💬｜general-chat

[astral goblet]

such a dumb term. if it's general intelligence, it's real.

a real oxy moron

[warm junco]

Controlnet with lineart Anime or lineart model

[gaunt pulsar]

I hope sd3 can generate rpg character sprites with some editing. Even Dalle-3 seems to be far from doing that

[charred mesa]

BREAKING: Stability AI an official partner with @rendernetwork

[charred mesa]

We are thrilled to welcome @EMostaque Founder of @StabilityAI as an advisor to the @rendernetwork to collaborate on next generation AI models, IP rights systems, and open standards powered by decentralized GPU computing.

[pure basin]

Anything we should know for new people?

[astral goblet]

Render network is crypto startup. Stability has partnered with a japanese blockchain company too, animechain.

not looking like a bright future now. Crypto is fraught with scams and shows very little real world utility outside of that.

while crypto often is built on ideals, the effective use of it is only ever scams and crime

[fervent thunder]

guys is there an alternative of Converting the PT file to SAFETENSOR ?

[charred mesa]

as long as we get SD3, finetuning tools and controlnets that don't suck then I don't mind

but yeah this sounds somewhat shady then

[dusk canopy]

They're gonna pull a mistral moment

"ip rights systems"

[astral goblet]

really sucks to see stability hitching their wagon to all these different blockchains. this is not the way

[dusk canopy]

Equivalent to selling your soul

[fervent thunder]

[astral goblet]

i used to love blockchain tech and really thought it woudl make a change towards a more decentralized internet.

here we are though. the legacy is crime and subterfuge

[dusk canopy]

Stable diffusion cascade is good at making mood images

[charred mesa]

I wonder if SD3 will be the last model like Mistral's last model being mixtral-8x7B

ehh

its a grim future

[dusk canopy]

Idrc if that's the case

Sd3 seems future proof

[astral goblet]

"We always intended stability to be closed and proprietary" - George Lucas probably

[charred mesa]

yes

[dusk canopy]

I know it'll run easily

8b parameter is nothing

[rough fossil]

hello world 😃

[dusk canopy]

Barely using 7gb ram

Vram

[charred mesa]

really?

at fp16?

how do you know

im happy if thats the case though

this would mean T5 and SD3 is possible on 12GB

[dusk canopy]

Yeah

[fervent thunder]

is there an alternative of converting PT files to SAFETENSOR ?

[dusk canopy]

I'm going off large language models and my personal experience with sdxl and cascade

[astral goblet]

you can unload t5 after building the embedding too

[fervent thunder]

The online tool is a pain in the coding ass

[dusk canopy]

Image models are easy to run

The most high end open image models run on consumer hardware

Unlike large language models

[charred mesa]

comfy does that automatically, but to RAM so it can be loaded fast again

this is how I can run Stable Cascade

its super epic

[astral goblet]

yup. i think we'll be fine here, but post sd3 i don't think theres much legs under stability. blockchain is kind of jumping the shark

hows that for a boomer reference?

[fervent thunder]

So

Any tool for PT to SAFETENSOR out there that doesn't have coding issues ?

[charred mesa]

try asking in#🤝｜tech-support

[fervent thunder]

it's a dead place

[charred mesa]

I have no idea personally

[fervent thunder]

people ask for pc ideas there instead of actual tech support

[charred mesa]

thats stupid

[dusk canopy]

Ask gpt 4

[charred mesa]

well its task support but still

[astral goblet]

i bet there's already safetensors of whatever .pt file you're worried about

[dusk canopy]

Or mistral large

Kr Claude 3

[fervent thunder]

ooo

mistral large ?

[astral goblet]

and also, .pt files aren't all bad. 99.9999% of the time they're fine. That .00001% is the possibility that somene coudl embed a script and you could potentially load that into something that actually runs that script. All major UI's don't run pickle scripts out of the box though

[dusk canopy]

Ye ask it

[astral goblet]

there's a ton of fear mongering around the ckpt file format tbh

[charred mesa]

when mistral medium came out we thought it'd be a nice way for the mistral team to generate extra income from a propietary model whilst preparing for other open releases

[astral goblet]

custom nodes and extensions are a bigger worry for security vectors really

they actually are scripts and people DO run them all the time

[dusk canopy]

All I hope for

Is sd3 gets released

If they close source after that idc

Because we'd have dall e architecture level anyway

[astral goblet]

pixart guys have proven that people can train base models a lot more efficiently now. Pony XL has proven that a small operation can put out something exceptional

we kind of don't need stability at this point so if they go full Roger Ver after SD3, we'll be fine. It just sucks to see the hero of the story turn into the black knight, y'know?

well actually, breaking bad was kind of awesome

a beautiful disaster

[dusk canopy]

Yeah

The papers are public

You can train a base image model with consumer grade hardware

Unlike large language models

[astral goblet]

if the sex bot addicts are dedicated enough, they'll find a way to weird science a new model

[dusk canopy]

They're very dumb

[charred mesa]

uhhh

[charred mesa]

I wish there was a very cheaply trained model that was like crowdsourced (like if that would work lol)

[astral goblet]

crowdsourced is just a buzz word that corporations created to farm donations without decalring themselves as a non profit

[charred mesa]

wasn't unstable diffusion crowdsource

[dusk canopy]

I just hope they release sd3

[astral goblet]

lol that was just a regular ol kickstartr scam. they bought themselves new gpus

[charred mesa]

they will

[dusk canopy]

TBF I don't give a fuck about open source licenses my bad to offend anyone but

I don't follow open source licenses

At all

If you open source something then technically that thing doesn't belong to you

Very controversial yes but it's a good thing

[charred mesa]

I was very surprised when I saw that Mistral-7B and Mixtral-8x7B were apache

[astral goblet]

kickstartr is the corporation the benefits off most of these grass root projects. they skim off every campaign. they've got lots of copycats out there too

[charred mesa]

gofundme

welp

[astral goblet]

starcitizen

[charred mesa]

It seems that SD3 is what Deepfloyd tried to achieve (prompt adherence)

except that SD3 will run on our computers and will have massive community support

[dusk canopy]

Do you think consumer gpus can finetune sd3

[charred mesa]

nahh

actually

hmm

I wonder if 24GB is enough for a lora

or Qlora for that matter

if qlora would work then absolutely

stuff like 12GB and less will be out of question

[dusk canopy]

What 💀

12gb light work for a lora

[charred mesa]

I honestly will mostly care about inferencing like I did all this time, especially with how good photography looks on the base model

[charred mesa]

8B model

in comparison

[dusk canopy]

You can easily use 12gb to train a lora

Oh

Still

For large language models light work

So it's probably the same for image models

[charred mesa]

Qlora will be an interesting story, cause if that would work then 12GB could possibly work

[dusk canopy]

You can finne tune 8b Param models with 12gb vram I believe

[charred mesa]

I want IPAdapter to work well with extracting subjects

(not just faces -> portrait)

[dusk canopy]

It would take a very long time though

[charred mesa]

but yeah photos looks great by default and I can't wait for Loras or Massive finetunes that will ENHANCE prompt adherence with extra actions, interactions, facial expressions, etc

those will be massive

[dusk canopy]

Cascade is better than sdxl

It can actually do complex prompts

[charred mesa]

no cap??? on god?

can you show me some

do finetunes increase prompt adherence

[dusk canopy]

#🏞｜general-with-images

Issue is it's harder to run

[charred mesa]

well on 12GB its not bad

[dusk canopy]

Takes 30 seconds for sdxl image cascade takes 1:20

[charred mesa]

???

damn

[dusk canopy]

Yeah

[charred mesa]

I don't remember that but I could be wrong

[dusk canopy]

90 seconds for cascade

[charred mesa]

wait dont you have 3060

[dusk canopy]

They say it's faster 💀

Yeah

[charred mesa]

thats BS

[dusk canopy]

Yeah

The guy who coded the gui I'm using used chatgpt to code it 100%

[charred mesa]

idk I remember some mentions about efficienty or speed

[dusk canopy]

It's really unoptimized

[astral goblet]

don't underestimate the ability of a coder to make their own inefficient code

https://thedailywtf.com/articles/The_Complicator_0x27_s_Gloves

https://twitter.com/search?q=%23sd3&src=recent_search_click&f=live ugh its all crypto spam

[dusk canopy]

Lmao

[astral goblet]

somehow i got stuck in the dumbest timeline. i need to get back to where the smart people forked off too

i'm probably just dumb and belong here

[fervent thunder]

Wtf is a pickle in coding

[nova zodiac]

a way of saving a set of tensors created by a ml model

[fervent thunder]

Safetensors or just tensors ? Thank you.

Damn I need to start leaning python

[untold herald]

Sorry for interrupting, any ETA for SD3 yet?

