🏞｜general-with-images

[celest sigil]

6k + 8k

[vale falcon]

this my setup

[pallid ruin]

With that specs you can run most models withouth problems 👌

[vale falcon]

even stable diffusion xl?

[wispy nest]

do you know how much vram

[pallid ruin]

Yes, but im not sure how much vram the gpu has

sdxl can run with at least 6 I think

[wispy nest]

with shenanigans you can get sdxl running on below 1GB

but speed starts to take big hits

[vale falcon]

6gb vram

[vale falcon]

i have 6gb vram

[wispy nest]

that's great you will run very well

[vale falcon]

thank

i will install when i ll be home

i installed only invokeai

where s stable diff xl

?

i will install them when i ll come home

[wispy nest]

I don't use invoke maybe someone else will know

[vale falcon]

oh

[vale falcon]

u have stable diff xl?

[wispy nest]

yeah but I use Comfy UI

[vale falcon]

idk much about that

but

i will talk with you later

pls can you send me a dm with a tutorial or sum

i ll have to go for now

i have to go in the city w/ parents

cya guys

[wispy nest]

bye

[still rose]

I can run Flux 1 on my 16gb m2 pro macbook pro with draw things app

[robust path]

I can run it on my 8gb iPad pro m2 and it even runs on iphone 14

[lavish garden]

Prompt: https://civitai.com/images/23188623

Flux is pretty good (understatement) 👀

[still rose]

Honestly for IOS and MacOs devices draw things is a massive W

[robust path]

Yeah it's really amazing what Liuliu is doing with it

[still rose]

Ye

Especially since on comfy UI my python script just task killed and said it was using 45gb of men

On draw things it takes 209 seconds on my m2 pro macbook

My frens 4070 super takes 140 ish seconds depending on his workload so I'll take 209 seconds

[robust path]

4 steps schnell takes about 132 sec for 768x1024 on mine, I'm happy with it

[still rose]

50 seconds for me

[still rose]

this is the image

[still rose]

i like dev model

[vale falcon]

whats the best site to get models

is this good?

imma test and tell y all

[devout dagger]

Civit ai, they have finetunes for XL and 1.5 and many more models

[vale falcon]

i tried this but uh

[devout dagger]

Are you using the base XL model without the VAE?

[vale falcon]

how do i know if im using it right

it s my first time , i have no clue what to do

wait

something is happening

but it s slow

like very slow

[devout dagger]

You have how many Gbs of vram?, 6?

[vale falcon]

6

[devout dagger]

Then running XL or any finetunes of it will be very slow

[vale falcon]

ow

[devout dagger]

1.5 is your real fast choice for the amount of VRAM that you have

[vale falcon]

;(

will 1.5 be able to make this type of stuff

let me show you

https://www.youtube.com/watch?v=9v4lBexN_Mg&list=PLqb9FyQ12ZJKcqw0P864OhhLlUnS1t4rc&index=1

this

?

[devout dagger]

Yes through img2img, but you may need to upscale the image afterwards since 1.5 only generates at 512x512

[vale falcon]

oh

it s kinda small 512

oof

[devout dagger]

If you can, i recommend you get a RTX 3060 12gb, since they only go for 200 to 300 dollars and they are great if you want to get a start in AI

[vale falcon]

im on laptop man

i cant afford to get a new setup

this setup i ve been using for 1 year

because im at uni

[devout dagger]

Oh, laptops are generally not very good for ai, so you will have to wait long times for generations then

[vale falcon]

oof

[vale falcon]

do you have a specific link

because i only see anime stuff

like look at this 😭

[devout dagger]

https://civitai.com/models/133005/juggernaut-xl?modelVersionId=456194 this is juggernaut xl, its very good at realism

[vale falcon]

thank

can it run on my laptop?

[devout dagger]

can you run SDXL if yes then you can, since juggernaut xl is a finetune of SDXL

[vale falcon]

okay

i ll download it

and try

so the think i downloaded stable diffusion

is that i want to make some cool sheet pfp for spotify

i want to make it sick asf

i want to change my artist profile photo

i ve been using a game character since i started, and i managed to make 500 listeners and want to make a sick stable diffusion thingy

it will take me a while because i have no experience

i want to make some headshots of me to look like yeat photos or any other artist

[languid pebble]

It will be fun, too!

[vale falcon]

in a mystery and dark style

[vale falcon]

true

[vale falcon]

some opium photos

[languid pebble]

😄

[vale falcon]

again

it takes a lot to generate

can someone help me make some settingws

?

[languid pebble]

Always read the instructions for a model on the civitai page. Some models don't need a lot of steps and you can generate faster with them.

[hearty violet]

my inpaint sketch is so weak

any advice to better it?

[vale falcon]

where do i find sampler, steps and cfg?

is this the cfg?

bottom left

nvm i found out

[hearty violet]

advanced

it is sdnext?

[vale falcon]

idk.

i need to find the sampler option

then all the settings are done

[languid pebble]

I'm not 100% sure but there should be a juggernaut model to use with less steps, too

[vale falcon]

i have no clue man

if anybody can give me a model

i ll download and use

i have no clue what these settings do

i just want to make some photos

of me

to be cool edited

it still works slow and i have no clue 😭

[languid pebble]

The Hyper from Juggernaut ...

