✨｜sdxl

[copper kraken]

you can try asking me 😄

[glad grove]

you should try to mass ping staff again for better results

[copper kraken]

stability staff too?

[stoic turtle]

Ok, hi 🙂 - where do i go?

[glad grove]

yea devs,comunity guides and even elon musk 🗿

[gloomy lark]

Jeez finally got a 13. Sdxl doesn't like generating "13"

[pallid path]

lol

[copper kraken]

just type your prompt, i'll make a few key changes and then send you the output 😄

[stoic turtle]

I was told you can create images in this discord channel though??

[copper kraken]

correct, i'm even more intelligent than GPT 4

[stoic turtle]

Ok, so you can't then?

[gloomy lark]

The bots are down. Should be up inside of a week or so with SD3

[pallid path]

Yeah probably after the testing period it will be an SD3 bot

cant wait for next week where they invite more people

I wonder if it will be 10 or 12 people

[gloomy lark]

I did scared animals looking at all the @ symbols, but it just made it an A

[pallid path]

lmao

having multiple people doing actions really isn't an SDXL/SD1.5 thing without regional prompting

okay I used highresfix which gave the mane a female face and vise versa

this is how it looked before

im inpatient, SD3 would take these prompts to life

[gloomy lark]

What's the original prompt for that?

[pallid path]

I lost it but its basically about a man holding a knife behind a woman

like some horror film

[ashen otter]

Gunife she’s holding

or guknife?

[pallid path]

exactly lmao

I don't even remember having gun in my prompt but it still puts it in there

but there was, then it would make sense why it generated like that

[uncut steeple]

@copper kraken Please dont mass ping all the staff, simply reporting a message is enough, thanks.

[cyan crown]

https://civitai.com/models/344192

[copper kraken]

Roger, sorry. What's the best way to report it?

[uncut steeple]

You can react to a message with ⚠️ or right click Apps->report to staff

[copper kraken]

Hm, when I tried hitting the exclamation reaction, it appeared then immediately disappeared

happened both on my browser and on the app

[slim wren]

(that s proof that it is working)

[uncut steeple]

Yeah that means it works, the bot takes the reaction and sends the messages to a sepperate channel for us to review

[copper kraken]

oh! ha, I thought that meant it didn't work.

good to know. thanks.

i thought the exclamation might not be working if it was just me sending it without it being in response to a specific message

[gloomy lark]

As usually, it's just 2 people there with it. Not one using it on the other. Maybe censorship of the model?

[pallid path]

idk DreamShaperXL

also I'm using lightning which might hinder prompt adherence further

[trim orbit]

some sdxl models are better at darker scenes, but cascade base model is exceptional

[trim orbit]

thats not how censorship of these models works. it happens in the training set, and it happens in output classification. they can't lobotomize specific knowledge out of weights.

I kind of expected you'd know better, but i'm often mistaken so /shrug

[trim orbit]

just basic token bleeding. classic attention problem from the text encoder

[pallid path]

yup

attention is all you nee-

[trim orbit]

i still like the ada lovelace sneaking up on alan turing images i was doing with regional prompts. probably could make her look a lot more murderous and knife wielding

[shy kelp]

Hello hello !

[shy kelp]

I'm trying to generate something veerrryyy specific with SDXL but it seems to be very limited in this regard

[trim orbit]

fun fact. Ada lovelace was the FIRST computer programmer. Since she wrote code that would run on Babbage's difference engine

[shy kelp]

||I want to generate a pair of slingshot bikini on a female model ,but the character either goes out nude or with simple bikinis no matter the lora or the model that I use , how could I troubleshoot this, do I need to train my own model ?||

[trim orbit]

lol i want to generate something VERY specific ||its porn||

[shy kelp]

Aaaa not really

something inbetween

[trim orbit]

you need to prompt for huge monkin bazonkas . might help

[shy kelp]

as in ? I'm a newb

[trim orbit]

type "huge monkin bazonkas" in your prompt

[shy kelp]

eh

[gloomy lark]

Hah anyone ever tell you that you have an abrasive personality?

[shy kelp]

do I need to train my own model ?

[trim orbit]

yeah but its usually after i tell them how wrong they are and while they're being overtly defencive.

[gloomy lark]

Well when you start off with insults out of nowhere, maybe that's why they get defensive.

[trim orbit]

the whole entire "it's a censorship problem" conversation has long past the civility lines. You're on the hostile side. I'm not sure why you're surprised that someone might be tired of it by now.

[shy kelp]

festivalman do you know a solution to my issue ? thank you !

[gloomy lark]

If you feel that civility needs to end on a happy place with silly pictures, you need to take a break from the internet.

[trim orbit]

I just expect honest discourse but that's not what happens during the "censorship" conversation. People are entrenched in hyperbole and throw out wingers like "they edit the weights after training". instead what we get is ad hominem in most cases.

I mean, if you can't handle a little jab that would be tame by BBC standards, maybe you're not the one cut out for the internet?

[shy kelp]

what is the img2img tool ? the input image one ?

this one is img2img ?

[trim orbit]

thats fooocus. Upscale/variations are img2img for them yup. The fooocus readme is really useful if you're first exploring this app

All these projects tend to have a readme and if they don't the author is being lazy (imo).

https://github.com/lllyasviel/Fooocus

[shy kelp]

thank you !

[pallid path]

im using SuperPrompt offline

it take 1-2 seconds at worst to generate prompts

[gloomy lark]

looks great

[pallid path]

thanks

[upbeat summit]

making some machines with Aetherverse Lightning by @icy brook

[pallid path]

wait a sec haven't I seen you masslevel

[upbeat summit]

I've been on this discord for quite some time 😄

[native knot]

Lies. You aren't even here right now.

😛

[pallid path]

yeah me too, that's why I'm asking

and I think I have

[gloomy lark]

ideogram should license SD and have this option on their website.

[copper kraken]

Yeah an ideogram to SDXL pipeline would be great

[static prawn]

1 thing where sdxl really doesnt shine is textures in clothes, they always look kinda messed up

[gloomy lark]

what in particular about clothing texture is messed up?

[rich olive]

Yeah I don't there's anything wrong with SD textures, SD is only as good as it's user

[gloomy lark]

I've noticed this especially with cascade. if you don't specifically tell it to give you good output, it'll give you an average non-sharp image

maybe that's true of sdxl base as well and i'm just too used to the finetunes where I don't need to do that.

[pallid path]

lol what did I just make

its a bit offscale its hilarious

[gloomy lark]

"ridiculously enormous"

[pallid path]

yes

[vital ermine]

https://civitai.com/models/355335

[pallid path]

oh its general awareness

[pallid path]

this goes hard ngl

[rich olive]

Removing unwanted artifacts really helps

[pallid path]

wow thats nice

[copper kraken]

that one is way cool

[supple wind]

perfect prompt example to generate everything ? who got an example

[trim orbit]

everything and negative prompt nothing

[ornate torrent]

Does sdxl inpainting work? Are their specific models for it similar to 1.5? I tried searching but couldn't come up with anything definitive.