[honest spear]

it's not even in public beta, and they want to release it with a set of tools. We'll probably see the weights in a month

[dusk canopy]

Is the open beta like you have access to the model

Or just generation

[astral goblet]

https://docs.python.org/3/library/pickle.html object serialization in python

[astral goblet]

its a web interface as i've heard it but its all under nda so no one saying much. maybe they got actual weights. i dont' really know

[dusk canopy]

Wtf

[charred mesa]

https://stability.ai/news/introducing-stable-video-3d

[reef pecan]

are there instructions on how to run stable video 3d?

is it similar to stable zero123?

https://github.com/Stability-AI/generative-models

[bright breach]

What settings would you recommend?

[fallen gale]

so if I buy a membership to stability ai will they host the 3d model for me or do i still have to download it?

[gray fern]

pretty sure it just gives you commercial usage

[fallen gale]

oh

ok

[gray fern]

you still have to self-host, which I think is pretty silly tbh

considering you probably need a very powerful GPU to get speeds beyond a crawl with the model

[pseudo bough]

stable video 3d have any workflows ytet

[static cape]

It was released like 3 seconds ago...

[charred mesa]

I will never get this about the community lol

[rough fossil]

Its like those job descriptions "must have 10 years experience in xyz language" but xyz language has only been out for 3 years

[astral goblet]

its copy written by some clerk / assistant / intern. maybe even automated.

just apply instead. if you've got a hireable portfolio/resume, it'll be fine

[patent sand]

gn all

[hasty hornet]

it's alright, they are just looking for time travelers

[karmic marlin]

Someone shared an image to turntable video/image sequence here: https://www.reddit.com/r/StableDiffusion/comments/1bi1v7c/introducing_stable_video_3d_quality_novel_view/kvhoe56/ I haven't seen any for the second phase to generate a mesh.

[crisp remnant]

hello does it really take 10+ minutes to generate an image on 3080ti?

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

[fervent thunder]

Not under normal circumstances; that's a pretty damn fast card.

[fervent thunder]

Привет брать

[thin plume]

I've been meaning to buy a 12 or 24 GB VRAM nvidia card, what will you recommend?

[fervent thunder]

Go big or go home, 3090 or 4090 (if you have the money for the latter). I think the ML boom will cause / is starting to cause, a GPU "drought" somewhat like what we saw with the Ethereum mining plague of 2019-2021. So, I would buy now if you can. Usually if you wait, you can get tech for much cheaper, but with all the "AI" stuff, that may not be the case.

[thin plume]

the 3090 is with 24 vram

do you have one with 12 you'd recommend

since I am running on an rtx laptop version with 6

[fervent thunder]

Also, I know I am "boring", but I don't want my computer to look like a gay disco, I'm not into RGB, I like the reference models. Sure you can probably get slightly more performance from some (re)manufacturer.

[fervent thunder]

Okay, unless you're itinerant I would go with desktop, your money goes way further.

[thin plume]

no no I already have a laptop it's just that the gpu is faster than my pc

and I've been trying to save up for a new pc gpu

[fervent thunder]

Ah, I'd go for a 3090 then.

[thin plume]

is there a 12gb vram one you'd still recommend

[fervent thunder]

No, go big or go home.

[trail lion]

12gb isnt enough if you plan to do any training

[fervent thunder]

Do you own a Pine phone?

[crisp remnant]

изза чего может быть?

у меня проц 2010 года

[thin plume]

just generation

[trail lion]

no, the Pine had other meanings, one of which was pine trees

[thin plume]

as a poor man's 6 gb vram I've done a lot of workarounds to even get some stuff to run

[fervent thunder]

Oooh, like the riots?

(in the USA, those riots)

[trail lion]

nah, I live in a pine forest

[fervent thunder]

Ah, very cool.

Pines smell great.

And the sap like, repels moths or something (?).

[fallen gale]

is there a way to load sv3d on low v ram (16 gb)

i cant find where in the code i would specify to load in fp16

[karmic cedar]

Stability missed an opportunity by not designing a web portal for SV3D and making it look like it’s from Willy Wonka

DEVIIIIIIN I NEED YOUR ARTIFICIAL CODING SKILLS STAT /s

[fervent thunder]

У тебя есть Эльбрус? )

[hollow fable]

Does anyone know how to get 3d Mesh's out of SV3D? Looking at the github readme, I see no mention of mesh output, only video. Yet many of the blog posts about the model mention 3D mesh generation. Just wanted to make sure I wasn't missing an obvious CLI flag for that output, or if they're using a separate model to turn the video into the mesh.

[buoyant moss]

@Vex I haven't looked at SV3D specifically, but for pixel-wise prediction models like Zero one to three, the approach is to learn an implicit representation then run marching cubes

@hollow fable \

[pallid flame]

does anyone knows where i should ask questions about Reactor extention ?

[robust wind]

https://discordapp.com/channels/1002292111942635562/1004159122335354970/1219113832987885588
[Stable Diffusion 3] Paper figure 3 (link above): Does anyone know why the FID score increases with a higher number of sampling steps? Shouldn’t it decrease instead?

[desert ravine]

anyone know any good videos on how to use the kohya gui

[opal hedge]

Stable vid 3d ui waiting room

[trail lion]

this isnt new, but it's pretty thorough https://www.youtube.com/watch?v=xXNr9mrdV7s

[astral goblet]

Blockchains are a virus

[noble star]

if anyone wants to try sd3 I have a code

[robust wind]

I want

[lilac cypress]

yes plz

[blazing garnet]

Shameless plug but just a recent paper I worked on for the last few months - https://x.com/__z__9/status/1769911791117578518?s=20
Tl;dr: We demonstrate how to utilize generative data in category only online CL framework. More importantly, we propose a prompt diversification module and a novel sample complexity guided ensembling technique that strongly improves ID and OOD performance in online CL benchmarks.

We show SDXL, DaLLE-2, CogView and DeepFloyd can vary in generated sample complexity for same concepts and same prompts.

Would love some feedback 🙂

[buoyant moss]

Nice work! The abstract is quite hard to read, it doesn't read like what a native english speaker wrote it

[trail pecan]

Now it remains to wait for the ComfyUI update https://x.com/_akhaliq/status/1769926279053103565?s=20

https://x.com/EMostaque/status/1762920673779802522?s=20

[buoyant moss]

Question is if the SD3 Turbo will be better than SD3 Lightning!

Looking at the SDXL Turbo vs SDXL Lightning results, I feel like the Lightning model looks quite a bit better

[trail pecan]

The PDF file said that the prompta's following is worse than that of SD3

But the rest is at a high level

[blazing garnet]

Thank you for the feedback. Yes indeed, not all the authors are native english speaker. We will improve the writing of the abstract further. Anything specific that caught your attention to be indicative of being written by non-native english speaker?

[pearl ocean]

SD3XL when

[buoyant moss]

Although prior arts -> Prior art suggest that
whole sentence: Prior art suggests that webly supervised training can be accomplished using web-scraped data.
this poses challenges such as data imbalance, usage restrictions, and privacy concerns -> However web-scraped data may raise concerns in data imbalance, usage restrictions, and privacy concerns
Addressing the risks of continual webly supervised training -> In order to address the risks of continual webly supervised training
The proposed G-NoCL -> The proposed G-NoCL method
generators G along with the learner -> generators G along with a learner
When encountering new concepts (i.e., classes) -> When a new concept is encountered (i.e. classes)
G-NoCL employs the novel sample -> G-NoCL employs a novel sample

The abstract and paper has many many grammar mistakes 😦

[blazing garnet]

I dont think so, we used both Writefull and Grammarly grammar parsing. Also webly supervised is correct -> its not weakly supervised.
But sure, we will check further, thanks!

[buoyant moss]

Yes, sorry about the webly thing, I removed it

And yeah, it can be difficult

[iron slate]

Is anyone in here knowledgeable about double gpu setups? Cause i need desperate help

[buoyant moss]

whatchu doing

I ran a quad 1080 Ti setup until last year

[iron slate]

So when i run my AIs my PC blackscreens and then restarts. Sometimes bluescreen

[buoyant moss]

I bet it is a power issue

I'm serious

When I ran the quad 1080 Tis, the whole room's light would flicker when I ran machine learning jobs

And that was with a 1600watt PSU

[iron slate]

I have a 3080 and a p100, i have a 1200ps unit

[pearl ocean]

maybe you simply just need a 4090

[buoyant moss]

I run a 4090 now

much faster than quad 1080 Ti

[iron slate]

I dont have enough kidneys to buy one lol

[buoyant moss]

Well, I bet 5090 is coming out soon

[pearl ocean]

I just have a simple 4080

[iron slate]

But i really need help with the double set up. If its not the psu

[buoyant moss]

I really think it is the PSU

[iron slate]

Its so fast when it did work but then it dies(it worked day one then not again)

[buoyant moss]

Reaching 1000 watts plus on a P100 + 3080 doesn't seem impossible with a CPU

[iron slate]

I have a rysen cpu, 3900 12core i think

I'll look into a psu, any other things it could be that i can test right now?

