✨｜sdxl

[hardy cipher]

well without the math stuff

[nimble heart]

I'd do 3 images:
base, complex math junk, just increasing the weight normally

with the latter being either text weight in prompt or the unclip weight

[hardy cipher]

yeah, watch it be the same, lol

[nimble heart]

not the same cause of the concat but I wouldn't be surprised if they're close

[hardy cipher]

well I can say that with each new thing I added I did it incrementally. it's not like I hit the ground running with any of it

add it with .1 weight or something, then move it up slowly

[nimble heart]

ah yes the slowly boiling a frog to death method

I always do the opposite and start at 1.0 then move it down

[uncut fiber]

speaking about upsampling, this is best but need lot of power for large images. And of course good model but there is too much. I posted today 8x

[hardy cipher]

I've taken that approach

[nimble heart]

what the fuck is that

[uncut fiber]

it is father of Comfui

for image manipulating with some A1111 support

[hardy cipher]

are you latent rotating?

[nimble heart]

is that like a very manual tiling formula or something?

[uncut fiber]

it rotate fliping images resizing them and then blend them. It is not my work, just understanding how it works
try search for ChaiNNer, it is from it. Default Full TTA

[nimble heart]

wait it's all in pixel space?

[hardy cipher]

hmm. how would that work out? never really tried it

[uncut fiber]

with 8xNMKD-Superscale

[nimble heart]

that sounds like a really complicated way to do one of those emulator pixel scaling algorithms

[hardy cipher]

interdasting

[nimble heart]

does it work on more complicated images at all

lines are easy to do if you smear the fuck outta everything else

[uncut fiber]

i think images are not limit, only proper model for it, and some time or powerful GPU

it can resize realy easy load image upscale and save image.

you can send me some small image because HW limitation on my side

i am testing it because pal need help with printing quality.

[nimble heart]

define 'smal'

[native knot]

This is the lyrics from Winds of Change by Scorpions

[nimble heart]

like 720p?

[uncut fiber]

512x512? @nimble heart

i can measure it. But will cant do 8x probably. Maybe better to download it and testing for yourself?
Going test something

[nimble heart]

idk there's an SD 1.5 image i was trying to make 4k. native is 1080 I shrunk it by 50% so it hopefully doesn't light your pc on fire

[hardy cipher]

let's see, here's one without the math things

with them

I'll try something else. also, I realize I would need to up the weights of things in order to get a more accurate gauge of things

[nimble heart]

damn running a 4x on a 1080p image is rough

only using 3 gigs of vram though so idk

taking years

[uncut fiber]

🙂

going for it!

[vital ermine]

It should due to the higher clock but for anything to the gpu more buffer is better

[nimble heart]

I'd rather save 30 seconds on a long code compile than .02 seconds copying the latent to the gpu on an SD render

[uncut fiber]

ppl with 4090 could have it done in eyeblick

[hardy cipher]

without math or adjusting weights

[vital ermine]

Training it is much worse

I don't give two shits about generating images it is the training I am all about

[nimble heart]

• imo 3d vcache is overpriced currently. For the price of a 7900X3D you can basically get a 7950X

[hardy cipher]

first one here is with the math. then second is bumping up the initial clipvision weight by just shy of 2.5x

[vital ermine]

well, for me it will be an X3D just 79 or 8k depends what is next year

[hardy cipher]

so the stuff I'm doing definitely brings something to the image

[nimble heart]

i bought my 7900X before the 3d came out so I guess the choice was easy lol

[hardy cipher]

first and third aren't doing much

[vital ermine]

Well, AMD has said that going forward X3D will be for all of them so we will see. All about the chiplets now.

[uncut fiber]

well it is noticeable that crown isnt best for 8x resize

[nimble heart]

crown?

[uncut fiber]

or hat

[nimble heart]

Oh

yea I used realesr for my actual one cause its a lot more aggressive with gradients

[uncut fiber]

but in eye is some reflection

[vital ermine]

I am having a hard time training on a person. Styles are not an issue with me but the person seems better but nothing what I trained

[nimble heart]

1080p -> realesr 4x -> unwindowed sinc 1024 lobe 50% downscale

looks okayish

but definitely pushing sd 1.5 outputs lol

least without tiling

[uncut fiber]

to fully see it one to have open in browser.

