

[nova zodiac]

Im telling you to grab any checkpoint from Civitai and put the safetensor file in your comfyui install in the models/checkpoints folder

@sudden ruin?

[nova zodiac]

Link here - https://civitai.com/api/download/models/121557?type=Model&format=SafeTensor&size=pruned&fp=fp16

[austere olive]

theres a lot of models there already around 13 I need to know specifically what is this model that is being download, thats the point

[opal hedge]

This is a warning to all parents: remember to show attention to your kids and not drop them on their head

[nova zodiac]

Fetal alcohol sydrome is real

[nova zodiac]

If there are 13 files in the models/checkpoints folder, what are they??

[austere olive]

other models , but not this one wich is being downloaded very slowly.

[nova zodiac]

Got a screenshot of the list?

[twilit mural]

would the only way to get characters from a specific show be to train my own model? not a mainstream show so models dont really know it well from what ive noticed

or is there like a broad general tv characters model that might be trained

[pale latch]

lol qanon nonsense

[twilit mural]

how is that acct not banned yet

[pale latch]

mlp profile pic is the secret key to discord mod hearts

[nova zodiac]

Train own lora

[pale latch]

can't use windows 11? skill issue

[nova zodiac]

Although a lot of common characters are available on civitai

[pale latch]

yeah i would just get a dataset of about 20-50 of the character and train a lora on the base model

i've been using it this whole time. when is pay to use coming?

i'll just use the 2024 version

[nova zodiac]

Dont engage with crazy!!

[pale latch]

steam dominates pc games. that wont happen

[nova zodiac]

They will drag you down to their level and beat you with experience 😛

[pale latch]

who says i'm not the one dragging ?

[twilit mural]

is it possible to have one model for multiple characters, or does it need to be only one? like if i upload a bunch of pics of various characters and make sure to name them in the description for the training will that work

[pale latch]

all you gotta do is seed a little doubt and they wind themselves up so hard. its like giving a cat some nip

[pale latch]

you can train multiple characters into one lora but it just needs more dataset organizing and larger datasets means more training time.

training a whole entire model is less needed for such things. loras can be applied to many different models too.

you've never once read any software license ever it turns out. wuhwuh

Steam also works better on linux than it does windows, so them caring about what windows is doing is like.. whaaaat? what planet are you on?

i've got 1000 game steam library uh oh. guess i'm just really skilled at steam and you apparantly are very low level

[nova zodiac]

I warned ya

[mild portal]

quick question

im having problmes with automatic1111, can i ask here ?

or is the a specific place¡

?

[nova zodiac]

Here or #🤝｜tech-support

[mild portal]

thanks

if anyone has any idea... pleace 🙂

[nova zodiac]

Much better 😛

[ripe bane]

gotcha do you know what its called or where it would be?

within the control panel i mean

[nova zodiac]

https://www.reddit.com/r/LocalLLaMA/comments/17kl8gu/psa_with_nvidia_driver_56401_its_now_possible_to/

https://nvidia.custhelp.com/app/answers/detail/a_id/5490

[ripe bane]

thanks i appreciate that. I assume the drastic slowing is from the act of swapping to ram? even tho itll still be limited at 6gb vram it will be more consistent i assume?

[nova zodiac]

always fast, but will limit resolution heavily when doing animation

[ripe bane]

interesting so then would you say to just upscale it using another AI after i get my final generation or

[nova zodiac]

You can upscale within, just need to use tiled vae

May not work with animation tho

[ripe bane]

by animation does that include deforum because thats mainly what im interested in yeah

i also dont know what tiled vae is haha lots to learn but dont want to give u 25 questions lol

[nova zodiac]

Not sure if deforum plays nice with tiled vae or not

Tiled vae is part of the multidiffusion extension - breaks the numbers down into smaller chunks that the gpu can process

[ripe bane]

do you have the multidiffusion extension link handy?

[nova zodiac]

Not to hand, but it is searchable in the extensions tab

[ripe bane]

oh yeah forgot about that

thanks

[broken forge]

where can i find ? <lora:Chocochip-1024-Prodigy-32:1

is lora

[nova zodiac]

https://civitai.com/models/100145/chocochip-artist-style
Maybe??