[pearl ocean]

I think I have a 1000watt psu

[buoyant moss]

You could always try furmark

Furmark will test power consumption

[pearl ocean]

you guys think a 4080 is fine?

[buoyant moss]

maybe you need a 5090

[pearl ocean]

doesn't exist

[buoyant moss]

If you steal the leather jacket of Jensen Huang, it has time travel powers

100% chance it will manifest a 5090 in your oven

On a more serious note

I'm actually pretty excited by the Nvidia Thor product (500 tflops of fp16)

[astral goblet]

its for cars though

[rich kestrel]

can someone teach me how to use a purchasing bot to ensure I get a 5090 at launch

you would think there are captchas to prevent this type of thing in the first place

[astral goblet]

i ethically cannot spread such knowledge

[rich kestrel]

[astral goblet]

its not like retailers are trying to prevent sales. they're not really in that kind of business space

[fervent thunder]

Guys are there any demos where I can try this? https://stability.ai/news/introducing-stable-video-3d

[iron slate]

@buoyant moss mind if i dm you

[charred mesa]

https://arxiv.org/abs/2403.12015

Sd3 turbo paper

Ridiculously good results

Basically sdxl lightning, 4 steps, highres

And idk how buts its still intelligent and coherent a lot of the time

It does lose out on prompt adherence a bit compared to SD3, yet its still better than midjourney and below

It even generates at 1 step, but of course lose a bit of coherence

[robust wind]

https://discordapp.com/channels/1002292111942635562/1004159122335354970/1219113832987885588
[Stable Diffusion 3] Paper figure 3 (link above): Does anyone know why the FID score increases with a higher number of sampling steps? Shouldn’t it decrease instead?

[keen rose]

is there a way I can set some tags to appear by default? like EasyNegativeV2, uncensored, masterpiece etc.

they're used in all generated images anyway so it's annoying to re-type them every time

[opal hedge]

Imo from the new SD3 paper it looks like SD3 slaps the bajeezuz out of sd3 turbo

The speedup might not be worth the tradeoff in quality

Then again, 4 steps vs 40ish might be worth the consideration, who knows

[pearl ocean]

Still waiting for SD3

[static cape]

That paper is also low key teasing that normal SD3 understands the token "no" now ("no cows").

[pearl ocean]

could cry till the cows come home

[wise bear]

Unfortunately, from what has been said, it might be a while before SD3 comes out, like weeks or months.

[brazen leaf]

Any local open source tools describing images and generating texts?

[honest mica]

Check out taggui

[fervent thunder]

where is the 18+ section?

[brazen leaf]

I think this is an automated flow tool.

[storm shard]

#🏞｜general-with-images message

[opal hedge]

Though I am curious, why use 'no' when you can just add what you don't want to the negative prompt or make a negative positive prompt

[honest mica]

TagGui not tagui

https://github.com/jhc13/taggui

[pine fiber]

the new blackwell chips are 1000x the speed of my gpu lol

[honest mica]

The energy consumption costs more than my card

[pine fiber]

yeah but watts per flop is prob a lot lower

jensen pls give me one

[tight lance]

hi

[frail copper]

Hey everyone, I'm curious about something. If I use specific keywords and elements to train an SD Lora for creating images, and then later change up these keywords to design clothes, do you think the designs and elements on the clothes will come out consistently? Has anyone experimented with this kind of thing before?

[static cape]

Might've been a test since even bigger LLM's usually don't work well with negative instructions.

[honest mica]

Imagine you get one and then being bottlenecked by PCI Express 4 and your weak power supply. I would appreciate an affordable 24-36gb consumer card in the RTX xx70 range.

[placid wasp]

I have seen on multiple ComfuUI workflows that show connectors with straight line and 90 degree bends. How do i set this up?

[pseudo nacelle]

how much the least VRAM to make 1080p?

[placid wasp]

From what I have read in the past few days, you will need at least 10GB, and basicaly the more the better. but thatsa generalisation across SD

People were saying even though some stuff was meant to run in 8GB, that they had to use 10GB to do it

[pseudo nacelle]

hiayaaa, ok :') in 4gb vram

I can still create images in it tho, but it's just not hires

[placid wasp]

local install of SD?

[pseudo nacelle]

yes

#🏞｜general-with-images

[placid wasp]

on the Civitai site it states -https://education.civitai.com/sdxl-1-0/

"Hardware Requirements
The official Stability requirements for local inference (generating images) are 16 GB of system RAM, and an RTX 20XX GPU with a minimum of 8GB of VRAM. Linux users may also use a compatible AMD card with 16 GB of VRAM.

Training requirements are a little harder to pin down, but we have confirmation from Stability that LoRA can be trained on an RTX 2070 with 8GB of VRAM. With an input resolution of 768x, training used 7.1 GB of VRAM and took ~30 minutes.

Note: Despite Stability’s findings on training requirements, I have been unable to train on < 10 GB of VRAM.

Training at full 1024x resolution used 7.8 GB of VRAM and 2000 steps took approximately 1 hour"

This was for SDXL. I dont know if its the same for other SD models

[pseudo nacelle]

So I basically can't train lora... haiyaaa

[placid wasp]

You could ytry online. but it may cost. Civitai has a portal

im not bigging them up, i just found the site very helpful

[pseudo nacelle]

I mostly download models from Civit

[placid wasp]

me too, currently about 180GB worth

[pseudo nacelle]

oh wow

mine is just... around 30GBish

I'm new to this, like for a week

[placid wasp]

Im just playinga nd I have a massive system to play on

[warm junco]

Do you have --xformers --medvram --no-half-vae in your webui-user.bat?

[pseudo nacelle]

hmm, I got a question... some LoRAs sometimes gave me broken output when I set the resolution too high.. is that true?

[pseudo nacelle]

what even is that xDD

how to check?

[warm junco]

Startup arguments for to best performance for your gpu

[pseudo nacelle]

Oooh

[warm junco]

Right click and edit the webui-user.bat

[placid wasp]

i know on that. I believe the models are generated on a square 512 x 512 for SD1, 768 x 768 on SD 2,2.1 and 1024x1024 on SDXL

[warm junco]

At the line COMMANDLINE_ARGS=
You add --xformers --medvram --no-half-vae

Then save and relaunch the webui-user.bat

[pseudo nacelle]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --medvram --no-half-vae

call webui.bat

like that?

[placid wasp]

what quantity of VRAm will that allow to work CS10, does it off load to cup or just slowly burn through processing on the GPU?

[warm junco]

With that it should be much faster and uses less vram, allowing higher resolution

[placid wasp]

CS1o even

[pseudo nacelle]

but, at what cost?

[warm junco]

At no cost

[placid wasp]

just time?

[warm junco]

Thats why its in my install guide

[warm junco]

No its even faster

[placid wasp]

nice

what install guide?

[pseudo nacelle]

bruh, then why that commandline is not in default xDD

[warm junco]

Automatic1111 install guide for nvidia:
#🤝｜tech-support message

[warm junco]

Because it depends on the vram and GPU which args you need

[placid wasp]

Ahh, i went ComfyUI install, and just installed InvokeAi to compare and contrast

[warm junco]

Comfyui usually already uses xformers and adds the right args. But can be also modified if needed

[pseudo nacelle]

does the RTX 3050 Laptop still needs the --xformers and --no-half-vae?

[warm junco]

Yes, xformers is the boost you want and need.
--medvram splits large models into smaller pieces when loading them into the vram resulting in faster usage and less PC freeze.
--no-half-vae is to load VAE files as fp32 for compatibility

[placid wasp]

Yeah, i just added --listen to my startup bat so I could generate images on my phone from the install on the PC. no real reason, just geeking out

[pseudo nacelle]

oooh, ok got it

[warm junco]

Yea thats a nice arg, using it too

[placid wasp]

CS1o, Question, would I benifit from loading models into a RAMdrive. i have redicylous amounts of ranm

[warm junco]

There is also --share for access from anywhere

[static tapir]

#1080946152318443610 camputer chip

[pseudo nacelle]

time to give it a shot lmao

[warm junco]

That only helps if you quickly switch between some models.
They would load faster. The image generation time isnt affected by that

[placid wasp]

i was looking to grant acces to my box from the t'internet, but was wary. Im a cyber guy, so setting it up is all fine and well following best practaces but having read about pickle, I only use safetensors, not sure what other wories i should consider with SD

[warm junco]

If you use .safetensor files and also mostly download from civitai your fine.
Keep an eye on extensions as they have more rights and can load stuff.

[placid wasp]

Good to know. currently im on about week 3 of 'playing' with SD. I have no creative background so just download models and load them as and when to see what diferent outputs i get, got juggernaught, dreamshaper, blankCanvas and about 10 others 🙂 maybe I should streamline

[pseudo nacelle]

btw, does the Automatic1111 updates stable Diffuision when the new SD3 open source code released?

