✨｜sdxl

[hardy cipher]

I'm pretty sure we had the guineas book of world records 1984 sitting around when I was a kid

For some reason

Had the fat twins on scooters

That image was etched into my brain

Absolute units

[vital ermine]

Funny thing is I seen a picture like that before a very long time ago so might be one they scanned in?

[hardy cipher]

Back when extreme morbid obesity was a rarity

[cyan crown]

can you restore a photo like that in sdxl?

[vital ermine]

Yeah, when people were more active and had real food

[hardy cipher]

Yeah. Controlnet, and now ipadapter would probably help

[vital ermine]

I saw a whole video where AI is now being used to restore photos that use to take me 100 hours

does a damn good job too

[hardy cipher]

I haven't really found a good flow for it, but it's something I want to learn more about

[uncut fiber]

deoldify probably need some different UI, otherwise desaturate image and scale it very only

[hardy cipher]

Satan's turntable

What's going on here?

[stone fossil]

xD

[vital ermine]

When I say restore an old photo I mean replace missing pieces and remove creases and tears etc... that shit can easily take 100+ hours that now can be done in seconds

[stone fossil]

C0ck blocker.

😄

[hardy cipher]

Seems like something that would still need photoshop or gimp for optimal efficiency

But the ai cuts out a huge amount of work

[vital ermine]

Oh, yeah but all the man hours I used to do in under an hour. NGL, turned limp seeing that.

[uncut fiber]

Restore is also remove compression degradation. It is wide fields of things.

[vital ermine]

I only dealt with real photos none of that digital shit. if it was digital with jpeg artifacts go find someone else

[cyan crown]

and oof photos

^

[uncut fiber]

@vital ermine check this

[hardy cipher]

Tried restoring the Mona Lisa

[uncut fiber]

i am about compression artefacts in videos. TV series.

[stone fossil]

Perfectly done.

[hardy cipher]

Thank you

[vital ermine]

Yeah, but still can be helped like on the back wall.

[hardy cipher]

I think the key is a multipronged approach

[uncut fiber]

it was done manualy everything with krita. And it was version 1. just i made comparison image.

[hardy cipher]

Stable diffusion, photoshop neural filters, etc

[vital ermine]

I hate having to switch back to Windows but no way in hell am I going to allow a wayward dev contaminate my linux when they should be using a virtual environment.

[hardy cipher]

Thicc hand

[vital ermine]

someone was a bit overactive

[vital ermine]

I think this training went well

Now to see what happens when I call a guy if his arms are up

damn, I chopped his arms off and now he has missing arms

[vital ermine]

at least it gave him 1

[stone fossil]

Great idea thanks. 🙂

[fierce compass]

it seems to me that sdxl is kind of bad at "people in a bigger picture frame". Bad face, bad hands. I can fix the face pretty well, but I didn't have much luck about the hands

This is straight from the refiner

[floral island]

oof... i can accept it, but it's a cloned face 😦

[cyan crown]

yes it is

[floral island]

if i'm looking at my generations, the moment i see clone, it's skip

tends to happen a lot on 1280+ resolutions 😦

lyrics prompts do tend to fail. at times. lol

[mossy canopy]

sdxl material test

[fierce compass]

any hints on how to fix small scale hands? I have them separated with a hands bbox detector, and I'm running a sampler prompted with "one hand". It fixes them, but the results are far from perfect.

[mossy canopy]

canny?

depth?

[fierce compass]

I'm not running controlnet in this workflow, I guess that would be the next step

[floral island]

@mossy canopy just throwing in orb + material?

[mossy canopy]

yes.. testing

[floral island]

that's testing models tho, we need an all-intelligent model XD

[floral island]

amber?

anime fail xD

[floral island]

amethyst?

[mossy canopy]

yep :V

[floral island]

^^

[mossy canopy]

have you tested?

[floral island]

no, just guestting by color and transparency

[mossy canopy]

lmao

[native knot]

heh

[noble shoal]

Bismuth?

[mossy canopy]

ammolite

[noble shoal]

How about Vanta Black?

[native knot]

or carbon fiber

[mossy canopy]

brass

[fierce compass]

Somewhat better with JuggernautXL

[mossy canopy]

bumpy

[native knot]

That's interesting

[noble shoal]

How about wrinkly skin? 🤔

[mossy canopy]

citrine

[noble shoal]

Looks more like brain

[mossy canopy]

corundum

[mossy canopy]

cloisonne

cotton

[native knot]

Try broccoli

After that, try broccolini just to see if it does anything different.

[native knot]

Damascus Steel would be another potentially good one

or maybe kaleidoscope

[cyan crown]

After applying a Lora of me

[mossy canopy]

embers

[native knot]

maybe also try Fibonacci

That stuff gets wild

[mossy canopy]

embroided

[native knot]

oh, nice

[floral island]

dudes. my verser prompt desires song lyrics

[native knot]

lol

[floral island]

give me your song lyrics so i can visualise them into sd!

[native knot]

drop in Purple Rain's lyrics

[mossy canopy]

epoxy

[floral island]

it generates with anime flavour. lets see if it gets any good output in 8 hits 🙂

[floral island]

ooof, this is hard.the model knows the songtext to purple rain so i get prince even if i dont'want to

[floral island]

this is my kind of orbs ❤️

[mossy canopy]

just fire

[floral island]

fire orb?

[mossy canopy]

yes

[floral island]

it's a good orb

❤️

[mossy canopy]

fractal

[floral island]

mandelbrot!

add mandelbrot

[mossy canopy]

girih

[floral island]

not doubting girih orb. that's awesome!

[floral island]

ooof... the model certainly has had prince lyrics fed at it at some point. the bias is too heavy to have creative outputs 😦

[floral island]

glo.ri.ous.

[native knot]

Bummer...maybe try The Sound of Silence.

[floral island]

oooh, i certainly think that's a very good choice

lemme run it

[floral island]

infernal

?

[cyan crown]

fire sphere + Lora

[floral island]

lora = cheat XD

show us ur pure prompting skills

[floral island]

still ❤️ output

[floral island]

this soccer ball will disintegrate the world if you kick it

i'd kick it

[native knot]

Makes me think you should try something like "Portal by Valve"

[floral island]

snow globe + ???

[cyan crown]

still cheating

[mossy canopy]

herringbone pattern

[cyan crown]

water globe

[floral island]

cheats are just there to have more fun ❤️ lovely

[mossy canopy]

furry

[floral island]

#🍥｜anime need this output too

[cyan crown]

this was the snow globe

[floral island]

i gotta run these without the anime flavour!!! too good output to be tainted by weeb shit

[mossy canopy]

iridescent

holographic

ironwork

[ionic dragon]

Try neonpunk and psychedelic

[mossy canopy]

ivory carving

[native knot]

How about an ebony carving, then?