[tawny jungle]

Ancient costume figures, True style, Play Chinese musical instruments, White background

[minor pecan]

[copper kraken]

Here is the image you requested.

[meager canopy]

It does work and doesn't need a specific model. I think it's better to use the model you used to create the image to start with.

[tawdry current]

Just search for posts of mine in this channel and use the new differential diffusion workflow. Works with any sdxl model and you don't need anything special other than making sure you blur your masks so they have a gradient and make sure you use 30+ steps

[meager canopy]

An inpainting workflow

[sterile basalt]

Good afternoon, my name is Maxim. I'm from Russia. For a school project, I created a short cartoon using neural networks only. Please take a look and support. https://youtu.be/ruANV24h0Dw?si=Wr7Nbyo0Wr2_QB3-

[tawdry current]

I like these better than my emoji in a desolate landscape phone wallpaper example. But yeah differential diffusion is game changing and super awesome. It really removes like 70% of a workflow that you'd normally spend un-fudging it with delicate resamples

Again, can't stress it enough to people that want to play with it: you HAVE to have gradients(blur works) in the mask for it to work right and HAVE to have decent step counts.

[orchid cobalt]

Did did you download it somewhere?

Or if you did it yourself, could you share it?

[meager canopy]

I created it, and shared it in that image.

[orchid cobalt]

could you share a json?

[meager canopy]

It's in the image, just drag and drop it

[orchid cobalt]

thanks it worked!

[meager canopy]

You can delete the fooocus nodes, or just leave them disabled. I thought they made it worse.

[orchid cobalt]

which nodes are focus nodes?

[heady vale]

going back to XL lightning is a shock to the system when you can create 8 images faster than 1 in cascade

[coral shadow]

a tiny plane with a huge banner floating after its tail with the colorful word "dahmane !!!" written on it, seen from the beach, many jumping women looking at it with raised arms, blue sky, sunny weather

[stone fossil]

You should prompt for a large plane, the tiny ones do not seem to work.

[coral shadow]

a plane with a huge banner floating after its tail with the colorful word "dahmane !!!" written on it, seen from the beach, many jumping women looking at it with raised arms, blue sky, sunny weather

a large plane with a huge banner floating after its tail with the colorful word "dahmane !!!" written on it, seen from the beach, many jumping women looking at it with raised arms, blue sky, sunny weather

[sturdy abyss]

a large plane with a huge banner floating after its tail with the colorful word "dahmane !!!" written on it, seen from the beach, many jumping women looking at it with raised arms, blue sky, sunny weather

seems ok to me

[sturdy abyss]

just a suggestion:

https://www.google.com/search?channel=fs&client=ubuntu-sn&q=how+to+start+using+stable+diffusion

if you intend to use a bot, that is i think currently not available on this server

[coral shadow]

a plane with a huge banner floating after its tail with the colorful word "dahmane !!!" written on it, seen from the beach, many jumping women looking at it with raised arms, blue sky, sunny weather

a plane with a huge banner floating after its tail with the colorful word "dahmane !!!" written on it, seen from the beach, many jumping women looking at it with raised arms, blue sky, sunny weather

[fallow crater]

Wtf is that bro

Something out of disposable me

[glad dust]

🍀 What if... St Patrick's Day had a Fashion Show? 🤔 🇮🇪

https://www.linkedin.com/posts/ramonguthrie_stpatricksday-fashionshow-irishfashionweek-activity-7175415783949238272-Cl3e?utm_source=share&utm_medium=member_android

[static prawn]

still wish clothes would look better 😄

[rich olive]

Not quite what i had in mind

That's better

[upbeat summit]

very uncomfortable - nice work 🙂

[slate patio]

创造一只小狗

[rich olive]

a puppy?

[gusty swan]

Beaches, beautiful women, bikinis, coconut trees, yachts

[slate patio]

Can you tell me how it was made

[slate patio]

Can you tell me how it was made

[gusty swan]

Beaches, beautiful women, bikinis, coconut trees, yachts

[gusty swan]

创造一个海滩

[rich olive]

Prompt : photo of a (baby french bulldog:0.5) that resembles a [(frog:1.4):2], cinematic lighting, dramatic lighting, high contrast, low key lighting, high key lighting, chiaroscuro lighting, Rembrandt lighting, split lighting, rim lighting, back lighting, side lighting, top lighting, bottom lighting, natural light, artificial light

[round cargo]

Hello,excuseme,where we write the prompt to make images here?

[slate patio]

You should ask them, I'm not sure how to use it either

[round cargo]

Ya,that i did,but no answer for now.

[rich olive]

The bot has been down for a while, you have to run SD locally

[rich olive]

You can use seaart or playground ai, just search on Google

[round cargo]

Ok,thx,i cant do it cuz i havent a nvidia or a amd graphic card of 4gb,but thx anyway.

[round cargo]

Ok thx

[rich olive]

Nether do i, i run Comfy ui on kaggle, you get 30 hours every week, i also have other alternatives

[round cargo]

Could you write link to that site pls?

[rich olive]

https://www.kaggle.com

[round cargo]

thx man

[rich olive]

https://www.seaart.ai

[round cargo]

Which one you use most?

[rich olive]

Comfy for best quality but i use my hours efficiently, seaart is also good

[rich olive]

https://playground.com

[round cargo]

Diffu,i cant find sd xl on kaggle.

[rich olive]

@round cargo you need a good notebook, just search for Comfy ui on kaggle and look for a good notebook I can't help you with that

[round cargo]

Ok,thx

[rich olive]

Search for this exact notebook in the code section, it's the one I'm using

[round cargo]

I wrote comfy ui 823f37 at code- search public notebokk but nothing hapes

[broken dove]

How is the VAE Encode & Inpaint Conditioning looking like? I always get a very harsh edge when doing inpainting.

[rich olive]

Import it

[round cargo]

Thx,now ho wi use it?

[fallow crater]

Upscale

[fallow crater]

You need upscaling

[rich olive]

On kaggle, make a new notebook then import the one i just sent, don't forget to enable the x2 T4 gpu in the accelerator, start a session then run all the code one by one from top to bottom, the last one is unnecessary, they'll give you a link to the Comfy webui

[rich olive]

Yeah that was from a mobile app I'm using called Artful, limited customization, I'm saving my kaggle hours😁

And it actually uses Juggernaut

[round cargo]

the x2 t4 gpu? i havent a grahic card.

[round cargo]

What i do whit the comy ui link?

[rich olive]

I meant on the kaggle website, have you used kaggle before, if not you need a tutorial, search for one on YouTube

[round cargo]

Nope

[rich olive]

Forget kaggle that's way too advanced for you, if you're using Android search for an app called Artful on your appstore, i donno if the app is on iOS

[round cargo]

I made the notebook,now i choose new dataset to upload the file you sent me?

