✨｜sdxl

[crisp owl]

Really it's all still just a few months old

[fierce hollow]

you could edit the metadata to insert some vaguely similar prompts and such instead of yours, there should be some tool online for it if you're not into python much, no idea other than that

[vital ermine]

Yeah, someone is bound to take it on. Backend is fine it is the front end that needs the work.

[crisp owl]

Yeah. I mean there is Swarm which McMonkey wrote. I'm certain there will be more, or is more.
I've just gotten so used to it now

[vital ermine]

I can't get used to it as it screws me over when I start to go complex.

Too much noise on the screen and my mind melts

[crisp owl]

I have my fair share of moments where I'm just staring at the screen trying to figure out what needs to go where next lol

[crisp owl]

Slowly in the process of building out a new 1.5 workflow to share

[vital ermine]

I would never be able to handle that

One thing I wish this has was radio nodes

[crisp owl]

There's a branch working on subprocesses

[vital ermine]

those help with screen mess

[vital ermine]

Abelton Live has had radio nodes for a very long time

[fathom tulip]

Forgive me if it's been asked, but I'm tryin gto learn the ins and outs of the prompt system. What I've noticed is that my attempts to replicate an SDXL style on my own (i.e "comic book") don't look anywhere as good as when I tell the dream command to use the comic book style. Is there any available resources on what underlying style prompts exist?

[vital ermine]

radio nodes are simply two nodes a transmit and a receive. 100% no wires between on the screen between those nodes and so easy to trace issues down.

I never used dream so not sure.

[vital ermine]

Is that all you got?

Schwifty

[vital ermine]

@crisp owlDamn, I changed your workflow to be more inline with what I know from the website and it stopped being the same.

[vital ermine]

Got it

[crisp owl]

noise seed was not on bottom

haha

think you saw that

[vital ermine]

Yep

I posted the pic and spotted it immediately

[crisp owl]

had the picture still open before you deleted it

[vital ermine]

works now

If I save either pic does the entire workflow come along for the riude?

*ride

[crisp owl]

Yup

[vital ermine]

Oh, sweet. I have so many workflows I was saving

[cyan crown]

well you're working on workflows...I'm working on styles. There're so many styles in base model that it's incredible

[vital ermine]

OMG, I did that last week, and WOW

I bet even more too

[vital ermine]

macro

[cyan crown]

by Slinkachu

[cyan crown]

there thousand of styles

I tried few hundred

[vital ermine]

Damn, I tried a few dozen then saw the onslaught of them and said that was enough

[cyan crown]

I also find an incredible Lora that powers all images and all styles

xl_more_Art

Kyoto Animation studio, as usual sdxl base

[lusty wolf]

Looking to the Future...

[cyan crown]

Toei Animation studio

[cyan crown]

Michael Carson

[lusty wolf]

Go pro camera style got taken literally... 🤭

[tender timber]

Apologies, which model are you referring to, is it the base stability 1.0?

[vital ermine]

w/o and with my lora

[pallid path]

Michel Jason

Canyon East

[noble shoal]

Bread Pit. Am i doing it right?

[pallid path]

So.. the joke is about replicating a famous person and then naming them in a weird phony/botched name, but that doesn't mean that the AI generated image matches what you name it as, so for example, Bread Pit isn't literally a "Bread Pit", but is just an image of Brad Pitt.

I am currently learning Yapanese, I hope you know the language

[pallid path]

why does this look blurry and weird?

what's the prompt

[vital ermine]

Beats me it is something SD decided to do. I moved on and didn't save it

[pallid path]

ahh

[vital ermine]

All kinds of weird though

[lusty wolf]

Who 💩 in my porridge!

[half cedar]

Pretty stable Photoreal (donald trump:1.3) photograph, (television broadcast, news footage:1.2), charging up, spiky saiyan hair,realism

[lusty wolf]

😬

[half cedar]

Japanese CEO trump

[crisp owl]

https://tenor.com/view/surprised-whoa-shook-windy-wow-gif-16482549

[normal moss]

Is there some sort of API endpoint that I can run a prompt through programmatically to check it for filtered words? Or is there a CSV of filtered words?

[nimble heart]

Added more latent manipulation nodes if anyone wants to toy
https://github.com/Beinsezii/bsz-cui-extras/blob/master/bsz-nodes/bsz-latent-manipulation.py

different latent colors @ 70% strength
left column: red/green/blue/black
right column: cyan/magenta/yellow/white

[crisp owl]

So many things to tinker with 😅

[nimble heart]

black/white is the most useful by far

its basically just an input version of what my offset node does

[crisp owl]

I liked those two the best

[nimble heart]

white @ 70 is a bit strong

[crisp owl]

though the yellow is a nice effect also

[nimble heart]

yea the scaling isnt even it seems. Blue @ 50% is mostly gray while cyan @ 50% isnt even close

same how you can set black all the way to 100% and still get a colorful image while white @ 100 will turn it almost into line art

[crisp owl]

Yeah but that's just in general, sometimes changing a color like cyan vs yellow in photoshop also yields near nothing to completely changing the entire image

[nimble heart]

its extra pronounced in latent basically

the more luminant colors are spicier

why the right column is more extreme changes

rough approximation of SDXL latent channels based on just kinda looking at it

the rightmost is like cream at 10.0 and dark navy blue at -10.0 so idkwtf to call that

it's like off-luma

hot/cold luma

[maiden gale]

hey guys

[crisp owl]

beats me, that's beyond what I've looked at lol

[maiden gale]

yo @crisp owl , can I ask a followup question?

[crisp owl]

I'll answer if I know lol

[maiden gale]

I correctly set the Inspire nodes, but there is no image at the end

[nimble heart]

there's a latent debug node on my repo that spits out the average values of each channel if you wanna play

[maiden gale]

The sampler finishes but nothing shows up

[crisp owl]

Does anything show up in your output folder?

[maiden gale]

nope, nothing shows up

It just finishes, and restarts

[nimble heart]

"unload clone 1"???

[maiden gale]

yeah idk honestly

[nimble heart]

do you have any addons

or extra stuff

I've never seen that log message

[crisp owl]

I've noticed that before in my cmd, never could determine where it came from.

But I still got images

What's your full workflow look like?

[maiden gale]

no idea, but yes I got custom nodes

[nimble heart]

mines super different lol

[maiden gale]

I can't really post an ouput workflow since it doesnt do outputs...

without the Inspire nodes it works

So maybe it loads all the prompts at the same time? CLip encode took a long time (3-4) minutes to finish before I cut down the file numbers

I tried both approaches, all in one file, and different files

Am I doing something dumb by any chance? Sorry if its obvious

[crisp owl]

First try just putting normal text prompts through into your ksampler to determine if it's the conditioning where it's breaking or something else.

[nimble heart]

is that an XL model or 1.5?

[crisp owl]

Start with it simple, then add another custom thing

1.5

[maiden gale]

yes, it's the conditioning breaking, works that way

Conditioning takes amazingly long , 3 minutes or so with a big text file, just what I want to accomplish with this setup...

