🏞｜general-with-images

[celest sigil]

I've used many different methods. Do you have any particular image(s) in mind? All of the ones from yesterday are using Cascade.

[languid pebble]

Good morning coffee. Sunny day here!

[celest sigil]

Guten Morgen! Es ist auch hier ein warmer und sonniger Tag, aber ich vermute, wir sind im selben Land 😉 😎

I don't speak German, just blame Deepl 😄

[languid pebble]

Moin! Da wirst du wohl Recht haben 🙂 Ohhh... you got me! 😛

One thing about online communities is that it doesn't really matter where you are from.

[grand finch]

После дождя в лесу. Солнце на закате. Сквозь листву видна радуга. Разрешение картинки 5600*5600

After the rain in the forest. The sun is setting. A rainbow is visible through the foliage. Image resolution 5600*5600

[nimble mason]

Here is the image you requested.

[languid pebble]

5600x5600 👀

[celest sigil]

Forcefield clothing

[languid pebble]

Ahoi!

[candid surge]

hoi

[nimble mason]

ahoy!

[wispy nest]

hello
is there any chance you could try this prompt for me?

a detailed photo of a star wars spaceship battle in space, multiple spaceships of various sizes and designs, iconic star wars ships like x-wings and tie fighters, laser beams and explosions light up the dark space, a large star destroyer in the background, debris and smaller ships scattered throughout, stars and distant galaxies visible, intense action with ships maneuvering and firing, dynamic and chaotic scene

[nimble mason]

sure, i'll toss it in for one of the next tests

[wispy nest]

its the one I use for testing the CADS stuff

it was written using the prompt from the Auraflow paper

[languid pebble]

SD3 on glif

[wispy nest]

wow thanks that's great

probably once it is open source we can fix up the ship anatomy a bit using PAG

PAG is amazing at fixing background ships

that double-snout on the star destroyer is a super common hallucination bTW

[languid pebble]

I wonder why there are these cuts on top and bottom

[wispy nest]

yeah SD3 medium does that too sometimes

[wispy nest]

that's really nice

I like it when they do the dark versions

ZavyChroma gives you the dark version about 10% of the time or so

[languid pebble]

Could be a prompt for a test like this ... I think there's a workflow for that on civitai ... https://docs.google.com/spreadsheets/d/1vViQh45B-txERc5wnDVh_aNfHWLuEIrOCz7_HGXpPZM/edit?gid=0#gid=0

[wispy nest]

everyone forgets about Lumina

Lumina has nice aesthetics though

Hunyuan DiT is currently just a level below the others as far as I know

[languid pebble]

I have worked with Lumina a lot, too ...

[past pelican]

Smaller face seems happy...

[wispy nest]

thanks this is a really cool look

[languid pebble]

NVIDIA is reducing the production of RTX 40 chips ... prices could rise ...

[royal monolith]

are they getting ready to start on the next version then?

[languid pebble]

sounds like they wanna empty the warehouses ...

Maybe they wanna sell 40ers more expensive first? And after that start with 50ers? I can't say ...

I think they wanna do the x-mas business ... so 50ers should be out for it ...

[covert nimbus]

Hi

[dense nymph]

anyone doing architectural rendering enhancements with artificial intelligence?

[wispy nest]

no but we can help you with it anyway

stable diffusion works the same no matter what the subject is

[clever oar]

right version

[past pelican]

Giga chin shark

[wispy nest]

nice colours

[royal monolith]

used the /style command in the artisan channel, really like how it works

[worthy birch]

What style is this? Looks great

A bit Burton ish

[toxic hollow]

cute cat

[royal monolith]

i uploaded this image for the style refrence and the prompt was "still life with roses and guitar'

[ember gull]

cute dog

[grave prairie]

how will generrate image?

[languid pebble]

https://discord.com/channels/1002292111942635562/1237461679286128730

[celest sigil]

🤣

147 days to go...

[turbid minnow]

Pon algo tio

[wild sorrel]

That happens when people buying less, which happens to every new gpu after some time.

[languid pebble]

Sure ... let's hope for better products coming 🙂

[languid pebble]

Good morning coffee ...

[polar wave]

Generate maps of india

[unique condor]

saw this this morning and remembered talking about it the other day: https://ai.meta.com/blog/segment-anything-2/ this is probably what you're looking for. it's the new SAM 2.0. Kijai is working on a node for it for comfyui https://github.com/kijai/ComfyUI-segment-anything-2 (he makes a lot of great wrapper tools like these) and it likely won't take him long to have it running with all the bells and whistles.

[wispy nest]

are the benefits mostly for video

[unique condor]

Yeah but it can track multiple things at once, which is what they were looking for.

They were wanting to track multiple players in a game iirc

[wispy nest]

ah okay yea

its a great technology

just not for image, which is fine

[unique condor]

It does everything that SAM does, just better now. SAM has been widely used with diffusion for a long time now for a lot of workflows

This will just help even more for those that also want to do video stuff as well now

A lot of "detailer" type add-ons run SAM under the hood

[wispy nest]

yeah what I mean is that the abilities of SAM for image are probably maxed out at this point

but video is way more challenging

so there will be a lot more improvement possible

[unique condor]

Ahh I get you

[wispy nest]

its really annoying to do but
the main thing with detailers is fine tuning Yolo

then it gets really good accuracy

its not bad out the box though

dining table is wrong, TV is wrong, cell phone and hair dryer are wrong

but its rly impressive still

[gray canyon]

its *general with images not generate image's jaa kar auto 1111 download kar bhaii

[royal monolith]

read the information here #artisan-faq

[wispy nest]

https://tenor.com/view/home-homelander-the-boys-milk-gif-9228418437453891441

[sterile kiln]

Oh, SD3 is finally working on A1111

[languid pebble]

yay 🙂

[sterile kiln]

working really bad on A1111 or its me?

[languid pebble]

Stopped working with it ...

glif worked much better

[sterile kiln]

glif?

[languid pebble]

Free online generation website

https://glif.app/@Oliveira/glifs/clw44qfbl0000m0zztwqk2tnf for example

[ebon kestrel]

Capture Lily's focused expression as she delicately decorates a batch of cupcakes behind the counter. Show the flour-dusted apron and the array of colorful frosting and sprinkles she's working with.