[rich olive]

No, theses a drop-down called File somewhere at the top, click it and then click on import notebook then import the one i sent

[round cargo]

Done,now?

[rich olive]

See the panel to your right that says add input, scroll down to where it says accelerator and select the x2 T4 GPUs, make sure internet on is enabled (it should be somewhere below where it says accelerator), now you're ready to run

[meager canopy]

The examples are literally below the workflow.

[round cargo]

...I scroll down but i dont see there some thing called x2 t4 gpu.

[rich olive]

Bro there's a separate panel to your right, scroll down to where it says accelerator (it will show none because it's not enabled) then you just click on where they say none and then select GPU T4 x2

[round cargo]

Yes bro,I see output sessions options add tags shedule a notebooks to run save code help,but i dont see nothing called acelerator.

[rich olive]

Bro i really wanna help you but you're like completely new to all of this, just download the Artful app i mentioned you'll be able to run JuggernautXL, they even have dalle-3, Kandinsky also the base version of SDXL, i highly recommend you just get the app

[round cargo]

Ok,but it is not mi fault if there dont appear anything called acelerator,i did all you said.

[round cargo]

And i wanted use it on pc cuz my phone is really slow.

[rich olive]

Lol

[round cargo]

Ya i see all that but that acelerator dont appear on pyton option

[fallow crater]

You don't have much prompt control

[rich olive]

Bro tbh I don't think you'll be able to run it, you need at least some coding knowledge to understand what's going on here

[round cargo]

This

[rich olive]

actually I have, i think they're running a1111 through cloud, all i really need is to be able to change the scheduler

[fallow crater]

Model In general

[rich olive]

Silly me, you need to sign in and also verify with your phone number, sorry i forget to mention

[fallow crater]

Prompt was "egirl" ☠️

[round cargo]

I will back later,thx for help.

[rich olive]

I make use of Syntax, it gives me close to dalle level prompt control

[cyan crown]

https://civitai.com/models/355518

[pallid path]

guys... 😔

[sturdy kindle]

/bangbande

[oblique cairn]

ComfyUI session from this morning. Some Final Fantasy-inspired scenes of wizards defending their cities from a hoard of invading dragons.

[copper kraken]

Here is the image you requested.

[normal pelican]

A young anime Dalmatian dog, different scenes, different actions, white background

@sturdy kindle A young anime Dalmatian dog, different scenes, different actions, white background

[sullen crown]

boys play ice hockey in a ice rink

[icy gull]

https://tenor.com/bQUV0.gif

https://tenor.com/r1hWXKj5OAX.gif

[abstract garnet]

Located in the city's underground black market, Chinese style, here to sell a variety of things, strange, winding, lively and mysterious. From ancient spells and exotic herbs to legendary artifacts and mysterious exotic animal specimens

[rustic garnet]

my DnD characters with your Lora:

great work as usual 😍

[meager canopy]

Through SV3D...

[gloomy lark]

Wow that's really neat

[civic gyro]

Happy Spring you all wonderful beings! - https://twitter.com/HikariUchu/status/1770116278075945328

Hikari Uchu (@HikariUchu) on X

Welcome the first day of spring with the enchanting lifecycle of a cherry tree🌸#SpringAwakening #SakuraSeason #Cherryblossom #cherry #tree #sakura #벚꽃 #dance #animation #AIイラスト #Aiillustration #美人 #beauty ＃自然 #nature #hikariuchu

Animation: @stabilityai #stabilityai…

Twitter · Mar 19, 2024, 3:53 PM

[round cargo]

Im back man,now what i do? i verified acc and now yes i see that gpu4 t2

[rich olive]

Ok select the GPU x2 T4 then you're ready to start running, the notebook i sent uses JuggernautXLv9, you can add any other checkpoint by copying the link from civitai or huggingface, run the code one by one by clicking the play button by the sections of code, as i said yesterday you don't need to run the last code, only the first 3 cells, they'll give you a link to comfy webui

[round cargo]

Wich one i must click?

[rich olive]

@round cargo bro not there, you need some coding knowledge for what comes next, go watch a tutorial on how to run code in kaggle

Bro where did you run SD previously

Just go to the seaart website it's beginner friendly, comfy ui isn't

[rich olive]

Germs beach

[gloomy lark]

Is Syntax an app or comfy node?

[rich olive]

You need a node to use it in comfy, i think its the style prompts node you'll find it on git

[gloomy lark]

Thanks

[steep turtle]

I hate this 🤣

[rich olive]

yeah it's meant to make you uncomfortable

[rich olive]

Sorry the name for the node is comfy ui prompt control i just checked

[gloomy lark]

Cool thanks

[pallid path]

lol the proportions

me when no SD3 news for 3 milliseconds

[copper kraken]

that's an epic number of traffic lights 😄

looks like the intersection from hell

[pallid path]

yeah lmao

[coarse kelp]

3072x3072 but it looks weird

this is 2048x2048 and it looks better imo

[rich olive]

I love liminal spaces

[wet nacelle]

Thanks brotha.

[rich olive]

How does these make you feel

[wet nacelle]

These are great!

[copper kraken]

@gloomy lark btw check this out, regarding upscaling...

the latent upscale is indeed vastly worse than the other method even with the vae encode decode sequence... so i bet if you're liking the outputs more with the latent upscale, the reason might be the extra noise it's introducing

bet there's a better form of noise you could inject to get what you're looking for

(or maybe not, which would be interesting in its own right imo)

[gloomy lark]

So here's a few from today using the 20 steps dpmpp-2m sampler to latent 1.5x area upscale to 20 step dpmpp-2m sampler with 0.50 denoise. I was reading online that it's pretty much the cleanest and fastest way to upscale. Anything better is going to take a LOT longer. That make sense if I'm in the interface doing 1 image that I'm working on, but for this particular purpose I'm doing 6 images 12 images against 4 models and want the results back quickly. Best bang for the buck.

[copper kraken]

yeah, very much depends on what you're doing

retaining the likeness of a person i've found def works better with tiled if you're using a lora etc

or ipadapter

but yeah the rest will take a hell of a lot longer

[gloomy lark]

I also realized that I can do dpmpp-2m-side-gpu in the same time, but it bombs the api render. Works great in the comfyui interface but errors when doing it via the api where all other settings are identical. That better sample takes care of the occasional not perfectly round eyeball etc. I'm still working on figuring out why it's doing that.

[copper kraken]

that or just straigh tup using one of those 4x upscale models which really aren't too shabby

huh

[gloomy lark]

Yeah this skips the upscale models which would be great, but it's just the speed thing. Perfect for figuring out prompts rapid fire while I'm out of the house. 🙂

[copper kraken]

yeah

what i'm curious about is alternate noise injection vs the area upscale

maaaybe that's law of first optimization stuff

[gloomy lark]

Doing the latent needs high denoise value. It's what I figured out anyways, but found lots of people saying it's needed or it ends up looking pixelated from the latent upscale.