[ionic dragon]

Dayum, that's great
What's the prompt

[mossy canopy]

kyanite

[ionic dragon]

Can you even try steampunk

[mossy canopy]

psychedelic

[ionic dragon]

I mean the entire prompt

[mossy canopy]

jasper sphere shaped orb in a white background

[ionic dragon]

Just that?
No tags?

[ionic dragon]

Try acrylic pour

[floral island]

very good output, check dm 🙂

[ionic dragon]

Are you using depthmap?
Because the position is kinda consistent

[mossy canopy]

nope

khokhloma

[ionic dragon]

@mossy canopy why don't you create an animation

It would be great

[cyan crown]

I need an help with controlnet for sdxl. What is the difference between diffusers_xl_canny_full.safetensors and t2i-adapter_xl_canny.safetensors

?

[mossy canopy]

great idea..

[floral island]

this orb is glorious!

[floral island]

<- has played path of exile to much. any glorious orb is ❤️

[mossy canopy]

linguini pattern

[vital ermine]

POE, sigh

[mossy canopy]

light particles

[ionic dragon]

Try cyborg style

[floral island]

shh... you know orbs are life

[ionic dragon]

Use depthmath

[vital ermine]

LOL

[ionic dragon]

So the spheres are consistent

[floral island]

any sdxl model will accept orb 🙂 *should*

[vital ermine]

I still have nightmares going deeper and deeper

[floral island]

the purest of orbs ❤️

[ionic dragon]

Try van gogh

[mossy canopy]

liquid bismuth

[ionic dragon]

Milkyway, dusty

[mossy canopy]

mandelbulb

malachite

[mossy canopy]

magma

liquid gold

[ionic dragon]

I mean both different
Milkyway
Dusty

[mossy canopy]

moldy

[ionic dragon]

Which workflow do you use?

[mossy canopy]

milkiway

dusty

[mossy canopy]

pure prompting

mesh wire

mesh

marble

[mossy canopy]

opal

[vital ermine]

https://civitai.com/models/143415?modelVersionId=159160

[mossy canopy]

onyx

ornate

polka dot

paper

pasta pattern

[ionic dragon]

Try Damascus

[mossy canopy]

sapphire

[mossy canopy]

?

spinel

[ionic dragon]

Damascus knife style

[mossy canopy]

sequin

[azure oxide]

These are nice, reminds me of materials in unreal engine.

[ionic dragon]

Try
Pointillism
Glitch
Maximalstic
Baroque
Sci-fi
Vaporware
Cybernetic

[mossy canopy]

tinsel

[mossy canopy]

The END

[ionic dragon]

@mossy canopy can I dm

[mossy canopy]

dm

[vital ermine]

Old skool vibe

[upbeat summit]

lots of posts - nice posts ;). textures and geometric shapes are always great. nice work!

[noble shoal]

The Skinphone

[vital ermine]

See I released two new loras? Just doing all the ones I tried for over six months to do on 2.1

[upbeat summit]

nice. I've seen them and will check them out.

[vital ermine]

7 months ago, sheesh

[upbeat summit]

at least now you can make some of your ideas happen with SDXL

[vital ermine]

Yes

Dunno why 2.x was such a bitch to train

1.5, and now XL easy

I mean all of these have been training the TEs too, oops, as I just noticed that. Still easy.

[noble shoal]

That are some impressive before/after pictures.

And here is some eyebleach

[crisp owl]

Someone's already taken a bite of that

[floral island]

well, interesting instrument

[floral island]

ketchup..... lots. of. ketchup.

[noble shoal]

"photo of a very disgusting thing" gave me whatever this here is.

[floral island]

it's disgusting.

your prompt did not fail.

[vital ermine]

sob, civit flagged another image for review.

[floral island]

heart for the environment?

[floral island]

is very nsfw... because of reasons

[vital ermine]

Did you see the first one they flagged instantly?

[floral island]

no

but i've had my share of images tagged too for review

[vital ermine]

My nsfw no longer works and I got hit with dogs, balls, and horse penises yesterday

this one was instantaneous

[floral island]

obviously, the way she is handling/handing her meat.

[vital ermine]

obviously

If I turn on the NSFW filter I lose even one of my images because I now have to be full on allow porn or full on no mature content. iow, the majority of the site goes dark

They broke something or it was intentional

[floral island]

they just full on embrace the meme

'the internet is for p0rn'

[vital ermine]

100%

[noble shoal]

Does it give you a reason why it is flagged?

[floral island]

more ketchup goodness

[floral island]

100% automated detection

[vital ermine]

I went to their civit discord and it threw me into a room with voice

[noble shoal]

Flagged

[floral island]

too much?

[uncut fiber]

@noble shoal it is and car tyre

[vital ermine]

They never ever give a reason you just wait until a mod checks it and passes it

[noble shoal]

Maybe it's falsely flagged because of an IP.

[vital ermine]

IP as in internet protocol or intellectual property?

[noble shoal]

Maybe for civitai

[noble shoal]

Second

[floral island]

you called my image flagged, if it's not ok, i'll remove

[floral island]

no problem

[vital ermine]

Personally, I have always despised that damn place

[noble shoal]

No, I am totally fine with that

[vital ermine]

problem is that is the only place since HF is not really meant for sharing outside of a university

[floral island]

civit obviously has to adhere to some 'community standards' too

which by definition was the reason civit came into existence

[vital ermine]

Oh, rly?

[floral island]

any big platform has to

as dumb as it is 😦

[vital ermine]

With the filth I see on that damn place I call BS to any standards out side of a cesspool

[noble shoal]

Maybe General Awarenesses images are too less p0rn 🤔

[pallid path]

too much drip

[floral island]

agreed

[vital ermine]

I can upload an image of a big titty bimbo and it passes

[vital ermine]

I seriously think if I did porn upload it would pass and no need for moderation.

[floral island]

i think it goes for moderation too

[vital ermine]

I am going to try it. Screen grab a porn hub clip and see

[noble shoal]

Problem successfully solved

[floral island]

got awesome image -> has female nipple, can't show here because of reasons. still awesome image

[floral island]

alphonse mucha was involved in this image

[vital ermine]

yeah, I see a lora for it

[floral island]

art through the ages -> nudity is good. now?

meh

artistic nudity should not be frowned upon, but something something conservative values.

[vital ermine]

I always found it more sensual than butt nakedness.

[floral island]

i mean, if you go for obvious exposition of genitals, yeah, that's not artistic nudity

sensual is skirting the line, and that by definition is art

i mean, aphrodite? good luck finding stuff of her that does NOT depict her in ways that would be ok on this channel

[vital ermine]

well, I know where my line lies.

