#💬｜general-chat

1 messages · Page 171 of 1

vernal ore
#

If anyone's interested
Please DM

#

Domain specific video generation is still video generation

fervent thunder
#

nice gan looks good

still hare
#

Hey i think this may be a dumb question. I joined this discord a few months ago and I forgot the website and or where to download this. I meant to try stable diffusion but I haven't gotten around to it till now lol any help pls

hoary gull
still hare
#

Gracias

hoary gull
#

Yeah 3.5 medium does much better 2MP images than 3.5 large, but above that it starts to break again :/

brave jetty
#

HI

#

I'd like to know if someone already created the GGUF Q8 version of SD 3.5 medium (maybe turbo/4steps) also?

#

I tried, but failed

tranquil lotus
#

I am using SD 3.5 medium on comfy UI on a M2 Mac Pro 18 GB. It's generating good images - but want to understand how to up my game :

  • How to make sure the model follows my prompt to the t
  • What settings do i use ? steps, cfg, sampler, scheduler, denoise ?
chrome herald
fervent thunder
#

could also unload and reload encoders

kind thorn
#

Does 3.5 have the same kind of licence as XL?

proper girder
#

can you run 3,5 on automatic UI?

brave jetty
dusty trellis
#

Does 3.5 Medium require 9.9GB VRAM or is that just a recommendation? Will it work on 6GB?

low moon
#

It's a friendly suggestion.

#

XD

#

If your friends with your GPU he'll do it for less.

rare holly
#

where can i get help with stable ai iam getting grey images ?

unborn hedge
#

so is 3.5 medium gonna be like the new SDXL?

low moon
#

could be if the community embraces it..

desert dagger
spark tinsel
#

I suppose by toaster you mean 3060 12GB? lol

quartz siren
lost pond
#

Haven't really been in the sd scene for a while, but me and my friend wanted to try to train off her artstyle. we got abt 15-20 images for a dataset. would it be better to train a model or a lora/ti? there aren't any specific characters, but lots of stylistic choices such as color, exaggerated limbs, etc that we want to capture, and i've only ever created loras for character styles like ocs

still sandal
#

hey guys, im using SD 1.5 and i want to be able to take an image i made, and keep creating that specific character in different ways, on a beach, etc and im wondering how i do that

Is their a really simple way to take an image you made, and keep pushing out new pictures in different scenarios

mystic siren
#

I just tried to do this days ago and it's kinda hard since it depends how much time would you invest into that

still sandal
#

yeah, like i mean im kinda willing to put some time in but i dont have time to wrap my life around it, but id love to sorta "build a character" for my avatar

#

i have a great image , just need to figure out how to get SD to read the image and keep making similar images

stuck sun
#

Hey if anyone has had the same issue as me with SD where loras/adetailer sometimes takes male faces and feminizes them and vice versa when you are making images with more than one person, i comissioned Anzhc to train a yolo8 model that detects only male or only female faces

mystic siren
#

The short answer is to create a lora with that character, I saw a complete video on how to do it, I save those videos because you never know when you need it, I will pass the link in a min

stuck sun
#

be useful to get feedback from other ppl, works well for me but until today I was the only one testing it

mystic siren
still sandal
#

because im on AMD

desert dagger
vocal lintel
#

been out of the game, what's the current best anime model? flux? pony? MJ/niji? is 1.5 obsolete?

low moon
#

1.5 never goes obsolete. It ages like fine wine, the older it is the better

#

probably the best for anime also in spite of technically better models that came out since

vocal lintel
low moon
#

u shld!

vocal lintel
#

do people mostly use comfy over auto1111 now? did auto stop getting supported?

low moon
#

comfy is the best indeed. A1111 is kinda of unsupported... If you want to use A1111 you try Forge which is a better fork of A1111

vocal lintel
#

ok awesome thanks for all the help!

void fossil
#

download sd3.5_medium_incl_clips_t5xxlfp8scaled right now anyone know if its supported on Forge ui yet? or just comfy ui?

dusty trellis
low moon
#

u mean comfy or forge?

#

im not using comfy.exe if thats what u ask

void fossil
#

im using forge and loading app right now

#

nope not working for me maybe cuz i got amd gpu

ornate remnant
still sandal
#

i still use A1111 because of the automation, a few batch command lines and my AMD card works beautifully, but its definitely not got flow and stuff and is just the standard ui

#

also, alot of safetensors i download are SD 1.5

desert dagger
fervent thunder
#

it used to be the only model that had IC-light also, however IC-light for Flux is coming out soon

#

its possible Hunyuan-DiT's new control nets are stronger but I have not verified that claim yet

desert dagger
fervent thunder
#

its the last of its kind
a model with training data mostly scraped and not curated

#

makes it sometimes the most interesting

untold dagger
#

Hey guys, anybody know a workflow to do svd image to video along with a text prompt of how you want the image video to actually be animated?

last bay
#

what is the difference between sd 3 and this new one

#

is it worth downloading it

fervent thunder
#

the medium or the large?

last bay
fervent thunder
last bay
fervent thunder
#

don't click the link it is a scam and likely malware
Stability AI do not have support tickets

sleek pilot
#

hi

gusty mist
#

anyone here got an easy workflow for sd3.5? Mine is somehow not working even tho i used it with sd3.5 l

keen isle
#

#free-generations

rare holly
#

If iam using python webui what models work with that Web ui ... very new to this .. any help is appreciated

#

Is it pony model ?

#

Thanks sorry if I made a mistake

#

@fervent thunder where was the link p

warm junco
#

Crypto scam...

fervent thunder
#

So common, means its profitable, need to get into it

silk condor
#

How much ram is auto 1111 supposed to use? Like RAM and not GPU Ram

uneven yarrow
#

hello

sonic wave
#

do you guys find inpaint tool useful to correct images that are faulty in some parts but otherwise good and cool elsewhere?

rotund girder
#

anyone know of any recent benchmarks for stable diffusion that include enterprise gpus? curious how much faster I could generate with a cloud service than on my 3090

still sandal
fallow veldt
#

I corrected and use several methods like instant-id for the resemblance, or to get the beverage he is holding, or to general correct things

sonic wave
#

interesting, nice to know that it has some merits, I don't feel like I'm wasting time anymore

#

instead I feel like I lack skills necessary to do what I want with these tools but that's a matter of time

fallow veldt
fallow veldt
elfin stump
#

Has anyone else noticed that the latest models have a bias toward making noses, cheeks, and ears pink (or reddened, or flushed)?

mint geyser
#

Hello, where can I find proficient stable diffuser hobbyists in the community, who could help me with realizing a promo video for a project?

solemn turtle
#

Hi

desert dagger
desert dagger
marsh trellis
#

Hi

elfin stump
desert dagger
fervent thunder
#

if this means Flux then yeah

#

that's what population average preferences tend towards though

desert dagger
#

it's probably because there is way too much anime in the data training set, at least if you're talking about flux and flux loras

fervent thunder
#

not sure, flushed look might match what people look for on average

#

population average tastes are not good LOL

mint geyser
elfin stump
desert dagger
bleak matrix
#

Good morning, everyone! ☀️ How are we all doing?

fervent thunder
#

hello ☀️

bleak matrix
#

How are you doing?

fervent thunder
#

not bad

kindred pendant
#

☀️ pretty good

fervent thunder
#

Hello

faint monolith
#

Hello, can anyone point me to the tutorial on installing stable diffusion on a machine with ryzen 5 4550g ?

lapis urchin
next sundial
#

ciao a tutti

radiant creek
#

bonjour, puis-je avoir une assistance en Français pour la génération d'images?

fallow veldt
#

It is different to make img2img, and then upscale, than to do img2img and upscale in the same go, in latent, so only one vae decode?

#

I see many workflows that have everything integrated, like generation, refiment, upscale, but I don't like working like that, I prefer to do those steps when I think the image is good for example

jade wind
#

Hey guys, just a little question: I have a little study exercise and for that I need some text-to-image pictures with the topic "big city and jungle". I have to do a little collage with animation. Mood isn't important, but I thought about some kind of "I am legend" type images. The images has to be big for the animations. Would be great, if someone could help me!

wary sparrow
#

G

worthy delta
#

Anyone have Workflow SD3.5+Reactor?

upper plinth
inner latch
#

Hi guys. I am a cameraman and editor. I want to learn how to work with SD. I have a Mac Studio M2 Max with 96 GB of RAM at work. I’m trying to find the best setups for generating with ComfyUI on it. Guys, please share your thoughts/experience.

desert dagger
heady lake
#

Is there any plan for fp16 editions of 3.5 models? fp16 performance is 8x on my hardware

desert dagger
fossil needle
#

Hi guys, I want to know how to sign in the SD in my private server after I add the bot to my private server. Only Sign up here.

mighty bane
#

How do I easily check what trigger words the Lora I am using has?