[copper kraken]

yeah, did you see those images in detail above

the latent area upscale one looks like hell

i did also check the NNLatentUpscale version, it looks a lot like the model upscale route with the vae vae sequence

except just a bit less detailed

so i'm guessing it's gotta be something to do with the shittiness introducing noise to denoise

[gloomy lark]

Oh yeah you need a denoise ksampler after the latent upscale. Without it, it's just a mess. 🙂 but even with that, only 7 seconds total

[copper kraken]

yep but what i mean is...

the upscaler models really are very good these days, if all you're looking for is blowing the image up without pixelating... but...

we usually want to "upscale" by adding extra stuff to the image in the process

which usually requires adding a bit of noise after upscaling and then denoising

so that terrible latent upscale is the source of that noise

what i'm curious about is if the structure of that noise is better than other sources of it for these purposes

[copper kraken]

and if that's the case, i'm curious about finding a way to add some parameters onto that style of latent upscale

[gloomy lark]

ah

i see a couple of "upscale latent with model" nodes, but then they don't have any options for upscale amount etc. obviously i'm not understanding how those are suppsoed to be used.

[copper kraken]

you sure that's not upscale image with model?

if there's a latent one i'm missing it and would def like to know lol

i've tried to grab every latent anything node and every sampler node, but i could've missed one

(not that i know wtf i'm doing with them)

[gloomy lark]

ah, yep upscale image. still can't use it though even if i did vae decode/that node/encode. it's missing settings, or clearly needs some kind of other node to help it.

I tried adding more latent upscales and ksamplers and i was reminded of my first tests with this. one 1.5x upscale and it looks awesome. more than that, and it falls apart.

[copper kraken]

oh wait is that workflow not working?

or do you mean something else

upscale image with model?

[gloomy lark]

this one. when I try to use the "load upscale model" node with it, it won't connect. so I'm clearly not using the right supporting node with it.

[copper kraken]

oh!

[gloomy lark]

cascade first, put through 1.5x latent upscale and then dpmpp_2m with 0.5x denoise with sdxl model. does a great job and it's fast.

[copper kraken]

what scheduler?

lemme know if install missing custom nodes finds it, if it doesn't load the correct node here

[gloomy lark]

yeah, you're right, that works. but how do I specify the scale_by?

[copper kraken]

you can't

[gloomy lark]

I think I combined 2 different gripes into 1 with that. 🙂

[copper kraken]

those models are neural networks trained to do 4x only

lol

so you gotta follow with image scale by

[gloomy lark]

ok let me try that.

[copper kraken]

i usually do lanczco s and 0.375 to get 1.5x

i haven't tested everything on the planet, but i did spend a while looking through some that were available and of the ones i tested i've found this to be one of the best

4xLSDIR and 4xLSDIRplus depending a bit on the image type, haven't been scientific about it

[gloomy lark]

hah yeah ok, doing it with the upscale with model now and it's taking an insane amount of time

[copper kraken]

most of them tend to smooth the image

huh

it should noly be a couple seconds

oh waiiiit how big is your image 😄

[gloomy lark]

let me remove the "upscale by" afterwards.

maybe the model is already upscaling to 4x.

it's definitley faster now, although still very slow. i'll see what the res is afterwards.

original is 1152x768, so if this is 4000+ res, then yeah...

[copper kraken]

wtf, of all the small irritations life has to offer... the order of the schedulers got flipped? lol

huh that shouldn't take long at all

what model are you suing for it

maybe some take longer, idk, i haven't timed them

[gloomy lark]

hah yeah.. the upscale with model upscaled it by 4x.

that's why there's no resolution input.

it actualy looks worse than the regular upscale image by 1.5x

let me try a photo image, not a 2d drawing.

[copper kraken]

interesting

are you using LSIDRplus or something else

maybe the resolution does something wonky too, idk

[copper kraken]

now that i think about it i've generally just done 1024x1024, 512x, 768x

[gloomy lark]

4k imaging will turn a 4090 into sludge.

yeah it totally borked it.

maybe i'm doing this wrong.

maybe i shouldn't be denoising...

at lal

all if i'm doing the upscale model.

actually yeah, need to do much lower denoise now that i'm not doing latent.

[copper kraken]

this took 6 seconds

yeah you need a lot less denoise

that or it's a good idea to use an ancestral sampler to get some noise back

there's a lil less smoothing with the base LSDIR (top was LSDIRplus)

[gloomy lark]

so this is cascade, with the same method you're using, although with the siax upscale model.

the problem with cascade is that it has stuff that definitely needs denoising to fix. i don't just want it bigger with more skin texture.

those wonderful cascade swirls on faces.

[copper kraken]

yeah for sure

it's an oversampling issue

aimingfall figured that out

i've had some success with tweaking the denoise schedule

if you stop stage B after something like 6-7 steps it's generally better than any greater number unless you've modified the denoise schedule

[gloomy lark]

cascade / upscale image by 1.5x - area method / 25 steps of dpm_2m_sde_gpu-karras / 0.5 denoise

hah even fixed the gap in her teeth

lol the original sdxl refine was the new proteus-rundiffusion that came out the other day, this is with dark arts images. everything is way more sinister!

[copper kraken]

awesome

the ones i tend to use the most are 0.45-0.6 denoise dpmpp_a, typcially 0.5

with karras

then sometimes dpmpp_3m_sde_gpu exponential when i want to preserve the original composition

with 0.5 or so

without the gpu version it changes it a bit more

i haven't been systematic in testing everything... so many fn permutations

[gloomy lark]

hahaha this is with playground 2.5

[copper kraken]

loving that you're setting up a mini gpu farm to crank out shit like this 😄

btw since you're in comfy or were recently... i realized that the concat conditions with multiplier node is awesome for combining weird shit

its from the inspire pack

[gloomy lark]

playground does really good bugs

[copper kraken]

is that 2.5? that's fuckin awesome

[gloomy lark]

yeah when playground 2.5 works, it's really impressive. be back in a few.

[copper kraken]

k

just play with the ratio in that multiplier concat node

if one prompt is dominating, then swio.h.pe

cat lol

then reduce its multiplier, and go back and forth till you reach equilibrium

[gloomy lark]

yeah I'm assuming sd3 is gonna be a good amount more taxing, so waiting for higher render time images is gonna get old real fast.

and turbo is lame when the whole point is better prompt adherance.

[copper kraken]

for sure

i just hope the 5090 isn't a disappointing pile of stale butt

[steep turtle]

*And the price is somewhat reasonable and accessible 💀

[copper kraken]

i also hope it turns out nvidia has already started production and accidentally mails me one

[steep turtle]

Imagine if Nvidia sells it for $4000

[copper kraken]

doubt they'll go that high

[gloomy lark]

they still haven't announced specs on 5090 though right? it was just that rumor post of 1.5x memory speed and the same 24gigs of ram?