[royal monolith]

read the information here #artisan-faq

[hearty violet]

what is the better way to me remove only the tree and the traffic lights on this image keeping the façade architecture?

[royal monolith]

photoshop, gen fill and carful work

[hearty violet]

ok, I will try with inpaint on sd too

[gentle blaze]

Or Photopea

The smoke is because they were burning the tree, traffic lights, the taxi... in the back of the building

[hearty violet]

hahaha sou good! inpaint? empty prompt and controlnet inpaint only+lama?

[gentle blaze]

photopea, select and remove....

[hearty violet]

remove is a tool?

oh, I see

ok thanks!

[hearty violet]

my photopea dont has this option, is the same than fill?

[hearty violet]

finally I got it

[wispy nest]

nice

[alpine pelican]

When your prompt isnt followed at all, but it has perfect hands and eyes
https://pbs.twimg.com/media/DT5p8ZEXkAAUK6k.jpg

[languid pebble]

Interplanetar Good Morning Coffee transporter 😂

[shut sinew]

Cell phone

[wispy nest]

LOL

the funnier thing is it thought the stormtrooper was a hair dryer

[shut sinew]

I mean tbf, they probably need one after wearing that bucket for a whole day

[orchid turret]

画一个屈家岭考古遗址内部照片，要求遗址内部的文物要能够体现出立春节气

Generate maps of india

[steep sleet]

SD3@ClipDrop

[raven sluice]

So, flux

[real lintel]

Flow actually looks really good. Even if SD3 fixed the anatomy i still dont think it would be as good.

[dense nova]

First try, 1/1 sample. It understands anatomy and physics. We might have AGI folks.

Prompt was: a hand holding a pair of scissors, casting a shadow on a wall

I added the boxes myself. This is the first new model in ages with full anatomy understanding. 🙂 (And the male anatomy is perfect, which I have never seen from any model ever.)

[fleet aurora]

if i upload a image can someone tell me what is the model or style?

[dense nova]

It's anime style.

[fleet aurora]

i know

i want the name

[languid pebble]

Wich model?

[fleet aurora]

@pallid ruin

[fleet goblet]

(Flux.1-dev)

[languid pebble]

5090? 😄

[fleet goblet]

nah, arc a770 16gb

they just updated it and now you can run it in fp8 via 12gb of vram

in comfy

[languid pebble]

I more need 32GB 😄

[fleet goblet]

oh you mean the image

LMAO

[languid pebble]

Yes 😄

[fleet goblet]

I was thinking you meant what it was generated with

yeah probs a 5090

lmao

[languid pebble]

We have been talking about 5090 here some time ago ...

[fleet goblet]

You think itt'l be another underwhelming card?

I feel like arc battlemage might be competitive

[languid pebble]

I don't really know that one ...

[fleet goblet]

Battlemage is the card that comes after the current generation of intel gpus were on

which is Alchemist

I'm referring towards gpu architecture btw

[languid pebble]

But A.I. is written for NVIDIA ....

[fleet goblet]

Except that isn't true.

AI is written for torch.

Torch supports AMD quite well now

[languid pebble]

Does Intel support torch?

[fleet goblet]

intel isn't too bad either anymore with the devs.

yes, via IPEX (Intel Extension for Pytorch)

After all, that's how I'm generating these images. On my Arc.

And they're simply beautiful.

[languid pebble]

Cool ... I'm a bit limited in my information cause I run NVIDIA ... but I often read about AMD Users looking for workarounds ...

[fleet goblet]

Well, you have the advantage.

You have the most commonly used brand of GPU.

Literally most coding things will center around either you or CPU

But compatability is getting way better

[languid pebble]

But first of all I am happy that it works for you, too!

[fleet goblet]

For example

I cannot use xformers

that doesn't exist on arc

something commonly used on nvidia to speed up inference time

[languid pebble]

I am waiting for special A.I. processors ...

To use GPUs can't be the best solution ...

[pallid ruin]

x2, yesterday I read that there was some type of asic or npu that could ran llama 3.1 405b, the same for smaller models (including images one) would be epic

[fleet goblet]

You mean Latent Processing Units, either that or NPUs.

LPU/NPU for AI acceleration

ASICs designed for the transformers architecture will do too

[languid pebble]

At the moment they sell everything as A.I. computer cause it can run Microsoft A.I. ....

[fleet goblet]

Which everything can do, because that is cloud-based.

Unless you mean their NPU laptops

Which....eh.
🤷‍♂️

Not really worth it compared to a dedicated setup

[languid pebble]

It's just stupid marketing ....

[royal monolith]

that's what they were created for, why isn't it the correct solution? you want the CPU to run your machine and the os, you want the GPU to process the, you know, graphics applications

[royal monolith]

not. you need to go do some research

[languid pebble]

I don't think they were created for A.I. They are just the best available solution at the moment.

[pallid ruin]

The GPUs were supposed to process graphics only, they are used in AI because its cheaper

[royal monolith]

GPU = Graphics Processing Unit - created to process graphics. however, because AI is written with python, and cuda only runs on Nvidia GPUS they are used for AI applications as well. to create a CPU that does that isn't an overnight deal, and you still have to split the processing up between the OS and the hardware - the machine - and any application processing you want done. that's inefficent. you want the cpu to run your machine, not your applications

what they are working on isn't new CPUS, they are working on AI chips

[fleet goblet]

Chips designed to run the very architecture they are designed upon

[fleet goblet]

faster than any cpu or gpu

[wispy nest]

no explanation

[languid pebble]

That's what I meant. GPUs can calculate A.I. better than CPUs ... but there will be a better solution ...

[royal monolith]

yup. and specific for running AI - those are being developed

[wispy nest]

btw the girl is his sister

gyat

[fleet goblet]

a hf transformers asic would be bonkers

i mean look at groq's lpu

[royal monolith]

I'm sure someone's got that on the drawing board

[fleet goblet]

THAT is basically a pci asic

[wispy nest]

bonkers donkers

ahh

[royal monolith]

there are 2 hurdles - designing it - and getting customers to use it

[fleet goblet]

well groqs lpu is 20,000 dollars lmao

[royal monolith]

not surprised.