[drifting flame]

is it possible to have GPU/CPU/RAM usage floating over all windows? kinda annoying to check task manager everytime

[warm junco]

Yes with Rainmeter

Its a tool for floating windows, after the install it already adds CPU, ram etc usage

Very cool Programm with a lot of overlay extensions and themes

[drifting flame]

rainmeter kinda complicated

doesnt mean I cant do it but

something like Glasswire maybe? it can show net usage panel

[warm junco]

Idk if there is something for that like glasswire

[molten tide]

Hello there! I've a question, what the best way to be able to use a really specific bottle into image generation?
Should i train a checkpoint, a lora, both or something else?

Thanks

[nova zodiac]

lora

[molten tide]

So Just by training a lora i'll be able to put in every images i want Always the same and specific bottle with no problems?

[nova zodiac]

well not no problems, but is best bet at maintaining shape of bottle.. ipadapter is another option with a high control strength

[molten tide]

Mmm, interesting. What Is ipadapter? Any tutorial or guide?

[lunar breach]

Hi there, im very new to this. Can someone here explain to me how i would get a person to look away from the camera when generating images while still making the character consistant? (using fooocus)

[nova zodiac]

Sebastian Kampf has a tutorial on it on youtube

[molten tide]

Ah ok thanks i'll check It out and i Hope It Will work

[nova zodiac]

Add looking away from camera to positive prompt, and looking at viewer in negative prompt usually is enough

[molten tide]

Any other suggestion Is welcome on how to photo insert a specific object and play with the scenario where around It
Thanks

[lunar breach]

still looks at the camera - could the reason be that im using the "face-swap" feature of fooocus in image-promt where that image has the character looking at the camera?

If yes, then how would i generate an image with the character looking away while keep the character consistant with previous images?

[nova zodiac]

Yeah thatd need to be off, no idea on how to do it in fooocus sorry

[lunar breach]

which other front end would you do it in? The first i tried was comfyUI but it was a little overwhelming for me as a beginner so i swapped to Fooocus

[nova zodiac]

Automatic1111

A middleground between the two, and theres lots of tutorials on yt for it

[lunar breach]

Alright ill install Automatic1111 - once installed - could you guide me to generating an image with the character i already have but looking away from the camera?

or link me a video or article?

[nova zodiac]

Again, Sebastian Kampf with the ipadapter tutorial is best bet, but youll probably want to lower the controlnet strength and set it to “the prompt is more important”

[warm junco]

For an install guide of Automatic1111 checkout the Pinned Messages of #🤝｜tech-support

[muted skiff]

Hello. Can I use SDXL with 6GB VRAM? Or it's not worth the effort?

[frank rampart]

ive noticed that some models are more "expensive" than others. like most models are 2-3GB but i got revAnimated recently, gorgeos model. 5GB. takes forever to switch to it and sometimes renders just will crash, I have to use smaller word counts and smaller resolutions for it. is that normal?

[untold flax]

hey guys - I'm new to this stable diffusion and wanted to ask

when you use a picture of yourself, how do you instruct the prompt to 're-imagine' the picture you provide of yourself

[static cape]

That depends on which of the trazillion different user interfaces you are using. Not everyone uses the same terms or even same techniques / Models for that.

[frosty swallow]

I want to get good at sd

how I do this

[untold flax]

i've just installed easy diffusion and trying to put my face onto a character or body

what is the easiest way to learn how to do this

[frosty swallow]

I am being silly in this question btw

Still learning but it seems like so much. I really want to learn control net well

to get as specific outcomes as possible

I get my new gpu today

[untold flax]

thanks fawks for replying

[untold flax]

what GPU are you using for this?

[frosty swallow]

rtx4060 ti

16 gb

[arctic spear]

I am a complete amateur when it comes to using stable diffusion. I am in the process of buying a new computer, and the options are a MacBook Pro with m3 or a PC with an NVIDIA RTX 4070 Ti Super 16 GB WRAM. Is one clearly better than the other?

[trim magnet]

the rtx 4070 ti super is better