Trying 4x ultrasharp

[shy kelp]

i think these turned out nice

[uncut fiber]

yes nice consistency

@nimble heart

[nimble heart]

yea I used realesr over ultrasharp because its better with flat lines

ultrasharp's good for high frequency details like photoreal images

if you side-by-side you can see the ultrasharp one you have smoothes out less stars and stuff but the linework on the helmet is way blurrier

[uncut fiber]

yes stars are nice, would be nice to have some mask and resize different parts differently 🙂

[hardy cipher]

I prefer 4x_nmkd to ultrasharp

[nimble heart]

least with XL it doesnt matter as much because you can straight up img2img @ 3840x2160 so the upscaler choice is just for mild detail enhancements

I have nmkd superscale idk if that's the same

but ultrasharp did better as a base you denoise on top of

for XL

[hardy cipher]

ahh yeah. I'll sometimes do 2x to 2x with ultrashapr first

[nimble heart]

so sample 1368x768 -> ultrasharp 4x -> sample 3840x2160 @ 30% denoise

is what i usually do

least for the few 4k images i posted in here

[nimble heart]

my gecko phone wallpaper was just regular bicubic scale I think

[uncut fiber]

check color of this version @nimble heart

[nimble heart]

what's with the ringing

also oversaturated

[hardy cipher]

I love burning images

[uncut fiber]

yes it is studio lite

bottom part intact

[nimble heart]

the tattooes and hair are ringing really bad

[uncut fiber]

it is not resized 😄

[shy kelp]

how do you make those comment boxes in comfy like how it's c key in unreal

[nimble heart]

double click search 'note'

unless you mean the big groups, that's just right click 'add group'

groups are cancerous so most people dont use them

[shy kelp]

yep, group thing

[nimble heart]

if you like snag one pixel anywhere you do the equivalent to your workflow as moving an image in a Microsoft Word document

[nimble heart]

so what was run?

[uncut fiber]

just studio lite colors only

[nimble heart]

idk what studio lite is

[shy kelp]

one of mine in unreal

[nimble heart]

looks like a bubba'd unsharp mask

[hardy cipher]

that's a map from 1988 nintendo power

[uncut fiber]

it is actualy 1x broadcast to studio lite pytorch based implementation, it doesnt mater what it do, it my not only resize but denoise, removing Jpg artefacts and so o.n.
This changing color on some pth based model rules

[hardy cipher]

lazy eye

[uncut fiber]

:)))

[nimble heart]

lol sdxl does that a lot

make everything a side profile so they cant have lazy eyes

7s/it oofie

[hardy cipher]

just bokeh everything so you only have to render one thing correctly

[nimble heart]

I also do that lol. "out of focus background"

so long as the subject turns out okayish rest doesnt matter

[uncut fiber]

@nimble heart have you tried lexica?

[nimble heart]

Don't know what that is

[hardy cipher]

better

[nimble heart]

I always post-process using my own filters I wrote for GIMP

[uncut fiber]

2xlexicaSwinIR

[nimble heart]

ah upscale

2x isnt even enough for 4k

[uncut fiber]

but realy nice, one can aply it twice. Nice stars

[nimble heart]

blurry helmet again

[uncut fiber]

one cant have everything

[nimble heart]

you notice the linework a lot more on a 4k screen than the stars

im not super worried about it anyways now that XL is out and I can direct render 4k

I finally replaced my wallpaper

[uncut fiber]

yes xl is great for large renders, just i can probably afford for me enough fullHD.

[nimble heart]

1.5 I could direct render 1080p which is what that midna image is. Maybe with an nvidia gpu + xformers you could direct render 4k @ like 60s/it lol

XL you can sometimes do 1920x1080 natively without any upscaling

and get not-gross results

768p compositions are better though

[uncut fiber]

i did such wallpapers directly but not handy as wallpaper actually 😄

[native knot]

Prompt: Lyrics from Africa by Toto

[nimble heart]

1920x1080 without any upscaling

runs pretty well, 1.21 it/s

[uncut fiber]

yes i did it as well native

[nimble heart]

?