#

I am using ComfyUI

desert dagger
weary socket
#

Boo! 👻

cedar salmon
low moon
white parrot
#

can i have some quest tion

sacred basalt
white parrot
#

i think it's a bit greedy but...

#

i have some fun time with sd 1.5 be 4 but never try this

#

idk iff it wwork and how?

#

ic ty

quartz siren
white parrot
#

i think my lap good to use meta stuffs, can you recommend me?

small haven
#

Hello Everyone,

I'm happy to be here and hopefully I'll get my questions answered.
I'm using Stability AI API for my image generation service called Francis AI. I'd like to make a subscription / buy credits as a company but I can't find any form / input that makes it possible, nor in the stripe checkout nor on the UI of the Stability AI client.

If anyone knows or faced this issue before, I'd really appreciate some further information 🙇‍♂️

Thank you guys in advance and sorry for disturbing peace with my silly question

past wadi
steel fable
#

Why is it so difficult to get the results you want from some Stable Diffusion Lora?

upper plinth
#

what sd version are you using

steel fable
#

Im useing pony version

upper plinth
#

I use pony too and the loras work for the most partr

#

if ur using clip skip 2, try using 1

sweet token
#

I don't like to use pony that much for loras, I much much prefer Illustrious XL, it's like pony but better and responds to Lora training way better then pony

upper plinth
#

and is that model good for anime

sweet token
steel fable
#

@upper plinth Mister can i send u one Lora for u to try to use please? This one idk how to make work

upper plinth
#

u can send me the link if u got it from civitai too

sweet token
quartz siren
heady lake
desert dagger
brave jetty
desert dagger
quartz siren
#

Going to be pretty slow as well, probably 5-10mins on a 8gb vram gpu.

brave jetty
heady lake
desert dagger
quartz siren
# brave jetty Is there a specific model that will fit my config?

There are 2: 5b and 2b.

With quantization, 5b should fit too and be better quality then 2b. 5b also should have more support too.

I would recommend the “fun” models since they support different res video generations, and different length video generation too(less then 6 sec)

white parrot
desert dagger
white parrot
#

wow

fervent thunder
#

can someone help me

#

how do i make sfx using stable audio

warm hull
#

Guys , where can i find a lora-based image styling comfy ui template . I want the image to be generated based on the Lora that i select. I could only find a template for using multiple loras.

desert dagger
warm hull
desert dagger
warm hull
fervent thunder
#

they do

desert dagger
#

they very much do - DM'd you the invite link

warm hull
#

Thank you!

desert dagger
#

welcome

floral umbra
#

Has someone made a "stable diffusion" version with "Stroke-Based Rendering models"? where the A.I with descriptive text is trained on concepts of mountains, rivers, people, and other descriptors, but instead of hallucinating images made with images themselves that "isn't yours" generates a random sensical painting by applying actual public knowledge of painting techniques, brush styles, ways to use the brush to create a truly unique painting?

desert dagger
#

if you've got a couple million dollars, and a datacenter, you could train a model like that

floral umbra
#

Of course i don't lol.

desert dagger
#

that's unfortunate. seems like a good project

floral umbra
#

Also, who's the mods here? Or is fruit here the only "mod"?

#

@karmic brook Where do i go/who do i dm to report a scam bot? As a account in this server spams phishing links to a discord pretending to be customer service.

Or do i just ping you here with their ID?

karmic brook
high orbit
#

What is better (for pony): 1024 resolution without highres fix OR 512px with Highres Fix?

desert dagger
fallow veldt
#

in 512 you are loosing quality and highresfix won't matter much

fervent thunder
#

you can get SDXL/Pony to work smaller resolutions but you need two things

  1. reduce Unet temperature with this https://github.com/Extraltodeus/Stable-Diffusion-temperature-settings
  2. Res-adapter, which was trained on SDXL to give it 256x256 to 1536x1536 range https://github.com/bytedance/res-adapter
fallow veldt
#

stable-diffusion-xl-1024-v0-9 supports generating images at the following dimensions:

1024 x 1024

1152 x 896

896 x 1152

1216 x 832

832 x 1216

1344 x 768

768 x 1344

1536 x 640

640 x 1536

For completeness’s sake, these are the resolutions supported by clipdrop.co:

768 x 1344: Vertical (9:16)

915 x 1144: Portrait (4:5)

1024 x 1024: square 1:1

1182 x 886: Photo (4:3)

1254 x 836: Landscape (3:2)

1365 x 768: Widescreen (16:9)

1564 x 670: Cinematic (21:9)

#

Pony is SD XL so

fervent thunder
#

doesn't have to be a specific number any more either

fallow veldt
fervent thunder
#

yeah its called a domain-consistent adapter

#

it was trained by TikTok

#

funnily enough

fallow veldt
#

but is something you have to install and takes resources?

fervent thunder
#

its tiny its like the size of a lora

#

its in comfy manager, the nodes handle it all

#

you don't need any other download

fallow veldt
#

It seems to prevent double generation on bigger resolutions

#

But I don't see the point really, better to generate with native resolution and then if you want resize and then upscale

fervent thunder
#

doesn't it seems easier being able to pick any number though

#

I agree its not a huge deal

dawn mulch
#

Does anyone have ComfyUI-3D-Pack And can you tell me the size of it because I am currently downloading and I have a 1 MB Internet speed and so far I am over 400MB and is still downloading. Downloading a two gigabyte file takes me about a day. So wanting to know the size is important . but The real issue is not the size since it will complete eventually but rather the issue is that I am using a script to download from Github which is supposed to auto resume any download when I get a connection issue and I do get connection issues quite often with large files so I don't know if the repository is being downloaded properly whenever it gets a connection interruption, or If it's not downloading properly I could just be downloading data forever.

desert dagger
brave matrix
#

is flux 1.1 available to train a lora with yet?

desert dagger
umbral pelican
#

hello everyone

#

may i know what exacly is this server for

#

i also saw a few images
did you people make them?

unborn hedge
#

i think its dev only

#

and tbh unless its a anime style or something, just stick to SDXL, i been having massive trouble with flux training for my specific character

desert dagger
azure crescent
#

Hello 👋 I'd like some help with setting up or joining a log in for this app, the thing I use is AUTOMATIC 111 and I don't know what the url or login and password are

#

Please ping me if you have a solution. Thank yiu

old lion
azure crescent
#

Could I see where to look?

old lion
#

when u run webui.bat

azure crescent
#

I got in

#

But got a error message when I went to test it

old lion
#

do you have a model?

azure crescent
#

Not sure whay thay is

#

Oh says default

old lion
#

need to download a model and place it in models/stable diffusion

azure crescent
#

Could I have help knowing what to download?

old lion
#

also in case i stop responding or u need further help there is a dedicated support channel #🤝｜tech-support

azure crescent
#

Okay might need to get on computer my last question is it available on computer

old lion
#

yes you need a computer if you want to do it local.

round trout
low moon
old lion
oblique elk
low moon
dawn mulch
# desert dagger did you do a git clone?

yes the usual way for any costum node. im at 1.8GB now. not sure if the download is screwed. i am able to download other custom nodes fine but they are small. i dont know if ComfyUI-3D-Pack is just silly large or my script is screewing up. if anyone can download it to confirm its size please help me confirm it. normaly it takes less then 5 minute for normal people to download 2GB but my internet speed is 1.mb

fervent thunder
#

took around 0.05 seconds on 3GB internet

#

your file is too big

#

something up with the script yes

brave jetty
#

Hi!

#

I'm trying to figure out why I don't have Python Embedded folder in my ComfyUI installation and how to install it
I want to use FaceID, InsightFace, etc.

midnight halo
#

hello

muted skiff
#

anyone know what Emad is up to nowadays?

merry atlas
#

hello

dawn mulch
unborn hedge
gray vapor
#

Im trying to deeply understand Stable Diffusion architecture, but i need somebody to help me understand something.

