🏞｜general-with-images

[smoky oak]

considering it took 5 seconds

wombo dream

vs SDXL with refiner

[oak osprey]

for these generators it almost even makes sense to cache the most common terms and just return one of the prev upscaled images

[smoky oak]

although, the top one is a very specifically made style for that exact look lol

[oak osprey]

"you want free img in 5 seconds? here u go, trash"

[smoky oak]

dream actually is just that fast, and I think its pretty impressive

[oak osprey]

pretty impressed with their progress

good job craiyon

[smoky oak]

now what happens when I remove the specialized style...

[oak osprey]

it improved as much as Bing Image gen did

a stunning portrait of jesters at the twisted carnival is the prompt i used for craiyon

[smoky oak]

oh I see...

[oak osprey]

not even anything spectacular

[smoky oak]

oh wait, thats not their diffusion model

thats the other type, I always forget the name

[oak osprey]

retrieval?

[smoky oak]

maybe?

its not stable diffusion

their finetuned ones are

[oak osprey]

Imagen?

[smoky oak]

no, not that, its been out for years

[oak osprey]

a GAN?

[smoky oak]

its like... clipping or something

it has the word clip in it

[oak osprey]

Clippy?

https://giphy.com/gifs/MicrosoftCloud-check-msbuild-microsoft-build-HZrx8kjIA7lyeTqXVM

Mini UnCLIP?

https://huggingface.co/docs/diffusers/api/pipelines/unclip

[smoky oak]

maybe thats it

I had a whole conversation about this before SD was even released

[oak osprey]

well that's Imagen

😄

[smoky oak]

I am trying refiner into refiner lol

ok so DON'T do that lol

[oak osprey]

do it like so many times it takes 24hrs for a single gen

[smoky oak]

it loosk bad lol

[oak osprey]

or not

can the base model do img2img? haven't tried here

[smoky oak]

it can

I do not recommend it lol

[oak osprey]

PFFT

[dense tapir]

is outpainting broken?

Apparently something has changed since the last time I used this feature so 6gigs is no longer enough mem.

yep, it is now rubbish. 768x1152 I want to add 192 pixels yet get a kublas error BUT I just did a full on 1152x1152 gen

128 works

[oak osprey]

@steel apex this is teh cornhole edition

in my prompt anyway

[steel apex]

I've been staring at this for a minute

It's a masterpiece

[oak osprey]

i have more

[steel apex]

Ford: If you're confused, so are our trucks

[oak osprey]

well this is a ram 6900

[steel apex]

"How much truck do you want"

Dodge: Yes

[oak osprey]

it has two truck cabs, so you can separate yourself from your shitty children

there's no back seat in the front cab

just two windows for some reason

[steel apex]

https://tenor.com/view/its-free-real-sate-gif-7215175

Truck guys be like

[oak osprey]

these were WIP validation images from fine-tuning

oh my god SDXL just finished its conceptualization of the ram 6900 cornhole edition

these are awesome

[steel apex]

Dooom

[oak osprey]

try this one

stunning++ photographs of jesters+ at the twisted carnival-

[sick pagoda]

opinions?

[smoky oak]

Man these SDXL weights are nanoying me, can't get images that look like the bots from them lmao

I wonder how massive their additional negatives and positives are

[split valve]

will it reach higher coherency? it cant resolve details like eyes or other small objects without screwing them up.

i just got this out of it

it actually nailed the eyes? how is that possible?

[reef field]

Could be RLHF effect too. Is the model they opened early up to date? The bots probably use the most recent iteration

I wonder if it's possible to finetune with GANs... for example, make a hand detector and train a GAN to tell SD what exactly is wrong with it

[smoky oak]

I have found a new workflow that seems to help with quality considerably

[split valve]

SDXL tigers.

[smoky oak]

I think I may be able to get some better ones with my new workflow in comfy

[split valve]

SDXL attempts fashion models.

[smoky oak]

SDXL run locally

[split valve]

not bad, except the detailing of the background... and the overall 1024x1024 detial limit

[smoky oak]

if you saw what it looked like when we first started, you would be amazed lmao

cause this model is shit if you don't baby the fuck out of it

this is what it looked like with tigers 24 hours ago

[split valve]

i cant comprehend how it can be improved... But that it can improve is a good sign.

its looking good now, it still is incoherent on things like hands, eyes, tails and such

[smoky oak]

Its the exact same model, its just a bitch and a half to get it to behave correctly lol

[split valve]

so like, getting it to diffuse correctly.

like it depends what we are aiming for, artistic or photoreal.. if it mixes elements of bad art and you're looking for photorealism then you will get crap!

[tropic shell]

Tigers are many things ❤️

Love the pics

[smoky oak]

No like, you have to use very specific CFG for each model, with a very specific amount of steps, with very specific prompting, with a very specific two model per single image workflow, and various other things

It's kind of insane how much of a pain in the ass using this model is, but I have hope that the official release will be far less dense and annoying to use lol

[split valve]

well its a complicated neuralnetwork right? so its going to take a lot to coax it to work correctly... Then do all that behind the scenes and you get clipdrop.io

Cat trio on the roof, surreal, masterpiece, octane rendering, focus, realistic photography, colorful

[tropic shell]

What are you prompting?

CATS!

[smoky oak]

That is true, but still, I've had to generate hundreds of images and try dozens of different nodes setups in order to get even remotely decent images out of it, luckily I did just finally figure out how to get the models to do 50/50 on a single image, and while it is a bit of a pain in the butt, my results are clearly improved

The main thing that really annoys me about all of this is that the bot makes it seem like you don't need negatives,

And then it will just be extremely easy to do, like just a few words required and you'll have great images, when that's really not the case

I really wish the team would share what kind of modifications they're different styles are doing, because if I had information on that, I would be able to recommend the base model weights considerably more

[split valve]

right... clipdrop must be doing negative prompting behind the scenes

[smoky oak]

Excuse my wonky grammar and weird messages, I'm using voice typing at the moment

[split valve]

those styles on clipdrop are interesting and quite powerful

[smoky oak]

Yes, interesting is one way to put them, extremely difficult to replicate the quality of is how I would put it

I have finally honed in on some form of negative prompt for decently realistic images, though it's still fails considerably with certain subject matter

[split valve]

i like their analog filter, and the isometric is interesting.

[smoky oak]

The artistic styles are honestly quite easy to replicate, it's just the realism ones that are a pain in the ass

[split valve]

oh right like photography and such

[smoky oak]

Yes, both the base model and the refiner are extremely biased towards digital art and painterly results, so you basically have to manhandle them in order to get realistic images out of either of them, let alone both of them in succession

And don't get me wrong, I'm not saying that this is bad, I'm actually very impressed by just how much quality I have been able to squeeze out of something that seems so bad on the surface, I just really wish that there was some clarity on SAIs side in terms of the positives and negatives they are cramming into our generation requests

[split valve]

tricks of the mind to get the brain thinking its photography... like DOF, glare, lensflare, framing, filmgrain

[smoky oak]

All of that is actually quite easy to get, the main problem is getting things that don't look like they were made through some form of painting or digital art, the post-processing style stuff is really easy, but the core imagery is hard

[split valve]

right i get you.