[uncut fiber]

#✨｜sdxl message

[nimble heart]

oh lol you literally sent that as I was uploading mine so I missed it

[uncut fiber]

would prefer 1088 probably will play with it, and see

[nimble heart]

why 1088

same thing 1088

[native knot]

Because you need to render at something divisible by 64.

[nimble heart]

no you dont

latent space is 1/8th of pixel space so it need sto be divisible by 8

that's why all the latent image sliders step by 8 in comfyui and auto1111

[hardy cipher]

well encoder divides each side by 8

[nimble heart]

yes

1080 / 8 == 128

wait

maths

[native knot]

no

[hardy cipher]

1024

[nimble heart]

135

lol

its still a whole number is what i was saying

[native knot]

You get greater incoherence at steps not divisible by 64.

[nimble heart]

prove it

I've never ever had someone prove it to me

[native knot]

results prove it

[nimble heart]

Show them

mine show no difference

beyond just seed changing slightly

[hardy cipher]

I only render in multiples of 7

[nimble heart]

ah yes 504x504

[nimble heart]

are you remixing my midna wallpaper I spent like a whole day on into 80's B movie posters

[hardy cipher]

yes

[nimble heart]

aight

[hardy cipher]

that's a good one

[nimble heart]

I guess good is subjective

my poor childhood

did you try the direct 4k remix?

[hardy cipher]

I'll do that now. got sidetracked

[nimble heart]

hide rick rolls in every image you post on here using that method

[hardy cipher]

good idea.

could be done discreetly using the mask input with a solid mask. put him in an out of the way place

like a where's waldo, but it's rick astley

alright, I'll do 2 with the same seed number and what not. with and without the pre-sampler upscale

[nimble heart]

a 2 pass will always wreck a single pass

direct 4k may be possible but you'd have to cherrypick it

[wide crane]

elegant workflow

make me bussss

[crisp owl]

I did that for a bit with 1.5 metadata in another server because someone said my prompts were trash, so I upscaled everything I posted with a 0% denoise with the rockroll lyrics as the prompt just so it changed the metadata of the image.

[hardy cipher]

alright, here's one upscaled 2x vs 4k input

I'll have to try a few

[wide crane]

absolutely

beautiful

workflow

[uncut fiber]

@nimble heart simply i believe in 64 🙂

[hardy cipher]

this is beauty

[nimble heart]

1920x1080 cleanly turns into a 240x135x4 tensor so im not sure what else would cause it to be "less coherent"

I'd probably have to use canny edge to make it more subtle lol

[uncut fiber]

was in A1111 it isnt dead at all

[vital ermine]

That is so damn hard to do

[nimble heart]

the actual ComfyUI source code for turning an image into a latent shows it just rounds to the nearest 8

[uncut fiber]

yes i believe it is

[vital ermine]

I wonder why SD refuses to do the vulcan salute?

[nimble heart]

its a hate symbol

[vital ermine]

ffs, lol

[nimble heart]

was trained out during diversity conditioning

[vital ermine]

Hahahaha

[hardy cipher]

I made my own so I could see how it works

[nimble heart]

he looks so miserable

[vital ermine]

I would be too if I had a giant hand

[nimble heart]

make the rest of him giant too

so its even

[vital ermine]

no idea how

[hardy cipher]

tell it to stop being a stupid idiot and make him bigger

[nimble heart]

big tall handsome Spock, broad sexy shoulders, thick masculine neck, instagram model, realistic hdr 4k 8k HD nikon photoreal make him hot please

[vital ermine]

I told it large hand in the neg it did this

[nimble heart]

now he's even more depressed

like he had to put Kirk down via fingergun after he was infected with a brain parasite

[vital ermine]

LOL

[nimble heart]

i should pitch that to netflix

[nimble heart]

see it worked exactly as intended. now he's bigger

[hardy cipher]

whoa

[vital ermine]

hard nips

damn I can't say ee reckt?