[noble shoal]

Yeah, I have an awesome image of a computer mouse that looks a bit too much like the female reproduction organ entry part. Still an awesome image, though.

[floral island]

sobs in acknowledgement that he cannot see it publically

[noble shoal]

You could try it with: "a product photo of a computer mouse made out of wrinkly human skin, wet, product photography" 🤷

[floral island]

i can imaginary see what this prompt would do

[noble shoal]

The mouse wheel does also it's part

[floral island]

lemme try it after this batch xD

if my eyes become cursed after this, i know who to blame

[noble shoal]

Yourself, your life decisions and the trust in others prompts.

[floral island]

stuff like this? XD i have a g403 and damn, does this look similar lol

[noble shoal]

Close 😅

[floral island]

i can imagine people not even wanting to be remotely connected to these abominations... i think i might be sick if i ever touched one of those

very good prompt ❤️

[noble shoal]

I would buy one.

[floral island]

i need to improve upon this

i should go to bed, but i have to make it more gross!

[noble shoal]

More wrinkles! And maybe strip the "wet" part.

[floral island]

nah, the wet is good.

[noble shoal]

Oh, and "computer mouse" is exchangeable. Think out of the box. Maybe a whole gaming pc

[noble shoal]

I am so proud of you

[floral island]

with each click you hear a "splish"

shudders at the thought

lol, it has some issues understanding keyboard lol

[noble shoal]

I think it should have non tactile switches. No click, just a soft touch like red switches.

[floral island]

you just have to splish hard enough

holy shit... just got one that's 100% horror certified

it's. glorious

these are all glorious 😮

imma post them in #1072015504870494359

[noble shoal]

You got my attention

[floral island]

i put them where they belong, they're lovely ❤️

[noble shoal]

Nice, if I find more prompts like that, I'll let you know! Ok, I am heading to bed. Good night

[floral island]

organic pc case lol

gotta go to bed too ❤️ sleep well

[uncut fiber]

GN

[zinc cargo]

eXistenZ vibes 🙂

[uncut fiber]

Does it make sense in English?

[eternal fog]

Not really no

[uncut fiber]

😦

[hardy cipher]

the words make sense but not necessarily in that order

[uncut fiber]

thats main difference in english and my language. English word order is must here it doesnt matter much 🙂 Thank you @eternal fog and @hardy cipher

[hardy cipher]

what's your language?

very nice. I might run that through ipadapter

the more I use it the more I realize how powerful it is

[hardy cipher]

it's effectively like using an image as a lora

I realize it's not a 1 to 1 thing and they're two different concepts

[hardy cipher]

neato

[hardy cipher]

I've been experimenting a lot with different combined approaches

[eternal fog]

When you add a woman into it

[ionic gulch]

LOL:
text_positive_g_styled: cinematic film still "American Psycho," directed by Mary Harron, Jennifer Lawrence
text_positive_l_styled: shallow depth of field, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy, "American Psycho," directed by Mary Harron, Jennifer Lawrence
text_positive_styled: cinematic film still "American Psycho," directed by Mary Harron, Jennifer Lawrence . shallow depth of field, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy, "American Psycho," directed by Mary Harron, Jennifer Lawrence
text_negative_styled: anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured

[eternal fog]

Think you broke it

[hardy cipher]

using the ipadapter for one image, then running the clipvision g on another image along with controlnet, then using one of those or a third image as the input latent

two many images gets a bit weird though

so kind of hard to mesh it all

[eternal fog]

I've been doing that sort of stuff

[hardy cipher]

red rectangle?

[eternal fog]

I don't use text prompts much anymore lol

[ionic gulch]

yes, and not only once, but twice with different seed. this is all the model knows about that movie i guess.

[hardy cipher]

I've been getting a lot of black boxes when I push the ipadapter parameters too far

seems like things need to make sense together for it to work properly

[ionic gulch]

this was just txt2img with that long prompts

[hardy cipher]

tried to improve this image

[uncut fiber]

czech

[hardy cipher]

I've been trying to get a better understanding of the flow of data from the ipadapter onward to try and figure out where the values turn into nans

[ionic gulch]

can you try if you also get a red screen with a prompt like that?
text_positive_g_styled: cinematic film still "American Psycho," directed by Mary Harron, Jennifer Lawrence
text_positive_l_styled: shallow depth of field, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy, "American Psycho," directed by Mary Harron, Jennifer Lawrence
text_positive_styled: cinematic film still "American Psycho," directed by Mary Harron, Jennifer Lawrence . shallow depth of field, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy, "American Psycho," directed by Mary Harron, Jennifer Lawrence
text_negative_styled: anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured
changed the movie to blade runner and it works again:

[hardy cipher]

and wondering if there could be some kind of band aid put on the data, maybe some thresholding or something if things are out of bounds. or if it simply doesn't output anything

[hardy cipher]

I'll try it out

[ionic gulch]

the movie might have had that red screen, but that shouldn't be the only thing sdxl knows about it (base model)

[hardy cipher]

I wonder if it'd work with juggernaut. I'll have to move the base model back onto my ssd

keep my extra models on my hdd, but don't load from there since it literally takes minutes

hmm, loaded that workflow and sooo many nodes I don't have, but then I installed yours. did you update a bunch of things?

well I installed them quite a while back

I need to update the description on mine

[ionic gulch]

this is what juggernaut gave me for that prompt:

[hardy cipher]

lol

[ionic gulch]

seems to be a problem with multiple models

[hardy cipher]

weird

well I'm going to go ahead and update things and then play around with it

with my painfully slow video card 😭

bought a nice laptop right before I got into stable diffusion. was on the fence between desktop and laptop

[uncut fiber]

what gpu do you have?

[hardy cipher]

I chose wrong

3060, but the laptop version which is 6gb vram

it works, but it's not fast

[uncut fiber]

i am bit better, just a little bit 3070

[ionic gulch]

all other movies work. this is "the departed" for example:

[hardy cipher]

I've considered getting an external gpu of some kind. but then at that point I might as well just get a full on tower of some kind

[uncut fiber]

if only AMD is working better...

[hardy cipher]

stupid laptop has 64 gb ddr4 ram. but does that matter? nope

[ionic gulch]

that model/lora is really great. gives such good pictures almost every time.

[eternal fog]

Yeah it's NightvisionXL, using IPAdapters and Revision

[ionic gulch]

do you get anything from revision that ip adapter doesn't do better. haven't had much success with revision. always creates some kind of artistic styles instead of the style of the input images.

[hardy cipher]

what the heck is revision even? is it separate nodes or just a different workflow?

[eternal fog]

I'm using it to sort of mix stuff. I can't really use 2 IPAdapters because I run out of VRAM and it starts going really slow.