[fleet goblet]

not even including the fact that you need hundreds of them

[languid pebble]

Public is working with A.I. for 2 years now ... so should be the time to come out with a product ...

[wispy nest]

man these mfs in general w images are so smart

[royal monolith]

that is not how hardware design works. that's not even how software design works

[fleet goblet]

https://tenor.com/view/think-about-it-use-your-brain-use-the-brain-think-brain-gif-7914082

[wispy nest]

https://tenor.com/view/blue-furry-im-not-like-most-girls-i-have-a-cock-gif-17727071

[wispy nest]

B)

[fleet goblet]

you are brainrot

brainrot gets banned here chief sorry

[wispy nest]

b-BRAINTOR

brainrot gets banned from life here chief sorry

[fleet goblet]

then why are you brainrot

[pallid ruin]

💀

[languid pebble]

Pride week here ...

[crisp stream]

which isn´t entirely true because pre AI they were already used for bucket rendering for example and it´s the way how GPU handle things like certain parallel processing they are very good at. You cannot really use it for linear stuff like audio processing in a DAW very well. Not a technical expert, yet at least that´s how it presents itself from my understanding

[crisp stream]

like said, where using the GPU for an application is working well it´s fine, yet there are areas where GPU simply aren´t the best choice, that´s where the CPU comes in.

[languid pebble]

And maybe there should come a better solution ...

[royal monolith]

not the best, but right now, sort of the only. they are working on new chips however

[unique condor]

The big reason why GPUs are so good for AI is because they have thousands of simple cores that excel at doing very specific types of matrix math in parallel. Operations like fused multiply–add. AI models are just big layered sets of spreadsheet values(weights) and they interact with basic math operations. So when you've got thousands of CUDA cores doing those operations in parallel, tokens per second or iterations per second go BRRRRRRRRRRRRRRRRRRRRRRRRRRRR

That same type of hardware is mostly useless though for regular CPU-like tasks though

[languid pebble]

I ony said I'm waiting for a processor (maybe as an addon) just to do thant kinda work.

[unique condor]

they already have them, a lot of cellphones and devices are starting to have a "neural coprocessor" built into them

which are basically just a big vat of "cuda" cores

[languid pebble]

My old mobile has been advertised with that ... it's bullshit from our point of view ...

[unique condor]

phones have to have at the minimum, a 24h battery life. you can't expect a 10w device to diffuse 2048x2048 images without draining your battery

or taking forever to do so

as the tech gets better and better though, you'll start to see more efficient hardware that can pull off more and more for the same wattage cost

[languid pebble]

I thinjk they have added kinda special CPU to do photos better needing less energy ... that's not really A.I. for us who wanna create pictures .. but for marketing ... it might be ...

[unique condor]

but if you want true efficiency, our human brains do everything they do with roughly a 15w average throughout the day

[fleet goblet]

That's why people are fascinated by the idea of NPUs.

Low-power, high throughput.

[unique condor]

yep, and we're working on them

[fleet goblet]

I'm still of course more interested in full power and not efficiency

[crisp stream]

wdym exactly? Except for the actual real-time graphics display or mentioned bucket rendering the rest is basically done by CPU only in 3dsmax for example. And there is lots of software that heavily relies on the CPU

[languid pebble]

We are waiting for them 🙂

[fleet goblet]

He has come

And he has a big guard

[unique condor]

almost all of the 3d apps have GPU rendering modes where they use next to no CPU after prep

but you get limited by what you can fit in vram

[fleet goblet]

Blender's embree is quite nice for rendering via non-CPU methods.

Especially for really fast, real-time RTX.

That's how I did my profile, actually.

[royal monolith]

you really need to start reading twitter and following the scientists insted of assumeing. people are working on desiging chips specifically for AI applications

[fleet goblet]

People don't go to twitter for research info.

[crisp stream]

Like said, rendering and RT-Graphics yes, then there is a whole lot of other stuff to be calculated in 3D-software that falls into the realm of the CPU

[royal monolith]

the entire scientific community, all the very well known and highly respected scientists, are ON TWITTER! and you'd know that if you bothered to do much research

[fleet goblet]

Yeah? You know what I use to find new stuff instead?

Huggingface

[crisp stream]

who´s talking about AI chips? I´m talking about the current use of CPU in software

[royal monolith]

and guess where all of huggingface staff, and Clem, post all the news first? Twitter

[fleet goblet]

News vs actual repositories

I'll take repositories

and not consistent reminders of trailers for models that aren't even out yet

[royal monolith]

https://x.com/huggingface

Hugging Face (@huggingface) on X

The AI community building the future. https://t.co/VkRPD0VKaZ
#BlackLivesMatter #stopasianhate

Twitter

https://x.com/ClementDelangue

clem 🤗 (@ClementDelangue) on X

Co-founder & CEO @HuggingFace 🤗, the open and collaborative platform for AI builders

Twitter

[fleet goblet]

I

am not going to do that crystal

[royal monolith]

you want to know what they are doing, read twitter, follow them

or get your news late, after everyone else has it

[fleet goblet]

Which doesn't matter btw

Because it's news.

Unless this news is urgent sure

[royal monolith]

bunch of wannabes

[crisp stream]

funky twitter anyway, then I still don´t see how it relates to what I mentioned exactly?

[unique condor]

yeah and i'm telling you that almost all 3d apps have a gpu compute rendering mode that will support usually around 95% of all the same features as cpu rendering does. after the prepwork is done, the rendering is almost purely done on gpu

[fleet goblet]

tells you that getting news fast doesn't matter
wannabe

👍

[unique condor]

here i'll render a demo scene in blender real quick to show you

[royal monolith]

the AI chips i mentioned, which you'd never heard of, have been being talked about, and information posted about, for quite a while. but i no longer care if you know

[fleet goblet]

gets his research from twitter and not actual papers/repositories

[crisp stream]

who´s interested in AI chips in relation to what I mentioned and whether I heard about anything I´m not even talking about?

[fleet goblet]

gets mad when people don't use twitter

Then he proceeds to insult you on your information

despite the fact that it isn't even at all what you were on about

[crisp stream]

that´s what I´m wondering as well 😄 Possibly he´s mixing me up with somebody?