[smoky oak]

I do not see there being an easy to use workflow for the base model into the refiner implemented in UIs like automatic anytime soon, as there is a ton of nuance in the relations between the base model and the refiner

A to an extent I do like the ability to kind of lean the model from realistic to artistic without having many problems once you do figure it out, but that learning/growing pains phase is going to be a huge pain in the ass for the community, and I know that the vast majority of users are not going to want to have to fiddle around and find out which 50 words you have to put in your negative to not get images that look terrible

That said, when you do start aligning the right variables, it is a very powerful model, and I have a feeling that the 1.0 release is going to be even better

[split valve]

negative prompting should be made automatic then or even just click icons that represent combinations of negative prompt words

dont forget not eveyone is looking for photorealism.. a lot use it for artistic

[smoky oak]

The main thing that I do have to say is that using it as an image to image workflow, where you have the base model generate an image, and the refiner run a image to image pass on it is not the correct way to use these

You will get significantly worse results doing it that way, the correct way from my findings to do it is to have half of the image steps done by the base model, and the other half done by the refiner. The quality, coherency, and accuracy to your prompt goes up considerably when you do it that way. I just have zero clue how something like that will be implemented into auto

[split valve]

interesting...

[smoky oak]

Oh for sure, for artistic stuff it slaps, it's very easy to get phenomenal artistic results out of this model. And that regard I would say that it really is as easy as just throwing in a couple positive words, a couple negative words, and having great artistic images in seconds

[split valve]

but with artistic stuff its easy for the human eye to ignore what are actually mistakes or incoherence

with photrealism you really see what the model is capable of

[smoky oak]

Not even just that, they fundamentally look really good

[tropic shell]

It's always been stated that negatives are included with the bot. Just keep in mind the bot is a preview, not a finished product, like 0.9.

[split valve]

oh right so it does digital art really well..

[smoky oak]

I wish those negatives were provided so the community could know what kind of tedious amount of negative refining we're going to be in for

[tropic shell]

There are tons of great words you can use

[smoky oak]

As it is right now, my research on it is pointing to very tedious for realism and semi-realism

[split valve]

whats its base resolultion though? the upscaling looks like it needs improving,, all details look like reptile skin or alegator skin

[smoky oak]

its 1024x

[smoky oak]

if you are using clipdrop, yeah, their upscaler that runs at the end is absolute trash and makes the images look worse

[smoky oak]

Any examples of words you would recommend for getting better realism? I would massively appreciate it, and include it in my findings for SDXL

[split valve]

it is a trash upscaler. Why dont they use like diffusion based upscaling..

isnt that possible?

[smoky oak]

for sure

it doesn't even need the post processing, the images look good without it lol

[tropic shell]

Give me something you want to prompt

[smoky oak]

If I want a realistic landscape

For example, a photograph of a park at sunset, in a photo realistic style

That has still managed to be something I can't get looking even remotely realistic

[tropic shell]

Any particular kind of park? Place?

[wispy nest]

what sampler do you suggest in Auto for 0.9 api Sunny?

[split valve]

landscapes i was struggling with... foliage and trees it fails at

[smoky oak]

Lets say a park during autumn at sunset

[wispy nest]

for photoreals

[smoky oak]

its not usable in auto, so there is no way to know

from my extensive searching in comfy UI, DDIM seems to be the best

[wispy nest]

im using the api

[smoky oak]

ohhh, my bad in that case, thought you meant local haha

[tropic shell]

Heh. Give me a bit

[smoky oak]

alrighty, I look forward to some insight cause I have gotten a lot of things realistic, but not landscapes for whatever reason

also, if anybody wants to send me some prompts to try in local run SDXL, I could use some new prompts to research

positives only, realism focused

[boreal falcon]

i preferred the local one while i still used it over Dreamstudio

i actually forgot for a while that i used local SD at one point lol

but yeah, i liked it more than Dreamstudio

[split valve]

same prompt on clipdrop io... one is with no style one is with photography style

[smoky oak]

local just has wayyy more control

[boreal falcon]

it breaks down to usage case ofc but i prefer more control too

[wispy nest]

how did you guys get local? you mean a ckpt?

[boreal falcon]

Github

[smoky oak]

Yeah, I have the base and refiner ckpt's

[split valve]

i would run it local if i had a good card.

[wispy nest]

whats the link?

[split valve]

delv into all the deep controls

[smoky oak]

You have to apply for them

[wispy nest]

ah

[boreal falcon]

my VRAM isnt the best either but usable definitelly

RTX 2070 on laptop

[smoky oak]

8GB VRAM is enough to run SDXL with the refiner in comfy UI

[wispy nest]

thought it was somehow available to dl

[boreal falcon]

i even use it for rendering and simulations which is still good workable

[smoky oak]

not at the moment, for now its still in beta/researchers only mode

[boreal falcon]

for large/larger projects i will upgrade tho

need a whole new PC

[split valve]

Nvidia has been stuck at 8GB since 2016

its a joke

[smoky oak]

I am conducting tons of tests to see which workflows work best in general, I have found the best CFG's, best samplers, best scheduelers, best pipeline, and now just trying to understand the prompting better

[boreal falcon]

i will need the 24 GB VRAM

[wispy nest]

how much will it need not that much I hope

[smoky oak]

8GB VRAM minimum fro SDXL

in comfy that can let you run the base and refiner, in auto I assume it will be much more restricted

in terms of gen speed, same as 1.5 and 2.1 at the same res

only slow down seems to be the text encoders, as there are 2 of them now, and they are huge

they can take a couple seconds to process your prompts

[split valve]

if you think about it philosophically... A photograph is not always art. not by definition. But diffusion AI art generators are supposed to be art... So that is what it tends to be good at.. and to get it to only make an image look the same as a photo is kind of missing the point?

but artistic photographs,, yes...

people want art.. they dont want a random cellphone photo looking image

[boreal falcon]

for photorealism i might as well use stock images

[smoky oak]

you can see here the difference between Base, Base with refiner in img2img, and then 50/50 diffusion between base and refiner

as you can see, the 50/50 pipeline works the best for sure, but it is also muchhh harder to implement in auto

[wispy nest]

so does the sampler not matter on the auto upi version? like is it just processing whatever the default sampler is on dream studio?

[smoky oak]

(by 50/50, I mean half the steps are done in base, then it gets sent to refiner which does the otehr half of the steps)

[split valve]

middle one looks best.