[maiden gale]

yes 1.5

[crisp owl]

Well, I haven't played around with that aspect of the pack yet, so really uncertain how it is required to be setup. Seems it should be pretty simple though if it's just pulling from the txt file, and even if it pulled nothing, it should still result in an image as an empty prompt still generates an image.

So it could be it doesn't like the 1.5 models?
Or something is connected incorrectly

[maiden gale]

Yes... The wiki says it zips the prompts, then it unzips it. Doesn't really say if the prompts are loaded all at once, or what the purpose of the node is...

I don't think the setup (other than the inspire nodes) are wrong, because it works when I remove them and use manual clip input

Do you know if the creator messages here?

[crisp owl]

He does, but goes by a different name if I recall, perhaps he doesn't want to be bothered here

[maiden gale]

ah makes sense, couldnt find him by name

It seems there is another solution to this, will try it tomorrow

Someone also said there is no control over the sequence of prompts in the inspire pack

[crisp owl]

Always more than one way in Comfy lol

[maiden gale]

Hahah yeah seems so, thanks again for the help @crisp owl

[maiden gale]

thank you as well, gn!

[heady vale]

used IPA to make the room darker

[nimble heart]

that'd probably be a good use case for offset latent

[heady vale]

Ive only played a bit with the offset lora. thought I would try with this

[nimble heart]

if you start with a latent closer to black instead of middle gray it darkens the image

ex: #✨｜sdxl message

[heady vale]

yeah theres a lot to play with there

[nimble heart]

you could probably ref a black image in IPA to achieve something similar

modifying the input latent is free performance-wise though. idk if ipa slows it down like cnet and lora

[heady vale]

tried a complete black image and result wasnt as good

I used an image that was generated with more shadows, less lighting

[nimble heart]

@vital ermine ayyy the blockers are gone in miopen. an internal testing build might be up on dockerhub by the time your buddy gets his 7900

[vital ermine]

oh, snap

[ivory blaze]

Oh okay mcdonad's, keep half filling my fries and giving this diabetic regular cola, one day people are going to say they're done, and then what...

[nimble heart]

major libs all have a 6.0 release branch cut now too it seems

5.7 was put together pretty fast so hopefully 6.0 doesnt end up like the first rdna 3 release where it's just internal testing for like a solid month before any news of the actual release

they're hustlin lmao

thats not even half of the libs with major activity in the last hour

[nimble heart]

dynamic range is squashed

the highlights should clip

but they level @ like 70%

[heady vale]

yep, that was reducing the ipa image levels in PS

[nimble heart]

never use levels for lowering values, always use a curve

so the clipping vals are retained

levels are for creating clipped values in the first palce

[heady vale]

I reduced the greyscale

this was just a darker image, not altering anything

its intersting that it actually changed the image based on the input image lighting

[nimble heart]

yea, all 8 of the color examples I posted earlier are the same prompt/seed

just different latent colors

[vital ermine]

Yeah, tomorrow it arrives and he is a Nvidia guy so knows nothing of the AMD world. He does know this will be "fun" to get all to work.

Worst part is he trains so I hope 6 fixes that

[heady vale]

this is a solid black image

doesnt darken it the way you would expect

[ivory blaze]

she asked me how too make beef stroganoff... I sent her this

[ivory blaze]

i think it turned out quite well

[static prawn]

i dont get which vae exactly use with sdxl

i always get weird artifacts for some vae

even the new fixed ones

like rainbow colores scanlines, cant really describe it

[upbeat summit]

https://huggingface.co/stabilityai/sdxl-vae/resolve/main/sdxl_vae.safetensors
this is the correct one. it's the SDXL 0.9 VAE and produces no decoding artifacts

[ivory blaze]

boy typos do some things sometimes...

[static prawn]

its weird i had some fp16 fix versions that work for me

all those produce scan lines

[ivory blaze]

0.9 makes scan liines?

maybe your model is an oddball,

[static prawn]

looks ok now

🙂

do u know if there is some full package of all the different upscalers? 😄

[upbeat summit]

you can probably find some collection repos on huggingface but I got mine mostly from https://openmodeldb.info/ and github

[static prawn]

oh ok allright 🙂

i mostly use 4x ultrasharp

[upbeat summit]

many are using 4xUltrasharp, lollipop or upscalers by NMKD. I personally use 4x SwinIR s64w8 https://openmodeldb.info/models/4x-classicalSR-DIV2K-s64w8-SwinIR-M. it gives me the best output quality compared to many others I've tested including ultrasharp

[heady vale]

+1 for that swinIR upscaler. I found it to be really good

[nimble heart]

i found 4xultrasharp the best for an in-between upscale if you're doing further denoising after

[upbeat summit]

yeah it's a bit slower than 4xUltrasharp but absolutely worth it in my opinion. the details are more clearer, less over-sharpening and artifacts - at least when creating realism images

[static prawn]

dunno how nice highresfix works for sdxl

if i have a model without refiner

[upbeat summit]

yeah as in-between it's great because it's really fast. but as a final or post-processing upscaler SwinIR s64w8 looks much better in my opinion and it doesn't add unnecessary details and stays very close to the source

[nimble heart]

ultrasharps' aggressive smoothing and accentuation also helps the 2nd denoise pass follow details better.

while the denoise fixes any ringing or other overtune artifacts

[upbeat summit]

yeah - it's great for that

I did start to experiment by injecting noise in between passes but it's quite wild if you put upscalers in the mix heh. so I mostly upscale the latent now without a model

[nimble heart]

latent upscale is impossible with UHD images

I've had 0 that've turned out well

[nimble heart]

I even tried an iterative approach where I latent upscale and denoise between 768p -> 1080p -> 1440p -> 2160p

[upbeat summit]

okay - I mean of course I get a lot more bad images but the ones that work have pretty good details

[nimble heart]

and it was still just a soupy mess

I can latent upscale up to like 1920x1080 without issue but every extra pixel after that gets exponentially harder

[static prawn]

is highresfix in auto doing the same as ultimate upscale ? like when i have 1024x1024 and do a highresfix on 2

is it the same when i do it in ultimate upscale?

it is a simple img2img , isnt it?

[nimble heart]

ultimate tiles and does other stuff doesnt it

[glad grove]

high res fix can do latent upscale, ultimate upscale cant

[vital ermine]

I never managed to get latent upscale to work and everyone said it is rubbish because they couldn't either.

[static prawn]

but in the end its a img2img? like when i send my pic to img2img, set everything the same, and set the res higher, denoise -> i get the same result?