[placid wasp]

I also cant type

[warm junco]

Yes

[pseudo nacelle]

oh wow, that's cool

[warm junco]

And which webui do you use ?

[placid wasp]

ComfyUI

I like it because its teaching me 'the process'

[warm junco]

Yea its nice for seeing the workflow and understand what its doing

[pseudo nacelle]

I wanna see how ComfyUI looks xDD

[placid wasp]

ill pop a screen grab into gen-with images

[warm junco]

I'm more the auto1111 user. I don't like the custom node thing. Would take me to long to build workflows.

[pseudo nacelle]

I asked Gemini, the difference of ComfyUI and Automatic111 xDD

it says, the ComfyUI is for beginners cuz it was easier to use and understand

while the Auto1111 is more complex for those who wants some fine tuning

[warm junco]

Lol xD I would say auto1111 is easier for beginners

But its different for everyone

[pseudo nacelle]

after I saw twelsh37's img, I think so too

[placid wasp]

I get that part. i was 'brave' yesterday just installing InvokeAi, but it had a seperate install location and allowerd me to reuse the already downloaded models which was a godsend. Im sure i could throw Automatic111 on my box. what port does teh webuo come up on by default? I take it ot dosent clash with others?

[warm junco]

Auto1111 doesn't interfere with any other webui. You can git clone it into a seperate folder and its fine, you can also use the same model paths. But it doesn't have such a nice ui like invoke. But much more features and Extension support

[placid wasp]

Thats what I thought when i rist saw it, but not knowing any better i have persevered, there is a guy on YouTube, Olivio Sarikas, who does awesome vids. learned so much from watching him

[pseudo nacelle]

I only watch youtube the installing progress, the rest, I learn it myself xDD

[placid wasp]

Yeah, i thought that. i dont want to say its the 'grand-daddy' of teh UI's but it seems to have a lot of support

[pseudo nacelle]

wow, the commandline really does the job...

[placid wasp]

You would definitly struggle with ComfyUI then 😛

[pseudo nacelle]

xDD

[warm junco]

It is in fact the first webui ^^
It has the best support for Extensions but comfyui is on par mostly

[pseudo nacelle]

heck, 1280^2 really looks good

[placid wasp]

And thast the reason forums like this are great for knowledge sharing

[pseudo nacelle]

that's the reason why I find this server

[warm junco]

Make sure to use 10 hires steps. And for 1.5 models dont go to high on the default resolution as then the deformations and duplicates start

[placid wasp]

same

[pseudo nacelle]

oh, okok

I wasn't use hires, pure 1280^2

[warm junco]

Oh okay

[placid wasp]

CS1o, any tips on fixing hands and feet? models?

[pseudo nacelle]

I use textual inversion

[warm junco]

1.5 models are trained on 512x512 so you better stay near like
512x768 for portrait or 540x960 and then Upscale by to for FullHD

Or 768x768 works too

[placid wasp]

That was one of my first learning points. model size and teh cube.

[pseudo nacelle]

sometimes, when I upscale it... the image is just... like, looks ugly

[placid wasp]

well square

[warm junco]

Upscaling, good prompt, and there are some loras and embeddings that should help. But I mostly just upscale and hope for the best xD

[pseudo nacelle]

so I have to use like, 1.5x upscale

[placid wasp]

use the ugly keyword in your negative prompt

[warm junco]

You have to set the upscaler to Esrgan4x for example and then 10 hires steps and denois 0.5 And it will look good.
The default settings of hires fix are not usable

You can also use latent bicubic with 0.5 denois

[pseudo nacelle]

which one is good for anime like arts? R-ESRGAN4x Anime6B or the one you recommend?

[warm junco]

There are a few good ones you can download from openmodeldb.info
Like Fatal Anime 5000
Or lolypop
And yes the Esrgan4x Anime 6b is good too

[placid wasp]

This is so informative but i really need to go and prep for my talk later. Good to meat you both. CS1o thanks for all the tyips. I will go take a looka t installing Automatic1111 later

[pseudo nacelle]

got it

[warm junco]

No problem 🙂 if you have any technical questions regarding the Installation, my guide, or how you set the model paths, feel free to post in #🤝｜tech-support

[pseudo nacelle]

also, I notice that, when I use higher res, the character is just getting smaller, due to the base of SD1.5 models which is 512^2

but that means, wider view too

[warm junco]

You can prompt for close up portrait

[pseudo nacelle]

well yes

the pic is still not closing up to the chara, but when I use 512^2 upscaled to 1080^2, it works

[warm junco]

That should enhance the character as full body on 512x512 (not upscaled) is bad

[pseudo nacelle]

hmm, yeah

[warm junco]

For any image related questions feel free to ask in #📝｜prompting-help

[dusk canopy]

Someone give me prompts to generate I'm. Bored

[pseudo nacelle]

Hyundai N Vision 74 maybe? :v

[dusk canopy]

Ok

I'll paste in "Hyundai N Vision 74 maybe? :v" and hit generate

[pseudo nacelle]

well, no

[dusk canopy]

I can't add models

Or Loras

I'm in college

Generating images through gradio

[true gale]

Made an offer Emad can't refuse. 🤣

https://x.com/nightgrey_/status/1770032627329777838

Edit: In case you don't see Emad's tweet, this is a reply to him writing "I'll give you an invite if you give me an invite".

On a more serious note, have you guys seen Infinite ID? I hope they'll release the code, looks really nice! ✌️

[glass lily]

Hey guys! Would you be kind enough to link me one of the masking tools in the Extras tab? I reinstalled SD and don't remember what it was called.

[bleak matrix]

Good morning everyone! How are we all today?

[rugged mirage]

is there a channel for the 3d model which Im not seeing

[pseudo nacelle]

you mean inpaint?

[true gale]

There's not one specific to it, but it is probably #▶｜stable-video-diffusion given it's a Video 3D diffusion model.

Not sure tho.

[jovial wraith]

for those interested, i just released my pixel art animation workflow for Comfy

[pseudo nacelle]

and it's like, felling all my RAM and VRAM...

I found out this issue when I searched about the commandline too, any fix for that?

[charred mesa]

SD3 Turbo with highresfix is gonna be awesome

I can literally do what I have been doing with SDXL Lightning so far

except better prompt adherence 🤩

[trail lion]

do those turbo models offer anything apart from speed? and why does that matter, unless you have some kind of quota

[charred mesa]

its called being impatient

[trail lion]

got it, hah

[charred mesa]

if I got a super complex prompt that wouldn't work with Turbo I'd just unload the models in comfy and switch to normal SD3 weights and I'm good to go

[shrewd jasper]

how many it/s is considered decent ?

[charred mesa]

well depends on your patience of course

imo its like 3-5it/s if its like a 20-25 step image

idk

[trail lion]

what about s/it

[charred mesa]

I generate slower but cause I'm using quite a high resolution so I think its justfied

its like 1-2s/it when doing highresfix for me

[shrewd jasper]

i do 20 step 512x512 for 1 it/s only :/

on an intel iGPU

[charred mesa]

Oh really?

thats not even bad

openvino or that fastSD repo?

[shrewd jasper]

openvino

[charred mesa]

wow thats good

[shrewd jasper]

actually im not sure if it is iGPU

the new intel branding is confusing

[charred mesa]

is it intel Arc or Xeon or just generic iGPU

[shrewd jasper]

it is listed as intel arc but it is actually the integrated graphics of my intel core ultra cpu

[charred mesa]

hmmmm

Supercharge your gaming and content creation experience with built-in Intel® Arc™ GPUs on select Intel® Core™ Ultra H-series processors1, or upgrade to Intel® Arc™ Pro GPUs2 for ISV certifications.

[shrewd jasper]

but i cant do ipex on it so i think it is still iGPU? perhaps?

[charred mesa]

wow built in intel arc-like igpu

thats nice

and probably consuming 1/4th watts of power of my dedicated gpu

[shrewd jasper]

yeah

[charred mesa]

SD3 Turbo is gonna be sick, I really thought it'd be some lower res thing again with heavy penalties to prompt adherence or image coherence

but its more like lightning

jesus christ stability devs and researchers were cooking so hard

[karmic cedar]

Turbo will open a lot of doors

Lots of flash marketing opportunities to say the very least

[charred mesa]

We compare SD3-Turbo 1024²-MAR to SOTA text-to-image generators.

Yup, so it's 100% 1024px just like the base model.

[karmic cedar]

makes sense

[stone latch]

Hey I'm out of the loop, there's SD3 Turbo now?

[karmic cedar]

as their focus has been on the context not the scale

Not yet but it seems to be Stability’s cadence

[charred mesa]

its SD3 but faster at lower steps

similar to LCM, Turbo and Lightning

the trade off is a *very* slighty decrease in coherency and a decent decrease in prompt alignment

but it generates in ONLY 4 steps!

[stone latch]

I can't wait to toy around with SD3, but if a Turbo model does come out it will be very convenient!