[copper kraken]

correct

also a rumor it's 32gb

also a rumor there's 50k cuda cores

all kinds of random guessing

[steep turtle]

I mean, the 4090 is like $2000 🤣 It's insanely overpriced in my humble opinion

[gloomy lark]

well, we'll see what the speed increase is. if it's something really noticeably faster I'll get 1.

[copper kraken]

in terms of performance vs other cards, it's not, but they are price gouging

[gloomy lark]

but it would have to be a lot faster than the 4090 for it to be worth it.

[copper kraken]

4090 just blows the other shit away completely

yeah def

[gloomy lark]

not like 30% or some such

[steep turtle]

I have to check the prices in Bulgaria 🤔 It's been a while since I've looked at GPUs

[copper kraken]

yep if it's double the speed and/or 32gb i'll be getting one

[heady vale]

moores law hinted at 60% better

[copper kraken]

but if it's 30% with the same 24gb, meh

[heady vale]

its all rumours though

[copper kraken]

yep

[gloomy lark]

it's 47,000 stotinki per gig of vram over there.

[steep turtle]

So 470 leva? 🤔

[gloomy lark]

give or take

[steep turtle]

That doesn't add up because the 4090 should cost around $5000 here then 😅

https://plasico.bg/komponenti/video-karti/filter-156846-156849 It will probably be difficult to read the Bulgarian but you can take a look at this store and compare the 4090 prices

[gloomy lark]

yeah I'm just making stuff up. 🙂

[steep turtle]

😂 That flew over me

[copper kraken]

do you have your discord bot set up with comfy workflows? i forget

i wish comfy was even remotely usable on android

[gloomy lark]

yep, it's all comfy at this point. took me a few days of troubleshooting, but I got the queueing system working. tonight I added cascade as well so it adds into the queue correctly. but yeah it's using comfy workflows for everything

just waiting on the hardware now

[copper kraken]

does it allow you to select a certain workflow, or alter any of the parameters?

[gloomy lark]

nope. 🙂

[copper kraken]

you weren't kidding about PG2.5 and bugs...

[gloomy lark]

wow

i take it that's native with no cascade?

[copper kraken]

yeah, that's playground 2.5 on comfy

res_momentumized with no momentum in ksampler

that is one of my new fav samplers, damn is that one interesting

it generates even wilder outputs than dpmpp_a

[gloomy lark]

there are many things that cascade and playground have that are similar, so it just modifies it a little. but then there's stuff like this where it completely changes it.

although I guess for the better

[copper kraken]

the other thing that's interesting about res_momentumized i've noticed is tons of steps does actually help

i'm starting to use 50+ out of habit

[gloomy lark]

ok lets try that again with 60.

[copper kraken]

notice anything interesting yet with the sigma values? i haven't palyed with them much

[gloomy lark]

changed it quite. a bit.

20 vs. 60

[copper kraken]

yeah for sure

a few times i've gotten amazing shit by going to 200+

it seems to keep adding details and complexity for a while

idk a damn thing about that sampler

[gloomy lark]

that said, it's clear that some stuff doesn't work with the api that works with the regular interface, so i can't just take anything and throw it in there. I have to figure out why that is.

[gloomy lark]

with 200 res_momentum

[copper kraken]

for the bot? yeah, been talking with someone who's been messing with that

apparently the api is a shit show

[gloomy lark]

with sd3, we should be able to have a wide assortment of bugs

[copper kraken]

i'm eager to push the boundaries of clown, shark, bat, manta ray, samurai, insect, and arachnid themed images

is the token limit still 77-2?

i've barely touched pg2.5 tbh

this is pretty great

[coarse kelp]

"me at the zoo", an 18 years old video from youtube reimagined:

it's my first attempt so it isn't particularly good

https://www.youtube.com/watch?v=jNQXAC9IVRw

this was the video i used for it

[copper kraken]

400 steps

[heady vale]

8 steps using the 4 step lightning lora

[gloomy lark]

lykon said it's 512 now in sd3, but that's one of those things I'll wait to see. we can do really long prompts in sdxl as well, but we know it's not that simple

[copper kraken]

oh, yeah, i was wondering about pg2.5 too though

i'i'd imagine it's the same unless i'm overlooking something

[gloomy lark]

Wow neat. Yeah playground still has the same engine limitations.

So these are all cascade put through dark arts images. I think playground is awesome by itself, but deviates too far from cascades native imagery. Dark arts is closer and adds that little dark arts flare if you go "dark".

Although I may have to go with juggernaut, dark arts might be still too overpowering for cascade e

[copper kraken]

yeah, playground has a hell of a style it wants to enforce for sure

[vagrant pebble]

#1047610792226340935

[copper kraken]

Beep boop. Would you like an image? Provide a prompt!

[copper kraken]

yeah gotta hand it to em this is pretty fn great

it's too bad it's not supported on a1111 or forge yet

playground deserves more attention that they don't get for reasons of platform compatibility and sheer inertia

[gloomy lark]

It has a very midjourney vibe to it. Lots of high contrast colors.

[copper kraken]

def

not as much control as you get with our fav sdxl models, but no denying it looks great

brownie points for being open source and beating MJ/dalle3 at their own game imo

it does have that mega trained feel to it - there's a lot of stability to the image outputs, not a lot of variation from seeds, and some prompts just MUST have a certain element

for example, not like my efforts were exhaustive, but it sure as hell thinks that a playground at night should have a light somewhere

[gloomy lark]

Well, have to be able to see it. 🙂

[copper kraken]

sometimes though you might just want to barely see the outline of the image

at least i do lol

https://civitai.com/models/295530/zavys-dark-atmospheric-contrast-sdxl

i discovered this last night... lifesaver

[gloomy lark]

Sure

[lilac nova]

?/dream

[copper kraken]

shit, even the lora didn't save me with that one, a damn light appeared

[gloomy lark]

You're right about the lack of variation on seeds with pg. barely any. Have to use llm to create variation between generations.

[copper kraken]

yeah that's not really a good thing

but yeah what it's aiming to do i think it knocks outta the park when considering it's working within the limitations of the sdxl architecture

cascade respects my deepest desires

fuck actions, i just want SD3 to be able to create images where you can barely see anything lol

[gloomy lark]

Same prompt but with llm varying it up.

[copper kraken]

you've got that llm stuff down

that's def the future of all this

[gloomy lark]

Hah that's a black square

[copper kraken]

yup

prompt" a barely visible outline of a playground at night, pitch black"

i changed "very dark" to "pitch black" and got that

changed to "almost pitch black"

[gloomy lark]

That looks like a barely visible comfy workflow

[copper kraken]

lol

it actually does 😄

[patent grail]

This is so cool, whats model generated

[slate patio]

Friends, can anyone help me make an image, I don't know how to use it

[tight scroll]

https://tenor.com/bWs2c.gif

[neon vale]

which clip does the refiner use, L or G?