[glad grove]

when it works it works but when it doesnt it just destroyes the image

[static prawn]

needs a higher denoise, at least 0.5

[glad grove]

u have to try different denoising strenghts to make it work and also some loras can break it or make the img deformed

[vital ermine]

yeah, for me 0.6+ and my image is so changed just not worth it

[merry shore]

whats beef strokingoff

[upbeat summit]

hmm... so right now I'm also using a lot out-of-spec resolutions to get those nice details. more bad images, but worth it when it hits. depends on the model. The Vision models by @delicate kelp have really great coherence and work quite well with higher resolutions.

so my current experimental workflow (ProtoVision 0.6.2.0):

• 1st pass: in-spec 1216x832 (optional FreeU to try to get more coherence)
• latent upscale scale_by 1.25 to 1520x1040
• (optional) injecting noise via Power Fractal Noise
• 2nd pass: latent upscale pass (hires fix)
• pixel upscale using SwinIR s64w8 scale_by 0.33
• final output resolution: 3040x2080

[nimble heart]

what's the denoise?

on the latent

[upbeat summit]

0.7 🙂

[merry shore]

man i dont know what gpus yall have but my 1070ti is sucking

[nimble heart]

1520x1040 is so small you could almost generate that directly

[upbeat summit]

but the sweetspot is 0.57 from my tests

everything below does too much latent blurriness

[ivory blaze]

i like to make homeless ronald mcdonald

with high detail

[merry shore]

give him a mccrack pipe

[upbeat summit]

yes, but you won't get those details without the latent pass

[merry shore]

hes like a joker ronnie hybrid

[nimble heart]

problem is upscalers like Swin/ESR looks like shit on a native 4k display

so you need a latent 4k pass

[static prawn]

oh i have no idea about those more passes, just using auto 1111 😄

[nimble heart]

cant just render like 20% higher res then SwinIR the rest

[merry shore]

What gpu do you guys use?

[static prawn]

gtx 1070

[ivory blaze]

3060ti

[nimble heart]

7900 XTX

I still have my 1070 in the garage

[upbeat summit]

I'm not a fan of pixel upscalers at all. but I don't think they look super bad when I use them in my pipeline and I got good displays I suppose

[glad grove]

3dfx voodoo 3

[static prawn]

i will probably upgrade to a 3060 soon

[merry shore]

im on a gtx 1070 but my upscalling can never go more then 2

[static prawn]

for sd 1.5 just use tiled vae

[nimble heart]

my old wallpapers were all rendered in Blender @ native 4k, moving to SD wallpapers and 1080p upscaled to 4k looks like blurry garbage

[static prawn]

gives u a lot of power

[static prawn]

and make sure u use medvram for sdxl

otherwise it takes hours haha

[merry shore]

yes medvram il grab my prompts

[static prawn]

but this time i upscaled with ultimate upscale

not sure if i can do 2x out of the box

maybe with tiled vae if its available for sdxl, im not sure

[ivory blaze]

what upscale model you use with UUSD?

[merry shore]

do you use sdxl models @static prawn

[ivory blaze]

Foolhardy works really well

that's what I used on the clowns above

[merry shore]

i call stable diffusion / stable confusion and i dont even mean to

lol

[ivory blaze]

0.18 denoise 30, @ tile size = resolution size of image

[static prawn]

this time it was 4x ultrasharp, i love the image itself but the quality isnt the best for the cat, anyway i like sd 1.5 way more then sdxl / im not so much interested in photorealistic stuff and i think thats where sdxl shines

[merry shore]

this is my only prompt i have --medvram-sdxl

[ivory blaze]

nah it does some GREAT artistic features too

it's model dependant

[upbeat summit]

@nimble heart check these in full size.

• base 1536x1024
• upscale before latent upscale to 2304x1536
• final output: 3041x2028

[static prawn]

i mean i like the cat but missing some details here and there, but i like the overall composition

[merry shore]

i need more power lol ai art is addictive

[static prawn]

but i use auto1111 so i dont have the power like u can have with comfy ui 😄

[ivory blaze]

See how artistic SDXL is?

[nimble heart]

ultrasharp @ 0.3 denoise vs latent upscale @ 0.55

[static prawn]

with sdxl

[upbeat summit]

yeah if you compare it to a real native 4K render - sure. I totally agree

[nimble heart]

it just cant handle 4k

the refiner breaks down too passed like 1440p

[merry shore]

beauty

[merry shore]

is that brittney spears

[nimble heart]

so I cant use the refiner either, even with ultrasharp

[upbeat summit]

native 4k? no it can't

[merry shore]

brittany

[upbeat summit]

they work in 1024x1024 and we are trying to push further

and there are limits

[nimble heart]

base XL does okay enough with a low denoise I can get some images to work

[ivory blaze]

but yeah photorealism is exceptionoal

and that'

is no refine no upscale, not sure where the upscaled version went it is ridiculous though

[nimble heart]

4k @ 0.3 denoise coming from Ultrasharp. dont look too bad imo

[ivory blaze]

so tthis is not so hot

[nimble heart]

0.3 is enough to clean up the upscaling fuzzyness

but yea refiner breaks even @ like 0.1

[upbeat summit]

checked the details and sharpness on my images? they are not native 4k sharp, but I suppose the fidelity is alright 😉

[ivory blaze]

there's a decent one

[merry shore]

any generative fill plugins for SD to only change certain sections of an image yet?

[upbeat summit]

it's very stylized. if your intention was to push it towards realism it does look blurry. it depends on the style you want to do here

[ivory blaze]

that was an SDXL artistic oone

[nimble heart]

clean edges really. I mostly dont do photoreal

so the ultrasharp's into 4k denoise has worked fine.

idk I can get up to like 1440p probably with purely latent techniques so maybe SwinIR or something will perform better with more input data

[upbeat summit]

yeah - it really depends what you are aiming for. as a stylized theme this looks very cool. but of course you are looking for specific details you want enhanced

[ivory blaze]

is this art?

i mean there is nothing realistic about him having fun, but , still

[upbeat summit]

do you see loosing a lot of details when upscaling and the first pass looks better to you?

[nimble heart]

but yea I have a few SD 1.5 wallpapers that i made using 1080p upscaled with whatever GAN looked best and the comparison to a 4k denoise isnt even close

[nimble heart]

wdym

[upbeat summit]

I mean how does the first pass image look - does it have the same problems you find in the upscale?

[nimble heart]

as the GAN upscale or the latent?

first pass looks fine. usually the best version for thumbnail sizes

I do 1368x768

[upbeat summit]

these are all the passes from one of mine

• starting with native 1536x1024
• latent upscale (denoising strength 0.57)
• final output (3041x2028)
• final output + post processing (some film grain...)

not 4k I know 😉

[nimble heart]

1024 -> 2028 just latent or is that also swin?