[wispy nest]

whats refiner

[smoky oak]

SDXL is 2 models

the base model, which is 3.5 billion parameters

then a secondary refiner model that runs ontop to fix it up, that one is 3.1 billion parameters

[wispy nest]

noice

[split valve]

what happens if you run the refiner twice

[wispy nest]

so you like get the image, then put it through the wash again somehow? or it does it automatically

[smoky oak]

its just like running any diffusion model twice, just changes the details a bit, but doesn't refine them

[smoky oak]

you can do it 2 different ways

one is img2img, like high res fix

take this image, run it as img2img with the refiner

you get better results, but not the best

[wispy nest]

gotcha

[smoky oak]

other way is to diffuse the image 50%, like this

then let the other half diffuse in the refiner, which gives much better results

for example again:

[wispy nest]

how big can you go with the images before it like dupcliates the subject or whatever

[smoky oak]

base

refiner img2img

split

[wispy nest]

im just on my laptop now, honestly they all look great cant see much difference lol

but this is a smaller screen

[smoky oak]

as you can see, the 50/50 workflow looks much better on fine details and colors and stuff

[split valve]

it nailed the hair on the split. Its doing something right

[smoky oak]

also, bonus with 50/50, is its faster

[smoky oak]

yup! The refiner is really good at smaller details, the base is really good at posing and basic blocking out

base@wispy nest

split

[wispy nest]

yeah i can see it now

[smoky oak]

not perfect, but a huge improvement

[wispy nest]

although so much makeup is added to the second it also almost looks like a new person

[split valve]

The limit is surely the resolution and quality of the training data.

it cant get beyond that because how can it diffuse to it

[smoky oak]

thats not true, you can get 1.5 which is base res 512x512 to theorhetically infinite resolutions

[wispy nest]

like the first with better eyes would almost look like a natural beauty type the second a more professional photo shoot

[smoky oak]

once we have right finetunes, we should be able to do 4k with SDXL like childs play

[split valve]

interesting...

[smoky oak]

for example, some of my upscale workflow images in 1.5

[wispy nest]

I gotta say though its pretty impressive just the stuff I can see on clipdrop and the api. I wonder how much more 'wow factor' full 1.0 will have

[split valve]

so its got the intelligence to work out how to get more resolution, so its not just simply diffusing to a trianing image and then stopping

[smoky oak]

these are all 1.5

oops, last is a bad screenshot lol

[split valve]

So is clipdrop going to improve further as they do stuff in the background??

[smoky oak]

no idea

[wispy nest]

do they own clipdrop?

i dont get why dreamstudio isnt like featuring the best looking stuff

[smoky oak]

SAI owns clipdrop, yes

[wispy nest]

maybe I'll drop some cash there was gonna on Dream Studio but it seems its taking a bit to get 0.9 up there or similar to clip drop

[split valve]

clipdrop seems so limited with its 1024 res and terrible upscaler

[wispy nest]

yeah things look better on the api

[sterile temple]

@smoky oak can you try this one in comfy? change the two models of course

[wispy nest]

but clipdrop is still impressive as hell to me, just to see what the potential will be you know

[smoky oak]

oh lawd he crusty

[smoky oak]

like I said, I am looking for prompts to run local, which is as close of a look as you can get haha

[split valve]

clipdrop is nice, i paid for it because its £8 a month which is way cheaper than midjourney and 1500 images a day, at a faster rate than mj

it shows how things are improving since the early days

[smoky oak]

oh god I broke something

ono

every so often, certain seeds for SDXL get these super super weird pixel perfect splits

like this, excuse him being shirtless, apparently "handsome" means no clothes lol

[sterile temple]

i've seen split images in 1.5 too, rare

[smoky oak]

I have gotten it a couple of times

[split valve]

cant imagine what is causing that error

[smoky oak]

seed might have some random rounding error thing

oh mannn, this prompt looks great so far

man, when we can finetune this, its gonna be so fucken over for MJ lmao

[sterile temple]

did you try the kitty pic?

[tropic shell]

A bench at a park, near a river, autumn trees in the background

On Clip, it comes out as:

[smoky oak]

yes, but what is the negative?

[tropic shell]

For specifics on 0.9 on what you should do, for research, I would recommend referring to the appropriate channels for such things.

That said, consider the following:

I don't know how you're prompting, but start small. Don't overprompt. And think logically.

If you want a tiger in the snow, you might say, "photograph of a tiger, professional animal photography"

I'm not prompting this rn, but as an example, I would start exploring more options.

"Photo of a tiger, walking in the snow, professional animal photography"
"Photo of a tiger, walking in the snow, view from the front, professional animal photography"
"Photo of a tiger walking in the snow, detailed fur, view from the front, in a field of snow, midday, professional animal photography, in the style of National Geographic"

(Yes or no on the detailed fur.) You can always use more positives, too, whatever.

In terms of painting, I would recommend trying different painting terms. What's in a painting? Using cartoon, etc, can work, such as my wip demonstration on hands in my vol 2 guide.

Honestly, I've written a ton about negatives already, so if you have questions about negatives, I'd refer to my documentation

[smoky oak]

That was the whole point of the question lol

[split valve]

eyes and fur looks pretty sharp except for the front legs look odd., the background foliage is nailed.

[tropic shell]

So there's plenty of phrases already in there

[smoky oak]

the positive prompt is easy, its the negative

literally 0 help on the negative, the BIGGEST bottleneck here lol

[smoky oak]

Slowly getting better haha

[tropic shell]

Heh heh

Really, though, check out my documentation

[smoky oak]

these don't look very realistic either, so I assume it may just be a limitation of the base in that case

[smoky oak]

where can I find that?

[tropic shell]

https://atypicalconsortium.carrd.co/

[smoky oak]

my nodes are such a mess to get this model running its best haha

[tropic shell]

It's a bit scattered, but all my links there

The Friendly Guide has info on negatives; vol 2 has some more info in the prompting section you might find useful

[smoky oak]

ok, biggest thing to say, 1.5 and 2.1 negatives do not work at all in this model

so that is all going to have to be changed for future things, myself included

I have had to completely change negatives to get this all working

oh also, negatives need to really be driven

positive prompt is all weight 1, negative is weight 1.4

[split valve]

im going to post a bunch of my clipdrop images.

[tropic shell]

Do it.

You know, maybe I should, too, lol

[tropic shell]

Gorgeous!

[smoky oak]

oh wow

[tropic shell]

Yes ❤️

[smoky oak]

oh man, ok

the img2img version did a way better job on that rose prompt

that one looks very good

roses look much more... rosey lol

[split valve]

alll... i havve tons more

no just clipdrop

[smoky oak]

feel free to not share that information ever again, thanks lmao

[wispy nest]

I love you too

[smoky oak]

[wispy nest]

Always make the best out of life

[smoky oak]

ohhh, looks like clipdrop has the same issue with being bad at water color

I had to fight to get SDXL local to do good water color

[tropic shell]

This is actually supposed to look like this (with the way I prompted)

It's a type of digital watercolor style; it's not bad--it's just a style of art

[split valve]

one thing ive noticed on clipdrop... versus midjourney v5... the images when taken to photoshop from clipdrop are already tuned, if you run auto tone, auto contrast etc nothing will happen they are already good.. But midjourney is washedout bad contrast and bad color

[smoky oak]

ah, did you ask specifically for digital water color? Cause normal water color shouldn't look like that haha

[tropic shell]

There are many kinds of ways to do digital watercolor, and various kinds of programs that do it

Some people may use overlays/brush textures

[smoky oak]

base SDXL with no refiner does real water color the best

[tropic shell]

Some people may use different brush systems

The look can vary from program to program, as well as the ability of the artist

Yeah, real watercolor is def beautiful for sure

Traditional art can be hard to replicate, for sure.