[fleet goblet]

@royal monolith People like you are the reason why there is such knowledge dysphoria between people in this discord.
You can't keep up a normal conversation with a person without getting mad.

What's really sad is this isn't the first time I've argued with you either...

This is like the third/fourth time in the years I've been here.

[unique condor]

see? the only reason it's really even using the cpu amount it is likely has to do with the fact im using sysmem since i cant fit it all in the 8gb vram

[fleet goblet]

aka mem offload

[unique condor]

yep

[fleet goblet]

trading inference time for convenience

[unique condor]

i could tweak the settings to bring down the vram usage

and it would use like 1% cpu

[fleet goblet]

I'm also using FLUX 0.1 btw

It's quite good.

[crisp stream]

like said, once again, not talking about rendering. Rendering naturally can be done very well on a GPU, there is simply a lot of stuff when working with the software (not rendering) that is done by the CPU for given reasons you mentioned more technically detailed+

[fleet goblet]

As we advance, I doubt GPU's will be a seperate thing forever. It would be more effective to incorporate it all into one thing for maximum performance.

We're already limited by the CPU to GPU's distance, REBAR moment.

[crisp stream]

Are you referring to the actual distance for communication/tranfer?

[fleet goblet]

Yes, I am.

[unique condor]

yeah sure, there's that too, but the shit in the viewport is still usually handled with the GPU for display purposes. the only well known exception is zbrush, which is entirely CPU, but they do some weird stuff to pull it off. it's actually pretty neat what they do under the hood. but there are plenty of non-render related things that are done on the gpu within a lot of 3d apps as well. like particle simulations, batch instancing, and so on. you're always going to need a CPU for some things, nobody is claiming you aren't. and no, cpu isn't always more technically detailed, it just depends on what kinds of features you're using. you don't always need 100% of the features if it's something you're not using

[crisp stream]

With "more technically detailed" I was referring to your earlier description of what sort of processing is being done well by a GPU and what not so much. Like said if you go into audioprocessing it´s the CPU doing the work, as well for rendering because it´s linear and not parallel (like what a GPU is good for).

[unique condor]

yeah but technically, even audio could use some gpu multithreading. but it's kind of pointless because at worst, you've got like what, 20 tracks with a few VSTs stacked on each of them and some filters at the end? most cpus can handle that pretty well since it's not something you need 2000 cores for

in realtime

(my family has producers in it and i've spent a lot of time in studios)

[clever oar]

🖐️

[crisp stream]

yes, that's what I ment, where stuff can be done in a parallel manner, many cores in parallel are preferable while with linear processing you would have to wait till one thing is calculated before you can process stuff further, so you would have merely "one" cuda core calculating at a time (for a possibly exaggerated example without me knowing technical details).

[crisp stream]

I once suggested/asked in a DAW-forum about GPU usage for audio processing and was told by the DSP-people it wouldnt really work the way I imagined for reasons you mentioned as well.

And while its been a while ago, I haven't seen any GPU audio processing in any of the DAW I know so far

[languid pebble]

Keep aware from AceMagic .... https://suno.com/song/2f746e8d-e515-4bcb-a0d9-7e601db54d99

[clever oar]

dont fool your Mojo Yates 😃

it’s clear that this is not just a song, but some kind of life situation for you

[unique condor]

yeah threading/task scheduling would be a bitch basically, there'd be so much stall time

[crisp stream]

😀

[floral flicker]

Illyasviel seems to be truly back in the Forge saddle. 6 hours to approve - I tried pushing this months ago and canceled it after > 1 month of silence

[robust path]

❤️ flux

It can do 4 panels

in various styles

[fleet goblet]

I am ready to woop

[past pelican]

How's flux? It's comfy only right now or does A111 have support yet?

[wispy nest]

flux is by a very long way the open source SOTA

really amazing step up

with the exception possibly of Cascade/Ultrapixel for high res

due to the latent compression

[nimble mason]

yep that's the only thing that's not kinda wiped out by this

already scheming on how to leverage them together

[eager wasp]

and style 😬

[nimble mason]

yeah

the style with cascade is just incredible

[eager wasp]

sdxl is good too, cascade seems better still, but all these new models seem to have made style a 2nd class citizen

[wispy nest]

for some stuff I also think Kolors is still in the running for aesthetics

but Kolors just baked that style in LOL

[eager wasp]

maybe 8b is still better in that regard than flux, but i used 8b a LOT so know it a bit better

[wispy nest]

SDXL base has really poor style
Leosam Hello World has really strong style, a year later

so there is time

[eager wasp]

yeaj sdxl has help from a lot of finetunes 😂

[hearty violet]

this is new?

[clever oar]

hazard

back yor icon

[hearty violet]

kkkkkk

I'm tryng remove the tree

the control map is withou it (I'm using inapint only+lama)

but the final image the tree is there

why? any suggestion?

[royal monolith]

what you need is to use object removal

[lost holly]

how to use it?

[wispy nest]

FreeU is so weird
if you make a picture of a bar
under certain FreeU settings the bar will be dirty and the bottles chaotic
and then under other FreeU settings the bar will be clean and the bottles nicely organised

[south temple]

wellllllll hi friends

[wispy nest]

hi

[south temple]

the updates.... hoooooo dang

[celest sigil]

Flux is looking good!

[south temple]

im back on a video trip

[south temple]

i dont think anything new exists tho

noting new with svd

[celest sigil]

The prompt had the word cardboard in it, and you can see the edges of the front are cardboard

[south temple]

now that

that I dig

cat bear owls of doom

lets go

sunflower head demon cheshire cat smile

[nimble mason]

lol

[south temple]

ice cream and 3 kittens

[celest sigil]

a storm in a teacup

Colour me impressed!!
a glass ball half full of red liquid is to the left of a glass square that contains blue liquid. To the right of them is a green glass bottle

Not a good hand, but maybe nobody will notice 😉

Every time! 🙂

[fleet goblet]

Making really small resolution gens is kind've funny in my opinion

oh no!

m-m-mario, where are we?