[upbeat summit]

there are definitely limits how much you can push it. at one point you will just get garbage

[nimble heart]

1368x768 is SDXL's native 16:9 aspect

[upbeat summit]

yeah and I work a lot in 1536x1024. it's out of spec and does give you deformations and mutations, but it still gives you nice images and the fidelity can be a lot higher - because more native pixels

[nimble heart]

first is ultrasharp4x into a 30% 4k denoise and the second is 1080p XL just upscaled with SwinIR. the high frequency details look jpeg-artifacty to me in the 1080 scale without the additional 4k denoise

seeds are a bit different ofc but you get the idea

[upbeat summit]

yeah these are upscale artifacts

[merry shore]

@jolly zinc is yours like this after enabling tiling it wont let me pick an SDXL model checkpoint

set COMMANDLINE_ARGS= --medvram-sdxl

[nimble heart]

1.5 couldnt render above 1080 so I never tried in-between resolutions

guess I could

actually mistake, that 2nd one was the wrong scaler

[upbeat summit]

I started doing 1920x1080 experiments with SD2 and got about 2 good images in 50 maybe heh. so lots of wasted compute. SDXL is much better at it

[upbeat summit]

I mean I can't say how swinir s64s8 does work with every kind of style or outlines, but it does work great for realism and also for renders

but there's a limit. I tried out how far I can push it before the look gets too artificial

[nimble heart]

real comparison now

let me try 1440

damn torch has to compile a 1440 kernel

[upbeat summit]

is the left image the 1st pass?

no it's not

[nimble heart]

left is 1st -> ultrasharp -> 4k 30%
right is 1st -> 1080 latent 60% -> swinIR

was to demo how GANS can't produce native clarity

im doing a 1st -> 1440 latent 60% -> swinir now

but the kernel compiles are killing me

[upbeat summit]

I mean as said it really depends on the style, but I don't know how coherent the base image is. if you look on all the electronics components, they are not very coherent in the upscale - lots of AI noise.

the upscalers might not able to handle that in a satisfactory way to retain the style you are looking for.

[nimble heart]

cyborg wires are always just random BS

so at least make them clear

[half ivy]

bro

[nimble heart]

alright these are the 3 results

[upbeat summit]

yeah I'm not criticizing the image, style or artwork at all. it's just that upscalers might not be good at upscaling textures like that

[nimble heart]

yea they do good at small repeating textures like dirt and anime lines and thats it

hence why I'm so determined to get a native 4k pass

[upbeat summit]

10240x5760

[nimble heart]

?

oh i forgot to rescale the SwinIR outputs lol

[upbeat summit]

so you are doing supersampling

this could give you sharpness, but also new artifacts

[nimble heart]

wdym "so you are doing supersampling"?

[upbeat summit]

first upscaling and than downscaling

or is it not what you are doing?

[nimble heart]

swin 4x's it, you're always going to have to downscale if its your last step

if the final step is just an SDXL denoise I dont supersample

i denoise at my native monitor resolution

i wonder...

[upbeat summit]

yeah I just do it like that

4x at 0.33 scaling ratio

[nimble heart]

yea idk i try lots of things but the 1MPX -> ultrasharp -> 4k 30% always looks closest to native to me. Its what Sytan developed and I just integrated it into my principled node.

4k runs @ like 7s/it for me so it kinda hurts but I mean it looks so much nicer when it works out it's worth it IMO

kinda wild 4k is 7s/it but 1080p is like 1.4 it/s

[upbeat summit]

yeah. that's why I work a lot in out-of-spec resolutions. it might not be the most productive workflow since you are wasting a lot of compute, but when it works the fidelity can be pretty good. these use just 1 sampler pass in native 1920x1024.

[nimble heart]

why not just add the extra 56 pixels to do native 1080p?

[upbeat summit]

because it will perform worse. you will get more in-coherent images when using full-hd. it makes a difference from my experiments

[nimble heart]

keeping the one edge native res?

[upbeat summit]

yep

or at least on a native in-spec res

[nimble heart]

tbh I tested that a bit when people insisted that mults of 64 performed better and I found it was more an area thing than an exact size thing

[upbeat summit]

1920x768 (~21:9)
1920x1024 (~16:9)
are both working quite well

[nimble heart]

like in this 1920x1080 image you can basically see where it loses itself. the approximate 1024x1024 square on the right with the girl looks OK then when you leave that it gets funky

and that's how you end up with dupes

like it only has 1024² pixels of vision to look at at once

so the shape of thos epixels isnt really the killer its more the area

[upbeat summit]

are you using all the SDXL text encoder values (text encoder width/height, target width/height)? it's not a 100% fix but it can greatly increase proportions in different specs

[nimble heart]

yes

well

not crop

just height/width + targets

they're set automatically by my Principled node

based on latent sizes

maybe you could fiddle with crop to center the 'square' more I guess

[south horizon]

evenin'

[upbeat summit]

I don't use it. does it set the target values correctly based on SAI's best practice? that is that the longest side is always 4096

[nimble heart]

Not at all

[upbeat summit]

but it should 🙂

[nimble heart]

I did my own measurements specifically for upscaling

[upbeat summit]

yeah

[nimble heart]

I found targeting the ratio as 1MPX did the best

personall

the difference is super tiny though

[upbeat summit]

of course. there's no right or wrong. I just adapted what we discussed with Joe Penna in chat and build the Preset Ratio Selector node with humblemikey (Mikey Nodes). I build a preset file with all official SDXL resolutions from the official paper and calculated all the target values following this method

[nimble heart]

tbh I found the official resolution spread kinda lame too. Long as it's 1MPX you're good

so I do 1368x768 since its closer to 16:9

instead of 1344 or whatever they listed

[upbeat summit]

well I wouldn't call it lame. it's the resolutions it was trained on

[nimble heart]

training ≠ mandatory

im a bit sick of people arguing with me that 1368x768 is litearlly satan

[upbeat summit]

https://arxiv.org/pdf/2307.01952.pdf page 17

yes, but there are reasons for it

I work in film so I want industry standards as well 😄

[nimble heart]

i mean if there's a reason to use the res for you besides "its what's on the list" then go for it.

[upbeat summit]

sdxl-raw-hdr please

[nimble heart]

but if someone's trying to make a 16:9 image specifically, use what's closest which is 1368x768

[upbeat summit]

yeah it's the closest to 16:9 in-spec resolution

[nimble heart]

I've never had one person produce compelling evidence that the tiny difference in aspect between 1368 and 1344 or whatever does anything outside of mild seed variance

[upbeat summit]

or 1536x640 for ~21:9-ish

[nimble heart]

Like I'm down to change my mind but whenever I ask for a clear demonstration it's just some random cherrypick like "look how cool this image is"

[upbeat summit]

yeah you can use whatever res you like. you do get more coherent images if you stay within 1024x1024. but since I mostly work outside of that it's rng anyway

[nimble heart]

its more you get more coherent the squarer it is I found

21:9 can be rough even at 1 megapixel sometimes

[upbeat summit]

since it's a 1024x1024 model that makes sense, but you can however, and a lot of work was invested in that, if the target values are used you can make very wide or high aspect ratios without much deformation

[nimble heart]

think I did 9:16 and cropped for my phone cause 9:21 was rough

yea my node straight up takes the latent sizes and feeds them into the clip sizes with the appropriate math so that shouldn't be an issue

[upbeat summit]

this isn't really a stable rule. this black box neural network thing does sometimes stuff that makes no sense because we have no tools to analyze it

[nimble heart]

my code for sizes in Principled

[upbeat summit]

yeah I think we talked about that before - you also looked at my json with your fancy shell viewer 😄

[nimble heart]

I used to have them set higher res but idk I found targeting XL's native res helped it not add random noisy BS

[nimble heart]

I've evolved since then. I switched from bash to nushell which can natively manipulate dataframes such as json

[upbeat summit]

it did fix many of my proportions issues when using 4096 on the longest side (target values) when working in a) native resolutions b) very high or wide aspect ratios

[nimble heart]

my first pass is always native res so I guess I havent had that issue. it was mostly for doing a 2nd pass denoise

more the a1111 way of doing things I guess

[upbeat summit]

in a1111 you can't even access all that stuff heh

[nimble heart]

I meant like hiresfix

my principled node is built to do an a1111 hiresfix

but better

i havent tested the 4096 thing as far as duplication goes but I couldnt imagine it'd help tbh

[upbeat summit]

maybe I'm misunderstanding what you mean, but the target values are not meant as your final resolution output values. they are configuring the target bucket values when pulling data from the SDXL model and are suppose to help with coherence.

it's sadly not really explained anywhere afaik how it really works. I only have the info from talking to SAI and what I picked up in chat.