[hardy cipher]

cause I don't have revision nodes that I'm aware of

[eternal fog]

Revision is the name that SAI gave to the UNClip Conditioning/Clip Vision G stuff

[hardy cipher]

ahh

I like to run that in conjunction with ipadapter

and then maybe controlnet as well

[eternal fog]

Yeah that's what I'm doing, you can run them together to get different results.

[hardy cipher]

clip g with the regular clipvision and then that other model with ipadapter

clip g with ipadapter doesn't seem to work that well

[hardy cipher]

well clipvision g

clip h works, but it's kind of tricky and don't really know what benefit it gives me

[eternal fog]

Yeah I can't see any benefit to using that one either

[hardy cipher]

maybe saves a marginal amount of vram

the one thing that eludes me is stringing the ipadapter model input/outputs together

I have tried so many different combinations of strengths and configurations, and nothing has worked

ahhh, these are mikey nodes. I guess I need mikey nodes

[hardy cipher]

look at all those lora loaders

[ionic gulch]

i love 2x ip adapter:
input / results (base / refiner)

[hardy cipher]

man, are you using nodes you haven't uploaded to github in that workflow?

[ionic gulch]

yes. many changes are coming. but i think i'm at the limit what comfyui can handle, because i have some perfomance problems sometimes. have to figure them out, before making the workflow public

[hardy cipher]

ahh, you apply one of the them to the refiner?

that makes sense I guess

also, Adapater, lol

[ionic gulch]

oh thanks. will fix that.

[hardy cipher]

what are positive and negative styled prompts? are those the refiner prompts?

[ionic gulch]

jennifer lawrence, that's all. still juggernaut model from my test before.

he got the whole style mix from the two images

the simplified blonde hair, the pink dress, etc.

[eternal fog]

Its interesting putting images of clothes into IPAdapter

[hardy cipher]

I mean. What are they though? Don't want to sound ignorant, but I'm not familiar with style prompts unless that refers to refiner

[ionic gulch]

it was just "jennifer lawrence" as prompt" and the two ip adapter images. everything else should have been turned off, including the style optíons

the new workflow - would be released for a while, if i didn't encounter those performance problems. sometimes it take 5-8 minutes to start on my 4090 - just doing nothing:

was able to reduce the 3rd party node packs by 50%.

[crisp owl]

I've found that organizing different flow paths in groups and bypassing the entire groups greatly helps the unnecessary node loading, thus reducing the desired paths generation time

[ionic gulch]

wish there was a way to do that with real switches instead of bypassing or cutting wires

50% of the time i spend on my workflow is because comfyui can't really handle that scenario and you have to find workarounds.

[crisp owl]

Yeah I posed that request in the comfy Github Discussions area the other day

To be able to have bypass connections

Impact pack did just release a new switch though

Multiple outputs

[ionic gulch]

and now i've hit a limit where wires get disconnected randomly when saving and some wierd performance problem making my system do nothing for minutes. it just says "got prompt" and then you have to wait from 5 seconds to 8 minutes, before it starts. i think it's the reroute replacement i had to create because of the lost wires. that system seems to cause comfyui go crazy sometimes. even if the seed is the same, it take 2 minutes to notice that, while chaging the seed only takes 70 seconds to create four images (incl. 2 upscaled versions). so it takes much longer to realize everything is in cache than just running the workflow.

[crisp owl]

https://www.youtube.com/watch?v=KLe1hShwVDs

[hardy cipher]

that sounds pretty goofed. have you tried to pinpoint what's happening?

lol

such beauty

[ionic gulch]

yes, but no success. guess i will have to try going back to my old pipe-system, but it takes an hour do do that and ususally after that another 2 hours fixing wires i've connected in a wrong way. :)

[hardy cipher]

just need to add a few dozen print statements to every node

[ionic gulch]

will have to try that. but those "any" switches have some problems too. was talking with lt. data yesterday and he confirmed some problems with detection of the type - that also causes problems at least in a complex workflow with many of those switches. those any-switches were part of my problem with lost wires.

same as reroutes. everything that detects the type in real time is problematic and may lead to lost connections.

[ionic gulch]

sometimes i had to reconnect 10 wires after saving/reloading my workflow. that problem is gone with the new solution, but now i have performance problems, i didn't notice before.

[crisp owl]

Interesting, I've just got one "any" switch in my main flow, luckily haven't had issues with it though

[ionic gulch]

that also explains why some people told me they had to reconnect some wires after downloading the workflow from github. that was already caused by those lost connections.

[hardy cipher]

that's strange

[ionic gulch]

the concept is great - that's why i had switched my whole workflow to those switches a few days ago. but after that the lost wires got much worse. i've used about 20 of those switches.

but just as x to 1 switches for all types, not to cut off parts of the workflow.

will have to check that feature. maybe it's worth it to use a few of them at least

[hoary saddle]

What sorcery is this? I gotta research ipadapter now

[hardy cipher]

ipadapter is next level

[ionic gulch]

oh, i just missed that question. they were generated by a tool, but you can just copy the text-part behind the colon - that's all the tool does. the rest is just added in the console output.

[hardy cipher]

gotcha. haven't went too deep into all that stuff just yet. I'm trying to understand conditioning data better atm. the deeper I go with this stuff the more I realize there is

[ionic gulch]

ip adapter + txt prompt "Alien on a christmas market.":

[hardy cipher]

I guess that's true with anything

have you messed with using masks to focus prompts or ipadapter output on particular portions of the image?

[ionic gulch]

i've tried masks once. but it didn't work - got some ksampler error. only used masks for inpainting regularly.

[hardy cipher]

you can use those solid masks

not my workflow

[hardy cipher]

damn. that's awesome

[ionic gulch]

ip adapter is magic. those aren't cherry picked. just 2 out of 3 runs come out somehow great or interessting.

[hardy cipher]

the paper on it came out less than a month ago

[ionic gulch]

the next one was even better:

[hardy cipher]

at first I thought it was just some clipvision mod or something. but it is not

[hardy cipher]

She's amazed too

[ionic gulch]

even her face came out great. that's not so common with background people

[hardy cipher]

It's really going to be a rabbit hole of mixing and mashing different images and concepts together

[hardy cipher]

Exactly

I might need to set something up on runpod or a comparable service so I can output this faster

[ionic gulch]

the 2x upscaler also works great for that kind of pictures.

[hardy cipher]

takes over a minute to render with my current workflow. I need more speed

[ionic gulch]

all 4 versions 2x standard + 2x upscaler also take 90 seconds on my 4090

[hardy cipher]

well yours are a bit more complex than what I'm doing at the moment. I'll sometimes work into longer flows like that, but don't start out there

[hoary saddle]

I'm not at the PC currently but is there a good workflow to test ipadapter I can snag in a bit? Very intrigued

[hardy cipher]

I just connected nodes until they worked right. just make sure you use the right models. it's a bit finicky

I'm sure there are lots of workflows out there. don't know of anything specific

[ionic gulch]

the old workflow on my github should still work for 2x ip adapter.
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

my latest one (included in the uploaded pictures) has unpublished nodes

[strong field]

IPadapter my fav

[strong field]

The demos on the GitHub are useful

[strong field]

https://media.discordapp.net/attachments/1089974139927920741/1149438111999348886/image.png

IPadapter mask, inpainting, etc.