[sterile temple]

@smoky oak hey can you try open a new tab of comfy, download this image and drag it into the interface for comfy to load the workflow. then change the models for the base and refiner to the sdxl ones, and try running it?

[smoky oak]

sure

what in the hell is this workflow lol

100 steps OMG

I will need to change a lot to get this working for SDXL

[sterile temple]

don't change 😄

[smoky oak]

okaayyy

[sterile temple]

I copied it from somewhere

[split valve]

just got these out of clipdrop

[smoky oak]

jesus christ this is slow as HELL

gonna be like 5 minutes of genning for this lmfao

first image is crap-

[sterile temple]

i'm missing a few custom nodes, but i did my best to copy a workflow from a image

[smoky oak]

I am actually hurting to understand what this is trying to do-

@sterile templeWhat was this uspposed to be? cause its... terrible lmao

[sterile temple]

😄

[smoky oak]

gore-juss

[sterile temple]

beautiful

[smoky oak]

alright, thats enough of that lmao

[tropic shell]

I love the ears

[tropic shell]

In a way, it's kind of an alien cat

[smoky oak]

@tropic shellAny idea why SDXL randomly just fucking kills itself? lmao

[tropic shell]

Like I said, report any problems/feedback to the appropriate channels!

[split valve]

now for creatures that dont exist.

[tropic shell]

Ahahahaha! These are fantastic! I love the bottom left--what fun.

[smoky oak]

witch cat!

[tropic shell]

With very, very soft silky fur.

[wispy nest]

Insane

[smoky oak]

what are the chances lmao

[earnest solar]

The character from mine is more wuxia than sakura romance

[smoky oak]

I was just trying to see how well it does different races

wonder how it handles people of color

pretty well actually, nice

[earnest solar]

depends on the model

[smoky oak]

its SDXL

[earnest solar]

some anime models have a really hard time doing African and Jamaican phenotypes, for example.

[smoky oak]

Yeah, but SDXL is a general model

it does everything

[earnest solar]

It also helps if you ask for actors with the specific phenotype you're looking for

[split valve]

yikes

[smoky oak]

starting to feel like my SDXL implementation is better than the one on clipdrop haha

[split valve]

Just came up with an idea... ask Chat GPT to come up with a list of extinct animals like megafuna, such as the saber tooth tiger. And get it to write prompts to describe them. Then feed SDXL the prompts.

[split valve]

DODO photographed in the wild.

Also these clipdrop images now seem razor sharp.. like its improved in the last few hours

[split valve]

Wow look at these!

[oak osprey]

that's fucking nasty

she's depicted as a 10 year old, pedophile

[stoic pivot]

Some time I think my old SD with the older models is better than SDXL, mostly when it come to realism, it lean towards the odd and boring look MidJourney have, the images have god quality but are Ultrarealistic.

Ultrarealistic is not the same as realism, it is how we want and think reality look like but with greener grass and so on, sometime also called Hollywood-realism by some, this is a great art style but sadly it started to be used as synonym for Realism in Ai and I thing that is why we nor get rather bad images when we use Realism or Photorealism in our prompts.

Tags for those images was, "photorealism, black and white analog photography, high ISO film, film grain, side portrait, photo of an older female grizzled rock artist in a leather jacket holding an electric guitar with wrinkles on face detailed rough skin tired look, dark room, raster paper print"

[stoic pivot]

Thank you, the point was not to say SDXL is bad, but that it now is better on illustration and fantasy, I often like a bit raw realism. I love those images, So I will try again.

[split valve]

right.. this was just a first try.. it seems they improved stuff behind the secens today, the gernations are looking better than ever

[oak osprey]

you know, it's interesting. i've got the weights in my local pipeline now and i've been comparing them. i do not use 1.5 anymore, because it's really poorly performing on zero-shot image gen with high details.

so i compare against fine-tunes of 2.0 and 2.1 that i made over months. like, very carefully fine-tuned.

but my 2.x models are heavily slanted toward realism. like, it's almost all they can do. trying to get fantasy stuff out of them really isn't easy, and when you do get it, sometimes it just looks silly

i also don't have the SDXL refiner model working. i'm fine with that because i'm interested in extensively testing the base model on its own, as that is what most people will be capable of running.

i think fine-tuning SDXL base for realism will take it a lot further than 2.1 / 2.0 are capable of doing, but as it is now, it is not there.

when you add the refiner to the SDXL workflow and tweak its settings properly like Sytan has done, the results are just beyond incredible, without fine-tuning

[split valve]

so SDXL has potential? do you think it will compete with midjourney toe to toe?

[oak osprey]

i don't get why people worry about that, lol

do you mean, will StabilityAI be more preferred as a platform than Midjourney for end-users that find Clipdrop/DreamStudio and compare it to MJ?

[green socket]

I think the majority of users will use the easiest program with the best outputs

[oak osprey]

a lot of people will simply avoid MJ because they don't see the point of using Discord, a clunky interface

[green socket]

Yeah I tested MJ before and the server lagged like crazy for me

[oak osprey]

sucks having to type the same crap every time you do your prompt, the --v 5.2 --ar .. blah stuff

[green socket]

Too many people in the room. It's probably worse now.

[oak osprey]

i think SDXL will be more popular over Midjourney for those who find it because MJ has more collective mindshare right now it's the one people assume you use

to be honest right now i think Bing image gen is more popular than MJ

like, stuff is changing for their dominance and not necessarily due to SDXL

[split valve]

I saw that google imagen is being released, but it looks like shit. Yet at the time they boasted that it was so powerful it needed to be kept a company secret and nobody could use it

[oak osprey]

seems fine

Imagen uses a large frozen T5-XXL encoder

sweet

no wonder it can do the absurd demo stuff.

We show that large pretrained frozen text encoders are very effective for the text-to-image task.
We show that scaling the pretrained text encoder size is more important than scaling the diffusion model size.
We introduce a new thresholding diffusion sampler, which enables the use of very large classifier-free guidance weights.
We introduce a new Efficient U-Net architecture, which is more compute efficient, more memory efficient, and converges faster.
On COCO, we achieve a new state-of-the-art COCO FID of 7.27; and human raters find Imagen samples to be on-par with reference images in terms of image-text alignment.

this is why it's good that they're releasing it

don't worry about "what it looks like", it's just preconditioned priors

you can fine-tune it.

[split valve]

how good is imagen realistically compared to SDXL?

[oak osprey]

probably depends what you're trying to get done

[split valve]

personally i like amazing quality output, like some of the stuff ive been posting today im impressed with for SDXL... But some people seem to think if it can make for instance a dolphin wearing a wizards hat surfing on a sea of candyfloss and smoking a spliff, then it must be really good

[oak osprey]

yes, the ability to generate new content that has never existed before while retaining coherence is groundbreaking. it's doing all of that with a small diffusion model / unet, and a very large text encoder.

[split valve]

So its technically impressive, but is it quality art that people want.

[oak osprey]

it does both..