[vale falcon]

you got a link?

[languid pebble]

https://civitai.com/models/133005/juggernaut-xl

[vale falcon]

thank

[languid pebble]

You have to select the hyper and download it ...

[vale falcon]

okay

[languid pebble]

Recommended Setting Hyper Version (VAE is baked in):
Res: 832*1216 (Any SDXL Res will work fine)
Sampler: DPM++ SDE Karras
Steps: 4-6

Should run way faster....

[languid pebble]

looks good ...

[vale falcon]

im waiting for it to download then i hope i can generate

if not, i ll search for some tuts

because i really want to try, some of my friends told me it s possible to get good results with this gfx card

but yea i ll see

[languid pebble]

Good luck!

[vale falcon]

it s still laggy asf

slowy

i managed to find another model maybe idk

[royal monolith]

how much vram do you have and do you have a gpu?

[vale falcon]

rtx 3060 laptop ver, 6gb vram

somehow i found a model that generates

[devout dagger]

Brother, don't use 1.4 its very outdated, use runway's 1.5 or any finetunes of 1.5

[vale falcon]

okay

this?

[languid pebble]

Better a finetuned one from civitai

[vale falcon]

okay

do you have a link?

[languid pebble]

You can use filters so you only see 1.5 models.

https://civitai.com/models/4384/dreamshaper?modelVersionId=128713 was pretty cool

[vale falcon]

thank you

do you have some recommended settings

?

like steps, cfg, resolution, things like that

[languid pebble]

It's been a while I used models like that. But you should find everything in the description. And the Images also give information how they have been generated.

[pallid ruin]

if I remember correctly I used 768x768 and around 5 to 7 cfg on sd 1.5 models (at 20 steps)

[vale falcon]

okay

i ll try

[pallid ruin]

And if an image is good you can use img2img with low cfg, like 5 to 6 and 30 steps (to have more details)

[vale falcon]

i want to make smth like this

[royal monolith]

https://x.com/fabianstelzer/status/1821166845828673887

fabian (@fabianstelzer) on X

FLUX is incredible at prompt following, but can't do Style Transfer yet

here's a neat trick: you can use FLUX Schnell's gens as Controlnet into an SDXL workflow with an IPA + Style Ref, boom, you now both style and prompt following :)

some examples + link to a glif workflow:

Twitter · Aug 7, 2024, 12:49 PM

[royal monolith]

not easily. you'll want to use sdxl

[vale falcon]

i tried many sdxl models and it takes a lot to generate

rn im outside but i could download to show you

i really want to make something really cool like in that video but yea.. 😭

[royal monolith]

you're probably going to need to generate online then, sounds like your machine might not have the power it needs

[vale falcon]

can i use smth free to make smth like in the video?

[royal monolith]

mage.space has free accounts, you could try them

[vale falcon]

and make stuff like in the video shown?

damn

why s nobody telling me

[royal monolith]

no idea.

[vale falcon]

also i knew in the past there was something from google

like a page where i had to put some links and download some stuff then i could host

from google colab?

is that good?

[languid pebble]

Used it a while ago ... no idea how it is today and whether they still offer free capacities

I have rendered a music video ... took me 4 days ans 3 hours every day ... but I wanted a to high quality 😄

[scenic yew]

yes you can use google colab or kaggle(both are from google and free)
I honestly recommend kaggle since its alot faster and can even run FLUX which is currently the best model and comparable to closed source. It gives you 2 t4 gpus and combined is 28gb vram for 30hrs per week.

The only thing with kaggle is that you have to provide a phone number to verify while you don't need to with colab.

[flint gust]

What is the difference of these 3 models?

[unique condor]

First one is sd3+clipG+clipL, second is the same +t5 fp16, the third is +t5 but in fp8 instead of 16. There really isn't much difference between t5 fp8 and 16, so I'd go with the fp8 version. If your system doesn't have enough ram or vram, go with the first option(identical to the clips sdxl uses)

Fp8 will save a lot of ram and some storage space

[flint gust]

I have 16gb of ram

and 12gb of vram

[unique condor]

Try either the first one or the third one. Not sure if 16gb is enough to hold it all when offloading without resorting to using the slow pagefile

[flint gust]

Got it, thanks for all the info!

[unique condor]

Just keep in mind that the t5 is what helps the most with complex prompts. If you're used to and planning on prompting like sdxl, just use the first model

[flint gust]

Got it

[minor basalt]

Once in a while, flux goes a bit weird.

[pallid ruin]

😋 Im hungry now xd

[tawdry marlin]

#😊｜co-creators "daytime city park, anthropomorphized dogs in firefighting gear training, relaxed community, vertical, 9:16"

[gray verge]

Hello. I tried to recreate this cool generation, https://civitai.com/images/8940291 on my PC, I need help. I copied all the settings, but something goes wrong, the picture became not so realistic. Maybe problems are with those settings that I mentioned by green arrows, I do not understand what they means. What did I make wrong? (image size I choosed randomly)

[languid pebble]

Good morning coffee

[visual hemlock]

Just got to think of it.. Is it possible to hex codes as color stuff for like eyes, hair and such?