[nimble heart]

nope

many random words are banned

but elegant ejaculation is fine

[hardy cipher]

pen is as well I believe

[nimble heart]

lingeri3 is too lol

[vital ermine]

wth?

[hardy cipher]

poop isn't banned

[nimble heart]

ohhhhhhhhh

[hardy cipher]

so at least we have that going for us

[nimble heart]

400 iq idea for making good vulcan hands

genrate a normal image

inpaint out the 6th finger

[vital ermine]

hehehe

I am trying to figure out why it makes them giant hands?

I love the one with the old car about to run him over

Look at his hands, ffs

[nimble heart]

so much emphasis on the hands it pulls them closer to the camera

like if you add lots of handy words to your prompt it does that

in 1.5 as well

[vital ermine]

weird

[nimble heart]

probably does it for other things

try like nose facing camera, vulcan nose, nose in frame or something and watch it just stretch into the middle

[vital ermine]

I don't go down those alleys.

[nimble heart]

make spock into a real boy

[vital ermine]

Not just the vulcan salute, shit

[nimble heart]

maybe the handgun is a Vulcan

so are you training a vulcan salute lora or something

[vital ermine]

I guess I better, shit

[nimble heart]

lol

what are you training on currently?

[vital ermine]

Vulcans are vegetarians, so lol

[nimble heart]

its one of those insect burgers

[vital ermine]

Notice the damn hands? shit

Andre the Giant hands

[supple knot]

try making spocks head small

because thats a thing in art ,head small is big person ,head big is small person

[supple knot]

Is that one of those quantum computers?

[hardy cipher]

correct

and it's dividing by zero

[supple knot]

Im trying to do giant spock

[hardy cipher]

same. but they keep turning out like this

very patriotic

[shy kelp]

basic setup (i am new to using this)

[uncut fiber]

what are those three boxes in left upper corner @shy kelp ?

[hardy cipher]

2 checkpoint, 2 loras, text encoders, empty latent, vae loader, decode, save image

[uncut fiber]

oh 2 checkpoints, didnt know. And lora make sense

[hardy cipher]

wait, one checkpoint. I typoed

[ionic dragon]

@hardy cipher can I dm you?

[hardy cipher]

why not? I remember they exist now. forgot about them and left a bunch of people hanging. so rude

[shy kelp]

checkpoint and two loras

[wide crane]

<t:1694853780:R>

[shy kelp]

love the colors

[noble shoal]

What a lovely resolution

[hardy cipher]

https://tenor.com/view/papito-ily-chefskiss-gif-20536366

[noble shoal]

I need to go loooooonger

[nimble heart]

the subject isn't duplicated so you're already doing better than most

[noble shoal]

I have set my target width to 256. Don't know if that does the trick

[nimble heart]

I guess 256 * 4096 is equivalent megapixels to 1024*1024 so it's actually somewhat feasible

[noble shoal]

I guess i need to go longer

[nimble heart]

can you? I thought the max size was 4096

[hardy cipher]

max size for what?

[nimble heart]

I guess you could always override the node's slider

edge length

4096 pixels

so a 512 latent

[nimble heart]

one of these is not like the others

[hardy cipher]

he just has a slower metabolism

[nimble heart]

are they wearing pink panties on their chests?

[hardy cipher]

not sure, they just have their own style

[nimble heart]

lol that one purple guy trying to be involved

[hardy cipher]

blue guy on the bottom left in the second picture is the best

he's all cocky cause he's half human and doesn't have those black orb eyes

[noble shoal]

9632x280 (with a highres fix)

[hardy cipher]

you should try for 8x131072

[noble shoal]

Then i need to overwrite the comfyui max width

[hardy cipher]

you're within your rights to do that

[noble shoal]

Ok let me check. I hope its "MAX_RESOLUTION"

[noble shoal]

I had an inference but no real output at 8x131072. "Best" i could do was 26.206 x 96

[noble shoal]

If you click on open in browser, you might see the fresh salad, tomatoes, onions and grilled burger patty.