[languid pebble]

Take a picture of your Good Morning Coffee ... 🙂

[wispy nest]

does anyone have the longer Align Your Steps sigmas?

its called AYS 30 or something?

[steep sleet]

W/Flow by ClownSharkBW

[shut sinew]

Flux

[languid pebble]

a glass ball half full of red liquid is to the left of a glass square that contains blue liquid. To the right of them is a green glass bottle (Model: AuraFlow0.2)

a glass ball half full of red liquid is to the left of a glass square that contains blue liquid. To the right of them is a green glass bottle (Model: Cascade Ultrapixel)

[hearty violet]

and wha is this? only+lama is to remove

[lost holly]

how to use it?

[uneven elbow]

Who can help me

Pls make a wallpaper

Of zaraki kinpatchi wallpaper 4k high détail

Dark ness mod

Black and red

Pls pls pls

[hearty violet]

I can use cloud points architecture in any way with stable diffusion? somehting interestind workflow, extension, way to transform the cloud point on model or good image?

[fierce ravine]

Homem aranha

Spider men

[fierce ravine]

Minecraft

[shut sinew]

Flux

[lone jay]

Hi all

What extension is that in stable diffusion

Looks really cool and wanna try it out

[candid surge]

Flux-schnell online

[scenic yew]

flux shcnell with 6 steps
prompt: an image of a woman laying on grass

[candid surge]

flux-schnell - woman laying on grass

but might be a bit cheating, not showing the full body

[native sedge]

A touching moment of रामू and मोहन sitting together in the fields, laughing and sharing stories.

[wispy nest]

remember to pass the test
they have to do it without overfitting
which means it doesn't look the same every time

[sick sierra]

Can you get non-detailed styles with flux? 🤔

Whenever I try something less detailed (impressionism, expressionism, idk) it still adds the same amount of details.

[native sedge]

#🏞｜general-with-images A touching moment of रामू and मोहन sitting together in the fields, laughing and sharing stories.

Soft, watercolor style depiction of Itachi, peaceful and serene expression, gentle colors, ethereal background.Itachi Uchiha, watercolor, soft, ethereal, peaceful:

[candid surge]

a lady on grass becoming one with the world in a serene and calm scene, anime style

that was just the title and caption. full prompt used:

modern anime style, 28yo woman laying on the grass, side view, white outfit, fractal style clouds, realistic dark lighting, shadows, satisfaction, satisfied

now that 1st caption used as prompt:

"Swim, damnit, just swim!!!"

prompt: "being throw into the deep water to learn to swim, hard learning style, confrontational method, anime style, dark realistic lighting, action scene"

[dire sun]

Atomic orbitals made of water

Prompt: atomic orbital made of water

[candid surge]

"clip missing" ... is this bad ?!

[rugged hound]

no, it's normal

[candid surge]

speed gone up now to about 37 s/it ...

I guess its mostly running on cpu instead of gpu, feels like the old days before a gpu

[pallid ruin]

Remember using --lowvram @candid surge

[candid surge]

yes i am

does it do some memory management in a smart way, like for example looking at which parts of the bigger model are used, and only load those smaller parts into vram ... ?

[candid surge]

here my very first local FLUX gen, it only took about 13 minutes or so

I wish there was some smaller low-res preview model, low-mem, high speed, low quality. But with snappy prompts, good anatomy, various styles, all the good things in backgrounds and scenes. Then if nice result is found, I can ramp it up in 13 minutes with FLUX

basically a FLUX-preview model, low-res, high speed

against good practice and all, ill try some 512x512 pixel stuff using flux. see how it handles low-res.

[languid pebble]

1:40 minutes here

[candid surge]

lol, about 2 minutes is what I get for 512x512

does it run on your gpu hardware?

[languid pebble]

Flux ...

[languid pebble]

Yes ... 4090 but only 16GB VRAM cause notebook edition

[candid surge]

512px is not TOO bad ...

and what cpu and ram?

AMD Ryzen 5 1500X Quad-Core, with 16GB normal RAM for me

[rugged hound]

I get around 1:15 for a 1024x512 image on a 3080 10GB with 15 Steps

A Classic image works pretty well

[languid pebble]

13th Gen Intel(R) Core(TM) i9-13980HX 2.20 GHz, 32,0 GB (31,6 GB verwendbar)

I'm running on EnergySafe mode 😄

[candid surge]

why? laptops need to fly!

[languid pebble]

It's fast enough for me 🙂

Need to type prompts 😄

[clever oar]

there my flux generation

[clever oar]

I decided to arrange a battle between Flux and SD3 Flux is more beautiful and realistic but SD 3 has brighter and more interesting colors

[languid pebble]

SD3 Flux? I only know this one: https://huggingface.co/black-forest-labs/FLUX.1-dev

[clever oar]

no i use sd3 and flux its different

from your link

i use same promt

and compared the results

[languid pebble]

Ahhh.... got it 🙂

[clever oar]

battle you know)

[robust path]

cheeky burger

[languid pebble]

lol

[robust path]

school project

Sharkman

[languid pebble]

lol

[onyx isle]

13 mins wtf? How much gpu etc do you have?

I was just going to install it, but think I'll pass 😦

[languid pebble]

Just try ...

[jovial tiger]

it's probably because they ran out of vram. firstly, update your comfy. secondly, try running the model in fp8. it'll drastically reduce vram usage.

[onyx isle]

I only have 8gb vram is why I'm hesitant

[languid pebble]

I was able to install it, so it can't be to hard 😄

[candid surge]

Hi Becky93 - do I remember you from civitai? anyway, I have 12GB vram and 16GB ram, and still it uses harddrive for swap memory. means it gets very slow with flux.

[onyx isle]

It's the running it part I winder about! Though I see it actually does nsfw, so I'll give it a try while I'm doing chores or something. I'm not patient enough to wait over 20 miss for a single image lol

[onyx isle]

Rom civitai? Really innocent look 😉
I'm jealous of your vram lol

[unique condor]

There are also some pre quantized safetensors on kijai's hugging face page (he makes a ton of wrappers and add-ons for comfyui) if you plan to only use fp8. Saves a ton of space and loading time

[languid pebble]

OK, I do understand you. Let's see how long it takes you to create a picture ...