[nimble heart]

yea

I typically target 1MPX cause it seems to do the best all around

in the appropriate aspect

but I meant I havent tried say 16:9 with a 4096 edge as target for things like direct 1080p

[upbeat summit]

I have not explored it further the last few weeks since AIT doesn't work correctly with many increments of SDXL. so only 1920x1024 works - not 1080

[nimble heart]

for 1368x768 using my current method it'd be

width: 1368
height: 768
target_width: 1368
target_height: 768

and for 1920x1080 it'd be

width: 1920
height: 1080
target_width: 1368
target_height: 768

I found by targeting low it seems to reduce the amount of random AI noise gibberish you start to get at higher resolutions

[upbeat summit]

  "1344x768 (AR ~16:9 / Near 7:4 / DEC 1.75:1)": {
            "custom_latent_w": 1344,
            "custom_latent_h": 768,
            "cte_w": 1344,
            "cte_h": 768,
            "target_w": 4096,
            "target_h": 2340,
            "crop_w": 0,
            "crop_h": 0
        },

 "[CUSTOM] 1920x1080 (AR 16:9 / DEC 1.78)": {
            "custom_latent_w": 1920,
            "custom_latent_h": 1080,
            "cte_w": 1920,
            "cte_h": 1080,
            "target_w": 4096,
            "target_h": 2304,
            "crop_w": 0,
            "crop_h": 0
        }

[nimble heart]

i've not really played with width/height I just leave those at the latent sizes because I'm not sure what they do tbh

idk. i'll leave it as-is and I'll test that later once I can get flash attention or something working

get more than 1.3 it/s on 1080p

[upbeat summit]

target values have been rounded by multiples of 16

[nimble heart]

you round by 16

I round by 8

[upbeat summit]

yea

[nimble heart]

other people round at 64

idk

if its just affecting what "bucket" its giving extra attention to rounding shouldnt matter since those are pixel-space

[upbeat summit]

I choose 16 because it was recommended when we talked about it in chat and also some experimental nodes by SAI used that

so I chose this approach

[nimble heart]

yea im too cynical to use things people in chat recommend without A|B|X testing it myself on 20 images first

man I was wondering why it took so long to render then when it finished I realized I had all the wrong denoising settings lol

[upbeat summit]

sure - but it wasn't just some random chatter 😉 also I did so many tests... and found it worked quite well

this all so new experimentation is the only way to go anyway

there's no right or wrong

[nimble heart]

yea

speaking of

do you still use the refiner?

[upbeat summit]

no

[nimble heart]

yea

I defaulted the refiner node to being muted in my workflow and honestly I've never felt the need to turn it back on

just way more consistent

[nimble heart]

okay much better

[upbeat summit]

I only used 1 sampler for the last couple of weeks and have recently included a latent upscale / hiresfix pass

I did explore the refiner and it does have it's uses. I have it included in my workflow but it's almost always bypassed.

I fix eyes and other details with the latent upscale now - the refiner sometimes changes a lot of stuff or fixes a face but gets rid of a lot of texture details

it can be used for more stylized stuff, but I tried a lot to make it work for me

[ivory blaze]

that feeling when you accidentally load a regular model instead of refiner and generate a hundred images and wonder why they all look off

sometthing I have not seen, i looked a little but, are there any refine refiner models anyone has trained? or is that not even a thing

[upbeat summit]

refiner only images can be great too 😄

[ivory blaze]

yeah I''ve done that too

or reversed the m

that's ...creepy , dont like the tall man

tall man scary

i wondered why my stroganoff looked a bit off

cos i was refining with dream shaper

[upbeat summit]

just imagine 2 people wearing the coat

[ivory blaze]

it clearly is, i see his arm, sneaky

[upbeat summit]

hehe

[west breach]

try creating an image with refiner starting with a 768x resolution and then upscale

[upbeat summit]

nothing wrong about it. many use other models for refining

[upbeat summit]

you know exactly what's going to happen, right?

[west breach]

you get a 2.1 image 😄

[ivory blaze]

lol

2.1 is best model

for keeping your friends disinterested in AI.

I wish I could share this image, I mean it is not NSFW but mods be like... cos it's a skin tone body suit thing, but tthis fairy looks like she has down's syndrome :\

that's what hhappens when you use dreamshaperxl as refiner

[upbeat summit]

looks fine to me

[ivory blaze]

for dalle-mini maybe

[upbeat summit]

that would be pretty insane for dall-e mini hehe

[ivory blaze]

well the face, the rest of the detail, yes

[upbeat summit]

true

[ivory blaze]

well maybe not, dall-e mini looks like it has down's syndrome and some crayons

or at least did, when my friend sent me some images a few years back

are they even still around on that project?

there itt is on hugging face lol

looks about the same I guess

ok no, DALL-e mini is sick powerful

Im' done with sdxl, this is tthe future

[half cedar]

What a sweetie

[west breach]

here's a image started with refiner at 768 and then 2x upscale at 0.4 denoise

[nimble heart]

refiner does mid-frequency details at the cost of high-frequency details. so pebbels in gravel will look better but grains of snad will look worse

so for skin texture and fabrics base XL usually does better I found

while face contours refiner imporoves

[nimble heart]

she ate the wrong brownies

[nimble heart]

Keter class

[ivory blaze]

SCP-1239123312932 Potato Hat

or maybe that was the random seed, not sure.. .either way

[upbeat summit]

yeah, I agree. same for me in my experiments.

[nimble heart]

if I run the refiner it's usually not the full 20% as a result

its more like 10%

but yea recently I've just left it @ 0%

[upbeat summit]

same - I used it mostly below 25%

[upbeat summit]

the details are pretty good

I mean the medals went a bit crazy, but overall good fidelity 🙂

[ivory blaze]

looks like mj about 6 months ago

that's bett'er than the BDSM KISS Army harpy last time I tried when i had my models mixed upp

finally made my ai waifu, took years... but now...

ok done with dallemini

so, do past runs bleed into repeat fixed runs?

[crisp owl]

no

[ivory blaze]

I mean yes they change, but, I ttried to make thiis, with fixed seed in comfy, over and over, with nothing suggesting this background, repeated underground / gold and treasures, in both L/G and tried to negate castle but this one seed only seems to do this very similar style

just strange it won't listen at all,

[crisp owl]

sometimes the SD gods just are not on your side

[ivory blaze]

yeah I switched seeds and now it is underground every time without the prompt, I dunno. that's just strange to me,

like it rememebered

[tardy jungle]

Is automatic1111 stablediffusion 2?