[split valve]

brb im going to test SDXL on the prompt i just mentioned

[oak osprey]

what, what do you mean "google is releasing imagen", as in, the weights for it?

DeepFloyd is an open source implementation of Imagen

[split valve]

i just heard about them realisting it, did not hear all the news just a quick blurb.. i recall at the time they were keeping it very close to their chest and saying no plans to ever release it

[oak osprey]

well Imagen is from last year and in that context, it is possibly one of the most impressive models

they're using something like a 900 million parameter unet on top of T5 XXL

SDXL uses a 3.5B and 3.1B combo pipeline when the refiner is involved

not clear how limiting that 900M parameter is when it comes to fine-tuning or vast capabilities but the text encoder really seems to contextualise so strongly that it could help overcome that

[split valve]

so it understands prompts better than all others probably?

[oak osprey]

it's not so cut and dry like that

at some point these models stop being comparable in those terms and it's mostly just subjective concerns or prompting differences

like a very good 2.1 fine-tune (text encoder not frozen) and SDXL (text encoder frozen) are almost neck and neck for a lot of things. but there's subject matter where they diverge, and either one or the other "wins" for my desired prompt results with my prompt style.

you can still get "winning images" from the "losing model" for that subject matter but it becomes a lot more effort / time required

not everyone thinks in the same way to write teh same style of prompts, so one model might work better for one person and worse for someone else.

[split valve]

But what is the future, will they just pair them with LLMs that are beyond GPT-4 in power so that the context of the prompt is super well understood?

[oak osprey]

i was tempted to just split the 4 images output by my discord bot between a top tier 2.1 model and SDXL

[split valve]

like i hear GPT-5 will be trained on images and video

[oak osprey]

so you always get both

i don't know, man. i don't worry about future trends, it's all just hype designed to attract investors

i also just don't know enough to say with certainty

you can grab embeddings for a given prompt from GPT4's API. and assuming cost and speed weren't a concern with this, i'd then wonder like, how much data do you need to train the unet on to adequately be capable of representing SUCH A HUGE EMBEDDING?

i don't know how that relationship works out. to a naive observer, i'd assume that more and more training data is required for larger models

when a prompt goes in to the text encoder (and it has seen it during its training) but the unet never observed that concept during its training .... it's not pretty. it's confusing and noisy outputs.

[split valve]

This is too complicated for me to understand.

[oak osprey]

text encoders capture "relationships" between terms in the prompt and what features appear in images that the prompts corresponded to

[split valve]

These AIs are getting too complex. They dont even know exactly how the nerual network is working to output some text. so who knows what it would do if they paired an LLM with diffusion image generator

[outer spear]

the hardware used to train sdxl costs 15mil?

[split valve]

oh text encoder sounds similar to an LLM then..

[oak osprey]

it's a component of one

[split valve]

and if they really do train a GPT-5 on images and video then perhaps by default it will be able to generate images better than all the current AIs

[oak osprey]

the text encoder is pre-trained on billions of image-caption pairs and develops some kind of internal "blocks" of weights that correspond to these concepts, and these weights are numeric. they're a huge array of numbers that get used as input to the unet for guiding the diffusion process.

[split valve]

Because of how it will understand contexts of thigns in the world

[oak osprey]

training the unet, teaches it how to be guided by this conditioning input

the unet sees billions of image caption pairs, but they might not be the same ones that the text encoder were trained on

the text encoder is pretrained and frozen, and is not modified when stability is training their unet

so this representation remains fixed, and the parameters of the unet are slightly adjusted again and again over millions of iterations. and when i say slightly, boy do i mean slightly

[split valve]

ok this is hurting my brain

so the text encoder is trained and frozen on image caption pairs. Then the Unet is like a neural network that learns stuff over millions of iterations..

[oak osprey]

yeah you'd have to teach the unet to understand prompts directly without this pretrained encoder

that'd take a lot longer and be less effective

[split valve]

and how does a human brain do it?

give an artist a prompts and they draw it how

[oak osprey]

we're also taught how language works early on before we're able to absorb many other concepts

[split valve]

true.

[oak osprey]

children are taught from completely useless little parasites into fully functioning adults through iterative training

when they do something wrong, slight correction each time will help them minimise their loss over the predicted result

can't get angry at them and try to make them learn too quickly. same applies to ML

honestly, it's not terribly different.

ever met someone who just isn't quite right and then you meet their family and you're like "ohhhh"?

training data.

[split valve]

you have a point.

[oak osprey]

i've been thinking about this a fair bit over the last month

lmao

[split valve]

This AI stuff though its really starting to take off.

[oak osprey]

yes

that is more than we can say for humans 😄

[split valve]

oh we are in decline..

[oak osprey]

all these years of evolution and i still have no wings

[split valve]

we're going away!

[oak osprey]

obsoleted

[green socket]

Planes exist

[oak osprey]

no they don't

that's something an AI model hallucinated

[green socket]

Are they just large gov surveillance like the bird

[oak osprey]

correct, the myth of birds mirrors this

don't you hate it when you get Deja Vu?

why can't they stop cutting the hardline

[split valve]

what is this matrix stuff

[oak osprey]

@split valve but with that context, you see how a huge text encoder helps with image gen?

it helps but i think that's only if the unet can be adequately trained on all of the features the text encoder can represent

[split valve]

oh yea its needed to make it efficient.

[oak osprey]

i've actually asked the SAI devs this question before, and did not get an answer

it's possible it's one of the things we don't quite know yet

also it's why it kills me that we can't see CLIP-L's training data/captions

[split valve]

probably a company secret.

[oak osprey]

seems to have been trained on less data than other encoders, but possibly higher quality images AND captions.

@split valve yeah that's "Open"AI

training seems to be going in the right direction, lmfao

[oak osprey]

can get some really weird stuff out of SDXL

[split valve]

man the stuff im getting now

[oak osprey]

ever wonder what SDXL gives for unconditional guidance?

these have no input prompt

[split valve]

strange

[oak osprey]

the landscape seems random af

[split valve]

wow no input so whats guiding it

[oak osprey]

ahahahaha

RNG

[split valve]

random noise?

[oak osprey]

yep just diffusion without text

[split valve]

oh it starts with a random noise seed and just diffuses what it was

?

but these are not actual training images

[oak osprey]

it does what it always does which is start with random noise and then remove it successively and get some features from the data distribution

[split valve]

yup

[oak osprey]

this just isn't weighting the results by the text embeds

it's interesting that the landscape came out as the 4th image in the batch, i didn't do anymore batches of that

some interesting results from my prompt randomizer

[split valve]

sharp like 3D renders.

[oak osprey]

Generate a fascinating image prompt with stunning quantum caustics experiments++ featuring suspended ferrofluids++ and fluid dynamic based solids--!

[split valve]

if only they unlocked NSFW images...

its limiting the art

[oak osprey]

it's a little hypersensitive on the filter considering how incapable the model is of NSFW

[split valve]

This stuff i could not generate on Midjourney because of the keywords

but imo this is amazing art results

so allowing keywords even if we still have to have a NSFW filter, allows for more artistic expression

oops added the keyword "melting"

what is the token input limit?