[tawdry current]

Pretty sure* it uses both like the base, but you could always check the huggingface page

[opaque socket]

dragon cyborg in gold and pink with H.R giger texture and oriental pattern

[rich olive]

close up photo of a lit candle:1 AND close up photo of water:1

[pallid path]

woowie generated an almost midjourney level image accidentally

[wet nacelle]

Are you using face reconstruction?

[pallid path]

nope

[wet nacelle]

WOW!!

[pallid path]

just Highres-Fix

highresfix is amazing, SD3 will benefit from it the same way

before and after

before after

instead of reconstructing the faces only, I reconstruct the entire image

epic finger count

holy crap those reflections are so pristine

[pine topaz]

https://tenor.com/btiUi.gif

[worthy orbit]

Cogito ergo sum right?

It seems I don't have enougth VRAM to use hi-res fix, I can only use it to 1.2 or so, is there some way to make it work with higher amounts? I have 12 GB VRAM

[pallid path]

I have 12GB too

Turn on Tiled VAE in Forge/A1111 or use VAE Encode (Tiled) in Comfyui

this was I never run out of VRAM

Tiled VAEs don't have seams to my knowledge

[worthy orbit]

I need to try that, ty

how much do you resize with Hi-res fix?

[pallid path]

believe it or not like

1.2x first with kohya deep downsample -> 2x highresfix

so like 2.4x upscale on 12GB?

wihtout Tiled VAE I am inching very very close to OOM

I can only imagine how it will be with SD3 8B (if at all)

[worthy orbit]

I didn't know about that, kohya deep downsample

[pallid path]

in comfyui

maybe its prebundled in Forge

[worthy orbit]

this? https://github.com/wcde/sd-webui-kohya-hiresfix

[pallid path]

oh yeah

that seems like it

but you can go for 2.4X highresfix straight, idk why I even do it this way lol

[worthy orbit]

I need to try that

Also there is this button in Auto1111 now, so you can hi-res fix something you already generate

[sand sinew]

how do I generate picture

[rich olive]

Here's the image you requested

[thorny hill]

Cinematic.Redmond is here!

This is a Cinematic model fine-tuned on SD XL 1.0!
The model has a high capacity to generate Cinematic, artistic images, cars, people, and a wide variety of themes. It's a versatile model.
I really hope you like the model and use it.
I recommend generating it in cinematic proportion like 16:9, 2:1 etc.
If you like the model and think it's worth it, you can make a donation to my Patreon or Ko-fi.
Follow me in my twitter to get acess before for all new models:
~https://twitter.com/artificialguybr/~
You can use it for free here:https://huggingface.co/spaces/artificialguybr/CinematicRedmond-Free-Demo
Download it here:
HF;https://huggingface.co/artificialguybr/CinematicRedmond-SDXL
Civitai: https://civitai.com/models/359999

I hope you guys enjoy :)

[pallid path]

nice

[pallid path]

these are so epic

[elder plume]

does this use any 3d animated content?

and did you standardize on 1344x768 or something else?

and did you use the film scripts at all for the contents of the captions?

your trained content is really excellent btw

did you use pivotal training? so many questions

[pallid path]

daaamn

[stray warren]

Who needs SD3 when we have SDXL? 😁

[pallid path]

Me

I need SD3

[rich olive]

All we really need is prompt coherence

[pallid path]

yessisrky

SD3 has the quality of a finetuned SDXL model with the prompt coherency of DALLE3

[rich olive]

Better than dalle-3

[pallid path]

well I consider the fact that I might use SD3 Turbo as a daily driver like SDXL Lightning

especially if somehow SD3 has to be CPU offloaded then 4 Steps will really help out

[rich olive]

Yoda x Gollum

[gloomy lark]

I thought SDXL Lightning was the best thing ever until I started putting lots of promtps through it side by side with the full version of the same model, and even though the visual quality was there, the prompt adherence was garbage. Great for portraits, not so much for other stuff.

So I see this stuff about SD3 Turbo, and I'm glad it's there, lots of uses, but I'll probably steer away from it.

[pallid path]

Prompt adherence is still going to be above Midjourney v6 apparently for Turbo so you might actually consider it

though the failure examples do make me doubt it sometimes

[gloomy lark]

I've got subscriptions to all the services, and I only occasionally put them through midjourney. Now that I see SD3's higher channel VAE, I suspect that Midjourney V6 has that already and that's one of the reasons why it looks so good, aside from their clearly higher res training set. The prompt adherence of MJ at this point is only marginally better than SDXL, maybe a little higher than Cascade. I still can't get mecha to actually be doing anything with MJ other than standing here with their arms at their sides. I have to use regional prompter in SDXL to get them to do anything as well.

[pallid path]

can you roughly tell me what the 16 channels improve? higher quality VAE decoding? therefore, less of a need for HighresfixM

[gloomy lark]

my 2 second understanding is that it's not a composition thing, it's a color quality thing

it's the differnce between shiny waxy skin, and a real photograph appearance for people.

[pallid path]

so it IS colours, since it's colour channels

but that's great news, thanks

[gloomy lark]

yeah that one thing will probably have the single biggest impact other than prompt adherenace.

i could have done more fixing/upressing on the sdxl to fix it up, but i think the color difference between the 2 is obvious

[gloomy lark]

midjourney v6 on left, sdxl on right

[pallid path]

yeah

[gloomy lark]

that said, midjourney goes way overboard on pastel colors and shallow depth of field, even when it makes no sense. with sdxl it's a choice, not a default.

[gloomy lark]

Just to round out those 3, dall-e's version of it.

[pallid path]

😔

[gloomy lark]

hah putting that through sdxl is so much better

[fallow crater]

Dall e is being carried by the llm

[gloomy lark]

yeah, if you look at what sdxl just made of it, it's so much higher fidelity

[fallow crater]

Sdxl isn't good at prompt following

[pallid path]

yeah

[gloomy lark]

sure it is. 🙂

[fallow crater]

You're using llm

Maybe

[pallid path]

no

just using regional prompting

[fallow crater]

Or you're not as stupid as me

100%

[gloomy lark]

hah nah, it's regional prompter. the prompt for that robot with the money is actually dirt simple.

smiling man's stone head, mecha hands in the air spewing money, dystopian future
ADDROW
chibi cute 3d mecha arms in the air , money everywhere, gritty, grimey, wires, sparking, dystopian future
ADDROW
worshiping crowds with raised fists, dystopian future

[fallow crater]

Only.upside of sd3 I see is prompt.following

You can use a lora for perfect text

[fallow crater]

💀☠️

Ros getting dall e level prompts

@gloomy lark are u using Loras

Or just juggernaut

[gloomy lark]

no loras.

[gloomy lark]

It wouldn't do your username correctly. 🙂

[fallow crater]

💀

Cascade can do names ngl

[gloomy lark]

yeah it's definitely better at text.

but no regional prompter.