Why are my pictures coming out weird

Should I add something

[crisp owl]

make sure sdxl models use the sdxl vae
and likewise for 1.5, make sure they use 1.5 vae.

[tardy jungle]

Ok, let me try, thanks

[crisp owl]

And for SDXL keep to around the 1024x1024 ratio, or similar

[tardy jungle]

So automatic1111 is 1.5?

[crisp owl]

A1111 is whatever model you select

1.5, 2.1, sdxl

it's in the selected model

[tardy jungle]

Oh Alr

Is this the right set up

[crisp owl]

yeah that's the standalone sdxl vae

[tardy jungle]

Yeah

Should I add something

[crisp owl]

You can manually select it in the A1111 settings

just select the button to see all options, then do a control+f search for vae and select it, check the box or whatever to make A1111 use that instead of baked in vae

[tardy jungle]

Which button

Where

Vae isn’t showing up in the checkpoint section

[crisp owl]

it's somewhere int he settings, I don't run A1111 anymore

[tardy jungle]

It’s not on the checkpoint section

[crisp owl]

in the settings menu, on the left there should be a button that shows all options, then just ctrl+f and search for "vae"

[tardy jungle]

Yeah I see it

I applied the safetensors one

Now how do I use it

[crisp owl]

It'll automatically use that vae when you process any image now. Just make sure you have an SDXL model selected "checkpoint"

If you use the SDXL vae with a 1.5 checkpoint, or vice verse, you will get wonky images

[tardy jungle]

Should I use SDXL with a 2.0 checkpoint?

[crisp owl]

like for like. SDXL checkpoint with SDXL vae

[tardy jungle]

Ok

Any you recommend?

[crisp owl]

What type of images do you prefer?

[tardy jungle]

Idk like ones of skylines

Cyberpunk

Cities

Bright colors

Neon

Idk stuff like that

[crisp owl]

The base model of course you can always go with, it's pretty good.
For more bright colors, vividness and such, I like to use ZavyChromaXL, it tends on the ends of realism but pushes more artistic colors

[tardy jungle]

I don’t understand, the what is a checkpoint really?

Is vae like an extension of a checkpoint?

[crisp owl]

checkpoint=model

[tardy jungle]

Got it

And what is vae then

[crisp owl]

vae is ultimatley what's converting the latent space into pixel space

[glad grove]

its like the decoder

[crisp owl]

so in more basic terms, think what the computer sees vs what we see

[tardy jungle]

So it’s a prerequisite?

[crisp owl]

model is required, vae is required

[tardy jungle]

I see

And controlnet?

[crisp owl]

optional

[tardy jungle]

What does it do

[crisp owl]

can be used to help direct an image towards something. Some controlnet models can force a pose, while there are others that will extract the hard edges of an input image and push that to your generated image to follow

[tardy jungle]

Ohh

Ok

[heady vale]

some models have the vae baked into the checkpoint, so separate file isnt always needed

[tardy jungle]

Where’s the download button for hugging face

On the files and versions tab

I cloned it but it didn’t work so I just wanna download it from the website

[crisp owl]

little down arrow

[tardy jungle]

I don’t have it

Would be easy if I did

[crisp owl]

screenshot

[crisp owl]

it's the ones below

base model here

[tardy jungle]

Damnn

Which thingy

Which model

[heady vale]

thats Nightvision

[tardy jungle]

Night vision is the model?

[heady vale]

https://civitai.com/models/128607/nightvision-xl-photorealistic-or-portrait-or-photography-or-hyperreal-or-architecture-or-interior-design-no-refiner-needed

[tardy jungle]

Nice

[tardy jungle]

What size should I use for sdxl

Nvm 1024

[crisp owl]

Any of these were specifically in the training for SDXL

all leading to ~1MP in size

[tardy jungle]

I don’t have any memory for that

[crisp owl]

Minimum size before you start getting too wonky results would be probably 768x768 equivelents

anything below and you'll likely get off results

[tardy jungle]

Yea

[delicate kelp]

If you're using Auto1111, make sure you're using --xformers, --medvram and also install the Tiled VAE extension. With all that, my lapop with 8GB of VRAM can run all of the SDXL favored image sizes, and can even run batches of 4 x 1024. Not the fastest thing in the world, but it works.

[tropic turret]

Feels like I created a new style lol

[nimble heart]

shelobmobile

[weary yacht]

are those spider webs?

[tropic turret]

They are more like a web of neurons 😛

[weary yacht]

well, I wouldn't drive it

[tropic turret]

[strong copper]

Return of Conan Biker

[grizzled dune]

Hello. Does somebody know if there is a model for NormalMap with SDXL ?

[dry crypt]

hey guys, i am working on using stable diffusion sdxl model for inference purpose for generating images using some prompts, What is the hardware requirement for fine tuning sdxl? Thanks

[stone fossil]

If you can I would scoop up a nicely used 3090, you can probably do with less but you wont have no regrets having 24GB minimal if you wanna go cheap else get a 4090 or A6000.

[peak dove]

Where is the room to generate AI in SDXL named Distillery?

[frozen sundial]

I have only seen the depth models, which work very well. I have not seen any normalmaps

[native knot]

Cursed Conan

[wet nacelle]

the only Conan

[inner kelp]

Threw together a quick few for Friday the 13th, Here's some "Jason" Voorhees art. Happy Friday The 13th!

[half cedar]

@wet nacelle have any prompt tips for the kind of photography you've been doing?

[wet nacelle]

First use this model ||https://civitai.com/models/133808/pyros-nsfw-sdxl||

Then use this negative prompt. (cartoon), 3d, render, low res, low resolution, ((text)), ((watermark)), ((logo)), tongue out, old, ugly, masculine, over exposed, vibrant, colorful, .com, ((tanlines)), (( ososedki.com))

You also need to have your sampler set as dpmpp_sde and your ksampler as normal.