[oak osprey]

77

https://giphy.com/gifs/deal-with-it-sunglasses-miami-vice-1jnyRP4DorCh2

[split valve]

how are you getting those images?

[oak osprey]

my discord bot

[oak osprey]

@split valve try Generate a fascinating image prompt with stunning quantum caustics experiments++ featuring suspended ferrofluids++ and fluid dynamic based solids--! as prompt

GPT 3.5 made that masterpiece

[split valve]

cant wait for the total realtime versions of this.

where a scene is defined by keywords and the AI drives the keywords

like so it can render every material with diffusion

That would be the future of video game technology

[boreal falcon]

if that one even comes

right now it doesnt even come close at all to any of the art programs

including render ones

and those that are advertised for 3D end up being a big mess for them to be even considered for game development for example

[split valve]

would probably need quantum computing

[boreal falcon]

the topology is horrible

[split valve]

oh yea.. text to 3D is just like garbled mesh of triangles or ngons

for now

[boreal falcon]

i mean there are some more or less click and go retopology programs out there and plugins for major softwares

but a software that will generate you a clean 3D model and even render and materials is something different ofc

[split valve]

i dont think the future is 3D models or topology for that matter.. i think it would be diffusion based

i mean it would be 3D but not as we know it

like a 3D fractal program that runs with no actual genometry

[boreal falcon]

i wouldnt want to gamble and wait for such a potential thing and im already a bit too far gone to do that now tbh

with too far gone i mean the investment i made for this hobby 😄

[split valve]

i see.. Well im just waiting for the total immersion video games..

not that im a big gamer, i just want to try the total immersion stuff if it comes out.

[oak osprey]

https://www.youtube.com/shorts/fNKBqx5mNjI

[split valve]

like black mirror

[boreal falcon]

when you spend up to 3.000+€ per year on creative programs and plugins you dont "want" to then just drop it for the sake of hoping for an AI to do it all 😄

[split valve]

omg

[oak osprey]

immersive video games are already here

[split valve]

not in my oppinon.

i want the matrix

indestinquishable from reality

Thats the next holygrail.... photorealism is here already.

[split valve]

i see the predicament.

Well i do some 3D graphics stuff on the side, like sell 3D models on cgtrader... but im getting bored of it a bit.

i've got a skill AI will not replace.

even when the jobs dry up

[boreal falcon]

nice 😄

im not that far yet, but i will eventually end up making money of this too

sooner or later i might invest into a 3D printer and print and paint stuff and sell them

but one step at a time ^^

[sick epoch]

rate this image please. i rly struggle with ram so its always a gamble if SD is going to finish the image and this is the best image i got so far

[oak osprey]

darth barbie

[split valve]

This stuff blows away midjourney... ive not seen this type of stuff with MJ

[oak osprey]

on a technical level i see some issues but on an artistic level those issues work for it

it's interesting how those two worlds don't always see eye-to-eye

[split valve]

its messing up the lips sometimes and other features but artistically i love this stuff its what i wanted

[steady skiff]

i need this prompt in my life

[split valve]

lots of erotic keywords

Midjourney would not be happy

is this taking it too far?

caprice

[oak osprey]

looks like what happens when you ask for forbidden NSFW tokens in local SDXL too

it's not pretty

[split valve]

forbidden = art

[oak osprey]

i'm kind of glad about that because it lets me confidently open up the bot's use to SFW discord servers and not have to check inputs

[split valve]

Its not once flagged NSFW so its self filtering somehow

This is the limit of its sexyness

[oak osprey]

pretty much lmao it can do more but i don't know how to ask it to, it just kind of happens when you don't want it to, and then it's gone like a fart in the wind

[split valve]

true.. its flagged NSFW before when i wasnt trying

[oak osprey]

it just flags pretty faces too

the NSFW scorer is so conservative it's pretty sad

[split valve]

erotic keywords but SFW

[oak osprey]

left = before SNR gamma fix, right = continued training after comfy helped me fix my code

that's zoomed way in

contrast is what it is, it's an ongoing battle but i like the improved details/blending

hasn't been long enough to really say what'll happen 😄

[split valve]

less noise on the skin, smoother more realism?

[oak osprey]

yes

[sacred fulcrum]

When you're too high.

[oak osprey]

prompt?

[sacred fulcrum]

The prompt is: A young woman, Martin Parr

[oak osprey]

my 2.1-v flex model

SDXL 0.9

have to do a whole bunch of text embeddings to make 2.1 that good

i'm not even using the SDXL refiner

ah, right, it needs a higher CFG

2.1 with higher CFG - the same prompt won't do as well between the two 😄 need a lot less prompting with SDXL

think it gave her a bit of a beard

[split valve]

these images are better than all my midjourney V5 stuff

midjourney cripple themselves by banning most keywords

[shy fern]

cancelled my MJ sub after they went crazy with it when v5 launched

[split valve]

it outputs more nude stuff but i cant post it here

[shy fern]

is everyone using the clipdrop 0.9 sdxl?

[sterile temple]

I've been using the dreamstudio api. Some people have research access to the checkpoint and have been trying to get it working locally

[shy fern]

are they using user response to fine tune it on clipdrop?

[sterile temple]

not sure about that one. it would have to be using different methods for the 4 images, otherwise it's just 4 different seeds

[shy fern]

they need to enlarge the text input box on clip drop, its suffering

[oak osprey]

local access

[oak osprey]

supposed to be a dumb 1950s baby welding without PPE

[shy fern]

I wish there was a paid option for running locally, I cancelled my MJ sub a while back due to aggressive censorship. I would rather pay stable foundation $30 a month for constant updates and stuff to run it locally

[split valve]

im sure all that will come soon with version 1.0

[oak osprey]

it will

for free

[shy fern]

they need to optionally monetize running it locally somehow to incentivize development

[split valve]

using banned words outputs way nicer stuff.. even if its not nude

[oak osprey]

oh nooooo

that happened to me once but never at that scale

i was in the grocery store

[shy fern]

one good prompt I have found with it is "intricate technical blueprint"

[split valve]

my prompt is crazy

its got some interesting artists in it though which i can control it with

its madness how much you can control it with the right words

[oak osprey]

baby elon dreaming of space

[split valve]

now they're asian

anyway ive got to go to bed its 3:40 AM here

[oak osprey]

this thing does that

[split valve]

yea i lost 5 hrs the other day

[shy fern]

3am best prompting time

[shy fern]

asking AI to make instruction manuals on how to make humans is funny

[livid ermine]

A friend took this helicopter shot of her house

Made a few adjustments (hopefully that'll impress her enough to let me sleep with her)

[shy fern]

lol

[oak osprey]

how to build a car out of meat

[sacred fulcrum]

Prompt: "A list of artists and styles, extremely detailed"

[cyan snow]

Is SDXL still gonna release in the middle of this month?

[terse slate]

Barack Obama in Ankara clothing, high resolution, bokeh

[copper lotus]

Inspiration from a recent trip to Japan

[fossil spire]

Anyone know how to get this style of hair?