[hardy cipher]

it's very high quality

[ionic gulch]

ghost busters, star trek, tron, lotr, dune, 2001

[shy kelp]

hmm, what has happened in this hires upscale

(the upscale latent is plugged up, wires don't show because it's over the save image)

[glad grove]

increase denoise strenght by 1 or 2

[shy kelp]

hmm ill try

hmm

[spark dock]

why does it feel like the sdxl bot is nerfed?

especially the anime one.

anime style*

so quick as well as low quality generation

:/

also for some reason some styles just refuse to do the generation or take a long/very short time to do so

[hardy cipher]

I don't know

[glad grove]

because its a bot,if u want high quality and near perfect pics install local version

[hardy cipher]

one of my faves

[stone fossil]

https://civitai.com/models/146424?modelVersionId=163063

[clever verge]

Solved for good now, borrowed some code from efficiency nodes. It's done client side as my initial code did.

[hardy cipher]

right on. I rarely if ever make things like sampling and scheduling extral to the nodes themselves. but I do it with seeds a lot

[clever verge]

It just rewrites any node type you have that use noise_seed so it works with whatever

But you have control over what nodes types you "overwrite"

[shy kelp]

for comfy when

[clever verge]

Sampler is probably not needed now since I mostly stopped using the refiner but when you do I find it handy to only have to change it once.

[clever verge]

I like Subway

[shy kelp]

the flow grows

[clever verge]

Where does the node types come from @shy kelp?

[shy kelp]

those are for unreal sadly https://www.unrealengine.com/marketplace/en-US/product/electronic-nodes

[uncut fiber]

upscaling 🙂 You learned a lot, me on other hand trying find good ballance with all models i have, about 50 and wait for each turn which take say 90 seconds, then correct a bit, or change model which i dont know...

[shy kelp]

yeah im too comfortable on this one model and i need to use others
plus there is probably new shit out that i am not paying attention to lol

[hardy cipher]

there's always new things

[floral island]

sigh, my model merge is a failure, after all 😦

have to check out all the block merges to see where it went wrong 😢

[zinc cargo]

who's she?

[spark dock]

why does it being a bot affect the generation output :\

i know it's a bot but how does it being a bot make the output worse?

[zinc cargo]

what's the prompt here?

[zinc cargo]

ask @visual glade 🙂

[hardy cipher]

3 images as input with blip interrogators, ipadapter, clipvision and controlnet. not even modifying prompts for these

[zinc cargo]

i forgot how much you love not typing prompts 😛

[hardy cipher]

no, I'm all about prompts

deliberate was one of my favorite 1.5 models for that reason

[shy kelp]

i found out you could duplicate nodes by holding alt and clicking but i expected it to be a shortcut to destroy a node wire lol

[hardy cipher]

let's hope it works out, guys

[shy kelp]

that snow looks so cozy

[hardy cipher]

quality frogs

I like these frogs tbh

messed with some old watercolor style images I did in 1.5. did some sdxl upgrades

at least 2 moons there

[hardy cipher]

now with ipadapter and clipvision updating old concepts works great

those ones might work out well

[woeful patio]

what lies below

[uncut fiber]

a woman with a man are going forward to happy tomorrows.
A1111-exactly what i imagine 🙂

[hardy cipher]

yummers

[ionic gulch]

2001, blade runner, the cell, dune, the fall, the fountain, ghostbusters, hero, the grand budapest hotel

for the 9 images above: prompt "a woman", ip adapter source image:

[stone fossil]

I quite silently dropped the LoRa for this woolen like stuff so if you missed it here is the link: https://civitai.com/models/146424?modelVersionId=163063
Give it a spin and let me know what you think.
Big thanks @dapper dragon

[shy kelp]

i always seem to generate buildings with lights on in them and can't get rid of it

[ionic gulch]

prompt: a w00len woman, lora: woolify, ip adapter (same as above), sdxl styler movie: the fall

[noble shoal]

Might not give the desired result, but i managed it with "abandoned" Pos: A beautiful fantasy artwork showing a little (abandoned:1.4) town on a river at night, illuminated only by moonlight, nobody is home, it is dark | Neg: shallow depth of field, DOF, bokeh, cartoon, drawing, painting, naked, Light bulbs, Tube lights, Lamps, Torches, Lighters, Candles, Fire, Glow sticks

Ok, its more a ghost town now

[stone fossil]

Good but prompt like so: w00len, yourprompthere.