I’m not at comp so can’t share actual workflow but that’s easy enough to copy from the ss

[ionic gulch]

it can even mix pictures with different angles (needs a few more runs to get a flawless one):

[native knot]

This has been how I've felt battling my son's computer for the latter half of the day today.

[half ivy]

yusss, IP Adapter is where its at?

[native knot]

https://github.com/laksjdjf/IPAdapter-ComfyUI

[glad fulcrum]

I updated auto1111 and reinstalled control net extension, but my ip adapter models aren't showing up...

[glad fulcrum]

Oh I forgot to rename to pth

[hardy cipher]

Cursed

[vernal spire]

tag yourself im second row second from the right

[sleek pulsar]

TRIEMIRP

and ZTRUMP is just an alien hiding in human skin.

[hardy cipher]

I like how regular trump showed up in the top left

[sleek pulsar]

the 4 pictures on the left all look ok

[hardy cipher]

this image is aliens only

not quite passable as humans

[sleek pulsar]

that baby cry mouth though on all of them. like copy pasted. You might want to create one image after another and create a collage yourself if you want good results. making such a collage just to see what the limits are does make sense though

did you use a trump lora?

[hardy cipher]

oh, these are just mid range images. they're not the finished product

no, this is ipadapter stuff

just trying to make cursed images

[sleek pulsar]

oh ok, so you have 'photo of Trump' in the style of 'newspaper with several images'

[hardy cipher]

not exactly. used a picture from some old yearbook or something along with a picture of trump

and it put them together

[sleek pulsar]

ip adapter stuff seems very interesting and promising. what kind of gpu do you need for that sdxl in general seems very resource intensive

[hardy cipher]

well, technically you could run it entirely on cpu, but it'd take forever

but realistically probably at very bare minimum 4gb vram to render things without hitting the cpu which very drastically slows things down

6gb will be a lot better

and then from there it gets better as you have more obviously

[sleek pulsar]

for SDXL?

[hardy cipher]

sure

[sleek pulsar]

and IP adapter?

are you using auto or comfy then?

[hardy cipher]

it's not like there's a set cutoff minimum for most of the rendering stuff. it'll just hit your cpu like I was saying. and that is super duper slow

comfy

[sleek pulsar]

yes i know i used CPU and it took me 50 minutes for one image in the past, but I will buy a new computer soon and i am scouring of what's possible.

[hardy cipher]

well there are guides and what not. but if you're trying to get something that's optimized for ai things and you're on a budget, just go for the most bang for your buck as far as vram is concerned, and I'd strongly suggest nvidia. don't skimp too much on the other stuff. you'll want at least 16-32 gb of ram as well. you could get by on 16, but 32 would be preferable

I always max ram out anyway since it's relatively cheap as far as computer components

[sleek pulsar]

yeah, i thought 8gb vram and 32gb ram, but not sure if that's in my budget.

i need a laptop, and those are already limits you easily reach when you try to stay affordable.

[hardy cipher]

yeah, I spent 2k on a laptop right before I got into stable diffusion. ugh. was on the fence between laptop and desktop, but figured it had all I would want as far as hardware. but turns out that was not true 😭

and so I'm really not that enthusiastic about putting out a bunch more money on a computer right now

[sleek pulsar]

yeah my budget is also under 2k (us dollar i assume) for a laptop, so i already know i won't be able to do everything.

but i am a foreigner, so i travel between continents and am unclear where i end up working. i don't want to buy a desktop and deal with shipping it around the world in a year and being unable to use a computer when visiting family for several weeks at a time.

so gotta be a laptop for now

[hardy cipher]

You could use something like runpod. Like 75 cents an hour to use a 4090. And 7 cents a month per gb for storage

They have h100s for a few dollars an hour.

[sleek pulsar]

yeah, might be a thought as well for things that i can't do myself.

[hardy cipher]

Not sure why you'd need an h100, but you could also use 8 at once if you want

[sleek pulsar]

i thought about using a cloud service if i want to use things what require a ton of VRAM, as i am sure future models will do, but i am always a bit worried about privacy for these services, especially if i build a dreambooth with personal photos or something like that.

[hardy cipher]

Well there's that. Definitely read the fine print. You could also consider an egpu

But those aren't exactly great on a limited budget

[sleek pulsar]

well i buy a new laptop anywhere, and if i can do most things with 8gb for now, then i won't need it

[hardy cipher]

Not sure what the bare minimum requirements are for training. I've always used cloud services for that. But 8 gb vram would be sufficient for most things. Might not be blazing speeds but it'll work fine

[sleek pulsar]

yeah. havn't looked into cloud computing a lot, it's also legally kind of awkward, i don't want to get banned after paying if it's US servers when I am in China.

[hardy cipher]

yeah. I'd been using colab for things, was never the best option but it works with my google drive. but they've decided to snuff out stable diffusion usage it seems. so I'm exploring other options

[lusty wolf]

When are going to release the new nodes for your workflow? Would love have a play... Gr8 work!

[hardy cipher]

I wonder if anyone actually uses the nodes I put up

[peak dove]

SDXL Bot, prompt = beauty in the style of BLK Art and tinga tinga art

[noble shoal]

Hmmm, Coffee

[ionic gulch]

there's a performance problem with the latest version of the workflow. i'll have to figure out the reason (size/node limits of comfyui, bug in comfyui or one of the nodes) before i can upload it to github.

as many older nodes also changed (this is the biggest change i've done yet, with almost 50% of the workflow and nodes rewritten) i'll have to update workflow and nodes at the same time, because an update of the nodes would break the latest public workflow. so i guess it could take up to a week until i can release the new workflow and matching nodes. unless i can figure out those performance problems much faster than expected.

for 2x ip adapter the old version of the workflow and nodes should work almost exactally as the examples i've posted yesterday.

[ionic gulch]

funny thing about the performance problems is, that once everything works, the workflow is fast. that's why i could post all those ip adapter examples yesterday. the performance problems are only at the first start or if you press "generate" with the same seed (so everything should come from cache). in that case it just sits at the "got prompt" message for up to 8 minutes. changing models or generation mode can also take some time, but not that extreme. no ram or vram usage when it's happening. just hitting the cpu enough that my fans are going up (but still only 20% in windows performance monitor), so maybe only one core is used to 100% (have to check that in more detail). to me it looks like some internal comfyui process goes mad and starts a lot of conflicting threads - so i guess i have to work around that behavior.