[charred mesa]

and its STILL more intelligent than previous models!

exactly

[karmic cedar]

You’re going to have a lot of reading with these models

Lots of words

(good thing)

[charred mesa]

and if you feel like the prompt you made is super complex then you can just unload the model and switch to regular SD3

[karmic cedar]

Love it

[charred mesa]

but like a sentence is still usable with SD3-Turbo

[karmic cedar]

because they have to equivalent of FaceID going on more or less

[stone latch]

I wonder if SD3 will run well on a 4070 Ti 🤔

I was left with the impression you need 24 gigs of RAM to run that

[karmic cedar]

I don’t think it will. I think they will have quantized versions probably that aren’t quite as dynamically performative

As SD3 itself is a shift in complexity and thus a shift in resources

[stone latch]

That's a shame! Time to start saving for a 4090 😂

[bleak drift]

How exactly does DallE-3 work? It has to be some sort of stable diffusion to be able to make stuff of that good quality. Does the machine take the text you give and make a better prompt with it?

[stone latch]

4090 is ridiculously expensive...

[bleak drift]

RTX 4090?

[stone latch]

Precisely!

[charred mesa]

SD3 will hopefully fit on 12GB eventually or even at launch, it will depend on how much 8B Diffusion itself will take. Models are automatically offloaded to RAM once they have finished generating (T5 and the SD3 Weights).

T5 at 8-bit will probably fit on 12GB

[stone latch]

And on the topic of Dall-E 3, I like that it has more facial variety, so I came up with this - Dall-E 3 output + SDXL Canny ControlNet to add realism, as Dall-E realism looks very airbrushed

[bleak drift]

I am guessing there is a difference between a gaming laptop 4090 and a Desktop 4090. Because I am going to get a laptop with one of those, and a intel i9

[karmic cedar]

That’s really cool—I just hope the scale doesn’t diminish the quality too much, as there seems to be a lot going on in this architecture

[charred mesa]

on 16GB I think it will fit just fine

[bleak drift]

True. Technically, I could take DallE’s output and use image to image and inpainting to make it better or add stuff.

[stone latch]

Yep! Or you can use a ControlNet like I did. May I send an example here?

Not sure if I can send photos in this channel 😅

Looks like there's #🏞｜general-with-images

[lapis cedar]

same q

[stone latch]

Sent my example there

The face changes only a teeny tiny bit but it's not such a huge deal.

Though I wonder if using a T2I Adapter can be better 🤔

I tried with one but the results were a little messy

[bleak drift]

I really need to look up these differnent controlnet things, they seem really useful

[stone latch]

They are!

It's very powerful! I figured you could add realism to faces at first, which I did. Later I also figured you could change certain features, then combine a controlnet with inpainting to make very precise edits! And finally, colorizing B&W images. Though that last one isn't as simple as generating an image, manual work is required - I recommend using image manipulation software, layers and blending to achieve results.

[karmic cedar]

looks really good

[charred mesa]

I hope we'll get Canny, Depth and Openpose controlnets on launch

[karmic cedar]

yeah, that would be nice

[stone latch]

Imma send a few more examples...

[charred mesa]

they promised controlnets on launch I just don't know how good they will be or which ones they are gonna give us

[karmic cedar]

maybe even an alpha controlnet?

[charred mesa]

SD3-Turbo can also generate 1step images which resemble early LCM attempts

[karmic cedar]

oooo

[charred mesa]

I wonder what would happen if we went above 4 steps

I suppose they might just converge at around 6-8 steps and not improve 🤔

[karmic cedar]

eventually these are going to be so fast and the temporal consistency will be dialed in enough, that youtube streams will be all over the place of SD-generated realities lol

[stone latch]

Sent some experiments over at #🏞｜general-with-images

😄

[forest turret]

hey guys, does anyone know a good way to generate large scenes? it seems to want to generate an image of just one person/character but i want to generate large landscapes with many people

[broken smelt]

you're probably better off starting with generating the landscape then inpainting the people

[fallen drum]

4060ti 16gb vs 4070 12gb , Im trying to build a new computer for AI and im struggling between choosing which ones to buy , could anyone help me pick ?

[true canopy]

i mean, its somewhat hard to say, if its the regular 4070, and not ti, then 4060ti is prob better

[glass lily]

I don't think so! It masked like the character out (if there was a character), and the background became transparent

[stone latch]

The regular 4070 has only 8 GB I believe 🤔

[true canopy]

maybe he means 4070 super?

[pseudo nacelle]

Ooh, idk

[teal pagoda]

So there are so many versions of ControlNet canny, depth, lineart etc. for SDXL 1.0 and from like 10 canny models, only 2 work good out of the box. I tested them all. Why is that and why the controlnets for 1.5 models were better?

[charred mesa]

I actually have no idea they are lackin so much

[trail lion]

going through the same motions with xl vs 1.5, I agree, maybe there's better ways to optimize, but apart from that I agree more or less...but it is better than not having it at all

I've been more or less sticking to depth_midas, and ipadapter, both of which work pretty well on xl

the canny, sketch etc, hit or miss

also, playing with the resolutions in cnet for the models that support it, sometimes yields better or worse results

[rich kestrel]

because 1.5 is chad

[charred mesa]

it stood the test of time

I suppose now its gonna be between SD 1.5 and SD3 and the other previous models will be left alone lmao

unless SD3 is proven to be good at corn at the lower parameter count models

[karmic cedar]

SD 1.5 was brewed longer, so it has a higher ABV

😛

[charred mesa]

hehe

[karmic cedar]

(but i mean…probably true in essence)

[charred mesa]

I just hope SD3 will get GOOD and LOT OF controlnets at launch

Canny, Depth, OpenPose are a MUST, like it should be default

and maybe inpainting and edit if those are controlnets or whatever, idk if those are separate models

[karmic cedar]

Context is king…they can architect it any which way, but the encoding needs to have sufficient complexity

[fiery cove]

Lmk if this belongs in a different channel: If Blackwell chips are slated to cost 30-40k per Jensen, does that imply the price of H100s / A100s could fall? Or not necessarily because Blackwells could search well beyond initial retail price due to demand?

[karmic cedar]

I think they could fall, but…probably won’t fall as much as people would like them to because $$$

[pale latch]

People are stuck on 15. They'll stay

It's really sdxl vs sd3 and the 15 gang off doing their own thing

[karmic cedar]

not to mention with super resolution capability increasing on the broader end, there’s little reason for people to feel compelled to use newer models unless they’re dead set on text placement. Otherwise they can continue to work with fine tuning and SR techniques of their own.

[pale latch]

Stability is going to fall. Too many crypto partners lately. These are signs of buckling support.

[stone latch]

Why was Stable Diffusion 2 such a massive failure anyway?

[karmic cedar]

That’s how it goes with open source. 😦

SD 2 showed signs of being,..neutered. Though it was never made super explicit.

[stone latch]

Neutered as in?

[karmic cedar]

Less complexity.

More homogenization, IMO.

[pale latch]

Harder to make Loras for sd2 is why I don't

[karmic cedar]

that too.

[stone latch]

I don't see why I would use an SD2.1 model over an SD1.5 or SDXL model 🤔

[teal pagoda]

Hopefully SD3 will finally outclass SD 1.5.

[karmic cedar]

I agree

[charred mesa]

well intelligence wise absolutely, no question about it

Quality wise (if not the base already), finetunes will excel at photos

the base model looks soooo fricking good with photos

[karmic cedar]

yes it does! that tells me that perhaps their datasets are lighter, but the encoding is more precise / rich.

in a walnut shell.

[charred mesa]

I wonder how the 800M model will perform 🤔

[teal pagoda]

If SD3 will require resources like XL, then you're right. The 1.5 group will remain there.

[dusk canopy]

💀

[stone latch]

Not to be that guy but is SD3 super censored 🤔

[charred mesa]

with a lot of finetunin it can probably look as good as 1.5 finetunes but with more intelligence

[dusk canopy]

I don't care as long as they release sd3

[charred mesa]

idk it looks like as censored as SDXL

[karmic cedar]

@Sufi yeah, kinda expected at this point 😕

[dusk canopy]

Brother you can uncensor it 💀

[charred mesa]

which isn't that bad

[dusk canopy]

You just have to train or use loras

Etc

[charred mesa]

ehh

[stone latch]

It's not that easy

[dusk canopy]

They just won't train it on bad stuff TBF

[charred mesa]

well if they figured it out for SDXL then it will work for SD3, this isn't SD2.X or Cascade

[dusk canopy]

From my own personal testing these image models have no guardrails

I've only used finetunes except for cascade

[stone latch]

Also by censored I didn't mean whatever they might offer online

[charred mesa]

well no built in harcoded guardrails

but it still somewhat depends on how hard the base model was censored

[teal pagoda]

Guess so.