That should enhance it properly.

And its zero not a o 0 / O

[ionic gulch]

this is w00len, a woman:

(fist lora strength 2, second strength 1)

[shy kelp]

ty but i might have to hunt through my old gens and see if i can find it, it's just not going away no matter what but i do remember i was able to constantly generate with it off so

[hardy cipher]

I can make anything cloth with this one little trick

wool, cloth

[noble shoal]

Time for jeans. Challenge accepted

[hardy cipher]

this with some clip vision. all you need

[ionic gulch]

without ip adapter:

w00len, a toy dog:

[hardy cipher]

looks like a wes anderson kids movie

[uncut fiber]

are you using same model?

[uncut gull]

Is this using a noise lora? Or did you photoshop it after to get it so dark?

[noble shoal]

Just image level adjustments node after the VAE Decode

[uncut gull]

IRL photography you have to time this perfectly, just after sunrise or before sunset, when the dimness of twilight matches the brightness of the artificial lighting. 🙂

[stone fossil]

Yes use 1.

[shy kelp]

this house light

[uncut gull]

So... I've been using an SD 1.5 generated landscape as my desktop background for ages now. I just decided to remake it today with SDXL. I was born in Alaska, so it just makes me feel at home.
But, is it strange that these feel more true to my memory than photos I can download online? There's something about the fact that... This is a landscape no one has ever seen. There was no photographer standing here, wondering about color, touching it up later. It's... untouched.
But also not real? Hmm... Probably stuff that belongs in the philosophy section though.

[shy kelp]

fluffy pig

[upbeat summit]

nice fidelity!

[stone fossil]

w00len, a hyper realistic macro close up picture of a ((Pumba made out of multi color wool and yarn)), using the colors red, black and yellow, bokeh woolen multiple color jungle backdrop

[uncut gull]

It is really creepy how it mixes the pine and birch trees though. 😨

[floral island]

new model merge 😮

i think this merge is the one bois

ezmode 😮

[noble shoal]

Guess who just trained a LoRA with 0 Learning rates...and no i don't mean Kappa Neuro

[covert viper]

How high can sdxl go on resolution without deformities/twinning/etc. Without upscaling and hiresfix obviously

[noble shoal]

Every multiplication of length and width that has a product of 1.048.576

[shy kelp]

my anxiety!

neat

[uncut gull]

It depends on the subject. I tried a 4k landscape earlier, and it wasn't absurd, just... A little odd.

Native 4k SDXL render (no upscale, no hires fix).

It still got the water below the horizon, mountains in the middle, and sky on top. No stacking layers of sky, at least.

[tepid surge]

the like one time i come back in the chat after however long cx

[tepid surge]

yooo! killen it! <3

[indigo carbon]

Will there be any multimodal interfaces in the future that use LLaMa and SDXL in the same time? I feel like that would be the next step

[uncut gull]

There's tons out there. I am literally running llama2 and SDXL on my PC right now. Llama2 writes some pretty good prompts. It never runs out of ideas. There's also a project called Llava that uses Llama and CLIP to give Llama pseudo-vision.
Also, that multimodality might seem patchwork, but it's really not too different from how human multimodality works. Our vision and language processing aren't "unified". We process vision and language very separately* at the bottom, then integrate them at a high level.
*probably

[indigo carbon]

Huh.. like, prompting isn't the issue atm.. SDXL is by far the best at turning text to images, but it doesn't have the ability to get instructions for editing images, and all complex text and image input stuff. I think if SDXL and LLaMa could be inferred together we could achieve perfect image blending and or text interpretation

[uncut gull]

You want SDXL to "think" intelligently the way llama2 does. I get it.