[uncut fiber]

cant it be memory issue, as what you described is what i got with AIT keep in memory enabled.

[ionic gulch]

i don't know, as there is almost no ram and vram usage and the low values don't fluctuate.

my guess is, that comfyui tries to "solve" the new pipe system with special nodes instead of redirects and tries to do all at the same time which overloads a single cpu core with hundreds of concurrent operations or something like that.

but as there is no debuging info in the console window it's hard to figure out.

[uncut fiber]

yes

[nimble heart]

just like add some print statements in a few functions

it's python so you can just add print(thing) anywhere you want and restart the server

[ionic gulch]

guess i'll add some print statements to the new pipe nodes - so i get some info when they are executed. at least i hope so - because the caching mechanism could prevent the nodes from running their code if nothing changed and it's more an internal comfyui thing.

[nimble heart]

make an anti-cache node that has a random seed input that doesn't do anything

just like input latent + seed output latent but all it does is pass through the latent unmodified

or model I guess if it needs to be higher up in the tree

[oak field]

Is there any indepth guide which can teach me a1111. YouTube tutorial r not working for some reason

I just started

[nimble heart]

play with it a lot and read the official wiki on the a1111 GitHub

[ionic gulch]

will try a few things when i get home from work.

[uncut fiber]

not sure what you are looking for @oak field
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features ???

[oak field]

For example in one video they tell me about civitai. How to create same style by simplying putting img in pnginfo and extract the data set to create the same exact image. After many trials and doing everything shown in video. I didn't get result which anywhere near to desired result.

[oak field]

So can I get to it by simplying googling "GitHub a1111 wiki"?

I m not a tech guy

[nimble heart]

it's what Bernix linked

[oak field]

Ok thanks

[nimble heart]

Also lot of guides and model pages cherrypick the hell out of their results so random seeds won't look as good.

[oak field]

I will keep that in mind. One more thing the sd1.5 pruned i got. On 512*512 giving me very very very bad result. Like it hasn't to do anything with prompt. On the otherhand dreamshaper for sd1.5 has some stunning result but it has very cartoonish style

For me Dreamshaper sd 1.5is actually showing better result than sdxl

[nimble heart]

Try Realistic Vision for a 1.5 model that isn't anime waifu.

[oak field]

Isn't sd1.5 and dreamshaper 1.5 should be same?

[nimble heart]

?

They won't give the same results because they're different models

[oak field]

I m trying to work on model which isn't trained and just sd1.5 hoping to get more tailored result.

[nimble heart]

dreamshaper tunes sd1.5 for things people commonly use stable diffusion for, like cartoon waifus

[oak field]

Yes so the artstyle will be different. But why pruned1.5 creating the worst possible, most unrelated to prompt

[nimble heart]

It takes time to sort of learn what words and phrases stable diffusion likes and doesn't

I'm assuming it's your first day. Like most other crafts you start off sucking at it before you become kinda good at it.

That's why in my initial advice I recommended just playing with it. See what does what

[oak field]

Yes I downloaded it yesterday. So pretty new to this. Ok so for sd1.5 prune I need to really hammered the nail to the home when it comes to promot. While on trained model I can use it more freely. Thanks for the info man

[potent flame]

look pretty great to have 10 minutes per pic

[noble shoal]

Ok, time to stop. I think nobody would ever eat this.

[potent flame]

oh yeah without the second pass

[oak field]

U haven't meet french people yet?

[ionic gulch]

could use a second pass in the oven . :)

and i hope the eyes are only decorated olives.

[tepid surge]

Hi eveyone hope eveeryone's doing welll

[tepid surge]

cx you prompt wizard these are amazing

[noble shoal]

Thank you

[nimble heart]

No offence but the way you phrased those messages gave me Vietnam flashbacks to those giveaway scam bots

[tepid surge]

I recorded a thing for a project i'm working on and was watching it back and literally just was like wow I never realized how charming I was

[nimble heart]

Damn where'd you get so much confidence can I have some

[tepid surge]

So that's was funny to hear but yee that hi everyone def like hr speak cx

[tepid surge]

https://tenor.com/view/adventure-time-tree-trucks-mr-pig-alien-high-strangeness-gif-27409669

it's... in you already! <3

cx ✨

i need to sit and chill xD cx

[carmine gale]

is anyone familiar with Kohya_ss full Finetuning module, specifically the metadata generation?

[nimble heart]

dang now Im remembering I still didn't finish adventure time

[uncut fiber]

how to use malfunctional Microwave oven 🙂

[uncut fiber]

Tried to be more aesthetic 🙂

[noble shoal]

Understand. Let me also put my stuff in a microwave.

[nimble heart]

When I microwave a McDonald's dollar menu burger

[uncut fiber]

😄

[uncut fiber]

@noble shoal rather not readable 🙂

[uncut fiber]

jugernaut?

[heady vale]

nightvision

[indigo carbon]

I'd like one ℸ ̣ ⍑ᒷ ʖ╎⊣ ∷ᒷ↸ ᔑ!¡!¡ꖎᒷ ⎓ᒷꖎꖎ ℸ ̣ 𝙹 ℸ ̣ ⍑ᒷ ⊣∷𝙹⚍リ↸. ℸ ̣ ⍑ᒷ|| ℸ ̣ 𝙹𝙹ꖌ ℸ ̣ ⍑ᒷ ᔑリ↸ ℸ ̣ ⍑ᒷ ᓭᔑ∴ ℸ ̣ 𝙹 ℸ ̣ ⍑ᒷ ⎓𝙹∷ᒷᓭℸ ̣

[oak field]

I tried to make an order and accidentally summoned the lovecraftian monster. Send help

[indigo carbon]

I asked the waiter one question and my face melted off

[oak field]

Lol

[uncut fiber]

@oak field from archive

[shy kelp]

is promting with XL more dificult?

[delicate grotto]

Its different for sure

Less realistic stuff feel harder to prompt

[floral island]

@noble shoal skelecar 😄

[noble shoal]

Isn't that the dish that is eating you instead?

[indigo carbon]

nah, that's minecraft enchanting table language

[native knot]

Last night's Randomizer run had some nice renders!! As I've done for the past few days, I've been adjusting The Randomizer a bit before I run it. Last night I incorporated ChatGPT into the mix.

[peak dove]

Very consistent look!

[uncut steeple]

What's your setup?

Like with the randomizer

[peak dove]

Where do I get a Randomizer? 🙂

SDXL Bot

[native knot]

Right now I have 4 parts being fed into the a concatenator.

First is a text file of about 30k prompts from sandlot posted here: #🌠｜show-and-tell message
That now gets fed into ChatGPT and is processed into the description for part #1.
(This uses the text file loader and the random line of text nodes from the WAS suite, and the ChatGPT Simple node from the QoL suite v2.)