[stone latch]

But for example, SDXL simply can't generate certain things

[charred mesa]

probably

[dusk canopy]

What things

[stone latch]

In A1111 for example

[dusk canopy]

Could be a technological constraint rather than censorship

[karmic cedar]

we’ll see. in the end, a lot of laws seem to be coming up that are going to retroactively target abusers of copyright, etc.

[charred mesa]

no he probably means stuff like in 1.5 that works

I wonder how hard new concepts will be

[stone latch]

I actually just meant nsfw content 🤣

[charred mesa]

ahh makes sense

[karmic cedar]

mm

[dusk canopy]

What stuff

[charred mesa]

so far anatomy looks good for a censored model

[dusk canopy]

Bro 💀 there is no censorship at all then

Zero

[charred mesa]

uhh

1.5 had censorship but it was ONLY for names

[dusk canopy]

Unless you go into some really messed up anatomical stuff

[charred mesa]

like if it has like idk "nude" or whatever then it was excluded, but no NSFW image detection

[stone latch]

Sorry I was hesitant to mention it outright because I don't know what can get me banned 💀

[charred mesa]

nah its okay we shouldn't talk about stuff like these that much

[dusk canopy]

Won't get you banned

[charred mesa]

point being, it's 1000% better than SD2.X

anatomy wise, censorship wise

[dusk canopy]

Cascade is better kinda too

I've not tried anatomical lessons

Or things

[charred mesa]

even if CogVLM doesn't pick up nsfw, half the dataset was left raw captioned so there HAS to be some remains

[teal pagoda]

but why can't cascade be run like any other normal model directly in A1111?

[charred mesa]

and of course the model is massive (8B) so idk if that plays a role in anatomy

[teal pagoda]

without additional extensions

[charred mesa]

cause 2 models

[trail lion]

Just use a non base model, why the obsession with censored

[charred mesa]

1 for a very small image then the second model upscales it to regular resolution

[dusk canopy]

I'm running it off some GitHub thing some guy made

Issue is he used chatgpt to make it 💀

[charred mesa]

in comfyui it works perfectly under 12GB

[stone latch]

The Stable Cascade extension?

[charred mesa]

idk man I have faith in SD3 not being lobotomized

I just wonder if how good it will be at certain stuff like games and arstyles (of deceased or people who didn't opt out)

previous models kinda sucked at games

I wonder if Stability decided to get games and other stuff out of the model

[stone latch]

I don't think I have ever tried games 🤔

[charred mesa]

DC, Dragon ball and Marvel seem to work

like Batman, Goku and Spiderman at a table

[karmic cedar]

I think they zigged one way with 1.x, zagged another way with 2.x, and so on. They’ve been finding a great stride lately and it sucks to hear about the crypto speculation, but…open source is truly good work because it’s open source work.

[charred mesa]

well if we still get SD3 and SD3 turbo I'd be fulfilled

for a long time

[karmic cedar]

yes, definitely. and as for the new laws, we just gotta hope that they are enforced to principle and that the principle ultimately boils down to something that doesn’t destroy the freedom of expression or of creation.

[charred mesa]

SD3 Turbo is such a positive surprise :), I thought it'd be heavily watered down to get 4 steps but the geniuses at Stability figured it out

[charred mesa]

exactly

[stone latch]

I'm hoping that SD3 will come out soon though, because SDXL 0.9 came out like a week after it was announced if I'm not wrong 🤔

Then a month later 1.0 followed

[karmic cedar]

probably 2-3 weeks if they were talking April..

[charred mesa]

yeah emad said April and I have faith in that

even if like very end of April

[stone latch]

Well I can survive until then for sure 👍

[karmic cedar]

Emad sure doesn’t get enough credit in all the buzz.

[charred mesa]

testing phase will probably take place towards the end of this month and maybe beginning of April

they said they'll be inviting more people this week

[stone latch]

Still learning new things about what we already have! Such as this workflow I came up with to colorize images and add realism to Dall-E 3 faces

[charred mesa]

(besides like 2-3 twitter AI research users lol)

yeah DALLE-3 has a weird painterly look to all of the images

or maybe mostly for faces I've seen so far

there are some realistic photos from DALLE-3

[stone latch]

If you missed that, I sent some examples in #🏞｜general-with-images where I added realism to male faces 🙂

[charred mesa]

you mean img2img or within DALLE3 prompting?

ah yeah sdxl img2img

[stone latch]

Not exactly

SDXL Canny ControlNet rather

[charred mesa]

yeah

they look nice

I've been using SD since august 2022 and this moment will feel so special

maybe I'll feel the same way as I did back in august of 2022

having an image generator offline is now taken for granted so easily...

[stone latch]

There's something so specific about the way ChatGPT prompts Dall-E 3 though, don't you think?

[charred mesa]

yeah

there's SuperPrompt-V1 by Brian (a dev, he's here on the server)

[stone latch]

I can't put my finger on it but it feels so AI 😂

[charred mesa]

I made some gnarly images using prompts from those

unfortunately SDXL and previous models only understand like 60% of whatever superprompt makes

[stone latch]

Generate a photorealistic image of a man with a beard, capturing his masculinity.

I can't 🤣

[charred mesa]

SD3 is going to be massive with this tool

[charred mesa]

hmmmm

[stone latch]

Like why does ChatGPT prompt like that?

[charred mesa]

yeah its adding a lot of story-like detailing lol

cause I guess its an LLM made for general purposes

maybe the prompt they give ChatGPT isn't agressive enough

[stone latch]

But the question is why have ChatGPT be the middle man?

I could just prompt it myself like Bing Designer allows me to, just that in the latter I have 15 credits per day

[astral goblet]

i wouldn't say lobotomized since they're not going in and chopping weights out of themodel. What i think more is like, you know early childhood development problems? like that one girl who was left in her basement tied to a chair and never learned language.

[charred mesa]

well I suppose they did the same as SD3 (captioned dataset with vision models)

but unlike SD3 they might have done it like 75% if not more

therefore they might require that natural language type of prompts almost all the time to get good results 🤷‍♂️

thats just my theory

AN AI THEORY

[charred mesa]

EXACTLY
*Tenor Gif: office thank you*

[astral goblet]

it has the original sd15 clip layer still so pepople will still lean hard on prompt salads

[charred mesa]

and clip_g (which is more natural language?) and of course the T5 (which is optional lol) 👀

[astral goblet]

tenor gif? wtf. just gifs!! tenor gifs are like, bugs bunny being the maestro at the opera

[charred mesa]

lmao

[arctic sedge]

Do you guys think we could fit the T5 on 10/12Gb vram cards? 🫠

[charred mesa]

think so

[safe shoal]

Hi. I’m looking for a freelance dev for building a pipeline for stable Pm me if interested

[charred mesa]

at 8-bit 100% even if alone on itself

if you know about the comfyui's offloading technique then you know that we're gonna be fine VRAM wise (if the 8B weight is less than 12GB)

[stone latch]

Us in 2034: generate virtual universes using AI where we can insert our consciousness and just walk around and do things

[charred mesa]

[stone latch]

Just a random thought I had 😂

I mean with the speed this stuff is advancing...

[charred mesa]

@arctic sedge just take this with a pinch of salt, we don't exactly know the exact VRAM requirements for each file

but for T5 we have way more knowledge (LLMs are easier to guess out of experience, take LLAMA-7B for example, the T5-XXL is only 4.7B)

[stone latch]

We went from SD 1.5's "haha lol look at this distorted face and this terrible anatomy" to, well, this server's gallery in like 2-3 years?

[charred mesa]

exactly

[arctic sedge]

Yeah. It's something like 19Gb of vram?
But SAI also added that it was (Unoptimized) Whatever this is alluding to.

[charred mesa]

well they said 24GB (unoptimized) (with probably everything loaded at the same time 🤷‍♂️)

I wonder if xformers will chip off extra VRAM from the SD3 weight, idk if they already implemented that

[arctic sedge]

Possibly will, but i can't imagine it would be considarable enough ammount.

[charred mesa]

yeah

[astral goblet]

when i tested the lavi adapter code , t5 on sd15 didn't breach 10gb

and that wasa just me winging it with their test.py

[arctic sedge]

There's "hope" Lol!

[astral goblet]

pixart model uses t5 too i think. you can play around with that in comfyui today on your 12gb i bet. maybe not. don't take my word for it.

[charred mesa]

and that's without 8-bit it seems, can't find load-in-8bit:true in the code

but of course that can still decrease VRAM usage further with very minimal quality difference and without any conversion needed

[charred mesa]

yeah some people figured it out

same with deepfloyd

on 12GB

[arctic sedge]

Yeah. I also heard about that too. Shame it's too difficult to train. Not seen anyone make any majorly good finetunes of it.

[astral goblet]

all my life i been wanting more bits. when the NES landed i was like HOLY 8BIT!! then super 16 bit then whaat 32bit processors!? WTF!? NINTENDO 64?! HOLY FUCKIN SHIT.

then we lulled out for a while and now just as 64bit is starting to become ubiquitous, now i want less bits

[charred mesa]

there's gonna be pixart SIGMA

which also did some vision model captioning

[arctic sedge]

Oh nice!