Instead of brown I can use #32414xas something?

[flat sand]

no, text text encoder won't know most of the hex codes

you can always manipulate the image afterwards, though. Like replacing colors in a graphics program and then run img2img

[smoky vigil]

What was the promt? XD

[wispy nest]

its always worth a try with each new checkpoint or lora
because it is reasonable to expect that some hex codes may have sneaked into some data sets
it is pretty unlikely to work well though as I doubt there would be a sufficient quantity of any individual hex code
training an embedding (textual inversion) could be a good option if you don't want to train a Dora / Lora
otherwise feeding the colour to IP adapter could be good

[fleet moth]

what is this art style called?

[languid pebble]

Waiting for peace ... ... ...

[vast wagon]

Anyone know why images are coming out like this? comfyui-ultrapixel

[wispy nest]

VAE

https://youtu.be/tUKxWXslPf4

some interesting takes on different models alongside flux

[earnest cargo]

My 3vram Pc is crying 😭

It already struggles hard with pony

[floral flicker]

My 12GB 4070ti is crying - I'm not even bothering with Flux

[earnest cargo]

I won't even try flux local

[floral flicker]

I don't have the patience to wait 2 minutes per image

[earnest cargo]

lol ye in my case I usually let my pc run while I watch something

but ye with pony it already takes 3 minutes a image

flux would probably take 30min if I tried

[earnest cargo]

(artwork in the style of Tatsuro Kiuchi:1.5), colorful, digital, flat-colors, landscapes, nature, street-art, urban-life, whimsy

was checking out some artstyles

[hearty violet]

with folder I must put the embeddings? A11

[smoky vigil]

Geez everyone craps on SD3 :))

What did SD3 ever do to them! Some ladies are not meant to frolic in grass.

[smoky vigil]

But ponies are tiny, they should fit in easier than big horses.

[earnest cargo]

:kekw:

[smoky vigil]

Think of your Graphics card as a barn. Pony goes in, pony goes out. The neurons handle the rest.

[earnest cargo]

tru that

Workers of the world, unite

[hearty violet]

@smoky vigil I found only this models, I'm doing it wrong, right?

[smoky vigil]

For what I said you will need custom node sin comfyui like SAM detector and then you have to get the cropped image and make it big and sample it and then restore it to original size and put it back in the image

So it s abit more complicated than usin adtelair but it gives you the option to specifiy whatever you want to work on not just the face

[hearty violet]

I'm watchng yb videos about adetailer... I understand what you said I think I can do this with SAM on a11 too

[smoky vigil]

Maybe

[frigid stirrup]

Flux?

[languid pebble]

Yes 🙂

[frigid stirrup]

I can't run Flux in A1111 right?

Do I need to install ComfyUI?

[languid pebble]

I don't really know. But I'd guess it doesn't work

[frigid stirrup]

Are you using Comfy UI?

[languid pebble]

ComfyUI is pretty easy to install ...

And there should be a video above how to install Flux in comfy ....

I'm running Comfy, yes ... Re-Forge sometimes, too

[frigid stirrup]

Thank you.

[languid pebble]

Good luck!

[frigid stirrup]

Do 8gb vram and 32gb ram are enough for FLUX?

[languid pebble]

I have no idea, running it on 16GB VRAM ... there are smaller versions you can try

[languid pebble]

A friend don't need to be beautiful ...

[candid surge]

testing some conceptual background stuff with my new custom sd15 merged model

[solid seal]

hello, new to this, why is my SDXL looks this bad? what am i doing wrong?

[candid surge]

idk, does your prompt and settings work fine with other models?

[solid seal]

trying sd3 rn, i donwloaded auto1111 and sdxl since its latest model.

[earnest cargo]

could be a couple of things honestly, my bet would be resolution

if you doing 512 x 512 then you doing it wrong

it's meant to use with 1024 x 1024

[solid seal]

oh yea, im doing 512x512

[earnest cargo]

512 x 512 is going to look bad

yup

ye I was having the same trouble the first time I used sdxl

used to make 528 x 528 gens and wondering why it looked so cripsy

[candid surge]

🤨

[earnest cargo]

I was new to gens 🤷‍♂️

[candid surge]

i get these results with sd3 at low res of 512x512 with 20 steps

not all the best ...

lemme try another prompt, also with 512x512 and 20 steps

hmm, nope. not the best either.

lets try that last one with 768x1024 pixels

better, not the best either. probably needs more prompt tuning and higher amount of steps.

good enough. 768x1024px, 32 steps euler normal, prompt something like:

sharp anime style, full body shot, black-white dress, colorful nature background, fractal style, 32yo lady, older woman, (happy:0.9), white socks, gorgeous scenery, attractive lady, thick hips, (realistic dark lighting:0.5)

[solid seal]

haha thats something

that's not a unicycle...

[charred dagger]

Does anyone have any idea what model or how these thumbnails are made?

[frank drift]

Beautiful!

[frank drift]

Kick ass

[frank drift]

And so well done

[languid pebble]

Thank you!

Doing fashion now 😄

[frank drift]

I could see it on a runway

I'm currently inching closer to using Flux for the first time!

[languid pebble]

Good luck!

It's pretty cool!