Part #2 uses a random value out of a big list of subjects which is then fed into the SDXL Auto Prompter all set to random, and that's added into line 2.
(This uses 2 primitive nodes pulled out from the concat random node from the QoL suite v2.)

Part #3 takes what I have in my regular positive prompt and adds that.

Part 4 uses the DynamicPrompts Custom node.

All of those are fed into tinyterraNodes 7x TXT Loader to concatenate them, then those are passed through into the Show Text node from pythongosssss just so I can see what the results are. It looks like this:

My pomegranate queen.

[native knot]

You build one. 😉

I found the basic ones where it's like one node doing randomization are very limited, so the complicated one I built above adds a massive amount of further randomization to the prompts.

That said, I think I'm going to shake some things up and move the ChatGPT node to the other side of the 7x loader so that everything else that's being fed in also gets passed into ChatGPT. This will still maintain randomization, but will probably help with the way the prompt is worded. toward the end of it.

That said, I got some really great stuff...like this:

Oh, and the other thing is that when I'm running The Randomizer, I've also got my resolution set to pull from Comfyroll's Aspect Ratio SDXL node. I pull the dimensions out into a Primitive, set the width/height to the super-wide 2048x512 (so that when it lands on "custom", it uses that aspect ratio which isn't in his drop-down), and set it to randomize after each run. Looks like this:

His dimensions drop-down has these choices:

So it'll pick one of those or my custom res.

[peak dove]

Cool!

[native knot]

Last night, I ran everything against JuggernautXL. The night before that, randomized what model. Later, I'll probably choose another model to do another run. Each night I've ran about 500 images. As far as ChatGPT, it only cost about $0.30 to run 500 prompts through it. I have my account spend capped just in case something goes wild, but at that price I could run 500 per night for an entire month and not break $10.

[peak dove]

What does ChatGPT actually do to the prompt?

[native knot]

Takes all the input you give it and turns it into a natural language prompt with some randomization of its own added in. I could feed it literally whatever I want and it'll spit back out something that expands and adjusts the verbiage to be more descriptive. Adding additional descriptors via natural language often helps with SDXL prompts.

For instance, that dragon one above, the prompt ChatGPT spit back was this:

Cloaked in magnificent armor adorned with intricate engravings, the dragon scales glint in the bright, warm lighting of the scene. Rays of golden light cascade through the dense foliage, casting an ethereal glow upon the hero, emphasizing their divine presence. Each detail, from the meticulously crafted armor to the lifelike texture of dragon scales, is rendered with unrivaled precision, thanks to the utilization of Unreal Engine 5, resulting in breathtaking 8K resolution.

This fantasy illustration transports viewers beyond imagination, as it captures the essence of a heroic tale set in a mystical world. The sheer quality of the image astounds, elevating the viewing experience to new heights, as if one were a witness to an epic encounter between heroes and mythical creatures.```

I added some stuff to the back end via the concatenation, but you can't argue with the results of the image.

I certainly wouldn't have prompted it this way myself, but that's why I have The Randomizer. 🙂

Another for instance here...This is what I got back from ChatGPT:
Description: The image depicts a vibrant digital artwork showcasing a futuristic server room. The prominent colors used are a combination of neon blue and glowing green, creating an immersive cyberpunk atmosphere. The image is of high quality, with intricate details and sharp lines that make each element visually distinct. Tags that could be associated with this image include "cyberpunk," "server room," "technology," "digital art," and "neon lights." The image captures the complexity of a server room setup, with numerous racks and cables neatly organized, displaying a futuristic aesthetic. The glowing lights and reflections add depth to the image and give a sense of advanced technology. It is a visually stunning representation that perfectly matches the text.
The image certainly supports this.

In fact, that's a rather impressive image considering the complexity of a server room and the added flair imagined by the prompt.

[oblique tartan]

do i need to update my python to use sdxl?

currently on 3.10.6

[native knot]

Nope

Pretty sure that's the floor for being able to run it.

[uncut steeple]

Lovely with all the description stuff, ill see if I can setup something similar later down the road, thanks a lot

[native knot]

You know, under normal circumstances, the hand orientation would make this a bad render, but considering the subject...

Plus, at least it oriented his thumb in the right position for his hand being twisted like that.

[uncut steeple]

Cool effect

[native knot]

Grogollum?

Chocolotter.

[mossy canopy]

no more colab

[strong field]

natural language prompts FTW

[vital ermine]

No way around it and why I hated colab so very much. There were some hacks but they never seem to have worked by the time I found them. The problem is the virtual environment is never persistent so each time it resets/closes down/new log in it has to rebuild that environment all over again.

[native knot]

About another 500 images have been added to the Random Diffusions album on my site:
https://lychee.soulctcher.net/

[strong field]

didnt know you had a site, thx for the share!

[native knot]

Relatively new. 🙂
Hopefully it inspires ideas.

[strong field]

wonderful stuff here

already is 😉

[ionic gulch]

i thought that sdxl/comfyui has a very small token limit for prompts. doesn't that prevent using natural, very descriptive language prompts?

[strong field]

I just used this prompt with success.

Color photo of an epic astronaut

, a fearless explorer of the cosmos, donning a sleek, futuristic spacesuit adorned with glowing neon accents. Their helmet reflects the brilliance of distant stars, capturing the essence of the vast universe. The astronaut stands against the backdrop of a swirling nebula, its vibrant hues of blues, purples, and pinks creating a mesmerizing celestial tapestry.

Surrounded by the infinite expanse of space, the astronaut floats weightlessly, their body language exuding a sense of awe and wonder. The silence of the cosmos envelopes them, creating a serene and ethereal atmosphere. The distant planets and galaxies twinkle like distant diamonds, adding a touch of magic to the scene.

The photo is captured with a high-resolution digital camera, utilizing advanced image stabilization technology to capture the astronaut in perfect clarity. The lens used is a wide-angle lens, allowing for an expansive view of the space around them. The camera settings are carefully calibrated to capture the intricate details of the nebula, balancing the vibrant colors and subtle nuances.

In this remarkable photo, the astronaut embodies the spirit of exploration and human curiosity, evoking a sense of inspiration and possibility. It serves as a testament to the indomitable spirit of mankind and our unyielding quest to unravel the mysteries of the universe.

Wall of text

[peak dove]

Most generative ai emphasises the earliest part of a prompt. The further from the start the less likelihood part of a prompt will be fully used. But use a prompt several times over, the whole prompt - albeit part by part - will eventually be used.

[graceful egret]

Does anyone know where I can find an SDXL Inpaint model that works with A1111?

[sullen laurel]

How do the two CLIPs in SDXL,CLIP G and CLIP L,works?