But the model is pretty good. Non of my pictures has been cherrypicked ...

[candid surge]

yes, good, but for many not really usable locally

I would really love to see a mini-flux or something, quick low res preview flux

[wispy nest]

the image quality is very good

[candid surge]

sd3 btw

[wispy nest]

flux dev

[candid surge]

modern anime style, scientist woman, white research coat, detailed research lab, glass tubes, shiny, intricate, highly detailed, mysterious

(flux-schnell, online)

[wispy nest]

five fingers

[candid surge]

hey, its not perfect. It still has the same holes, but just less deep. and not everywhere.

wait, I mean the model's interior landscape of good images, to be found with the sampler.

[candid surge]

or well, the virtual neural pathways through the model to result into the images, rather

[languid pebble]

Looks like it can't do Text ...

[candid surge]

sd3 does text ok, but it seems to have an issue with pony legs. there are FOUR pony legs, not 3 and 1 different leg!

[clever oar]

i think flux do hands better

[languid pebble]

Flux

[clever oar]

ok

not see mutations

[languid pebble]

Yes ... damn good

[clever oar]

its strange why in new sd cant fix this problemn

[candid surge]

ok, now the sd3 model is playing with me. perfect text, bad image.

sd3 hand

1st lucky shot

ill also share the next 2

[clever oar]

you win lottery

[candid surge]

I guess I did ...

[robust path]

😂

[languid pebble]

Finally with Flux:

[jovial tiger]

your prompt with dev. I started using pony recently to reliably get that modern anime styling. flux is pretty good but not 100% on that style

[past pelican]

I've got a weekend game going on in #🔆｜dailies if anyone here wants to participate 🙂

[wispy nest]

i mean

[robust path]

Expectation

Reality

[languid pebble]

Hahaha!

[candid surge]

modern anime style, ceo woman, beige business outfit, golden expensive necklace, tight up hair, intricate hairdo, strict look, red lips, powerful lady, dark wooden interior, huge boss lady office, dark realistic lighting

flux-schnell, online

[wispy nest]

man i cant come up with an interesting motive

[onyx isle]

I downloaded the model, what next? Does it take its own specific comfy workflows or something? 😦

[languid pebble]

Yes you will need some stuff ... one second ...

[languid pebble]

https://comfyanonymous.github.io/ComfyUI_examples/flux/

For you maybe use the smaller t5xxl_fp8_e4m3fn.safetensors instead

[onyx isle]

Thank you so very much 🙂

[wispy nest]

hehe

okay the model is cool

[onyx isle]

Weights, what? No other model needed those... this better be good 😄

[royal monolith]

those ARE the model

be aware it's massive

[onyx isle]

Only 22gb that's ok, but... I do worry for my poor GPU lolol

[royal monolith]

that's the file size....

it's - 12 billion - parameters

[royal monolith]

sd 1.5 is 2.5 billion

how much vram you got?

[candid surge]

86 billion neurons. human brain.

[royal monolith]

neruons aren't parameters - we probably store trillions

[candid surge]

i know

just for comparison

neuron are much more than simple weights

[onyx isle]

only 8gb grumble, will never forgive bestbuy, grumble lol

[wispy nest]

okay i underestimated how good it is

[languid pebble]

A Penny for your prompts 😄

[wispy nest]

huh?

[languid pebble]

Just kidding ... they are pretty cool!

[royal monolith]

oh you're gonna be rendering one image for an hour at least

[languid pebble]

Thanks for my computer ...

[onyx isle]

You should definitely create a lora (or whatever they are with flux)

[wispy nest]

about what

its just ophanim

[onyx isle]

Amazing sets of eyes in various scenerios 😄

[royal monolith]

and call it EyesHaveIt

[onyx isle]

Flux via glif since my computer is vram deprived

Perhaps flux needs it's own vae instead of just plugging in sd3 vae? 913 seconds, and I got this 😦

[languid pebble]

It's all in the link I've send you ....

[onyx isle]

Ty 😄

[onyx isle]

Flux: 626 seconds
To bad it always does cartoons. I forgot to prompt photograph.
Btw it won't do NSFW on my own computer.

How that is better than SD3 is beyond me, SD3 is way more awesome!

[languid pebble]

I only tried one NSFW prompt and the result was

SFW for many countries ...

I hope I'm not beeing banned for that picture ... it's art and no porn!

[onyx isle]

But can he lay on grass? Kidding 😄

[gleaming marlin]

can anyone help me out and list the top models for photorealism

[fleet goblet]

iced hedgehog

garden gnome hedgehog

[astral snow]

Is there a chanell to ask for help? I installed StableDiffussion 3 but when I drop or load the test image I receive this

[wispy nest]

that doesnt look like sd3 workflow

you didnt drop the right workflow apparently

i can't help you much cause i dispise sd3

sd3 is bugged out deformed junk

i suggest you look into https://comfyanonymous.github.io/ComfyUI_examples/flux/

[astral snow]

okays ty

yes I think it is bugged it ate up all my ram to generate an image

[wispy nest]

its not that

sd3 can't render human anatomy properly and most of the time it will render abominations

this is flux

[astral snow]

ok thanks for the information I will look

[astral snow]

I will use it to mainly generate monster ideas for my game

[wispy nest]

try kolors

kolors is not as resource hungry as flux but far better than sd3

i should add there are also many sdxl models that can do great job

[astral snow]

okay I'm going to find out all that

[wispy nest]

txt2img landscape has expanded greatly in the last several months.. its not just about sd1.5 and sdxl anymore, if you can invest on good rig there is lot to explore with the options available now

[astral snow]

What do you recommend for Stylized monsters? if you have an idea. otherwise websites?

[wispy nest]

do your homework from the ton of examples at civitai and models shared there

[astral snow]

okays ty

[clever oar]

wow flux know game logos

[wispy nest]

look at the fingers 🙂

as they say flux the new king of AI

[clever oar]

why only now ai can do fingers

[wispy nest]

lol idk, but i love how its working out

[onyx isle]

We need a Flux section.
Does flux count as SD/SAI?

Obviously needs work but pretty good for coherent text

[astral snow]

how can I get a result like that on Stable? (Dall-e image)

with dreamShaper as a model I come across this as a result

[onyx isle]

Flux does even better skateboards than sd3!