[ivory blaze]

what do you guys use to get directional facing, I can't seem too get that down at all they just look forward, may be an issue of the model but, I can't imagine it is just that.. .

a closup photograph of a {profile left|profile right|forward|slightly-left|slightly-right|upward|
slightly-upward|downward|slightly-downward|upward-left|upward-right|downward-left|downward-right} facing woman

definitely does not seem to have much effect

[wet nacelle]

side angle

[ivory blaze]

hrmm

photorgraph of a woman whose face is angled {side|forward|slightly left|slightly right|up|
slightly up|down |slightly down|up left|up right|down left|down right} and has a {happy|sad|neutral|bored|excited|upset|unhappy|coy|intense|neutral} expression on a black background

tthat works well

the direction> angle was not catching like this modification does

a photograph of a woman whose face is angled down and a unhappy expression on a black background

but it still is angled, and not directly forward so may need to adjust down and up

[wet nacelle]

https://www.google.com/search?q=yaw+angles&source=lmns&bih=875&biw=1760&client=firefox-b-1-d&hl=en&sa=X&ved=2ahUKEwi6j7nA0_OBAxX-zckDHbkJAggQ0pQJKAB6BAgBEAI

[ivory blaze]

I don't even know why I use prompts

I could just do a head movement openpose model in blender and tween it for looking in all the directions and animate the hell out of that

[ivory blaze]

thiis is magnificent,

https://toyxyz.gumroad.com/l/ciojz?layout=profile

[shy basin]

[ivory blaze]

and this will be much easier for controlnet openpose animation in comfyui, does anyone know if there is an image loader that will increment, instead of hand loading like the load image node, or will i need to make such

i dunno much about automation and animation in comfy

[ivory blaze]

I have not laughed so hard, I mean I do nott know why it is so... just.. other things have been funny, but something about this truly hit me somewhere..

has to be his pose... it's just.. campy af but... so right

[shy basin]

lol thanks, inspiration came from this ad
https://youtu.be/BqpJvey-7-s?si=7-ZXW8rtYmlAqPEs

[shy basin]

[crisp owl]

Here's a couple open pose resources I've used before
https://openposes.com/
https://app.posemy.art/

[tardy jungle]

I have officially cleared my pc

300gigs of storage

Now I can get like 273834 models

[crisp owl]

almost

[tardy jungle]

My sdxl yesterday bugged out

Every picture came out blurry

I wonder why

[crisp owl]

image ratio or vae generally

[tardy jungle]

Yeah

Does image quality depend on GPU or is there a way to change quality manually

[crisp owl]

GPU won't have any difference on the quality of the final image

[tardy jungle]

Is there a way to switch the quality

By quality I mean 4k

Like image quality

[crisp owl]

no
upscaling from the initial image is how to get a larger scale image

[tardy jungle]

Not larger scale, image quality

[crisp owl]

no, image quality is not a toggleable item.
Final image output depends on a couple factors. Your prompt, the seed, and the image ratio.
If you attempt to use a ratio outside the trained parameters, likely gonna get bad results.
The prompt is pretty forgiving, but can of course make a difference.
And the seed is a luck of the draw. Some seeds just work better than others, so hitting "generate" again is all you can do for that.

[tardy jungle]

Alr

Control net just pushes an image in the right direction?

[crisp owl]

yes, like this is a canny representation of the input image.

The final output would resemble something akin to that outline

[peak dove]

Distillery@Discord #SDXL free while Alpha testing

[tardy jungle]

So that’s what control net does

You give it an outline

[crisp owl]

you give it an image

[tardy jungle]

And it helps use that and the prompt

[crisp owl]

it then transposes it to whatever model you select.

canny is an outline basically
depth creates a depth map
openpose extracts the pose of the character
etc

[tardy jungle]

Oh so it takes something from a photo and implements it

[crisp owl]

yes

[tardy jungle]

Ok

What GPU do you have

[crisp owl]

2060 super

[tardy jungle]

I have 1060 super

[crisp owl]

6gb should be enough to do full size images at least in ComfyUI, not certain about A1111.
COmfyUI is more resource friendly

but ComfyUI is not as "user friendly" persey

[tardy jungle]

I think I have 32 gb ram

Oh it’s memory isn’t it

Experimenting with DALL-E 3

[tardy jungle]

Anyone know the process that was used to create this

[crisp owl]

nope

[tardy jungle]

How is it so photo realistic

[crisp owl]

closeup hand, upscaled

maybe inpainted

[tardy jungle]

What does inpainted mean

[crisp owl]

mask a section, generate an image to go into that masked section

in A1111 you see that in the img2img tab

[tardy jungle]

Ohhh

[lusty wolf]

Happy Friday the 13th...

[tardy jungle]

What model did you use

[clever verge]

I do the same here. I think I get within a few pixels of your precalculated values. Not figured out your math though so can't understand why we differ. 😄

[ionic dragon]

@vapid roost can I dm?

[crisp owl]

was wondering why this image was taking a while
102 faces being corrected by the facefix node 😅

[crisp owl]

heh

[soft bone]

idk if you wanted it to do that or not but there is a way to filter the max number of faces that facefix will inpaint

[crisp owl]

Nah, I'd rather wait the time than have some wonky af looking faces next to fixed ones lol

[soft bone]

true, although with that method its giving everyone the same face

if you dont have wildcards

[crisp owl]

Buncha angry dwarfs, close enough lol

[vapid roost]

Thanks!

[hoary saddle]

anyone know who built SD Doodle? official team member?

[crisp owl]

Think I saw them post it in this channel after they made it, but can't find in search (at least not while mobile)

[pure crystal]

anyone running a comfy custom node wiki yet?

[weary yacht]

a wiki?.. not that I am aware of

[hybrid fjord]

Where can I find this clip vision ip adapter?

[vale eagle]

you should able to find them on their git

[hybrid fjord]

Not there

found it though on hugging face, nope i didn't lol

[hoary saddle]

wll search for it, thanks

[crisp owl]

ClipVision (For IPAdapter AND Revision):
Clip Vit_G:
install below in ComfyUI_windows_portable\ComfyUI\models\clip_vision
https://huggingface.co/stabilityai/control-lora/blob/main/revision/clip_vision_g.safetensors

install below in ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\models
https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/ip-adapter_sdxl.bin

Clip Vit_H:
an available lighter-weight version of ONLY IPAdapter (will not work with Revision) below:*******
install below in ComfyUI_windows_portable\ComfyUI\models\clip_vision
https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/image_encoder/model.safetensors

install below in ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\models
https://huggingface.co/h94/IP-Adapter/blob/main/models/ip-adapter_sdxl_vit-h.bin

Clip Vit_H PLUS:
an available heavy-weight version of ONLY IPAdapter (will not work with Revision) below:****
Use save clip_vision as above Vit_H

install below in ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\models
https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/ip-adapter-plus_sdxl_vit-h.bin

[hoary saddle]

T2I is a better way to go over controlnet for a doodler.

App results are a lot cleaner now

[vital ermine]

I wonder if Comfy is going to add glora to his comfyui?

[shy kelp]

yummy

[icy brook]

Aether Bubbles, Foam & Beyond.

[icy brook]

nice! unusual toned down scifi stuff 🙂 love it

[fierce hollow]

have you folks seen this?
https://github.com/YingqingHe/ScaleCrafter
4k x 4k gens with sdxl in 1 pass possibly, wonder how hard it would be to implement in a111/comfy

[zinc cargo]

is it out?

[icy brook]

Not yet. Need to have a break this weekend 😅

[half cedar]

Chic Sofa in a form reminiscent of a pumpkin, stylish loft apartment works very well, making me rethink my prompting style

[rustic garnet]

uuuh, this is so cool!

[vestal ivy]

this mf is stuck generating things that look like 2022 stable diffusion

but talking like he's generating some god-tier stuff lol

[wet nacelle]

I think what he was saying is that he inpainted for four plus hours

[vestal ivy]

and the result is bad

[rustic garnet]

to be fair: it's much easier generating good images "by chance" than generating an image where you already have specific constraints and requirements

[rustic garnet]

hm, tried it. It works okay for photography. For anime and paintings it gave me totally unusable results, though. I would say that tiled upscaling still works better :/

but the idea is cool

I guess it could work if you fine-tune SDXL on super highres images

[fierce hollow]

oh, guess it's not as good as claimed

[rustic garnet]

at least not in my tests. It still has artefacts, in particular for 4k x 4k resolution, while texture is very low detail.