[candid surge]

testing ongoing, the merged sd15 anime model was too cute for its purpose. It's called Sexy Lunar. New merge done, mixing in some of a previous real-anime merge. Should make the age a bit more flexible.

Current test is throwing female names as single word prompt, and see what the model comes up with. seems to lean heavy towards nsfw...

Full prompt: Yuki

hmm, ok. gonna put some words in the neg prompt to tame it down.

Full prompt: Annie

Full prompt: Demelza (with some added stuff in neg prompt)

[pallid ruin]

Nice images!

[earnest cargo]

Thanks! I gotten addicted to flux ngl

[mystic marlin]

I love Flux. lol

It should be obvious what 1996 movie this references.

[candid surge]

night chat!

[nimble mason]

random attention guidance for this one

[nimble mason]

cascade

[wispy nest]

reversible bogacki shampine sampler with smoothed energy guidance

[sterile kiln]

😂
@nimble mason spiders got really bad eyes, and prolly don't see anything in the dark x)

[languid pebble]

Good morning coffee ... 🙂

[unique condor]

You know, I'm something of a waifu myself...

[shut sinew]

Flux just knows

[smoky vigil]

AI will never do mechanical objects right. Never.

Hi. My name is a car and a spaceship and a gun. And you AI will never render me right because you have 0 clue about industrial design. You will turn me into a mush.

[hearty violet]

why my image are been generated so blurry? 😭

sdxl with controlnet ip-adapter and lineart
only happens when I apply lineart

[smoky vigil]

Give me 100% consistency. I make Hollywood cry. Meanwhile they laugh at us. They laugh.

[wispy nest]

cool pic.. is that flux?

[shut sinew]

Yeah

[wispy nest]

sweet!

[shut sinew]

I mean what isnt flux these days

[wispy nest]

yeah, i havent been fiddling with anything the last 4 days but flux

[shut sinew]

Are you me?

[wispy nest]

hah.. its the biggest thing in a long time.. everyone is absorbed with flux

[shut sinew]

Absolutely!

[wispy nest]

Brilliant set! 💯

[shut sinew]

Took a good while to figure out the best settings for sampling and HRfixing

[wispy nest]

have you looked into this https://civitai.com/models/607985/image-gen-fluxsd3sdxlponysd15-with-controlnet-supir-tile-upscale-face-swap?modelVersionId=695166

im about to set that up

[shut sinew]

Looks overly complicated

[wispy nest]

lol yes

[shut sinew]

I prefer setting up my own stuff

Then I know what does what

[wispy nest]

yeah good point, im just curious about that workflow

i have alot of missing custom nodes, don't think i will update them

just trying to figure out what the author did

[shut sinew]

Ill make a workflow once we get more CNs and Loras maybe

Until then ill tweak the one I have right now

[wispy nest]

i have two workflows that i use atm, grabbed the basic from comfy examples, and edited them a bit

but this complicated workflow is bit crazy

im gonna see if i can prune it

[shut sinew]

Ive played around with the negative prompt stuff and everything, but its not reliable enough right now

[wispy nest]

negative for flux?

[shut sinew]

Yeah

[wispy nest]

i see, but it doesn't work

[shut sinew]

It does work actually

[wispy nest]

i've done some basic testing

[shut sinew]

But the quality degrades and its slower

[wispy nest]

i tried to remove some of the watermark but didnt work :p

[shut sinew]

It does work for basic stuff like removing colors for example

[wispy nest]

the flux is designed to ignore negative prompt, hence cfg 1

[shut sinew]

Yep but the workaround has some impact at least

[wispy nest]

interesting

can you throw in a couple of image examples?

[shut sinew]

There was this post that kicked it off and a few others https://www.reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/

[wispy nest]

ty, looks interesting, gotta try it out

so increasing cfg can significantly slow down the render speed

i tried higher cfg with the workflow i made but didnt think of how much strain that can put on gpu until i read the user posts on that reddit link

btw my conclusion on #🏞｜general-with-images message can lead to a bad crash

[shut sinew]

Well

[fallow marlin]

Hello, guys I am working with SD upscaler "stabilityai/stable-diffusion-x4-upscaler" and find some unexpected case.
I am trying to upscale different types of images and when I upscale images with large amount of black color model produces some random artifacts, maybe some of you know how to fix it?
I am also providing example of it 1st is basic image and 2nd upscaled image totaly filled with artifacts.

[wispy nest]

could be a factor with denoise

[unique condor]

It degrades it because it isn't actually working the same way. Real CFG is slower because it's essentially making two completely separate things per step(hence it being twice as slow). The positive and the negative. It uses the negative to steer the positive away from things. With flux dev/schnell, there is no CFG.

What's likely happening and why it seems to kind of work is you're literally just subtracting the noise from the positive noise in the 16 channel latent space. Like when you inject noise into a latent for complex workflows you'd see with sdxl. If you shut the negative off early, the positive can contine refining it, but it won't be working with enough noise now to make a true image that it normally would have because it has to "recover" from the damage.

That's why they come out with lower quality and why it only works with broad negatives like red or mountains or w/e, since they can be shut off early in the process

[shut sinew]

So in theory the best would be shifting the CFG lower until 1 with each step, so the final image isnt impacted too much?