#

I understand that the VAE codifies the input image into a kind of "smaller" image, with less pixels or spatial regions, but with more channels that encode the formal patterns and therefore allow the decoder to recreate a full size image.

#

My doubt is in Self Attention and Cross Attention (with the prompt)

#

How is Self Attention performed? Is each "Pixel/Region" of the latent image treated as a token? if thats the case, How are the Key, Query and Value vectors generated for this region? The initial emedding for a single region should be too small, only composed of 4 values for the 4 channels.

echo kite
#

any general eta for sd 3.5 support in automatic1111?

ripe rune
#

anyone in this chat can help me out troubleshoot why after installing rgthree comfyui node still won't work?

peak tendon
#

Dods anybody know of a good lora for design sheets for pony? Like, I wish to get reference sheets of a character, it would be awesome to have it divided in pieces because I want it to model it, but I'm okay with different views of a character t-posing

I try searching civitai but even with the filters on there's so much porn and unrelated stuff that damn

And I don't know where else to search

spare kayak
#

Hello! I was curious whether or not Stable Diffusion 3 has been released and if/where we download it and where to put it in the folders

#

😄

coral orchid
#

I finetuned a Lora, but when using it made very small difference althought the weight is 1. What can be the reason?

hasty gazelle
#

Can someone point me in the right direction? I'm planning to train a model using images and a list of tags associated with each image. So far I've only been creating images and listing tags for each one, I haven't really looked ahead to see what to do with that information.

heady perch
#

How do I generate images?

quartz siren
# spare kayak Hello! I was curious whether or not Stable Diffusion 3 has been released and if/...

Yes sd3 medium released and it was very very bad at humans. Now there’s sd3.5 large and sd3.5 medium.

Sd3.5 medium isn’t that great either imo, but ok. It’s basically base sdxl with better prompt following and better text rendering but worse aesthetic and similar humans, but honestly I prefer finetuned sdxl models.

Sd3.5 large is great, it had much better prompt following, text rendering, and humans(not near flux level but workable, better then sdxl)

It’s a solid alternative to flux, it’s much more creative and knows more art styles, less biases, but worse humans, text rendering, prompt following.

spare kayak
#

is Flux Automatic 1111?

fervent thunder
#

I think FP8, NF4 and GGUF killed the need for models smaller than 8B or so

quartz siren
#

and yeah neonninjaastro is pretty right, with nf4 you can fit 12b model(flux) in 8gb vram and sd3.5 large(8b) in 6gb vram. You can fit it in even lower vram gpus if you can wait long enough.

spare kayak
#

I thought Stable Diffusion was Automatic 1111?

quartz siren
# spare kayak I thought Stable Diffusion was Automatic 1111?

Automatic111 is just an UI to load and use models and so is comfyui, swarmui, forge. Stable diffusion itself is a series of models by stability ai(sd1.5, sd2, sd3, sd3.5).

Flux is a series of models by Black forest labs(has many previous stability devs) which currently include(Flux.1 schnell which is 4step like sdxl turbo but for flux, and Flux.1 dev which is normal step range like 25). Both are 12b params in size and similar to sd3/sd3.5 architecture and open source.

spare kayak
#

Ohhh thank you for explaining that! So how do I install SB3.5? And download it?

#

I really appreciate it

quartz siren
native heart
#

why i cant generate brown person on flux

#

like i put "latino brown man"

#

and i get white dude

desert dagger
# native heart and i get white dude

well - is EVERY latino guy a dark shade like he's been in the sun too long? sometimes you have to use terms that the AI understands whether you really want to use those terms or not. geotarget the guy. use the name of the city he's from or something. and use a common male name from that area. like, maybe: Juan standing with his arms crossed, leaning against a tree

peak tendon
cedar salmon
#

maybe dont use latino

upper plinth
#

it is particularly hard to get non-black dark skin girls in 1.5

#

the models are biased towards white

desert dagger
delicate bear
#

i have two gpu's one with vram 12 gb on with 16 gb. is there any possibility to run stable diffusion video using these two. its would be a great help. i am new learner .

dusty trellis
#

I am trying to create images like the following, but they always come out blurry. I don't have any problems with other images with all the same settings. I also sometimes get blurriness with photos of white sand beaches. Is it the brightness? What is it about this prompt? "flat white background, 3D Unreal Engine render of an old green wooden chair"

silk condor
#

Does anyone know any other option compared to Dynamic Prompts? I had one installed but kept getting a lot of errors and it was aiming to the dynamic prompts and other people had the same issues so I removed it. But I like the way I could use it and wonder if there is other options?

silk condor
dusty trellis
silk condor
dusty trellis
silk condor
dusty trellis
silk condor
#

So I've done many restarts 🙂

dusty trellis
#

Sorry, above my pay grade, wish I could help. Wildcards are amazing. It will be worth it to find a fix

silk condor
boreal galleon
#

Was anyone already able to train SD 3.5 on KohyaSS ?
He released a new repo named 3-3.5-FLUX ?
Anyone ?

fervent thunder
#

in the simple tuner discord (its called terminus research group) they have made some progress fine tuning SD 3.5 already

#

I'm not really a big fan of using pre-made scripts though

normal pagoda
#

is it easy to make a transition from SD 1.5 to SDXL (including Pony)?

fervent thunder
#

almost everything about a workflow changes

normal pagoda
fervent thunder
#

oh that might be fine

#

A1111 has barely any features aside from just running the model straight as it is out of the box so it's probably about the same

normal pagoda
#

good - I plan on doing said transition
will search for all the LoRA's I'm missing on XL when I can

fervent thunder
#

you can often keep the same sampler, schedule and CFG settings, which is nice

hard flax
floral umbra
#

Is there a way to make python be multithreaded when loading up comfyui?

#

@still glacier Don't know if you have mod/admin powers. But we got a scam bot/compromised account

jaunty jackal
#

Sjjsksaksjsks

obtuse sapphire
gusty ore
#

Long shot here, but anyone have docker-compose for AUTOMATIC1111?

mystic siren
#

anyone here that uses controlnet can tell me if using the colored option of depth anything v2 is better than without it? because I'm using it on comfy but I don't see that it's using the colors of the depth map

hasty gazelle
#

Is there no mod role or admin role to @mention? No staff role?

desert dagger
hasty gazelle
#

That's some ridiculously basic server functionality that is missing, to be honest.

Anyway, use huskyeth_alp is DMing people with links to another discord server under the guise of it being a problem support server.

copper crystal
#

not everybody needs a direct communication line to (often volunteer) mods on a popular server. spamming people in DM is against discord's TOS, so report works there too

hasty gazelle
#

It rarely works fast enough to keep them from DMing a ton of people, when they could at least be banned immediately from the server. What would mods be volunteering for if nobody can get a hold of them?

At the very least, a server this big should have some kind of modmail/reporting channel given the amount of spammers and scammers that hover around accessibly tech communities.

prisma agate
#

does 3.5 work with a1111?

copper crystal
hasty gazelle
#

No, theyre here to moderate and keep a community safe. The user that DM'd me has only been in this server for a couple days, has already been posting links to a random server IN the server, and was never banned?

copper crystal
#

with mod mail, people tend to "editorialize" the reports. Right click report makes things more objective.

hasty gazelle
copper crystal
cedar salmon
#

you your all sheep to be fished

hasty gazelle
#

Looks to me like their message was a link, before it was edited to say what it was now.

cedar salmon
#

i can confirm that, i told the guy to not go there

hasty gazelle
#

^
But the person was never banned, they just started DMing people instead.

cedar salmon
#

not the only oe ive seen also, not sure if it was the same guy

copper crystal
cedar salmon
#

ya this discord is pretty much abandonware

copper crystal
cedar salmon
copper crystal
#

you don't need to write a big essay

copper crystal
hasty gazelle
#

Anyway, back to what I'm actually here for. I'm planning to train a model using images and a list of tags associated with each image. So far I've only been creating images and listing tags for each one, I haven't really looked ahead to see what to do with that information. Can someone point me in the right direction?

AI seems to move so fast that I dont really know how to tell if a tutorial from a year ago is even still relevant

desert dagger
desert dagger
copper crystal
desert dagger
#

no, they don't. those that were are now community guides

copper crystal
#

you misunderstood that change if you think that means this server is now unmoderated and lawless

desert dagger
copper crystal
#

weird

broken field
#

Anyone here use comfyui or know a good related discord server?

placid hatch
#

is foolhardy remacri still a good upscaler to use with sdxl?

cedar salmon
desert dagger
broken field
marble meadow
#

Hello

gusty otter
broken field
#

I do not.. let me try to find him.