[onyx isle]

Dreams hyper isn't the best model imo, try some others. Relevant loras also.

[astral snow]

how do you know if it's a good model or not? and do you have any names to recommend to me?

[onyx isle]

Trial and error.
Just use ones with similar artwork that you are looking for.

Try some WoW ones or? I don't know anything about that genre.

Or just try SD3, it's goid at a lot of things!

[astral snow]

ok thank you I will do some tests. I look on civitai (it is down atm)

[royal monolith]

here. this is my SD3 workflow. people that actualy know what they are doing get very good results with SD3

people that don't know what they are doing talk smack about SD3

[astral snow]

Ty

[onyx isle]

That particular Workflow makes everything awesome with sd3!

[raven mirage]

Civit is back up!

[royal monolith]

welcome. for this workflow, you've got access to each encoder seperately. give each one a different prompt. give clip_l the fine details, background, artsy, ambient stuff. give clip_g the straight up, black and white, concise description of the scene, and give t5xxl the longer, more narrative prompt and the text

[pallid ruin]

DMD2 is still goated, in some kind of images its still the best (or one of the best)

[onyx isle]

How many free images does fal/flux pro give? Anyone noticed?

[royal monolith]

1

[royal monolith]

or you can go over to mage.space, go to the home page, click on the General Fast V2 concept, and play with flux forever for free

[onyx isle]

They are using flux there now?! Awesome!

More fun than glif even!

[royal monolith]

you have a DM

[steep sleet]

Flux-Schnell, 8Gb VRAM, 64Gb RAM

OMG - Flux creates such beauty!!!

[chrome burrow]

High-end British restaurant logo

#😊｜co-creators High-end British restaurant logo

[steep sleet]

https://www.reddit.com/r/StableDiffusion/comments/1eix7l6/flux_on_8gb_vram_rtx_2070_64gb_ram_2tb_ssd_t5xxl/

[nimble mason]

Here is the image you requested.

[steep sleet]

How do you specify such a unique text style?

Flux on 8Gb VRAM (RTX 2070), 64Gb RAM, 2Tb SSD t5xxl_fp8_e4m3fn, Flux.Schnell

[royal monolith]

you can try using an actual font name, however for this i just said: the words "flux overlord" scrawled across the image

[steep sleet]

Thanks, I will give it a try!

[royal monolith]

@steep sleet you can prompt the color of the letters, the drop shadow, the outline, the glowing edges, the posistion of the text, etc

[steep sleet]

Wow! I'm impressed!

[royal monolith]

you can probably also prompt saturation and other things you're used to in photoshop - but i haven't tried yet

[languid pebble]

Good morning coffee 🙂

[steep sleet]

I was thinking that Flux-out-of-the-box (like SDXL 0.9) is oversaturated/oversharp/contrasty etc

[royal monolith]

that would be an issue with your settings, not with flux itself

[steep sleet]

Got the text colour, d/shadeau colour right!!!

[steep sleet]

Small adjustment in Paint - there were two "the's"

[steep sleet]

Is this flux? Nice dof

[nimble mason]

prompt: IMG0004.JPG

prompt: IMG0005.JPG

[vast wagon]

I don’t know anything about stable diffusion in general but I’m trying to make a large overhead map with details that I can zoom into. I saw your post on the safetensors @nimble mason . I couldn’t find them in the repository I downloaded but I saw they haven’t replied to your post? Anyways, I’m wondering if there’s anything you’ve made custom to use different samplers, schedulers, nodes, workflows, etc.,. I was reading more of the posts you’ve made with ultrapixel, since I’m making a map I want it to have some of those really intricate details that you mentioned people do a lot. Is there any way to generate them with some modification to ultrapixel? I want to upscale it after, but if it doesn’t have the small intricate details there might be a little less purpose to using it at high res, because I want to zoom in a lot.

Also, I posted some more questions in the general-chat. Main question being, do I need to download ultrapixel, cascade diffusion, and all the models and requirements listed under those separately from what's included in the comfyui-ultrapixel repository? Or what steps should I take to ensure a solid install?

[steep sleet]

What is UltraPixel? Am I missing something by not having it?

[nimble mason]

yeah that i'm wokring on

the official code from the paper authors was really bad

disastrous mess

[languid pebble]

A Cascade method to render very biiiig pictures

[nimble mason]

took me multiple evenings to rip out all the trash and consolidate it, so now i'm gonna be working on a rewrite

not just big, but more detail

[vast wagon]

Ok. Does the paper kinda just describe what it does though? I tried loading it, when it took too long I didn't bother anymore though.

[nimble mason]

the paper does give a decent overview

it's not very detailed about the implementation but the overview is decent

it's definitely a very effective approach, basically what it does is allow you to make much bigger stage C latents

[steep sleet]

UltraPixel is almost counter-intuitive in Cascade! It starts with as mini a latent as possible to reduce processing time ...

[nimble mason]

i think that was just a selling point

i always figured it had the potential to blow away other shit for detail

[vast wagon]

Ok, I figured it'd be more about the technical stuff the actual machine learning is doing. I'll try skimming to see if it details any installation instructions. While I'm doing that, I'm still curious on whether different samplers or schedulers are possible and if you have any nodes or workflows you'd recommend. Don't even know what any of those things I'm asking about mean or are used for... hopefully I can find some explanations online though.

[nimble mason]

with being willing to spend some gpu cycles on it

[steep sleet]

When do we get both Flux and Cascade Discord Rooms? 🙂

[nimble mason]

yeah, they're possible, just not implemented yet

lol flux is a competitor i doubt that' going to happen

cascade they deleted, idk why, a bunch of ppl complained and they just refused to discuss it

so that is that i guess

[steep sleet]

Bah!

[vast wagon]

Are you mostly talking about the schedulers or samplers or all 4?

[steep sleet]

Flux is astonishing - it appeals to me as a non-gamer artist

[nimble mason]

live look at Fort Sai

[royal monolith]

No

[nimble mason]

i got schedulers working

the remaining problem is getting specific samplers in

[steep sleet]

Just going thru dozens of back-catalog prompts - its amazing the new life which Flux gives!