[unique condor]

it's always going to affect it if you want it to actually do anything. the most changes happen at the start of the sigma curves

[shut sinew]

I wonder if someone comes up with a solution

I mean a neg prompt is not mandatory but its nice to have

We obviously can make awesome stuff without it

[unique condor]

like here's a bullshit mockup example showing what i mean. this is at step 15 out of 20, so it's not fully finished. obviously, this happens at every step, but i don't want to do this 20x lol

just using sdxl to demo because flux is slow as balls on my pc

note how both negative prompts are empty, the bottom ksampler is mimicing the negative portion

actually hang on, i can just use a hyper lora and do this in fewer steps to showcase it better

[shut sinew]

I get what you mean and the process behind it, no worries

Just wondering if there will be a different approach maybe or if were stuck with as it is

[wispy nest]

@unique condor does the photorealism flux lora have any good result on your end?

[shut sinew]

Guess what lol https://www.reddit.com/r/StableDiffusion/comments/1enxcek/improve_the_inference_speed_by_25_at_cfg_1_for/

[wispy nest]

switching to my checkpoint workflow with cfg option to try it out in a bit

[unique condor]

yeah i got a little ways into making an example but realized it was going to be a bitch without making some custom nodes, but the point is, the trick people are using to hack CFG into flux isn't actual CFG and it has some side effects.

now what they can do with the semi-decent output after the negative stuff, is then resample the whole image in an extra pass with the same positive prompt and it might rectify any of the quality issues

like with denoise around 0.5 or so

[shut sinew]

Thats what I tried but with medium success

[unique condor]

maybe higher

[shut sinew]

Higher denoise changes the image too much imo

[unique condor]

yes and no, seems to mostly only be centered around people and seems to break text

[unique condor]

right, but at least the base will mostly have what you want

like if you dont want red in it, it likely wont add it

[shut sinew]

Current loras for Flux all seem to be more conceptual

[wispy nest]

ahh some key points ty

[shut sinew]

Trying odd resolutions

[wispy nest]

very slow on my 12gb vram

[lavish garden]

Decided to join the Flux hype train...

[wispy nest]

@wispy nest

[wispy nest]

whoa... animatediff?

this workflow in comfyui

[wispy nest]

ok, will load in a bit

[lavish garden]

Prompt: https://civitai.com/images/23493208

"The Visionary" - Ziros Violette (OC)

[wispy nest]

lol ... do this

cool stuff 🙂

this is nice, you can find examples ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-AdvancedLivePortrait\sample

install advanced live portrait node first

how much vram does it require?

not much its a simple image to video expression.

[vale falcon]

yo

i have a question

where is img2img?

i found a model that kinda works with what i want to do

[vale falcon]

i explained days ago

but

i dont see the img2img tab

did i missed something while installing invokeai?

hey someone

please help me understand

[analog panther]

@dry crow lesgooo

[lavish garden]

Not sure what happened 🤷

[dry crow]

Nice!

Is that native Linux or WSL?

[analog panther]

native

couldnt get wsl to work

pulled up an external ssd

lmao i connected with a usb 2 cable

UI is still not slow af , cuz you know , linux

windows would have absolutely died

[analog panther]

its not detecting my amd?

it does detect in neofetch?

[candid surge]

that is correct, cuda is for nividia. it will not use amd.

you probably need ROCm, but I dont have direct experience with Fooocus

check https://www.reddit.com/r/Amd/comments/16kzjo0/amd_gpus_can_now_run_stable_diffusion_fooocus_i/ which is already a year old, so I guess nowadays it should work fine

[robust path]

tourists

[wispy nest]

made this beautiful song with suno ai

[languid pebble]

German ... for those who understand 😄 https://youtu.be/vxLnqgTCloY?si=CruhcsZnYdbKtq5z

[heady crane]

Flux dev

[fleet goblet]

bruh bowser pissed bout somethin

[nimble mason]

used random attention guidance for this

[wispy nest]

what could be considered an extraordinary subject to put into an image in the age of flow matching models?

[wispy nest]

what about a humanoid fish with a fishtank helmet so he can breathe, and he has some kinda weapon or something

fishman

https://media.discordapp.net/attachments/733762994525306972/899670992023076926/image0.gif

[wispy nest]

a photo of a person that has a fish face
finally

oh the horror i've seen lol

[wispy nest]

how human cyborgs look to other animals

[nimble mason]

v

[frank drift]

YOOO I DID IT! I GOT FLUX WORKING! FIRST EVER FLUX IMAGE!!

[frank drift]

I still have much to learn but I am hyped AF

[grand sapphire]

lightmodes cursed asfuck

[languid pebble]

Good morning coffee

[candid surge]

morning

[native gyro]

long

[dawn seal]

I'm liking the Anubi. (Anubises?)

[spring mist]

one Anubi 2 jackals

[dawn seal]

I dunno, I saw several images of Anubis, so...