#

thanks

astral pebble
#

upscale to 4k, and don't change the face make it real and same and upscale it HD image

can we change this for better results. now the results are very poor.

pallid yarrow
#

anyone has a better extension then Openoutpainting?

wispy cloak
#

why Monetico so popular on huggingface?

copper crystal
#

it's an interesting demonstration of efficient model training. peopl take notice.

little gorge
latent timber
bold aspen
#

hello

placid hatch
fervent thunder
#

how do I prompt for "waving hair" without making the person wave lol

fervent thunder
desert dagger
empty patrol
desert dagger
#

we have nothing to do with runway

delicate meteor
#

👀

marble meadow
#

a1111 or comfyui for a complete novice whose only experienced with dall e 3?

desert dagger
#

actually, install SWarmUI - let it do all the technical stuff, and run comfy inside it more information in #🐝｜swarm-ui the swarmUI cahnnel

marble meadow
#

is SWarmUI the most supported webui

desert dagger
#

it's designed to make your life a whole lot easier. just run comfyUI inside it. scroll down the page and you'll find install instructions

grand rain
#

is there a comfy node to adjust chroma?

#

as in hue

stuck sun
grand rain
quasi stump
#

Hello, is it adviced to host Stable diffusion on a Google Cloud server?

still glacier
#

adviced ? not really, it s harder than doing a local instance and google fight back against it as it costs them money. But it can be done

quasi stump
#

Do I care to download stable diffusion v1.4 model/checkpoint? or I should ignore that part and only look for LoRa and Checkpoints from Citiv.ai?

still glacier
still glacier
#

get some sd1.5, sdxl or sd3.5 (probably too soon, to run those on collab. I haven't tried) based models from civitai are those are the most used architectures. Therefore that s what tend to get the more support

quasi stump
#

Where do I find 1.5?

#

Ah so most civitAI LoRas are 1.5?

still glacier
still glacier
quasi stump
#

That's if I want the barebone/raw model I guess but if I'm looking for specifics I believe a Lora+checkpoints on civitai are enough

hard swift
#

gm all

abstract quarry
#

SD 1.5 is very old, though

#

SDXL is much newer and better and has as much support as the old sd 1.5

#

SD 3.5 and Flux are the two newest and best models, but there are also less tools for them

fervent thunder
#

control nets, brushnets and powerpaint, and IC-light are the main things I still like SD 1.5 for

#

apparently a big new HunyuanDiT may be coming

#

didn't realise they are backed by Tencent also

warm junco
#

For cloud there are multiple services. Websites directly for generation or cloud GPU hoster like vast.ai where you can rent a GPU and setup Stable diffusion yourself

gentle marten
#

Anybody know where I can learn more about all this? Especially the technicalities of it all

misty anchor
#

Hello, whatsupp guys

gentle marten
#

Hello

#

Trying to learn what is best

misty anchor
#

I see, welp I'm here to ask someone about webui_forge

gentle marten
#

Can't be me, hope you get help 😄

misty anchor
#

Thanks, you are new here also?

desert dagger
desert dagger
misty anchor
#

oh, I see got it, thx

gentle marten
gentle marten
desert dagger
gentle marten
#

Well, rather... I installed the SD a few years back with the pytorch and bash thing, but I've fallen off of it. However, I've got time on my side and a decent set up to work with 😄

desert dagger
gentle marten
desert dagger
gentle marten
desert dagger
gentle marten
desert dagger
quasi mauve
#

@desert dagger When you have a moment could I trouble you for one of those as well?

desert dagger
quasi mauve
#

Thanks!

desert dagger
#

👍

radiant tusk
#

Anyone got a good tutorial for animatediff?

desert dagger
desert dagger
#

okay, insults get reported to discord

cedar salmon
#

reported for suggesting having virginity is insulting

upper plinth
#

💀

desert dagger
desert dagger
trim thunder
#

Can someone help me with installing stable diffusion in windows11?

warm junco
trim thunder
#

Can someone suggest some good checkpoints ?

turbid spire
#

Hey 👋

#

I’m new here

bleak mortar
#

guys how can i make an image here?

#

on which bot??

#

arent they free

winter canopy
#

Hello - I’m interested in developing rendering techniques with SD. Can anyone recommend any pre-processors for controlnets (comfyUI) which can interpret depth and normal maps (exported from a rendering program).

I’m looking to improve image interpretation and rendering accuracy - while using SD to play with style etc.

kindred radish
#

Who is using premium gpt 4 dm now

rare holly
#

Can you make gifs in forge ui ?

half lichen
#

hello

#

newbie here

trim thunder
#

Can we now get the text in the image correctly?

desert dagger
hot hound
#

hey guys

#

i just started using stable diffusion

#

so just wanted to ask, is it normal to get bad images

#

on your first run with around 20 steps?

#

im using the SDXL model and yet getting really bad images

#

even tried using better prompts

#

is there something im doing wrong?

trim nymph
#

check your setup image size

#

also worth checking cfg amount

#

do you use auto111?

hot hound
#

yeah

hot hound
#

i have rtx 3050 6gbvram not sure if thats whats causing this issue

trim nymph
#

i think you need higher sizes for sd xl

#

personally i barely played with xl

hot hound
#

hmm

trim nymph
#

try using euler A as sampler

#

30-40 steps

hot hound
#

ok ill try that

trim nymph
#

cfg around 5-8

#

if possible, try larger image size

hot hound
#

whats that

#

oh

#

right

#

yeah i keep it 7

#

usually

quartz siren
hot hound
#

oh

#

ill try that

quartz siren
# hot hound oh

Also I would either recommend some sdxl finetune(juggernaut, dreamshaper, realvisxl), maybe the lightning versions(need 4 steps so 6x faster).

You can also try the new sd3.5 medium, the above finetunes are right now definitely better in image quality but sd3.5 should have better prompt following and text rendering, worse image quality tho.

hot hound
low moon
fervent thunder
#

Здарова

trim nymph
#

you have a point, might be time to put some thought into my avatar

fiery wasp
#

have anyone tried controlnet-union-sdxl? is there any difference between this and separate controlnet models in terms of quality (same, worse or better)? i guess they should be the same, but... if someone tried both, please share your thoughts

warm junco
#

Sdxl is trained on 1024x1024

austere relic
#

3.5 is still far away from red panda, flux pro or Midjourney, hope v4 is going to bridge the gap

desert dagger
oblique elk
#

Could you explain "expands". Are talking about outpainting?

fiery wasp
#

can you post a 3-4 screenshots for better understanding, or share the link to video

#

that is tiktok ai filter

#

you want somthing similar with sd?

#

will this help?

#

or this

desert dagger
#

that is video leap and that's the app. and that's what is used for those infinite zoom in and out

#

ah. the images on that clearly say capcut

#

they have a lot of video tools

#

go look at the top right corner of the images in that video

#

you really know how to tell people you don't want help, don't you

junior merlin
#

hello

steel notch
#

Hello

coarse dome
#

yo guys whats up.......!!!!! good morning

#

i bet you all are sceared of my chat........????

desert dagger
broken field
#

I need to generate an image of a gryphon standing on a supply crate, and it needs to look like line art/drawing.. anyone recommend a model and/or LORA thats good for that.

inner frost
#

hello

coarse dome
desert dagger
coarse dome
#

3 pm here

#

where are you now

desert dagger
coarse dome
#

wow thats great

#

what about weather

#

hot or cold

desert dagger
coarse dome
#

same here

stark pelican
#

Greetings, everyone.

#

Recently I acquired a computer with the necessary specs to run stable Diffusion and I would like to run it locally. I used it a lot back in the day when you could run it using Google Colab for free.

Now, I would like to go back to using it locally. Problem is, I'm an absolute noob at doing that.

Does anyone know a simple guide I could use to set things up?

ionic wraith
hot slate
#

Can i get advice on which model to train to produce the most life like images of my self for my linkedin?

stark pelican
kindred mango
#

Hello! I am trying to add "external captions" to images to train the using fast-stable-diffusion collad. Where can find more info on annotating correctly?