[shy kelp]

What is even the advantage of sdxl? I don't really get it?

It says you can make pics over 512x512, but i can already do that with normal models?

[glad fulcrum]

2 controlNets at the same time is not working on auto1111 on SDXL? or is it just me?

[sweet ore]

hey hi, where can i ask for a model sugestion?

[rustic garnet]

???
SDXL is MUCH better than SD 1.5

[shy kelp]

In what way?

[rustic garnet]

the naming might be confusing. SDXL is a new version of SD, basically SD 3.0

[shy kelp]

I have some 1.5 models that give me much better results than sdxl

[shy kelp]

Yeah it doesn't works for me either

[rustic garnet]

yeah, you can always pick a model that is better in one particularly thing

such models exist for sdxl, too. Wanna have better photorealism then use custom models for photorealism

but these custom models are always overfitted as hell and don't work as good in general

SDXL is the better base model which can do more or less every style. From that you can improve with custom models

but in contrast to SD 1.5 which is shitty without custom models, you don't necessarily need custom models for sdxl

[strange mist]

nice name 🙃
you'd probably get banned in several discords.

[shy kelp]

Why would i

[strange mist]

you know why 😏

[shy kelp]

Not really

[strange mist]

that's ok. me neither

[half cedar]

The depth of detail in sdxl is much higher, prompt with it for a while you'll see.

[shy kelp]

You can recomend a good xl model?

The 1.5 model i use right now can do all famous people, im using it on the discord bot on our server right now

The xl model i use not so much

[stone fossil]

Pablo Picasso ∞ SD XL 1.0: https://civitai.com/models/162136/pablo-picasso-infinity-sd-xl-10

Introducing ∞, an innovative LoRa (Text to Image) model that embarks on a journey through the captivating world of Pablo Picasso's artistry.

Pablo Picasso ∞ utilizes advanced deep learning techniques to immerse your images in the essence of Picasso's artistry.
What sets this model apart is its unique ability to interpret specific prompts.
By including the year of a Picasso artwork and the name of the piece in your prompt, you can guide the model to transform your images in the style of that particular period.
Each transformed image carries the year of the artwork and the name of the piece in its captions, offering a unique connection to Picasso's timeline.

Use trigger word: p1c4ss0

Trained on 28000 steps from a highly detailed by hand captioned large dataset.

• Special thanks the people who help me and who I care a lot about:
  @MarkOREZ
  @upbeat summit
  @osiworx
  @mix
  @Thibaud
  @Kamikaze(Elon Musk)

[sinful falcon]

Thanks to the help of folks here and elsewhere I've published my first LoRA ❤️ Its for creating stylized storybook sketches. I hope it comes in handy for someone: https://civitai.com/models/162118/storyboard-sketch

[indigo carbon]

"as Donald Trump"

[shy kelp]

What model?

[indigo carbon]

I'm using the model I made by calculative merging with IPA for image input

[shy kelp]

You using a lora for trump?

[indigo carbon]

nope

for the first one it's just pure txt2img optimized with AIT, the second uses the first as input, then it remixes it with a prompt using IPA

on paper it can seem extremely complicated, but it's compressed into a simple workflow(s)

allowing SDXL to blend and/or remix images was a difficult task due to CLIP being a bottleneck for SDXL, but it was possible

pure txt2img is turned into a science by SDXL, but the rest of the capabilities are limited due to CLIP imo

if SD3.0 would be a thing, I'd say the goal should be upgrading to a better text/image encoder like BLiP; while still going crazy with the UNET like done with SDXL

[half cedar]

So great. Top right has amazing attitude

[zinc cargo]

i enjoy this Lora so much!

[zinc cargo]

blanka?

[static prawn]

what does that exactly means?

am i doing something wrong?

[noble shoal]

A bit unsettling, but here is an image of the tooth fairy. Not planning to release a children's book any time soon.

[shy kelp]

Nano-Virus

[vital ermine]

feels dejected now

[soft bone]

Loving Vincent (van gogh) finally making a comeback in Lora form for XL. coming soon

[ivory blaze]

Why do previous generations bleed ino completely new generatioons new prompts lol which one of you said this is not true lol

and i think it does so, only when attached to a seed maybe?

cos changing the seed tends to clear that weirdness, but switching prompts with same seed, does some goofery

no way robocop is ''judy alvarez" when NOTHING in the prompt suggests this, and my t2i adapters all have robocop images for depth and color transfer lol

I had made this one, and then it shat this out after,

i refuse to believe that it is not bleeding previuos promptts.

or something in the workflow is caching oddly

yep changed the seed and suddenly more robocoppy

[shy kelp]

I noticed the same thing when using MidJourney

[ivory blaze]

hmm weird, i wouldn't' think MJ would have single user cacheing of anysort, but yeah it was weird with same randomseed, this was automatic1111 not comfy so not sure how that translates

this is the not bled version that it made

[ivory blaze]

this makes me uncomfortable

not sure why

[crisp owl]

I said it, and the only way you could get similarity of images is similarity of prompts. Some words are harder weighted than others

Perhaps there's a lora turned on.
Or a custom prompt node not properly updating at new image generation.

[vital ermine]

Hehehe

[shy kelp]

How common is it for the image to have the prompt text legible in the image?

[ivory blaze]

depends on model,

your best bet is using inpainted external masks you make of the words you want

[shy kelp]

I was just curious because it happened and I'd never seen it do that

[ivory blaze]

oh yeah that'll happen, natural is spelled wrong lol

I did a bunch of wanted pposters they were like WWANИИED

[shy kelp]

prompt was "naturepunk"

[ivory blaze]

фтвand it was kind of weird mix of greek and cyrillic and latin alpha

[shy kelp]

which is interesting, it decided where to break the word

[ivory blaze]

well there are probably images trained, that have Punk and Nature separate in the image

but not "naturepunk" so not having it separate, i even less likely

[shy kelp]

I've only been messing w SD for a couple of days but did a month of MJ and never saw anything like this happen so it was a shock

[crisp owl]

sdxl can be pretty good with text if called for, especially if you "quote" your text.
It'll still of course mess up, but it's come a long way

[shy kelp]

Thanks for this info. Just to clarify, I didn't use quotes in the prompt in this case

[shy kelp]

How can I get it to fill in more detail? Currently using sdxl refiner and 4x-Ultrasharp

[shy kelp]

Also, is 512 too small of an empty latent to use for sdxl?

[crisp owl]

yuup

[shy kelp]

Do I want 1024?

[vital ermine]

yes

[shy kelp]

Thank you

[vital ermine]

Welcome

[crisp owl]

these are the recommended

[vital ermine]

LOL, I was just about to screen cap that for him too

[shy kelp]

That is likely my issue with details

[vital ermine]

100