[spring mist]

ya just nitpicking Anubis is a singular god not multiple gods/entities/people

[vale falcon]

i generated this

with img2img and i need some more refinements like to add elements of my face

and i guess i will be done with my photo

cause this is epic

but whenever i use inpaint

to choose eyes, mouth or smth else, the whole picture change and rip all stuff

[wraith pond]

King

#👥｜roles king

[deft bison]

Anubises! Because artificial intelligence makes the creative artist! ... 😄

[dry blaze]

@languid pebble Doesn't seem to use just 16GB vram :P

This was with ram usage before it settled

[languid pebble]

I have no idea ... it just works here ...

[viral frost]

Flux:

[jaunty nest]

ai

[spring mist]

nah that cat is really writing

[old trellis]

美女

[mystic marlin]

Still trying to figure out what upscaler I want to finalize on, but for Flux I'm pretty satisfied so far.

[wispy nest]

@mystic marlin

[mystic marlin]

That is hilarious.

I love it.

[fleet goblet]

woodbot

Dev does do realistic images quite nicely too.

[brazen sand]

SD3 vs Flux (Flux got it in the first try)

SD3

Flux

SD3

Flux

SD3 omitted the lighting, Flux just put everything I prompted

[brazen sand]

that is some perfect shot

I mean it seems super realistic, it's perfect!!

I even wondering if it is not real lol

[candid surge]

Mask By Text workflow included

[wispy nest]

thanks 🙂

I posted this before but this is what Yolo World thinks objects are

it got most of it right 😄

but it needs fine tune to tell it what gun or stormtrooper is

because its censored so it doesn't know weapons, and it didn't train on hollywood so it doesn't know cinematic stuff

[untold zodiac]

this is my attempt at freddy fazbear from fnaf as a fortnite skin

[wispy nest]

sorry I had to give up on your image
I think SD 1.5 is not strong enough to really work well with openpose control net

I tried adding PAG, SAG, FreeU, res-adapter, rescaleCFG and tonemapnoise in order to help the model understand, but it still wasn't enough

[untold zodiac]

alright, thanks for trying

[wispy nest]

I forgot what SD 1.5 is like 😄

there is a new tool called Smoothed Energy Guidance that might be able to help
but the comfy node seemed buggy to me

[viral frost]

Flux:

I'm having trouble getting Flux to output in photographic style. The prompt contents seem to firmly bias it toward CGI.

[candid surge]

openpose controlnet works fine with sd15 in comfyui, it's just that sd15 still cant do text right.

[candid surge]

yes, that is how it is. Flux is not very good in realistic photo style.

it can do it, it's hidden in there somewhere. If you only prompt for realistic photo style, it works. Add some subjects, actions, other topics ... it's gone.

[lavish garden]

Prompt: https://civitai.com/images/23770977
Flux Art Style Ideas: https://enragedantelope.github.io/Styles-FluxDev

"Dream Team: Chaos Control" - Contest Submission (Civitai)

[candid surge]

fooling around with some Image Blend nodes in comfyui workflow

[wispy nest]

a beautiful Korean female wearing a colorful trendy contemporary outfit, out in the street at night, fashionable, perky, street lights, bokeh.

i love the simplicity and precision of prompts with Flux and its' outputs

[candid surge]

its like ghosts are wandering empty city corners - who is she?

[languid pebble]

Cool!

[wispy nest]

btw flux isn't as good with anime as realism

[candid surge]

🤨 wait. how many tries? did you tick the ai somehow??

I thought flux was pretty good with anime? I think it kinda sucks for REAL realism

[wispy nest]

lol the opposite

flux has to improve on illustration a lot

[candid surge]

nooo, the realistic photo stuff is always too smooth with flux

its really SUPER reality, as in above it, something else, a bit fake

[wispy nest]

those are easy fix with some lora maybe

yes well that's almost the same thing with all txt2img

[candid surge]

sure, idk how many loras there are now

[wispy nest]

just one for now

[candid surge]

flux is too big anyway, cant use it local

[wispy nest]

i am using it locally .. the schnell version

can work with 8gb vram

some people reported to using it with 8gb vram and 16gb ram

[candid surge]

i tried to use it with comfy and 12 gb vram, 16gb ram. didnt work.

needs to much swap memory, and I have an old slow system

[wispy nest]

i have 12gb vram too works fine here

takes me less than 1 min to generate with schnell

[candid surge]

are you using comfy? what settings? what resolution?

[wispy nest]

you can take a look at this page for fp8 schnell checkpoint https://comfyanonymous.github.io/ComfyUI_examples/flux/

[candid surge]

for me, all the loading in between is very very slow. the actual generation is quite fast.

check! done that

[wispy nest]

when it first loads up might take a bit of time, but it caches that into memory and runs faster afterwards

dont play youtube when you are running flux 🙂

[candid surge]

no, after the first gen it seems to hang

i dont

[wispy nest]

hmmm and you have 16gb ram?

[candid surge]

yes

[wispy nest]

that could be a reason

32gb is min rec

[candid surge]

yes, it needs swap. and swap is very slow, since both the swap file and model files are on the same drive

yeah, in that case it might run faster. but i have 12gb vram and 16gb ram

[wispy nest]

your vram is ok

its the ram thats bit low

buy a 16gb more

[candid surge]

i might consider that. i was already thinking about a new system

[languid pebble]

with 5090 😄

[wispy nest]

5090 is out of my reach

[candid surge]

ehm, my budget is limited

[languid pebble]

I have won a 5k€ coupon for the ASUS Store ...