#

I think i found it

austere wolf
#

"Hey everyone! 👋 I’m looking for a fine-tuned Stable Diffusion model that can generate web user interfaces (UI) based on prompts. Does anyone know of any good models specifically trained for UI or web design, or have tips on how to set this up? Thanks in advance!"

abstract quarry
#

use chatgpt for that

#

stable diffusion is for images, not webuis

desert dagger
sleek otter
#

So @karmic brook runs SAI's social-accounts, am I right?

karmic brook
#

I'm the server owner and verified as staff if you click into my profile

fervent thunder
#

I always thought fruit was an awesome username 🍌 🍎

desert dagger
sleek otter
#

Yummy

desert dagger
sleek otter
#

🙃

bright ember
#

hello, i need your help. i downloaded the ComfyUI tool for stable diffusion. But when i click on it it gives me an error: Connection denied from 127.0.0.1. Can you help me?

desert dagger
bright ember
#

my antivirus: windows defender XD

desert dagger
#

might be in yoru brower's security settings

bright ember
#

i use google chrome

rain portal
#

is there an easy way to rotate an image in automatic1111 for inpainting? if not is there an extension for that?

devout vapor
#

When i run a batch job with a set seed, it increments the seed from the one I set for each gen. I'm testing wildcards and i want the same seed for each gen. How do i keep it from changing the seed during a batch job?

hasty dust
#

hello

gentle cairn
#

hi there

timber wren
#

hello!

old lion
old lion
rain portal
#

is there any way for me to prevent adetailer affecting other parts of the image? when i put an image into img2img that has previously been inpainted and use adetailer on a different part of the image the part of the image that was previously inpainted loses detail.

#

it seems to affect the whole image but the inpainted parts are the most noticeable

warm junco
#

Then it shouldnt affect the whole image

rain portal
#

ok so on adetailer there is a setting called "inpaint only masked 2nd" is that the setting you are talking about? its already on

#

hmm i think i figured out what the problem is. with adetailer even with denoising strength set to 0 on the img2img settings it is still redrawing the image before adetailer starts

#

seems like i might have to manually inpaint to avoid this and not use adetailer

warm junco
#

There is a setting called: inpaint area

#

Oh wait I get it now you have already Inpainted it

rain portal
#

right ok, but i will still have to manually inpaint the area right? which defeats the point of adetailer

warm junco
rain portal
#

ok thanks man

warm junco
#

Np

warm crystal
#

hi

bright ember
bright ember
old lion
sinful fox
bright ember
old lion
bright ember
fickle siren
#

Are people using local stable-diffusion installs with Python and GIT, or Cloud GPU's?

#

I've got SD working locally but on an AMD card with 8 gig VRAM it's results are poor.

still glacier
fickle siren
thin oak
#

Guys, who will help to generate one image for the track, it is very necessary, please help

fickle siren
thin oak
#

Yes(((

fickle siren
#

One sec

#

I sent you a message so you can make SD images.

silk condor
#

Is it possible to run Auto 1111 and Swarmui on the same PC at the same time?

fickle siren
#

I think so. They might share some resources but should be independent of each other now.

#

I am not sure if SD needs different versions of Pythin for example.

#

*Python

#

Probably be the same

#

Even if they can't, it's pretty simple to reset and get one or the other working again.

vestal dew
#

This is quite confusing. When I DL a SD 3.5 model safetensor it isn't the model? I have to also download some text encoders and make a call to a huggingface API?

#

seems ridiculous

fervent thunder
#

sounds like diffusers format

#

in diffusers format all the bits are seperate

vestal dew
#

Good grief

#

I guess 3.5 is a no go for me, then

fervent thunder
#

well its just a case of matching the format you need with the software you want to use to run it

#

not sure which format you need since you didn't mention how you want to run it

vestal dew
#

It says I have to allow gated access, set quantization parameters, set offloading settings

fervent thunder
#

where does it say that

vestal dew
#

then also download the text encoders and a vae

#

for SDNext

fervent thunder
#

oh I see

#

not sure about SDNext

vestal dew
#

it does a bad job making images anyway

fervent thunder
#

don't really know about that one

vestal dew
#

for some reason, the images I get on SDNext are awful compared to those on a1111, so I guess Ill wait for 3.5 support there

fervent thunder
#

would recommend Comfy or Diffusers

#

or Swarm which is based on Comfy

low moon
#

Have a tremendous weekendfriday.

vestal dew
#

Comfy is impossible for me with all those windows and connectons

low moon
vestal dew
#

I cant see it

low moon
#

Its eays once u get the hang of it

#

the possibilities are endless

vestal dew
#

there's roo much to have to see...my eyes are bad

low moon
#

u can make it look clean

#

u have control!

#

unlimited power

#

comfyui

fervent thunder
#

sadly we don't really have a well-packaged "easy option" for people that has the full set of model support and features

#

it essentially doesn't exist yet

vestal dew
#

Yes...SDNext has a lot of hoops and manual installation to get 3.5 to work

low moon
#

Yes. It will be a bit more time before an easy option like Photoshop emerges but comfy is the closest we have...

#

or just use forge bleh

#

If you are a total newcomer I do recommend Forge.

vestal dew
#

I have been using a11111111 for about 2 years...tried comfy , it was no good with the spaghetti

low moon
#

Yeah so you can try forge. It;s like A1111 but a bit better. And A111 has no support for it anymore from what I gather.

vestal dew
#

does forge support 3.5? Does comfy without having to DL the encoders etc?

low moon
#

I think it does. It supports Flux for sure

#

its updated pretty regularly.

#

and comfy supports everything so yeah...

#

dats why i reccomend it

#

it took me almost 6 months to embrace it

#

but once i did i never go back

#

i only use forge for casual image gens

low moon
#

All hail! Make $500,000 dollars in 2 minutes doing nothing!

fast sage
#

Whats currently the best base model to use with comfy at the moment?

abstract quarry
#

I would say Flux base

#

but it depends on the style you want to achieve

slim quartz
#

I would like to discuss this:
ago I had some hopes that one of the video generator platforms would implement a software chain similar to the following.

Step 1: Generate image T-pose with AI image generator.

Step 2: Turn 2d image into 3d simple mesh with AI.

Step 3: Auto-rig and animate 3d simple mesh in an open game engine or one such as Unreal Engine or Unity with something similar to Muse animate and Cascadeur (including physics from game engine for water, particle effects, gravity and lighting to give a reference point for the AI. It's important to note that the 3d object itself need not have a detailed or extremely accurate mesh because it wont be viewable in the final visual output. Locations of humanoid/creature body parts, environmental objects is most important.

Step 4: Take frames of that animation and use something similar to Control Net to place New AI generated images on each frame to give any art style you want (Realism, Anime, 3d model etc).

Step 5: Since this is in a game engine you can move the camera and objects around at will to get the angle you want. Also object to object's relationship in distance and size will stay the same, so the user can create a more permanent scene and have longer video clips in the same environment, (usually AI clips are 10 or fewer seconds only and the locations are not easily reproducible).

#

Note for digital infrastructure: (A lot of video data could potentially be turned into "suitless" motion capture data for training a 3d animation tool. Synthetic data from video game assets/characters created inside the game engines would add to the pool. Within the game engines different angles could be recorded simultaneously with the segmented parts of characters and objects auto labeled for training data. The camera angles/ camera movement, the environmental objects, the character/creature type, the characters starting location and ending location, the type of movement (running, jumping, crawling etc) and the positions of all components, could be semi randomized to vastly increase the amount of quality multi camera angle data.)

quartz siren
#

Sd3.5 large will also use less vram I believe

slim quartz
#

I think it would be helpful to have a video game company like unreal or epic games paired with a video generator team.

desert dagger
slim quartz
fast sage
#

Dumb question, downloading the pruned of that, will that checkpoint "just work" if I drag it into my checkpoints in comfy?

slim quartz
#

@gritty scarab what are your thoughts?

desert dagger
fervent thunder
#

in their current states I would base a fine tune on flux
maybe in a few months there will be an amazing Realvis or Jugger for SD3.5

desert dagger
fervent thunder
#

however if you need apache license the choice is more like Schnell, Lumina, Auraflow

fervent thunder
desert dagger
fervent thunder
#

TBH I spent the whole year thinking a new Pixart with Apache license was coming and now it seems it isn't

#

cos they made Sana instead

fast sage
#

I assumed I was on SD 1.5

desert dagger
fast sage
#

Thanks!

#

Last question lads,

STOIQONewrealitySDXL_SD10 Anyone know what version this is?