[charred mesa]

In the SD3 research paper:
Mem is the memory required to load the model on the GPU. FP [ms] is the time per sample for the forward pass with per-device batch size of 32.

And T5 is listed at 19.05 GB for some reason? (Clip-G + Clip-L + VAE take up like ~3GB in total)

[astral goblet]

I dont know why all these alpha male types think sigma is a bad thing. none of them know latin clearly

nor realize that the alpha wolf theory was all bunked and made up

[arctic sedge]

Dab

[astral goblet]

it's got as much basis in real world as phrenology or humors really

[astral goblet]

in training they're probably keeping them loaded at full precision

https://deepmind.google/discover/blog/sima-generalist-ai-agent-for-3d-virtual-environments/ the fuck is this? i was just talking about how this stuff needs to come to game ai yesterday

too bad it's an alphabet model and they probably won't release weights

[charred mesa]

oh yeah right

this could be fp32

[charred mesa]

oh shit imagine some people buy this as a service and set it up as a stream

[astral goblet]

the last bastion of real people online, twitch, has fallen

we're all bots now

[charred mesa]

😔

its so over guys..

[astral goblet]

i type way too much so some people i can't convince i'mnot a bot

i fool so many others though

i mean.. i um.. i'm not a bot

[charred mesa]

👀

[rich kestrel]

yea the whole alpha male theory is a bunch of beer-chugging fratboy nonsense

but tell that to the roe jogan dude-wipes using population

[astral goblet]

don't forget hormone pill poppin dudes that are going to have broken prostates by the time they're 40. ||tate||

new nvidia driver today. remember when there was a quick minute where driver updates meant disaster or boosts for this pytorch stuff?

holy shit what a news day! theres even a new stardew valley patch! Who needs SD3?!

/copium

[forest turret]

why do all the SD models want to draw sexy females?

i want to draw some male characters, and even when i put negative prompt "female, woman, sexy" it still generates female characters

[stone latch]

I'm sick of seeing just females on every single page 🤣 When you go to Civitai, 99% of what you will see is women.

[forest turret]

it just defaults to drawing a lewd image of a girl/woman

[stone latch]

But about that, what model are you using and what's the prompt?

[forest turret]

look at this

oh i can't attach pictures here

go to general chat with images

i'll post there

[stone latch]

#🏞｜general-with-images

Yes

[low moon]

Anyone here into fooocus? any way to edit styles on the fly and refresh them without having to restart Fooooooocus all the time ?

[karmic cedar]

Has anyone bothered running Stardew Valley through a photo realism img2img controlnet because why not

[jovial wraith]

i vaguely remember seeing something like that before. i'm gonna try it myself tho, just as a challenge

[karmic cedar]

Could be really funny

[stone latch]

Oh no...

😂

[astral goblet]

ai artists have a type

[astral goblet]

yup i seen stardew sprites turned into real people

[karmic cedar]

I sEe Yo MaMa TuRn InTo A sPrItE

real talk, that’s cool tho.

one of my favorite activities lately has been img2img. i go through phases

[astral goblet]

https://new.reddit.com/r/StableDiffusion/comments/1anqmub/magnus_rasmodius/ was my favorite

[karmic cedar]

that’s where Jim Morrison would’ve landed had he not died so young

[astral goblet]

if anyone is a legitimate wizard, it's morrison

[pale latch]

nah. 1.5 gang will expunge about how sd3 is so censored and stay where they are, regardless of requirements

[charred mesa]

yeah this too

[grizzled void]

Where can I download SD3?

[jade mason]

When SD3?

[charred mesa]

no

👍

[grizzled void]

https://tenor.com/view/pc-man-crush-rage-gif-20192828

[charred mesa]

they gonna invite some more people this week to the secret discord server where they test SD3 using bots

[static cape]

I was really hoping for some news today...

[grizzled void]

Do you mean that it will not be open source?

[charred mesa]

bruh what are these conclusions

they will be downloadable and everything

code will be open source

and there will be an open release of the models

SD3 and SD3 turbo + controlnets

possibly in april

[karmic cedar]

the buzz has been more and more lately, and with models seemingly dropping at random I’m having a harder time gauging things

[charred mesa]

well there's SD3, SD3 Turbo

and from SD3 there are 3 confirmed model sizes (800M, 2B and 8B)

[karmic cedar]

the image model ecosystem approach is still pretty fresh, but i’m not at all complaining. it’s done when it’s done

[charred mesa]

if its actually in april I'd be glad

I wonder if it will be pushed to May

[karmic cedar]

me too—they have a nice rhythm

[charred mesa]

the model's quality looks good to me, like it has matured a lot since february

[karmic cedar]

4/19 is my birthday sooooo maybe 😮

[charred mesa]

thats epic

[karmic cedar]

fingers crossed

[charred mesa]

SD3 as birthday present lmao

[karmic cedar]

YOU HAS TEXT AND DOGGOS MADE OF BACONES

I’d have to try generating a certain Stability CEO riding a dinosaur equipped with an exoskeleton

Emad? His nationality is Bangladesh I believe

[trim nymph]

source? thats what i thought 3-4 weeks ago

[grizzled void]

it is him

[karmic cedar]

I’d never read his wikipedia before, doing that now 😮

[pearl ocean]

At this rate it will be out after GTA6

[charred mesa]

cant find the reply cause he has a couple hundred and twitter f*cking sucks and I cannot search from a user

[rich kestrel]

Want that shit to come out just to stop ppl from asking over and over "wHen sD3!??!"

I hope it's a bigger disappointment than 2.1 tbh. Will make it even more worthwhile

[charred mesa]

ok

[grizzled void]

https://tenor.com/view/who-me-huh-what-wait-gif-25560426

[fallen gale]

does anyone know how to simply convert sv3d videos to 3d mesh

without hassle

[astral goblet]

use a 3d model instead of sv3d for it?

[fallen gale]

but then what is the point of sv3d

[karmic cedar]

if a game studio really wanted to get nasty, they could get the rights to the DUNE II game from 1992 and redo all the gfx to Villeneuve’s aesthetic

but keep the pixel art style

[grizzled void]

meshroom ?!

[fallen gale]

thanks... will try

[grizzled void]

I don't understand what you mean, but search and you may find the answer

[fallen gale]

I will download meshroom and attempt it

[karmic cedar]

i wonder when we’ll see a Stable NeRF model of some sort

when/if*

[astral goblet]

EA has that i think

[astral goblet]

video maybe

[karmic cedar]

eh

I mean EA

i mean yay

…just give me another Shovel Knight game. I’m good with that.

or a proper follow-up to FTL, because I’m veeeery biased to that game

[rain flint]

In X/Y/Z Plot mode, can I have loras on one axis and the strangth on another axis?

So I get this:
LoraA:1 LoraB:1
LoraA:0.5 LoraB:0.5

[astral goblet]

silk song when?

oh thats hollow knight, mb

[karmic cedar]

i get those two confused a lot

would be interesting to see a hollow knight cameo in a shovel game tho….hmm

[astral goblet]

EA gets a lot of shiz for being EA. They're huge and have good parts too. I like Dice and Criterion. Who joined forces on 2042 which is actually great and i love it and haters gonna hate

[karmic cedar]

Nope, you’re right. But they have set a large target to be painted.

lol

[astral goblet]

free 2042 weekend this week

gonna have a lot of fun when newbs are ez pickens. Skill floor going to lower even more!!

[astral goblet]

I mean, yeah.. they've done it to themselves a lot of the time

Mass Effect Andromeda

[karmic cedar]

…by David Lynch, pretty much

lol

[mortal delta]

what can even be done with ai video besides memes, because you can create 3-5 seconds of video before thing usually turn out bad.
ive seen people create trailers or super short films but that must take hundreds of generations.

[trim nymph]

it isnt far away for one to create movies with this instrument-how long are scenes in movie productions nowadays (atleast in mainstream)-couple seconds. internet fried brains, couldnt hold out an tarkovsky/bela tarr movie anyways with scenes without a cut taking several minutes. character consistency & gens that are coherent like 20 seconds & everyone can build movies, theoretically for hollywood

[karmic cedar]

what we are seeing is a syntactic convergence of abstract and matrix.

the sky’s the limit; the exponents are resources and time of course

Look at what Sora’s doing—that alone should speak volumes.

And it does 🙂

[mortal delta]

I guess im trying to find a cool free ai tool that i mess around with that can help me create something but im not sure what to use or how to approach the field of ai. ive always wanted to make some sort of creative media but with ai i believe it might be helpful.

[karmic cedar]

Right now, folks are builders. People who get the most out of AI are the ones who approach their ideas as plans. If you can break the idea down into smaller pieces, you can start to think about the different tools you’ll need, etc.