[wispy nest]

i mean i'd have to save up for months to get 5090

i have other priority expenses

[languid pebble]

A bit sad that A.I. is such an expensive hobby ^^

[wispy nest]

yep

wish it wasnt

[wispy nest]

in average it takes me 36-40 seconds with images on flux

[candid surge]

for comparison, here's about the same prompt with flux-schnell (online)

[wispy nest]

i don't think flux anime compares with even sd1.5

[candid surge]

yes, the generation for me goes fast. it's just all the loading of models in between that clogs up the system

[candid surge]

sure, sd15 finetuned anime models are better. I use my own custom merged model with some proud!

but flux is more prompt snappy, can handle more subject and better composition descriptions

I need to gen a few low quality flux gens, then do img2img with my sd15 model 😄

[wispy nest]

true

that's the part i love about flux

[candid surge]

makima sitting in a corner between concrete buildings, some weak green plants growing in the cracks, orange hair, yellow eyes, white blouse, black tie, half body shot, long braided hair, strict powerful look in her eyes, sloppy graffiti on a dirty wall, face up close, 2d simple anime style, thin lines, her hand is touching the wall as if the is searching for something

check that key shadow on the wall! 😄

[wispy nest]

yeah cool

have you tried with thicker lines?

it feels flux anime is missing substance

[candid surge]

it's almost like pony anime, but without the empty backgrounds

[wispy nest]

no way

pony has come a long way to look very artistic

a young woman playing piano in a bright hall room, creating an immersive visual. 2D anime, thick outlines.

[candid surge]

That's it! Flux has much better segregation, which places each subject from the text prompt into it's neat little area on the image. that's why its so clean!

that's also why it's so good at prompt & composition!

amost two sides of the same coin, but how to separate those?!

[wispy nest]

i'd very much prefer flux had similar anime texture

[limpid lichen]

that feels more western i'd almost say it's bluth style. reminds me of that movie anasthesia

[wispy nest]

that image is from sd1.5 model

its more toon themed than anime

[limpid lichen]

lol i mean anastasia

[wispy nest]

yeah i was wondering what you meant at first

[real dagger]

hey guys,, sorry to bother u, but can someone help me get started with making flux work on stable diffusion ?

[wispy nest]

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

[real dagger]

comfy isnt working , i dont know why, cant flux work on web ui stable diffusion automatic 1111 ?

[limpid lichen]

apparently, anime's styles are mostly inspired by old donald duck comics , which were before bluth's time and sort of a project off of disney's main animations

these models tend to have a lot of art history knowledge. it helps to know and throw some history or lore at your prompt. people say "it doesn't know artists" but it really does help to talk to t5 about all the inspirations that you want in a style

[wispy nest]

well no i dont want such toon for anime look and feel, but i think i like western themed anime better than asian origin

[limpid lichen]

i'm not tech support please don't dm me for personal 1on1 assistance.

there's tons of information about how to do it on various websites. tap in and read. we're not here to answer when you smash zero

[real dagger]

THANK YOU

[brazen sand]

indeed, I was trying old prompts were I struggled to get something from SD, and flux just did all the prompt in the first try

[wispy nest]

https://civitai.com/models/633553?modelVersionId=708301

[limpid lichen]

https://civitai.com/models/638000/arnold-schwarzenegger-1990s-flux-lora?modelVersionId=713426 a worthy lora

[steep sleet]

Flux.Schnell

[wispy nest]

how did you make it move

[candid surge]

just generated a few images, and make it into an animated gif

[limpid lichen]

lol i had a voice in my head saying that like, dude wheres your car style. because i know it's just a gif. but then you posted that and i nearly spat my coffee

thank you for that ninja

[wispy nest]

👍

[steep sleet]

Flux.Schnell

[royal monolith]

Big mac and optimized Big Mac

[candid surge]

flux-schnell, online

[limpid lichen]

flux dev [nf4] in forge newest release.

[sterile kiln]

Flux works with forge???

[limpid lichen]

main branch forge new release yup

all new memory management options in forge. gonna change the game again

[viral frost]

Flux:

[limpid lichen]

nf4 version of flux dev in forge

[sterile kiln]

Wow

@limpid lichen there is no 22 GB version with nf4?

[limpid lichen]

nf4 is a quant of weights from the 22gb version. it's the same weights but in a smaller data format. in theory.

[sterile kiln]

I see, great.
Dunno why i have this error with forge 😦
TypeError: 'NoneType' object is not iterable

[limpid lichen]

i was getting that, deleted my forge venv, then restarted

its a heavy duty update

[sterile kiln]

yeah, and i didn't use Forge since long time, where is located venv on Forge?