#

is that 1.0?

fervent thunder
#

not sure but sometimes there is SDXL 0.9 and SDXL 1.0

#

it doesn't matter much

#

it will run using normal SDXL inference methods

desert dagger
fast sage
#

I need that specific version as I am testing a workflow

desert dagger
fast sage
desert dagger
#

so they've merged a bunch of checkpoints

#

and: Base Model
SDXL Lightning

fervent thunder
#

oh these numbers are the version numbers for the Stoiqo people

fast sage
#

Does that mean that the checkpoint the workflow is looking for wont be the same?

desert dagger
#

and lightning models are 'optimized' to run fast

fast sage
#

I should just go with the SDXL 1.0 and try.

fervent thunder
#

look at the model card under "details"

#

and it will say base model

#

its occasionally wrong on Civit but rarely

desert dagger
fervent thunder
#

there is an issue on Civit sometimes with knowing exactly what is in a checkpoint

fast sage
#

Yes I'm doing that, but from what I Can tell there's no checkpoint with the same name, so I'm assuming unless I have that exact "old" checkpoint I just have these to use.

desert dagger
#

to me it looks like they made a horrendous mess and tried to merge all sorts of models.

fervent thunder
#

is what I would do

desert dagger
#

that is what you're looking for

fervent thunder
fast sage
desert dagger
fervent thunder
#

I wasn't very clear
what I meant was I would have agreed that training fresh models is better than making merges
however certain merges like Realvis seem good to me

fast sage
fervent thunder
#

this checkpoint has FLAN which is kinda controversial

#

its not clear whether FLAN is a good idea or not

#

the official download is here https://huggingface.co/black-forest-labs/FLUX.1-dev

fast sage
#

Looking at this now: https://civitai.com/models/617609/flux1-dev

I don't understand the "- Should use VPS or if you are not afraid of burning your laptop (or if you are rich, it depends on you :slight_smile: )" 2 paragraphs down

Is flux sending telementry data? Or do they mean the GPU?

abstract quarry
#

yeah, base model usually works best. There are many Loras for Flux out, but quite often its surprising that the base model beats them easily if you just improve the prompt

fervent thunder
#

I really struggle to find good Flux loras on Civit
that actually improve image quality and don't burn the image at high strength

#

there are some, its just hard to find

fast sage
#

Actually last question, do you think it's worth it to upgrade from 1.5 to flux?

desert dagger
desert dagger
fervent thunder
#

SD 1.5 still has the best control nets, brushnets, IC-light and powerpaint etc

#

and it is the model that has the most Arxiv papers too

low moon
#

SD15 ages like fine wine.

fervent thunder
#

it really does yeah
tools keep coming out that make SD 1.5 better

desert dagger
fervent thunder
#

SDXL is catching up in papers/tools slowly

desert dagger
#

it really does take time ;)

fervent thunder
#

a lot of the papers are just done on base research models anyway
like there is a DiT just called "DiT"

slow dragon
#

hello

zealous furnace
#

i just downloaded it from some yt tutorial and everything i try to create looks so bad bro :(((

#

like if i go to a website find a cool ai photo and then use the exact same prompt mine turns out all distorted and low quality

#

some1 pls help

fervent thunder
#

could you try to describe what you downloaded

zealous furnace
#

Install Stable Diffusion Locally (In 3 minutes!!)

#

that was the title of the vid

fervent thunder
#

okay this is A1111 I can't help with that

zealous furnace
#

what do u use?

#

if there anything better bc this one sucks

fervent thunder
#

I use ComfyUI or Diffusers

split compass
#

/Factors associated with changes in subjective well-being immediately after urban park visit

#

#Factors associated with changes in subjective well-being immediately after urban park visit

desert dagger
idle axle
#

Well, damn. I just realized where probably much of the completely weird word salad originates from: people running clip interrogator on images and copy pasting bits of that to prompts (and other people in turn copy pasting them further). Just a few (non-ai) images and I'm seeing exactly same pieces of word salad as you find on civitai.

old lion
#

Yes. That’s the legacy 1.5 left us with.

fervent thunder
#

seems that prompting barely matters for image quality anyway

idle axle
#

Eh. I consistently find there’s a big difference in results from a decent negative prompt. Positive is then fighting with the text encoder to get the composition and elements I want (in reasonable computing time and without butt chin).

fervent thunder
#

prompt following and image quality are a bit different

#

there's not really been a prompt engineering technique that really moved the needle on benchmarks so far

idle axle
#

Which is why I specified positive and negarive differently. No negative prompt = outright bad image quality in my uses.

fervent thunder
#

feels unlikely since Flux doesn't have a negative and looks very good

idle axle
#

That’s the thing. Flux doesn’t look good to me and is ridiculously slow on top of that.

fervent thunder
#

the base flux doesn't look good to me either

#

needs some loras etc

hardy nexus
#

Any wip fine-tune of sd3.5m?

arctic coral
#

Do you know why when using refiner at any strength, tried switch at .10, 0.50 and 0.90, always corrupts the image at the very end?

The otherwise nice looking character in the preview suddenly turns into odd colors, low res and such

abstract quarry
# zealous furnace what do u use?

if your new then InvokeAI is probably the most newcomer-friendly UI for stable diffusion. You also don't want to use SD 1.5, its super old and looks really bad except you add 1000 additionals tools and custom models to make it look okay-ish

#

if your gpu is good enough, use Flux. Otherwise you can use SDXL Lightning or similar models. Or you look up tutorials how to run Flux on old hardware, but I'm not sure if you can then use invokeai

echo marsh
#

Question.

Does SD3.5M [mmdit] even use a value that signifies a set maximum % the model can alter the initial noise?

If so I want to try and apply a higher value however I don't know if rectified flow even uses such a value

#

(The value is sigma max in SDXL and SD1.5)

high orbit
#

Whats the angle called thats not from the side but also not from the front. the one inbetween

fervent thunder
high orbit
#

i mean where the character is not looking directly at the viewer but diagonally away. i cant find a prompt that ponydiffusion understands

fervent thunder
#

GPT 4o gave this standing with a focused expression, looking diagonally to the left (or right) of the viewer.

#

however since its pony this probably won't work]

#

you need an expert for that model

#

there is likely a tag

high orbit
#

hmm ok

last bough
#

Hey everyone,

I'm looking to turn a real-world video into an anime-style video. I've heard that Stable Diffusion can do this, but I'm a bit overwhelmed by the number of models available.

Could you recommend the best Stable Diffusion model for this task? Any tips or tricks would be greatly appreciated!

Thanks in advance!

broken smelt
#

best bet is to go to civit.ai and see which models you prefer. no such thing as best since its all subjective in what you think looks good

#

as for videos, tokyo_jab on reddit has a very good method for doing it, most consistent that ive seen so far. he posts a step by step guide on his profile. there's definitely other methods and a huge range of stuff you can try out but if i were to do that, I'd try his method first

last bough
#

Thank you I'll check it out

steel fable
#

Besides negative prompts, what more i can do to avoid double hands, body imperfection on SD 1.5?

mighty bane
#

what's a good custom node for resizing images in comfyui?

arctic fulcrum
#

is anyone bored?

#

could someone generate me a pr image of someone crawling into a man sized croissant thats like a bed but all steamy and delicious

#

lol im not at home and i need this in my life rn

brave jetty
#

Hi!

#

Do you know a way to run Omnigen with low vram (GTX 1070 8G)?

gray vapor
#

Is everybody using Flux/SD3.5 by now?
Im still too comfortable in sdxl, and i feel like flux and 3.5 are still not very optimized for quickly iterating in 8 to 16gb of vram

#

Should I switch? 🙂

craggy kindle
#

Hey anyone know any good prompt helprs or generators?

#

I hvve likeideas but idk howto promot for them correctly

desert dagger
desert dagger
desert dagger
craggy kindle
desert dagger
# craggy kindle Damn, I only use Pony Diffusion models

um well - pony was captioned a different way, so it may or may not be useful to use the descriptive text from an LLM with it. what you should probably do is look at how individual terms work with your models and learn how they think. start by just giving them simple terms like apple. or pear. or bunny. generate just that one word for a prompt several times and study the images. then give them prompts that are a few words like an apple on a table. or a bunny eating a carrot. learn how they think, how terms work with them, and then you'll have a much easier time crafting good prompts for them

gray vapor
desert dagger
#

don't switch - just add more tools to your toolbox

gray vapor
#

Also, is prompt following that much better?