[gloomy lark]

wow midjourney actually did a really good job with this one, which is so hit or miss. sdxl is consistent, midjourney is luck of the draw as to whether it does what you ask

[wet nacelle]

IT GIVES YOU A RANDOM ART STYLE EACH FUCKING TIME!!!!!!!

[gloomy lark]

yeah exactly. I've tried messing with the style settings. it does nothing

[wet nacelle]

I used MJ since it launched actually. The moment I jumped ship was when Stable was able to output consistent hyper realism. I was there up until the end of V5. V5 was shit too. fucking garbage mods and people too.

[gloomy lark]

I think they saw what SD3 is going to do with text and they upped their game. text with them has been meh until recently

[wet nacelle]

For me it was always really good at it.

[fallow crater]

💀

[civic gyro]

More spring magic - https://twitter.com/HikariUchu/status/1770589661251141657

Hikari Uchu (@HikariUchu) on X

Experience 'Blossom and Bee: Nature's Ballet' 🌸🐝 #NaturesBallet #SpringDance #SakuraSeason #rhythm #balance #spring #Cherryblossom #cherry #tree #sakura #벚꽃 #dance #animation #AIイラスト #Aiillustration #美人 #beauty #自然 #nature #hikariuchu

Animation: @stabilityai…

Twitter · Mar 20, 2024, 11:14 PM

[rotund surge]

Yoo I can only tell that´s AI because of the cat face and other minor details but it looks so good!

what model were you using? 👀

wait...is it ai, right?

😁

[spring fulcrum]

Hey everyone. I haven't been on in a few months. I was wondering what is new with SDXL? I keep hearing about Stable Diffusion 3 and Cascade. Are any of those out and available on ComfyUI?

[sleek pulsar]

cascade is

[sleek pulsar]

That cigarette one is so cursed, messed up anatomy

[jolly nexus]

i wanna generate some pics, How??

what command should i use

[rich olive]

Bot is down bro they're busy with sd3

[cosmic compass]

Is there a new solution now on how to upscale watercolor paitnings without actually loosing the texture of watercolor paintings ?

[worthy orbit]

I was just working on something like that, I got good results with Ultimate SD Upscaler, SD XL model zavychromaxl_v50, Denoising strength: 0.16, upscaler: 4x-UltraSharp, upscaled x 4, Ultimate SD upscale tile_width: 1024, Ultimate SD upscale padding: 64

Got some little aberration (a guy head on a post) but other than that I was pretty pleased with the result, it respected the painting (I was trying to resize in img2img but it always messed the good details of the painting)

[cosmic compass]

the result looks good but i was hoping to be able to keep those textures as of the wall on the left below the balcony, the upscaled ones allways kept losing those irregular patterns and became to clearly structured, was hoping that by now there is a solution

[worthy orbit]

yes that's true, I haven't noticed that, and I'm in the same boat. I was more focused on the subjects and light poles/trees.

It seems with upscale you always have to compromised something. Sometimes with several tries, trying different denoising strength, one hits the spot.

Yes SD Ultimate Upscale usually kind of wipe some nice textures, result may be decent.

[fickle axle]

Wan Gog

[fickle axle]

[worthy orbit]

Brute method I might think is to create several ones and in Photoshop merge the parts that has the desired textures, as sometimes some areas needs more detail, like the subjects, but other area needs less denoising so it keeps the original texture.

Then some inpaint if the merge isn't seamless... yeah some work. But afaik there is no other method than the denoising gamble.

[worthy orbit]

Woskar Wowoschka

[worthy orbit]

oh yes definitely it made a wan gog head there

[pallid path]

juggernaut

[stray warren]

Who's a good boy?

[minor pecan]

Dude..

[meager canopy]

These are great! Was it a lora or all prompt?

[errant arrow]

#🏞｜general-with-images

[vital ermine]

https://civitai.com/models/361974

[slate zealot]

interior design TV background

[primal rapids]

https://tenor.com/bbqHT.gif

[copper kraken]

Here is the image you requested.

[unkempt wagon]

A pod of dolphins playing.

[orchid ocean]

Im training a style lora for sdxl, which preset should i use Im a beginner T-T

[copper kraken]

Here is the image you requested.

[tawdry current]

There is no one size fits all preset but I'd start with the sdxl now prodigy one. Prodigy is a good optimizer. Also, be prepared for a lot of failures in over and under training. There is a ton of heavily conflicting information in various guides you'll find, so be prepared to not take every guide as the best

The reason why the information is so conflicting and vague is because you have a lot of people that lucked out with some settings on a specific train, so they think it works for all. Oh and graphs, you'll do a lot of looking at tensorboard graphs lol

[burnt shadow]

cascade?

[gloomy lark]

This was sdxl dark arts images. Prompt was hurricane of gang members. Kind of a combination of the 2 big things happening to Haiti. It's always interesting to see what the model comes up with.

[ashen otter]

Very interesting

[burnt shadow]

ah, yes interesting, thank you

[glass forge]

i wish so much we could somehow lock these characters and make coherent stories with them they look great

[lucid rampart]

做一个马

[dark crypt]

I've been away for a month, give or take, are there any big advances I missed? I'm using Radeon.

[minor pecan]

is that shaggy!

[gloomy lark]

if there a prefered comfyui node that takes an image input and proportionally resizes an image so that the largest dimension stays under a maximum that I've set?

[native knot]

Use the Image Resize from ComfyUI Essentials.

[gloomy lark]

this one?

[native knot]

You can set to keep proportion, set the height/width so that there's a max.

Yeah

[gloomy lark]

Thanks, I got it going

prompt:a cat drinking coffee at a coffee shop. with and without clipvision, which I finally got setup for sdxl (only did cascade before). The difference in "prompt adherence" is night and day. incredible.

and this was the input image. it didn't take much to set it on the right path.

[visual kelp]

Very interesting

[rough field]

a 22 year old boy with black hair, from india, indian skin color, wearning a computer glasses, looking into camera, background is dark brown plain, 4k resolution profile photo

[gloomy lark]

I'm figuring out that clipvision in sdxl is nothing like that in cascade.

for some reason the input image affects the final output WAY more in cascade.

[fair tinsel]

/ca

[dim steeple]

what is the best controlnet extension for pose control

?

[gloomy lark]

openpose

[dim steeple]

then do you load that openpose skeleton into controlnet?

which one of these is it?

is this right?

[gloomy lark]

sure you can choose All or select openpose.

then hit the little firework red thing to see the preview to make sure it worked.

that said, if you already have the open pose model, you don't technically need the preprocessor, because that's just creating that openpose mannequin out of whatever image.

so if you already have it....

[dim steeple]

Pixel Perfect?

does this mean it is not working?

something isnt working

am i using the wrong openpose model?