[outer spear]

If i wanted to make a lora of a person, would i make it on the base model or a finetune like absolutereality?

[reef field]

I tried many times and the best results, suprisingly, come from training on the very base 1.5

When using realistic finetunes the faces become weird and deformed, on 1.5 they're always clean

Not sure why that is, I'll try more but so far it was like that consistently. I use tags for description, they apparently work better than more natural prompts

[reef field]

What I noticed is that when I train on a custom model the output on other models is noisy, like if it's a poor mobile photo. No such effect if trained on SD 1.5. Maybe because most models are merges and training on them doesn't transfer well.

So far I've always had the best result for a realistic LoRA on base 1.5 and for cartoon/anime on NAI. The information I found is very contradicting, some suggest to train real people on realistic vision, I did that and the result is very noisy and blurry while when trained on base model it's crisp on any photorealistic model.

[wispy nest]

if I wanna mix Chris Evans with Ben Affleck its (chris evans:ben affleck:0.5) right? im getting decent results but want to make sure its the ideal prompt so to speak

[sterile temple]

in A1111?

if it's A1111 I think the syntax uses square brackets

[wispy nest]

I think either work, but just wanted to make sure thats still the best way to try and merge two people

[reef field]

Easier to interleave them with [chris evans|ben affleck] though it's a different way

In this case it would change the prompt on every step while with ::0.5 syntax it will render the first half of steps with one guy and the rest with another

[split valve]

is there a reason it disfigures fingers and hands every time?

[wispy nest]

cheers ill try this later

[reef field]

There's an official emoji here about that

Few reasons, I guess: low latent resolution and poor descriptions

[shut sinew]

Also the complexity of hands

a head will look like a head from every perspective, but hands man

[reef field]

Try upside down pose, also a nightmare material

Still a head but unusual angle, and it all falls apart

At least some typical hand shapes could be generalized such as holding/grabbing objects or making common gestures

But either there's not enough resolution or parameters or both

I rather wonder why the number of fingers isn't consistent. Rarely you get a second nose or third eye (basically never) but getting 4 or 6 fingers is a norm

[split valve]

fingers ruin some great images every time today

[split valve]

I might have an idea on how to fix the hands problem

if the model can be weighted or biased towards certain artists, then would suggest bias towards all artists known to have painted realistic hands and fingers..

figers of doom.

[sterile temple]

like Andrew Loomis

[sterile temple]

the model needs to read this book 😄

[split valve]

SDXL more sexy than midjourney...

SDXL has some serious art skills

[sterile temple]

sketch of hands by andrew loomis 😄

[split valve]

noooo. erase andrew looms from the database then

[sterile temple]

supposed to look like this

[split valve]

sketches of hands and fingers probably confuse the model. Because it sees it as kind of transparent

needs to be high quality photos of hands fed to it. The community should all take photos of their hands and send them in.

until the model finally learns

[sterile temple]

I think they've taught it certain poses where the hands look okay

[split valve]

it draws everything else like a god.. but when it comes to hands...

[slow fern]

hey
what the hell is happening on the subreddit?
bots? there was like no activity whatsoever and that is a ridiclous amount for any server , even the biggest subreddits dont even come close to that
the usual ratio is like 200 joined to 1 browsing not 3 to 1

[split valve]

?? because reddits API is screwed up ? i heard its going away

But it could just be because SDXL is so awesome and its images more sexy than Midjourney

[wispy ether]

google earth is scam

[earnest solar]

why, you paying for it? cos then you really got scammed.

[wispy ether]

its free????

[earnest solar]

lol

[wispy ether]

https://tenor.com/view/nerd-emoji-nerd-emoji-avalon-play-avalon-gif-24241051

[stoic pivot]

I'm so sorry for this, I just asked for a random meme that make all happy.

[split valve]

seriously

[split valve]

how did you achieve the seasons? did you use the same seed>

[iron field]

I used Canny in ControlNet

[iron field]

https://media.discordapp.net/attachments/862083427775021106/1125480234196684840/00052-190077392.png

[split valve]

amazing that its combining so many photographers, illustrators and painters so well. including Roberto Bernardi and the term melted candy.

you can probably also see David LaChapelle influence.

The origami filter as impressive as ever

If this is just the start for clipdrop then its going to blow midjourney out of the water.

same prompt with line art filter

[boreal falcon]

MJ has its own niche of consumers tho

While ofc it shares part of the community with SD (and Firefly and some others)

[split valve]

how do you mean

[green socket]

Probably people who just want stuff that's easy

No install, easy prompts usually work

[split valve]

midjourney you have to type a lot of stuff. where as clipdropIO is like super accessible and dumbed down... the prompting seems to be even better than MJ.. in terms of how much i can put into a prompt

and it will keep understanding it very well no matter how many artists or styles

AND, SDXL is less locked down and less censored

and its way cheaper and way faster

[green socket]

Yeah, but it's more recent than MJ.

[split valve]

it has limitations at the moment,, but if it keeps evolving then it willt ake over

[green socket]

Hopefully, it will take off

[split valve]

Stability AI say they will have continuous improvement on SDXL

And new models are being trained like SD3

[green socket]

I find clipdrop to be a lot easier than MJ is - especially since the website doesn't lag like the MJ discord chat does

But it hasn't been around long enough to get the name recognition that MJ has

[split valve]

i've seen some quite major AI youtubers putting vids out about it. Saying MJ is in trouble

for some reason the youtubers tho always show clipdrop in its unapaid version saying " oh theres a que" and stuff like that... yet they log into midjoruney to compare it. (even though midjourney is 4 X the price

[green socket]

"oh no this free thing is slower than the paid service!!!"

[sacred fulcrum]

Prompt: "codex gigas"

"codex gigas girl"

...

[boreal falcon]

MJ has its advantage that some users capitalize on

Instead of using SD

Sometimes they combine some of the AI art tools tho

As someone already mentionedy ease of use and also no need for a decent PC and they use MJ on mobile too

Stable Diffusion definitelly wins in terms of controls and customisation

Also its free

[oak osprey]

@marsh wadi

my issue actually isn't hands

it's just that characters look like a mule kicked them

bob eggleton vs darryn eggleton

willem dafoe vs greg egan

aqua teen hunger force version

steve buscemi style

bubblegum vs halloween

easter vs christmas

[oak osprey]

christianity vs buddhism

[oak osprey]

africa, canada

[split valve]

using a mix of photographers, illustrators and hyperrealism artists then using the clipdrop filters

[oak osprey]

i just use bob eggleton cuz he thinks AI is cool

peter gric is in sdxl

[oak osprey]

peter helck is

try peter max

[split valve]

im addicted to the pretty pink girls at the moment

will try other topics tomorrow. when my credits renew

[oak osprey]

then Peter Mohrbacher

[split valve]

I think SDXL can combine more styles together than MJ

[daring ginkgo]

Beautiful

[smoky oak]

I am now confident in saying that my local made SDXL images can competently outperform the clipdrop bot results for realism

and thus bitch slap MJ V5 in realism lol

specifically for human subjects, but not limited to just humans

[proud dagger]

now imagine what it can do once the wider community starts training variants of it owo

You find the perfect settings?