[ionic dragon]

can we train memojis?

[eternal fog]

Clip L is basically the same text encoder as sd1.5, clip g is larger and so in theory can better follow text prompts, it also seems better at natural language.

[crisp owl]

There's many differing opinions about this subject on how to "best" utilize the two clips in SDXL.
Clip G is as Arron stated, I use it as my more verbose description of the scene, written more like how I would be talking to someone.

I personally use the Clip L for my supporting prompts. But I also will use it if I can't quite seem to get the emphasis I need on a main prompt item, then I'll include that into the Clip L prompt as well.
Example:
Main Prompt: Keeping the fire lit with a blowtorch using liquid candy as fuel
Supporting Prompt: High quality, detailed, 8k, masterpiece, cinematic, blowtorch, liquid candy

[little stratus]

HI all, I've been struggling with this, I'm creating all my images programtically. I'm trying to image to image using SDXL but there doesnt seem to be clear way to do it. Right now I'm using SD1.5 for the image to image and SDXL refiner to imrpove and upscale it. I've also tried using ControlNet for sketch to Image but I get strange results with the outline of my sketch, seems like it uses it much to literally. Any help would be most welcomed

[crisp owl]

I have an img2img flow you can try if that's what you're looking for.
https://civitai.com/models/123048?modelVersionId=158153

[little stratus]

Thank you so much, I'll have a look at it

[little stratus]

What is the trick to use the SDXL base model, I stuggled to find a way

[crisp owl]

I'm not sure what you mean by trick to use it, that's mostly the model I stick with in my img2img by default

[little stratus]

Thanks but you do you use the sdxl pipe with an image, there is no option to use and image, only the refiner has an image option?

[crisp owl]

huh?
Your input image would run through the base generation first, using the base model. Then that pushes to the refiner using the refiner model, then that upscales using the base model.
Exactly the order it shows in

[little stratus]

I've not used compy ui, I'm using the python scrits because it all needs to run on a server backend for me

[uncut fiber]

and what is your problem with sdxl? You cant run it?

[crisp owl]

oh, well I've not played with any of that, so unsure on that.

[hardy cipher]

https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLImg2ImgPipeline

[little stratus]

Thnks for the help though

[little stratus]

I'm not sure how I missed this

[hardy cipher]

forgot to scroll down

[little stratus]

Lol I went through it, really intence reading, I'm enjoying the project but this part has been difficult for me, does that work with the base models

The example they give only uses the refiner, I think I tried it with the base and it didnt work?

[hardy cipher]

well I don't know why it would only use the refiner. that really doesn't make sense

[little stratus]

pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-refiner-1.0", torch_dtype=torch.float16

[little stratus]

Thats what I thought

I also couldnt find any ref to code for image to image using sd2, but I've seen people do it in AUTOMATIC, I'm guessing its still uses SD1.5 for the img to img part

[hardy cipher]

do you understand how it works?

you're just turning the image into latent data

that's the only difference

rather than using an empty latent that you put noise on

[ionic gulch]

the usual workflow for sdxl is to run 80% of the steps with base and 20% with refiner, without converting latent to image between and with keeping noise. for img2img you start the base not at step 0, but at a higher step (between 10% and 70%) and use the input image converted to latent as source instead of an emtpy latent

the later (higher step percentage) you start the more of the source image is kept intact. usually it needs fine tuning down to a few steps/percent to get the optimal result. so you will need to create a lot of images before you get the best mix and desired result.

50% - 60% is a good starting point

[little stratus]

Thanks, from a code pov how do I use sdxl to start the first part of the img to img, to go to latent and not start at 0, do you a a ref point that I can start from

This is what I've been referning to but cant seem to find the correct what to handle it https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl

Here is an example of what you are saying but I cant seem to find the way to start with an image ```
prompt = "A majestic lion jumping from a big stone at night"

image = base(
prompt=prompt,
num_inference_steps=n_steps,
denoising_end=high_noise_frac,
output_type="latent",
).images
image = refiner(
prompt=prompt,
num_inference_steps=n_steps,
denoising_start=high_noise_frac,
image=image,
).images[0]

[hardy cipher]

https://huggingface.co/docs/diffusers/main/en/api/models/autoencoderkl#diffusers.AutoencoderKL

[stone fossil]

Punched him right in the nose nice.

[hardy cipher]

giant bear or tiny man?

[stone fossil]

Man seems to punch bear. 🙂

[native knot]

This beast!

[vital ermine]

I dunno but it is awesome

[quick wedge]

'merica

[native knot]

Fuck yeah!

Elmo has started to let himself go as he gets a bit older.

Damn...this guy figured out how to do overalls with his own hair holding them up.

[native knot]

thicc...in the right ways

[stone fossil]

Oh my.

[hardy cipher]

well they seem happy at least

[stone fossil]

No doubt. 🙂

[lusty wolf]

What is the trick to get a good duality image with ComfyUI?

[vital ermine]

No idea wth that is in his pecks

[uncut fiber]

interesting reflection

[stone fossil]

To relect or not to reflect.

[cyan crown]

I should send this to Nike 😂

[uncut fiber]

why only nike? Will try as well some shoes, when upscaling done

[stone fossil]

I love nike.

If you give me options Ill pick em.

[lusty wolf]

In my search for a duality type image (prompt)...

[uncut fiber]

i also like nike.

[cyan crown]

i prefer asics for shoes

[stone fossil]

Rebok?

[uncut fiber]

Reebok yes

oh wait

[stone fossil]

Noice. 🙂

[steady grove]

British Knights

[uncut fiber]

now not sure

asics maybe?

[stone fossil]

Anything goes. 🙂

[empty nova]

anime ¯_(ツ)_/¯

[stone fossil]

Anime is a stupid for of entertainment.

But it goes. 😄

[empty nova]

[cyan crown]

Asics

and they are pretty real

[uncut fiber]

Lichtenstein 🇱🇮

[vagrant wasp]

Lora experts ?

[ionic dragon]

dayum, it got those stripes perfect

[cyan crown]

https://www.youtube.com/watch?v=vdZ9exqDFTc&t=508s

this was made with Midjourney and SDXL

[ionic dragon]

ultra realistic product photography of nike branded symmetrical glass gundam helmet, highly detailed, HUD face, deep black background, octane render, vray, shimmering, glossy, Fvckrender, geomerty, prism highlights, C4D, ray tracing reflections, prism shadows, diffraction, macro, flickr, 500px, photography, atmosphere, depth of field, grading, lumen reflections, golden ratio, hyper realistic, incandescent, rule of thirds

it got the nike logo and text perfect

[indigo carbon]

Also, when Pytorch 3.0 releases, will it be required to retrain the models? Or will it just get more efficient

Because 2.0 has been around for a while