#

Thanks btw!!

desert dagger
deft horizon
#

Does it display what you're asking more precisely?

gray vapor
#

Yeah thats the question. If you can prompt it more naturally, describing relationships between objects in the composition, etc

desert dagger
#

and THEY have the comprehension needed to actually use prompts like "a white man wearing a white tee-shirt and a black man wearing a green jacket running away from a small sports car. the car that is fairly far in the distance, in the background, is on fire."

#

or "on the left of the image is a green dog. on the right of the image is a blue cat. in the middle of the image is a yellow triangle with a frog sitting on it"

deft horizon
#

Hum... What?

desert dagger
#

this is not the way to actualy get your message across. all you do by doing this is make people angry and make them ignore what you have to say

gray vapor
#

Also, what are the most recommended methods for Inpainting today? It seems overwhelming.
Ive tried Fooocus inpaint nodes, Differential Diffusion, Controlnet Promax Inpainting

desert dagger
gray vapor
#

Promax Inpainting seems to have the most coherence, since it recreates the whole image

desert dagger
#

the AI should read the entire image, but only act on the selected areas

gray vapor
#

since this is never absolute, some pixels end up changing

desert dagger
gray vapor
desert dagger
#

okay

gray vapor
#

What is your favorite method?

idle axle
#

On the topic of SD3.5, is it significantly faster than flux? (which is too slow to be usable to me)

quartz siren
quartz siren
# gray vapor Also, is prompt following that much better?

They are similar but flux is usually slightly better at prompt following. Text should be better too along with humans which are considerably better. But as crystal said, sd3.5 large is more creative and has less biases, knows more art styles as well.

desert dagger
#

@quartz siren what's wrong with medium?

quartz siren
desert dagger
#

medium is designed to be artsy, anyway, so it's a really good refiner

bleak flax
#

Hello! I am new to stable diffusion and have come across this article which talks about the ability to use it/AUTOMATIC1111 and controlnet extensions through Google Colab (https://stable-diffusion-art.com/controlnet/). I am thinking of doing this, but want to make sure that the notebooks on Google colab are safe and the images i create remain private/my own as i want to use it to create my artwork. Does anyone know if these notebooks are legit/private?

pseudo nacelle
#

Does having more RAM while having 4GB of VRAM still helps to avoid memory out?

#

For instance, I upgrade my system ram from 16GB to 32GB

idle axle
#

You should probably upgrade to 32 GB in any case. It’s cheap and very helpful even outside SD.

pseudo nacelle
#

👌

desert dagger
lucid bobcat
#

I have been using Flux with Forge UI for the past week and wondered why my images looked so bad compared to Civitai gallery examples, even if I tried to recreate them with the same prompt. Now I know: Forge UI creates terrible looking images compared to ComfyUI. Now that I'm using ComfyUI my images look just as superb as those masterpieces on Civitai. Which is weird because I'm using the same model, clip, text encoder, vae, sampler, scheduler, prompt, steps, cfg and resolution as with Forge UI. Could it be that this is just a bug that will get fixed some time later? I would prefer to use Forge UI because it's way more user friendly. But right now I can't justify it.

median arch
#

hi

idle axle
zealous furnace
floral umbra
#

hoi, do you guys know of a workflow or 2 that upscales and sharpens images without the use of trained models? As the upscale models i have tried to use for sprites just gets too artifacty.

bleak flax
young acorn
#

Is it normal for the installation to get stuck with Easy Diffusion?

fervent thunder
#

Thousands of companies maintain databases of discovered security vulnerabilities through their bounty programs. These databases, while currently separate, represent a massive collection of exploitable security flaws. The crucial realization is that combining these databases could create a training dataset for AI language models specifically focused on generating remote code execution (RCE) exploits.When a company puts out a bounty for a security flaw, and one is discovered, they don't just write it down in a database, they patch the flaw.

#

Secondly, LLM technology in general is available to both people on the penetration side and on the security side.

#

But more importantly, you skipped the explanation of how someone is going to access and combine the private databases of security flaws. How are they able to access them in the first place?

cedar salmon
#

well you see the internet is like swiss cheese

fervent thunder
#

yeah the real issue is what if a mouse comes and eats its way through

cedar salmon
#

certainly a concern

fervent thunder
#

if its open source then the people on the security side can go and read it and patch it

#

you can't just skip to the step where you have an rce llm

#

if its based on open source security flaws those are already patched

#

but they close the flaw once it is discovered by the bounty hunter

#

they don't just leave the flaw open

cedar salmon
#

whats your answer or just uh proclaiming our doom

fervent thunder
#

why would Microsoft not close a known flaw?

#

why do Microsoft pay bounty hunters if they are not going to close them?

shy axle
#

hello

fervent thunder
#

again you can't skip to the step where you have a working RCE LLM

#

exploits don't have patterns they are often very different

#

besides which, the security side also has access to LLMs

#

in fact Microsoft has the unreleased prototypes of GPT 5 right now

muted skiff
#

What's Emad up to nowadays?

fervent thunder
#

I don't understand your point about bounty systems
because the companies close the loopholes found with the bounties

#

that's not really true though, they tend to be very different

#

but this is missing the broader point
the security side also has access to LLMs
which they can use to patch their systems

muted skiff
#

security exploits against popular software are like finding a needle in a haystack and there's no way AI is gonna be able to do it with their current level, when GPT4 is nowadays struggling to understand basic instructions

fervent thunder
#

to be specific
if the future of security is essentially a race between LLMs
then the large firms will win that race

#

given their larger budgets and compute supply

#

you're describing organisations that don't have a security focus, and then extrapolating that to organisations that do have a security focus

#

again you are skipping to the step where you have an RCE LLM

#

you seem to really want to skip to that step
but I don't agree with that initial premise

#

there is an upstream/downstream system for security patches, its not the case that every software company would have to do the work themselves

#

PS5 isn't the most secure system in the world its hilarious that that is your example

cedar salmon
#

applying patches from vendors aint gonna cut it

fervent thunder
#

you're doing the thing I described here ```you're describing organisations that don't have a security focus, and then extrapolating that to organisations that do have a security focus

lucid bobcat
# idle axle It’s almost certainly some generation setting that has a different default value...

I couldn't find clip skip in Forge UI. The quality problems my Flux generations have in Forge UI are two-fold: First, the quality of fine details like hair is really bad. For example hair looks jagged and crooked (instead of straight or curvy and smooth) and there's JPEG-looking artifacts around them. The second issue is the overall appeal. For example a squirrel I made with ComfyUI has a more expressive, cuter face and nice fluffy fur and the overall image looks very professional, whereas the Forge UI generation looks generic and uninteresting and definitely not production ready.

fervent thunder
#

your initial story involved hacking critical banking infrastructure
which is drastically more secure than Sony

#

Windows isn't the secure part of Microsoft
the secure part of Microsoft is their Azure datacenters, which so far have never been breached

idle axle
fervent thunder
#

there are companies with very little security focus
but those are vulnerable already so its kinda the same as the current situation

cedar salmon
#

Attackers breached Microsoft Azure by using credential theft through phishing methods and cloud account takeover, targeting mid-level and senior executives for financial fraud and data theft.

fervent thunder
#

this is still skipping to the step where you have a working RCE LLM

fervent thunder
#

its a separate category

cedar salmon
#

ya just leave them on the system lol

fervent thunder
#

I might be wrong about this but I was fairly sure there has not been a known software breach into azure data center

#

its always phishing

#

I've already given like 5 arguments as to why that isn't the case
I don't really want to keep repeating so I'm going to end it there

copper crystal
#

as far as LLM's remotely executing code, that won't happen.

The library that reads the LLM though....

fervent thunder
#

ah yeah its phishing again though

copper crystal
#

Personally whenever I see someone using "attack helicopter" as pronouns or a name, i put them squarely into the "needlessly contrarian" category though. I wouldn't invest too much energy into a discussion with this bad faith actor

fervent thunder
#

same with icloud, many phishing ones but never one via code, as far as I know

#

I always assume good faith personally

copper crystal
fervent thunder
#

yeah I agree there are phishing breaches

#

I consider phishing breaches a separate category to code breaches

copper crystal
#

yeah you can't program around someone giving away theiir credentials

fervent thunder
#

if someone was making the argument that LLMs raise phishing risks, I find that argument a lot more persuasive

copper crystal
#

it's like if you were to dress like you were getting ready for soultrain, but then acted like people were obnoxious for saying they expect to see you on soultrain.