[gloomy lark]

pixel perfect is when your openpose jpeg with the skeleton matches your render resolution you have set.

so it's a 1:1 pixel accurate

[dim steeple]

no preprosesor fixed it

[dim steeple]

it refuses to render this shape

is that a known issue?

[gloomy lark]

flossing isn't possible with sdxl.

🙂

[dim steeple]

lol

really?

[gloomy lark]

all this stuff and regional prompter is finnicky. different models, different cfg and step counts. have to fool around with all those settings to see if that's what's stopping it from working the way you want. in a1111, there's an XYZ plot at the bottom. do your controlnet stuff, then set the number of steps and the CFG values to walk the ranges and see if there's one set that will do what you want.

[dim steeple]

thanks

any tips for consistency?

character consistency

to make an animation

same seed?

[gloomy lark]

there's some ipadapter models like faceidplus v2 and instantID that can do repeatable faces, but they usually only work at very low CFGs, which is only possible with Turbo models.

other than that, you'd have to train a lora.

there's a ton of youtube videos that go over all that.

[glass forge]

And even that is just 80% accurate. No 100% concistency yet.

We need total control! Like a 3D program. I have to be able to film my characters/objects/backgroudns from any angle and lightign condiiton concievable!

Until then AI video is a gimmick

[ashen otter]

Might as well go full manual then lol but I get you

[glass forge]

heh

Its not that simple because AI could render things a million times better than other 3D programs and fiddling with photo editors and so on.

such a shame to get these amazing images but really we cant do anything with them yet

[rotund surge]

👏 Do you know if its possible to use clipvision on sd forge?

[gloomy lark]

don't know, I've only used clipvision with comfy

[rotund surge]

😁

[frank pivot]

Chinese painting style drawing of a tortoise

[full cradle]

rs image

[minor jungle]

Differential Diffusion (using zho_zho_zho's workflow)

[copper kraken]

They actually do work well with cfg = 6 with regular models

The issue is you tend to have body horror hell

I've found the best thing to do is use them on low strength for a first pass, then upscale the face, unsample and resample it, then patch it back in

I have some nice workflows that automate that

[minor jungle]

tried some different prompts fusing together

[copper kraken]

Link to workflow? Curious cuz I have some fancy ones I've come up with

Would be happy to share once my Internet connection comes back (on phone here)

[copper kraken]

All diff diff, workflows should be in the image still

[minor jungle]

wow, beautiful

[copper kraken]

I used a variety of different masks for these

I had these on my phone still so just pasted them in, the masks I can share once I get my Internet connection back

Diff diff is incredible

The one with the steampunk machine in the center, I was using diff diff for inpainting

Comfyui is just fine for inpainting, the stuff about needing a1111 for that is bs imo, you just need the workflow set up right

[copper kraken]

more diff diff, look close and see which way is up

[gloomy lark]

So what's your process for these? Inpainting? Because there's more going on here than what regional prompter can give you. There's a level of chaos that I haven't been able to achieve normally. 🙂

[rich olive]

Before and after upscaling

[native knot]

The upscale fucked up the sun rays.

[frozen swift]

Crazy sh!t

[peak dove]

InstantID, with the right LoRA(s) included has incredible facial consistency. My LoRAs of choice for I.ID are Ethereal Grace, Rainbow, and Shaun Tan ... but that's not an exhaustive list.

Without LoRAs, faces can be weaker in their consistency.

[rich olive]

yeah i was trying to see just how far i can push it, took about 4 min on a T4 gpu, a 3x upscale using pixelksampleupscaler with 0.3 denoising, i was using ultimateSDupscale before but couldn't get rid of those tile seems

[native knot]

Funny...I'm trying to push a PixelKSampleUpscalerProvider through Iterative Upscale right now on a 1152x2048 image just to see something. Taking forever on my 3080. 🤣

2x upscale over 3 steps. Probably gonna die again here soon and I may have to give up on it.

[rich olive]

Lol yeah it can take forever, i always use 1 step but i try to at least get a 3x upscale

[native knot]

Well, I would do 3x, but I started at a 2MP resolution, so 2x is enough. 🙂

[rich olive]

Upscaling a 2k x2 takes forever on the T4 im using, so i only upscale 1k images

[rich olive]

What upscaler would you say is the best rn, i tried ultrasharp and i hate it, way too sharp, i found a website with loads of upscalers but im still not sure which to use

https://openmodeldb.info

[native knot]

Honestly, it totally depends on your source image, if you're trying to add detail or not, and the style you're going for. I'm not as versed with upscaling as some of the others here, but my current favorite tool is SUPIR...but I was planning on trying something with this iteration that I'm working on with SUPIR after the iterative upscale.

[rich olive]

I haven't tried SUPIR yet but it looks promising, I've only been using ultrasharp and some other lesser known models i found on that website

[mellow tendon]

The power to bend light!

I find an upscale usually gets rid of most sun rays, so that one is actually not too bad.

[native knot]

?

[fiery garden]

old fashion lady

[copper kraken]

freaky one on the left is my kinda style 😄

[keen field]

you're so vain, you probably think this song is about you

[copper kraken]

layerdiffusion?

[vital ermine]

https://civitai.com/models/365328?modelVersionId=408256

[prime bison]

Using ICBINP

[wet nacelle]

Is there a way to swap the model while generating? I'm using Forge.

[minor jungle]

I think no

[meager canopy]

That's similar to refining. Just set the denoise to whatever percentage you want each model to work on.

Not familiar with Forge, but if it can refine, you should be able to.

[copper kraken]

yeah, i think you can just use the refiner checkbox, and tell it to use a regular checkpoint instead of the refiner

[wet nacelle]

Oh interesting

[pallid path]

This grandma found a giant sized rat in her garden! 😲 I think it's a good idea! 💡

[stray warren]

It's her best friend now

[crisp owl]

Mouse needs a cookie, that's all

[minor pecan]

rodent friend will be alone soon

[stray warren]

Hush, he's sad now

[minor pecan]

two lightsabers

double the skill?

[minor pecan]

what!

[worthy torrent]

how do you generate images where is that dream menu?

[minor pecan]

#1047610792226340935

[worthy torrent]

i see

[gloomy lark]

v

[wise relic]

a big car

[hoary saddle]

anyone know where to find this?

[minor mason]

6

[crisp owl]

quagh quagh

[minor jungle]

thats a horse + duck, so we can call it a houck

[wide basalt]

Generate an illustration of a little girl in a forest style, characterized by watercolor transparency, aesthetic, fresh, and elegant features.

[minor jungle]

#1047610792226340935 bots currently offline, you can run it locally or use online services like https://seaart.ai

and stop posting it in all channels

[peak dove]

How do you register via email at seeart? It gives no options?

[minor jungle]

there is a discord option

[burnt shadow]

Dorse

[minor pecan]

[stray warren]

Nobody expected this when they said the aliens have landed

[stray warren]

cute

[cyan crown]

https://civitai.com/models/355154