Instead of a toolbox, though, it’s all code repositories and UIs. lol

The best way to start is to look at the functions that are most available right now, i.e. GPT agents, or text to image, image to video, or image to 3D

And so on 🙂

[mortal delta]

I do want to be a builder bu tnot a complex one, like i have to stare somewhere and work my way up, also your right about the whole digital toolbox thing.

[mortal delta]

i see the thing that annoys me is that most ai is behind a paywall or something.

[karmic cedar]

It’s always going to start with the idea, and you can never spend too much time thinking about it. It’s the only thing that’s truly yours—of course there are copyright infringements, trademarks, etc. but that’s another world. I’m talking about the beginning of the generation itself.

Well, these tools are resource-intensive. They cost a lot of carbon to run.

[mortal delta]

i like stable diffusion because it feels open in a way.

[mortal delta]

you have a point.

[karmic cedar]

SD is definitely geared towards openness, but it’s prone to the same forces of change as everything else.

[hidden drum]

that's how stable diffusion was originally demoed for testing, discord bots

[mortal delta]

your so right.

[karmic cedar]

There’s that whole thing about not knowing what you were missing until it’s gone, etc—right now there’s a lot of folks who are blinded by AI, and it could very well vanish in a heartbeat with enough regulation. That means we have to start being more actively engaged in what’s going on politically, etc. in order to stay in the realm of open development.

It sucks, but…it’s our future. 😦

Don’t mean to preach tho

[mortal delta]

there are so many things i want to make, but im debating between a short film and video game, also you are good.

[karmic cedar]

my advice: write your ideas down, like with an actual pen or pencil. Get a special book. Call it “The Future”. lol

sit with them awhile, even after you’ve found a good resource to test.

[mortal delta]

i should probably do that or find a way to organize my ideas.

[karmic cedar]

AI is inspiring me to do just that lately so i’ve been trying to, and it’s actually been paying off

[mortal delta]

thats good.

[karmic cedar]

If you treat image generations like they’re coming off of a roll of film, that finite, limiting nature can provide more creative potential

[astral drum]

so i was trying to run stable diffusion on my cpu locally and it all went well and then when i ran the prompt and when it finished loading this is what i got

it doesnt let me send a ss

[karmic cedar]

switch to the images channel

[astral drum]

ok

[karmic cedar]

general-with-images

[mortal delta]

also someday i hope to upgrade my hardware but electronics are not very cheap.

[teal pagoda]

Because 99% of the SD 1.5 models are designed by h0rny people. I got the same issue when prompting for something else. Imagine that I prompted for a cat or a dog and I still got a woman.

[trail lion]

1.5 was just easier to overfit honestly, it's not just the user culture, though that's certainly not untrue

people can certainly sit home and generate cat pictures if that's a bucket list item

[proper marsh]

Hi everyone, I am new to stability AI platform, just looking at the pricing page, is there a free sandbox for dev to use? or just 25 free credits and have to start paying after that?

[dusk canopy]

how fucked up is fucked uo 💀

whats the best model for prompt listening

[trail lion]

no idea what prompt listening is

like feeding it audio?

[dusk canopy]

i mean listenin to prompts

[trail lion]

prompt recognition, prompt adherence...

[teal pagoda]

As 1.5, I think that dreamshaper 8

[trail lion]

that would be SD3

[dusk canopy]

if they release it 💀

lets not get ahead of us

it doesnt exist until its open sourced

[trail lion]

ok, SDXL is the next best, by a wide margin, some argue 2.1 was good....but no content so who cares

[charred mesa]

2.1 was okay

better than 1.X in prompt coherence a little

[trail lion]

1.x isnt on the map

[teal pagoda]

Hope the SD 3 will just have better controlnets out of the box. So I have tested like 10 canny controlnets models for SDXL and only 2 worked decent out of the box. Now I'm testing like 20 depth controlnets models for SDXL and everyone seems to work just good. Why is this difference between them? :)))

[pseudo bough]

stable video 3d workflow? got one but two nodes are undefined missing node install doesnt work

[charred mesa]

yeah I hope they won't mess up controlnets

they are teasing it to be a big release and stuff

[teal pagoda]

I simply have all the controlnets available models for SDXL and I'm trying them to find the best working ones. At the end of the tests, maybe I'll release the list with the best ones.

[charred mesa]

that's be great

[teal pagoda]

From CVL-Heidelberg, diffusers, kohya-ss, stabilityai, SargeZT, TencentARC and so on

Testing them all

Finished with all the canny and found only 2 good with the default settings (weight 1, start step 0, end control step 1)

I'm at the depth right now and every single one worked decent from the ones I already tested

[charred mesa]

yeah depth is usually okay with sdxl

[bleak drift]

Do I have to have the SDLX base installed if I want to use a sdlx checkpoint?

Like am i required to have both?

[trail lion]

nope

[bleak drift]

Sick

So that saves space

[astral goblet]

as far as the ai animation space goes, i'm refiguring out things like touch designer, to masks and controlnet frames for animatediff stuff

[trail lion]

does instantid not work well with profile faces? kind of struggling with it

[astral goblet]

from the side? yeah it does better with front on. where it can see all the features

profile photos do good for creating other profil photos maybe though. and i think you can load multiple images into the instant-id model

[karmic cedar]

i keep calling instantid faceid

[astral goblet]

you wanna get sued?!

[karmic cedar]

truly no lol

but my poor brain

[astral goblet]

wait until you go look at the LLM scene and they use tools like ooogabooga

[karmic cedar]

yeah, that’s a kind of stirring the pot that i’ve settled on a ‘no’ for

I mean, if you have a distinct creative vision or concept you’re trying to go for that doesn’t infringe on X, Y, etc.

then go for it

[solar knoll]

it's been a long time since last time i used this thing
did it stop being free? or something else

[formal sphinx]

so I've been out of the loop for a bit. for general purpose image generation I was using SDXL, but now there's Cascade, Lightning, Turbo... does anyone have a simple resource or overview of what all these models are good for and how to use them?

[nova zodiac]

SDXL and Cascade are base model architectures like 1.5 and 2.1 were.

Lightning and Turbo refer to methods for editing models so that they run in 1-8 steps rather than 20-50

Lightning and Turbo only apply to sdxl

there is also LCM which is similar to lightning but also works for 1.5 models

[formal sphinx]

ah cool thanks. So Cascade is the newest base model? I'm hoping to get some better prompt interpretation. The visuals are less of a concern because I can always run it through a series of older models in a ComfyUI network for fine tuning

[nova zodiac]

Cascade is, but I wouldnt dig down that rabbit hole as there isn't many resources for it and sd3 is imminent

[sick mountain]

A ZIP-like one that gives Latent more compression, but also requires a decompression model. So it becomes a three-part structure

[formal sphinx]

gotcha

[sick mountain]

Can I ask about those online models which generate text exceptionally well, how do they do it?

compared to SDXL models

[astral goblet]

do you think any of the major ui's will have lavi-bridge support before i'm able to sharpen up my coding skils and do it myself? or should i just wait like pedro?

[astral goblet]

better datasets

[karmic cedar]

Just fire up StarCoder2 and like…manifest it, i dunno

[sick mountain]

ideogram and SD3 , their text , title, words are all accurately, that is really amazing. But as you say, it just better and larger datasets ? that "easy" ?

[verbal osprey]

I'm not too excited for SD3. I hope I'm proven wrong. I didn't like 2, it didn't adhere to the prompt like 1.5 did. You could type something simple like "red shorts" and the shorts would be a different color.

[sick mountain]

#🏞｜general-with-images message

[astral goblet]

that happens on 1.5 still to this day

xl too

[verbal osprey]

To a certain extent, but I found that 1.5 generally gives you what you want, more than 2 or xl.

[pearl ocean]

what a time to be alive!
https://www.youtube.com/watch?v=5U_Q2Lmnq_c&list=WL&index=4

[astral goblet]

is probably just confirmation bias

[verbal osprey]

Ehh I don't think so. Whenever I tried to use it, I never could control it as good as 1.5. I eventually just went back to 1.5. If 3 works better, I will be very happy. The visuals definitely look better in 2 and SDXL.

Could be my issue, that I'm prompting wrong...but I gave up after a while

[astral goblet]

i think you've decided already and will find the same evidence come 3 release.

[sick mountain]

because SD1.5 include NSFW database, which incress a lot of Generalizability and Singularity

[gaunt pulsar]

SD3 seems so much better that I think ppl will stop using other versions soon after it releases, but I could be wrong

[verbal osprey]

I hope so. Eventually, there shouldn't be a question as to which version is better. I want the newest version to be objectively better in every way.

[sick mountain]

for me SDXL seems more better in quality.

[verbal osprey]

I agree. It does look nice.

[opal hedge]

It'll still need some time for finetunes to catch up but I agree

[pearl ocean]

in the end of the day SD3 should be the best 🙂

[honest spear]

SD3 is superior, you can simply notice it looking at Thibaud's posts on X where people posted their sdxl images and he replied with the sd3 version