The whole attack helicopter thing screams "bad faith"

copper crystal
fervent thunder
#

out of the points I made I think the strongest one was
if security is going to become a battle of LLMs
that's gonna be won by the big firms and not rogue entities
because LLMs are rapidly moving towards requiring a trillion dollar datacenter

copper crystal
#

If i came into a chat server with the name "Rittenhouse Swagger" what do you think that would say about me?

#

100% the security landscape is going to shift to a lot of LLM's engaging each other

fervent thunder
#

most big companies have an internal LLM of some sorts now

#

of varying quality

#

there are levels of vuln though
the high end vulns that are a risk to critical secure locations tend to be very convoluted and weird

copper crystal
#

a lot of those rely on old software.

and an LLM isn't required to deploy those at all. Scripts that hunt for open ports on old CMS versions already exist

fervent thunder
#

I also think its important to not oversell how good LLMs are at coding

#

I see GPT 1o errors all the time

copper crystal
fervent thunder
#

yeah I use them for most boilerplate stuff now particularly scripts

copper crystal
#

problem with current coding LLM's is they have no code validation in them. The library that loads the LLM can do it, but that's not conduscive to writing effective code

#

if you're suddenly 25x more productive, a 2400% increase, you should aim that productivity at creating a robust module testing framework because OH BOY you're going to need it

fervent thunder
#

would love it if you made a video, if you have a good methodology

#

I do worry that I am using them wrong

copper crystal
#

you have attack helicoptor 4000 as a name. All your flexing about clients is surely roleplay

fervent thunder
#

okay thanks

copper crystal
fervent thunder
#

hmm that's godwins law yep

#

the meme was from Obama era

#

when things were different

#

a bit like that green 4chan frog

copper crystal
#

i basically have zero respect for self proclaimed bigots, and figure they're always grandstanding some fictional version of themselves

#

if you believe this guy's bragging, you're a chump

cedar salmon
fervent thunder
#

he didn't make a particularly big claim he just said he is a programmer and claude helped him
I believe that at least

copper crystal
#

i believe he thinks it has made him more productive but hasn't used it in any code that is really involved yet

fervent thunder
#

I think LLMs made me more productive :shrug:
at the very least for code completion and stuff like wget scripts
you don't have to actually deploy the code they give you if its bad

#

GPT 4o can clean up a shell script pretty nicely

#

it can't do anything helpful in the area of statistical modelling or quantitative modelling which is a shame

#

it doesn't really even have the relevant software or languages in its training data much

#

and if you try to ask it the abstract math it can't do it in the abstract either

#

been trying to get LLMs to work with partial differential equations for like a year lol

#

ah yeah I prefer linux to windows

cedar salmon
#

build it a reference library rag

copper crystal
#

it's a pain in linux too. it's just the pain people prefer

copper crystal
copper crystal
#

so you used it for boilerplate code that you would've gotten faster at using regardless. gfy

#

if that mono code project is what you call productivity improvements, i guess

#

good example of how LLMs don't replace experience imo

#

ai didn't produce it. you did using ai to help you with vue.js and library management.

Faster doesn't mean better either. "functional" is not the only goal. manageable is a better one for long term library development

#

other tools could've given you these productivity boosts too. code templating is a big deal. What AI isn't giving you is a quality boost

#

that'll come with experience

#

i'd say that in your project's case, the core library was the MVP prototype already.

#

you just implemented a library meta made with vue js. 6h is normal for that, and it seems AI just brought you up to the mean

#

if you worked in the dev houses i had, your code would be under regular review and wasting the company's time. tbh.

#

lol, implementing sam2 with vue js is impossiblle in 6 hours?

i doubt you've got the kind of experience that qualifies that statement. your name is "attackhelicopter4000"

#

companies at that level are stacked wiht trans or furry software engineers, so it's unlikely you're one of their colleagues when you openly run with an alias like that

#

we're in 2024. a discord username with it is a social statement

cedar salmon
#

only for a minority

copper crystal
cedar salmon
#

i dont know, but you are saying true /shrug

copper crystal
#

i mean, noboyd has to argue against baseless speculations.

cedar salmon
#

word sensitivity is a new emerging trend

copper crystal
#

is it though?

cedar salmon
#

not new i suppose, maybe growing would be better

copper crystal
#

strawman helicopter is a strawman helicopter. one blazing arrow will down that tumbleweed

cedar salmon
#

odd tactic

copper crystal
#

i tend to not care about what bigots think of my intollerance of them

cedar salmon
#

pretty big leap there out of nowhere

fervent thunder
#

I think the strongest argument was that Microsoft et al. will have LLMs too
to use on the defence side

#

and they have more compute + budget than the rogue entities

#

this is stuff like Windows and .NET apps though
the secure and important servers in society aren't running things like this

#

what it is, is that you're underestimating how secure the more secure systems are
critical secure systems don't have a random .NET service that can be a vuln

#

what do you think about the fact that so far no one has ever breached an Azure datacentre using code

#

there aren't patterns to the high level vulns though

#

high level vulns for critical systems tend to be very convoluted and weird

#

its true that some are simple yeah

#

but the other side, the security side, also have access to LLMs that can race through assemblies

#

to find vulns to patch

#

its not really possible to hold a proper conversation with you about it because often I make an argument and then you just carry on and don't address it

steel fable
#

Anyone here is good on training Lora, i need a little help please

fervent thunder
#

its okay to get excited about stuff

#

the main reason its probably going to work out okay is that most of the "scary" scenarios involve LLMs breaching core key critical servers
and those servers will be secured
your typical person's domestic devices, and most company or government departments will not be okay, however they aren't okay now so there is no change there

cedar salmon
#

kiss your bitcoin goodbye, the plebian internet is doomed, maybe

lucid bobcat
# idle axle Yeah, some setting is just outright wrong (clip skip doesn't apply to flux - it ...

Could it be that maybe the samplers of my Forge installation are outdated or something? I noticed another quality issue: my images have high sharpness (sigma), for example between skin and blue sky there's a very dark line separating the two regions. I tried to reproduce this image: https://www.reddit.com/r/StableDiffusion/comments/1feibuv/guide_getting_started_with_flux_forge/
I got a different image even with the same seed. And it had all the same quality issues I always get. Could it be a difference between torch231 and torch24? On the other hand, I also noticed an issue with ComfyUI (although that could simply be related to the workflow I'm using): The shapes of the body are over emphasized. The ribs and collarbone for example looked protruding. The skin overall looked computer rendered whereas the skin on my Forge generation looked natural.

hardy sedge
#

Is stable Diffusion a GPU Killer when you use it every day 2 hours Long?

rain flicker
rotund wind
#

hello there

#

I wonder if processing an exhisting video with generative AI in order to get more realistic details is already something we can use in production.

I.e. providing a movie done in 3D and get a more realistic version of it . better looking people and vegetation, keeping a temporal coherence - much like we do on images for a while.

Anyone here has experience on this ?

#

Or maybe can point me to a link on this topic ?

patent jackal
#

Black horse man fusion with anime effect

#

Black horse man fusion with anime effect

desert dagger
fervent thunder
#

I'm not sure anything is worse for your GPU

#

apart from a handful of numerical computing tasks

young bronze
#

not anymore destructive than playing a game on ultra instead for those two hours

#

hey guys, hear me out.
Pokemon themed power rangers

fervent thunder
#

problem is people are running stable diffusion for many hours rather than two

#

and often leaving it running overnight as well

young bronze
#

there are extensions that monitor GPU temps and throttle it when it gets too hot but I have no idea if they actually work

fervent thunder
#

they do but at the expense of speed

young bronze
#

I feel like if the reduction is not too bad it's likely worth it for some folks

fervent thunder
#

it is sensible yeah

#

sadly many people go the other way and overclock and overvolt for more speed

young bronze
#

does that even work for image generation

fervent thunder
#

not sure

#

it may well not but people do it anyway

ionic wraith
#

is the 3090 24gb worth it for 575 euro?

fervent thunder
#

100% yeah

#

best deal there is

ionic wraith
#

Even if it is second hand?

young bronze
#

yooo, at that price?!

fervent thunder
#

yeah I knew you meant second hand

young bronze
#

it's a four years old card