[limpid lichen]

mine in stable-diffusion-webui-forge\venv

[sterile kiln]

I don't have this path 😦

[limpid lichen]

root of the forge folder

[sterile kiln]

`@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--medvram --xformers --api --port 7861 --ckpt-dir "D:\SD\MODELS" --embeddings-dir "F:\Stable Diffusion\stable-diffusion-webui\embeddings" --lora-dir "F:\Stable Diffusion\stable-diffusion-webui\models\Lora"
--clip-models-path "F:\Stable Diffusion\stable-diffusion-webui\models\Clip"

@REM Uncomment following code to reference an existing A1111 checkout.
@REM set A1111_HOME=Your A1111 checkout dir
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
@REM --ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir "F:\Stable Diffusion\stable-diffusion-webui\embeddings" ^
@REM --lora-dir "F:\Stable Diffusion\stable-diffusion-webui\models\Lora"

call webui.bat`

[limpid lichen]

oh you got the portable. run update.bat ?

[sterile kiln]

already did

set VENV_DIR=
Empty, its supposed to be where x) ?

I searched a folder called venv, nothing

[limpid lichen]

its a portable version so theres no specific venv. it'll use the embedded python's envinrment.

i'm not sure if it's safe to go into that embedded folder and delete stuff

[sterile kiln]

why do you tell its a portable version, I don't remember but I think I took it from Github

@limpid lichen

[limpid lichen]

the versions on github from the "release" menu are portable installs. myself , i just git clone the repo and run it on my system's python install

not really any difference other than yours has python embedded

[sterile kiln]

Found the problem. I was starting FOrge with run.bat, just started with webui-user.bat, as A1111, creating a brand new venv

I can't imagine how many GB I've eaten for nothing, somewhere in C:, with that other useless Venv...
@limpid lichen

[viral frost]

Flux:

[delicate wharf]

I can't run FLUX on forge.. help?

I have RTX 2060 with 12 GB VRAM

[lunar sonnet]

I'm having an issue with eyes when generating images. During generation (including inpainting), the eyes in the preview looks fine, but when the gen finally completes, the eyes become a mangled mess, and the region of the face is notably lower resolution compared to the rest of the image.

[lunar sonnet]

Okay so apparently the low resolution was due to CodeFormer restore faces, I'm using ADetailer instead but I'm still not getting great results.

[delicate wharf]

I think ADetailer has its own resolution also in its settings

[lunar sonnet]

To be fair it's the first time I'm using ADetailer, but I've found the setting you're talking about. I'll try to bump the resolution up.

[oak kettle]

充电桩，插画，蓝色

[still rose]

Flux 1 dev FP16 doing logos

[exotic tartan]

does anyone have any idea how do i fix this image with impaint?

[nimble mason]

sketch a rough outline of where the arm should be in MS paint with the pencil tool

then use the paint tool inside that

[fleet goblet]

itsa' me....

mario.

[sterile kiln]

Photoshop + Stable.art + inpainting

[humble topaz]

#Modify this picture to make it more dreamy, technological, blurry and free

帮助我用ai修改这个图片

[languid pebble]

In China they are selling Geforce RTX 4090D and Geforce RTX 4080 Super with 48 and 32 GB VRAM ...

[steep sleet]

The Bible has some astonishing descriptions which can be turned into prompts

This one is from Ezekiel

a stormy wind blew from the north, a great cloud with light around it, a fire from which flashes of lightning darted, and in the centre a sheen like bronze at the heart of the fire. In the centre I saw what seemed four animals. They looked like this. They were of human form. I heard the noise of their wings as they moved; it sounded like rushing water, like the voice of Shaddai, a noise like a storm, like the noise of a camp; when they halted, they folded their wings, and there was a noise.
Above the vault over their heads was something that looked like a sapphire; it was shaped like a throne and high up on this throne was a being that looked like a man. I saw him shine like bronze, and close to and all around him from what seemed his loins upwards was what looked like fire; and from what seemed his loins downwards I saw what looked like fire, and a light all round like a bow in the clouds on rainy days; that is how the surrounding light appeared.

Flux.Schnell

[celest sigil]

Yup. NF4bnb

[languid pebble]

Good morning water 😄

[wispy nest]

flux?

[celest sigil]

Yes, nf4bnb

[wispy nest]

ugh trying ultimate sd upscaler with flux didnt work

[celest sigil]

You don't need it

[wispy nest]

what if i want it 😄

[celest sigil]

Then you need to try harder 😄

[unique condor]

Hell yeah, I have a whole folder of very similar images with the red olor theme. I'm a sucker for baren landscapes with cities or ruins in them.

[mellow veldt]

is stable diffusion better than flux?

[celest sigil]

https://tenor.com/view/yes-man-no-man-no-gif-7347715096894588969

[steep sleet]

Flux.Schnell

[wispy nest]

I'm still using SDXL for what its worth

but I am waiting for full SD3 release

[raven mirage]

生成一张宣传Midjourney的图片

[raven mirage]

@celest sigil 生成一张宣传Midjourney的图片

[steep sleet]

Excellent that the faces are so natural and not "perfected AI"

[nimble mason]

Here is the image you requested.

[steep sleet]

Bust this!!!!! 😄

[celest sigil]

Busted

[steep sleet]

Boho gypsy look - fetching!

(Embonpoint!) 🙂

[steep sleet]

Did she spray those on? 🙂

Jules Verne/Montgolfier - Flux.Schnell

[wispy nest]

oh nice it has Jules Verne

[steep sleet]

Burning Man Festival Flux.Schnell

I'm d/loading checkpoint bnb-nf4 for Flux - will it be a good thing? Or a mere disappointment at all?!