[smoky oak]

new workflow lets me go to 2048x easily, these images are all crops

[smoky oak]

my workflow is getting insane lmao

getting a little silly goofy

[proud dagger]

as a professional AI data analyzer, I posit that the top row gets better results than the bottom row

[smoky oak]

yeah, I was just comparing 4 different workflows at the same time

or rather, 5

there are 5 chain of events going on there, and thus 5 final images

but yes, the top one is the success row

2048x in SDXL

mucho detailed

[wispy ether]

Why so big

Is sdxl slower or faster than normal sd

[proud dagger]

in raw time for image, SDv1 is faster... to give you a very small, relatively low quality image.
in a fairer comparison, at 1024x1024, SDXL is equivalent or slightly faster (when running just the base) to produce a higher quality image.
If you toss in the refiner stage, SDXL becomes slower again but at even higher quality

[smoky oak]

SDXL is close to the same speed as 1.5 at 1024x1024

tho, if you use it fully, with all of the different clip models and their included nodes, it does slow down considerably

but the output quality is well worth it

[smoky oak]

god these are cursed lmao

[split valve]

reminds me of my Blender node systems.

[smoky oak]

pingy pingy

oh yeah, my blender nodes were insane when I still used it, but I don't use blender anymore

[split valve]

I still use Blender but im currently bored of it

its like whats the point

its a great software but i have to have a creative thrist and be into it.

They say that SDXL will continually imrpove... Does this mean that coherence will improve and it will draw hands and fingers better?

[split valve]

And absolutely no jpg compression artifacts... Thats the beauty of Generative AI

[smoky oak]

1,5 can already do some very good hands, so I am certain SDXL will get even better at it

[split valve]

So you're telling me theres a chancec?

[smoky oak]

funny you should say that, cause I compressed it to a JPG lmao

[split valve]

lol must have been at quality 11

[smoky oak]

oh for sure, if 1.5 can do really good hands (which it can) then SDXL will be able to in no time

[split valve]

i just have no idea how the development of the model works so i was worried the bad hands was just baked in and would never improve

[smoky oak]

Have you ever seen base 1.5?

[split valve]

yea ive used it i just cant recall specifically what the hands were like

[smoky oak]

the whole model is bad lol

this is base 1.5 trying to do a wolf

and we were able to finetune it to this level, and even beyond

so if SDXL is as good as it is now, and we can improve it even half as much as we did with 1.5, we will have perfect hands in no time

[split valve]

i notice though if an object is further away it cant get the detail

[smoky oak]

That has been a big problem now for a while, though SDXL is 4x the res, so you have to get further to lose coherence, and even then it can do pretty damn good

[split valve]

some random old 1.5 stuff of mine

[boreal falcon]

Wdym whats the point 😄

Some people even use that SD plugin for Blender ^^

[split valve]

Sometimes when i see all the AI generated stuff i think whats the point of doing it the hard way in Blender... Even though i will admit AI has not caught right up yet

[boreal falcon]

Oh i could write a novel on why someone would use the "harder" one ^^

Gotta respond later again my pause is over, maybe in between 😄

But you can write

[reef field]

One of the reasons might be that some people love the process and some love only the results. AI takes the process of creation away and mostly replaces it with selection and fiddling with settings. To each their own I guess!

[boreal falcon]

Both result and process are a reason why i stick to "the hard" one

I do use generative AI partially tho

[split valve]

Nah i just have like a creative burnout currently for blender... Probably i feel its not getting anything achieved and not making much money so i think whats the point...

Dont get me wrong when i was learning it and getting back into it it was a lot of fun

[boreal falcon]

Its not for "everyone" ofc and you might end up having no use of it

I spend thousands on this hobby per year before i even earn money out of it

[split valve]

yea because its fun

[boreal falcon]

Thats how worth it is to me

[split valve]

But things dont stay fun forever... usually

i mean i dropped 2 grand on a guitar pedal and a grand on a guitar at the same time... cause i value it

And heres me with a GTX 1070... because i dont value current nvidia cards or video gaming

[foggy jay]

I've been envying these types of images for months that midjourney creates almost effortlessly, to the point of analyzing many of them. Apparently there is nothing special, it is what is called low key (I am not even remotely an expert in photography), however, first of all if you look for this type of images you will see a significant loss of detail and high contrast, these are They are characterized by having almost 70% of the pixels in the blacks and shadows with a magnificent preservation of detail and without excessive jumps in the gradients. I am surprised and happy to see something similar, which I have been unable to reproduce no matter what type of lora light fix I have used. Would someone be so kind as to tell me if this type of images can be created consistently in SDXL?

[split valve]

can you simplyfy that i dont get it.

In my experience using midjourney V5 every image i had to take to photoshop and fix the contrast and color as they were washed out and needed more clarity.

SDXL you can take the images to photoshop and if you do just an auto tone nothing will happen because they are already perfect

[foggy jay]

Well it is obviously a matter of personal appreciation, from my point of view MJ's lighting is amazing, in any case, regardless of personal taste, as I said this type of lighting does anyone know if it is easily recreated in SDXL?

[split valve]

well i like i said i dont rate midjourney they make washed out images.. as of V 5... SDXL i think shall be very capable.

its hard to answer you exact question though, as i dont know the inner workings of the model, its some vastly complicated neural network with weights, and with training data that i have no idea of the quality of...

But it does understand concepts very well, like i was using keywords like octane render, or subsruface scatter or translucency or camera terms like "macro", or lighting types and it will produce it very well

so its understands context and concepts from its training data as far as i know.. This should allow it to be able to produce anything that you can describe with key words

[boreal falcon]

I liked art since childhood

Too bad i was told its only for talented people and too hard to learn

So i gave up for a long tome

Time

Then when i came back generative AI art came and i had a crisis for a short time again, but now i appreciate art way more

[split valve]

i've never really been into art.. im more of a technician.. like i create blender models. and like to render..

but i like to appreciate art.

[boreal falcon]

GAI just gives me more reasons to continue with art

[foggy jay]

Ok thanks for your interest sono, we are all waiting to be able to test the model in depth, in any case I leave the question open in case someone with access to the beta is so kind as to inform me. In another order of things, it is possible to paste external links here, specifically towards deviant art, to better exemplify the type of lighting I am referring to...

[split valve]

i don't see why not.

[boreal falcon]

There is too much too important things that just make me stick to art including modeling instead of relying on AI

I have fun with GAI tho

And it has partially a place in my workflow

[split valve]

Something happend to my brain from zonking out for months using SD and midjourney... My brain must have been learning.. because when i randomly pick up a pen and tried to sketch something,, i was amazed.. i never used to have the talent

SDXL cant draw hands and fingers correctly like in that gallery yet... I believe they are still working on refining that area for V 1.0

[foggy jay]

Yes, I know, it's understandable, the hands will still be a source of problems, I mean more those images with low light but with ultra-defined details.

[iron field]

Could I post a woman in a bra here, like an underwear advertisement, or would that be NSFW?