[vast wagon]

Ok.

[nimble mason]

i'm screwing up the noise scaling or something with the RES sampler (best one there is right now)

cuz anything with sigmas over like 5% the max just burns to a crisp

[vast wagon]

I don't know what that even is but can you dm them to me? And if you can keep me updated on progress that would be awesome.

[nimble mason]

i'll grab some links tomorrow, gotta sleep soon here

[steep sleet]

I think SAI doubled-down - and threw good money after bad - and are now scrambling to pick up the pieces! But I hope they do it: SD3 holds a lot of promise.

[nimble mason]

it's not really release ready atm

yeah that would be good

[vast wagon]

Ok.

[nimble mason]

having some crashes

[steep sleet]

Flux on 8Gb VRAM (RTX 2070), 64Gb RAM, 2Tb SSD t5xxl_fp8_e4m3fn, Flux.Schnell

No crashes yet ...

100 seconds/image

[nimble mason]

main problem has been model swapping with this node

serious problems with the implementation that gotta be fixed, it was such a mess lol

it's caused my system to freeze like a mac 20 years ago

mouse just stops and that's it

[steep sleet]

Memory leak?

[steep sleet]

color photo of a head surrounded by dark fears, pulling down to the bottom where nothing is visible, and the struggle to resist with a feeling of powerlessness. The photo captures the essence of a profound psychological battle, rendered in a black and white pencil painting from the year 1800. The monochromatic tones add a timeless and haunting quality to the image, evoking emotions that transcend the passage of time. The delicate strokes of the pencil bring depth and texture to the subject's expressions of fear and desperation. It is a captivating portrayal of the human struggle against overwhelming darkness and the relentless pull towards the unknown abyss. --chaos

[languid pebble]

Love the Questionmark 😄

[steep sleet]

So Flux ignores the first instruction of the prompt: Colour Photo!

But takes up the later "black and white pencil drawing!"

[languid pebble]

Same prompt in SD3

[steep sleet]

SD3 using smudging and smearing because it has no idea of what detail to use!!!

🥳

I know this person ...

... scary! 😄

[languid pebble]

My neighbour ...

[languid pebble]

Don't post private photos 😄

I've got the feeling flux tends to do to big hands ...

[languid pebble]

Backpacker ...

[steep sleet]

They're very robust, it's true! Her hands are huge (for a woman!)

Something like Wolverine's smaller sister!!!

Do that backpacker "swimming out of the pure, uncontaminated waters of the River Seine!!!" 😄

[steep sleet]

Urbex

[languid pebble]

Any errors?

2 thumbs ...

[steep sleet]

Flux on 8Gb VRAM (RTX 2070), 64Gb RAM, 2Tb SSD t5xxl_fp8_e4m3fn, Flux.Schnell
Prompt = beauty in the style of basquiat, rauschenberg, warhol, hambleton, burlet, blk arts, haring, jessie makinson, john caple

[steep sleet]

One major fault of Flux - gigantic hands!!!

[languid pebble]

"Two perfect hands"

Flux Model is trending on Youtube now 😄

[steep sleet]

Olivio Sarikas video is cool!

I'm just watching the "All Your Tech AI" video ...

[wispy nest]

you've got to be careful with diffusion youtubers

some are good but

some of the worst comfy workflows I have ever seen came from youtubers

[languid pebble]

And we have played around with Flux first 😄

[candid surge]

flux-schnell

[robust path]

Flux Schnell

[celest sigil]

Why is everyone using schnell and not dev?

I thought dev was the better of the two

[wispy nest]

they cant fit

and don't like cloud apparently

[celest sigil]

Have you used the fp8 versions?

https://huggingface.co/Kijai/flux-fp8/tree/main

Half the size

[wispy nest]

I use cloud so I dodged this issue

[robust path]

We'll be getting Schnell in DT so this is what i want to get used with @celest sigil

[celest sigil]

DT?

[wispy nest]

apple

[robust path]

Draw Things

[celest sigil]

oh

[void sorrel]

One message removed from a suspended account.

[robust path]

Flux Dev @celest sigil

[celest sigil]

Also

[robust path]

1 is dev 2 is schnell?

[celest sigil]

All of mine are dev

[robust path]

A few years later

[cunning mesa]

I want a picture of a model wearing the bracelet

[wispy nest]

guys is she real ?

shes on facebook , i thought shes real cuz how the body looked.

there one video that has pixalation on the edjes of the face.

[gentle blaze]

Flux.1 DEV fp8
Prompt: Golden hour, Warrior portrait face with shiny armour, orange rim light
ModelSamplingFlush Node (Default 1.15 = Without ModelSamplingFlush)
Shift: 0.10, 0.25, 0.50, 1.15, 2.00, 3.00

[royal monolith]

videos are quite capable of pixilating the entire thing, not just the edges of something.

[pallid ruin]

Is this flux? 👀 👀 👀

[nimble mason]

Yeah

[jovial tiger]

This is another one of those where the defaults look best IMHO

[wispy nest]

what do you think ? is she ai model or a perfect human being ?

i mean she look like one of the generated faces in civitai.

[royal monolith]

I think you now know why non-AI people are so scared, don't you?

[frail pasture]

mate, how do u do the prompts

[wispy nest]

i am lazy prompter, i use civitai to grab data... i just change the words and have fun.

[royal monolith]

do the prompts for what? an image like that one? i'd probably start the prompt with: cinematic photograph of...

[royal monolith]

that's a good way to not get good results, better you learn how the AI thinks and how to communicate with it

[frail pasture]

😦

cinematic photograph of cat

[frail pasture]

just like that?

[frail pasture]

now what?

fuckme:

[languid pebble]

You can
't generate pictures here use -> https://discord.com/channels/1002292111942635562/1237461679286128730

[frail pasture]

😮

[wispy nest]

they are as good as they look in civitai cuz everything is imported even settings.

i just change the colors and poses and play with words.

i do video to video prompts , i am impressed how ai can change an entire human body.

yet keeping the org video motion.

[pallid ruin]

@frail pasture in case you want to try flux (another model) you can use this https://replicate.com/black-forest-labs/flux-schnell

bread hat :3