[indigo carbon]

IPA is a step in that direction, but it loses a bunch of SDXL's insane quality in the process

[indigo carbon]

Yeah, we NAILED the quality of the images, now the next step should be capabilities

[uncut gull]

In the same way that learning coding makes an LLM vastly better at logic, theory of mind, and creativity; I expect learning visuals would make it better at something too. 🙂 It's beyond my ability to create, and there are no open source projects that really meet this criteria.

[indigo carbon]

SDXL is way more than enough when it comes to the images and speeds themselves. I just think it would be insane to have an inference where you can give an image then give instructions on how to change the image, what to blend the image with, etc.. there was actually something called "IPA" that released recently that allows it to do something similar, but it makes SDXL lose much quality

So if there would be an SDXL+LLaMa inference, it could potentially enable SDXL to do all those things

[hardy cipher]

How does ipadapter cause the image to lose quality?

[indigo carbon]

Honestly, idk. It's really experimental and chaotic when you don't provide text (which providing text isn't ideal for image blending)

IPA also doesn't support the new optimizations like AIT

[uncut gull]

So does controlnet, and so does high CFG. The more you control SDXL, the more glitches you get, especially on small faces.

[indigo carbon]

Which also isn't ideal.. I think the right approach shouldn't be modifying SDXL's components, it should be implementing other components (such as LLaMa)

[uncut gull]

I disagree. The "glue" connecting these can't be hard-coded algorithms. It would have to be an AI at least as advanced as SDXL and Llama.

[hardy cipher]

I just set up a blip model to interrogate the image find into ipadapter on mu latest flow. Rhythm send thst to clip g

It works

[indigo carbon]

I don't see SAI training something from scratch just to give it more capabilities. There must be another solution

[uncut gull]

Tesla tried originally integrating many neural nets with heuristic code. It hit an absolute limit, and couldn't reach human-level driving performance. Running AI full-stack, from input to output, worked.

In our own brain, the connection between language and vision is not something simple. It's the same kind of intelligent network that the rest of the brain is made out of.

[indigo carbon]

That could be a decent way to blend the images, but not instructively modify them

[hardy cipher]

I haven't written a prompt in days

Have it combine with an empty text box for adding to it if I want. Then through a styler

[indigo carbon]

So you think SAI is going to be making a new type of model that has those capabilities? Because I don't see that happening any time soon..

[hardy cipher]

Well I don't know why you'd need a new model to do more. I don't think monolithic all in one models are the way to go anyway. At least not in most cases

[indigo carbon]

I do remember hearing about something called "CM3LEON" that has those capabilities and it's made by Facebook, but they are awfully vague about it and said it's going to be closed source

Idk if the only way to achieve instructive image editing is making a new model from scratch though. Maybe the way is to combine LLM with diffusion and something like IPA, but idk man.. we'll find out eventually

[uncut gull]

Super multimodality.
https://arxiv.org/pdf/2309.05519.pdf

You use SD and Llama as an initialization base, but you train one giant "LLM" on top of it. You only need 1% parameter adjustment over those base models to start seeing evidence of a model that "understands" language and vision together.

[indigo carbon]

Interesting. Y'all think SAI would be doing something similar to this?

[uncut gull]

I don't think they've announced anything like that. Last I heard was their stable audio release. (Haven't tried it out yet.)

This paper released 3 days ago, so they might not have started yet if they're planning on it. 🙂

BTW it used Vicuna, Stable Diffusion, and ZeroScope. Definitely something SAI could tackle with their own toolset.

[indigo carbon]

I took a look at that paper and it does seem like they're talking about inferring diffusion with LLMs to achieve crazy multimodal like this. Idk how close we are to having something like that available as an open source

[uncut gull]

Using AI to "glue" together these amazing models like SDXL and Llama2. It doesn't take much glue, and the results are as powerful as SDXL and Llama2 combined. (Hopefully!)
They might release their code. The paper is still a WIP.

[indigo carbon]

Awesome! Speeds with that shouldn't be too problematic also, since we have optimizations like AIT and OneFlow just for stuff like that

[uncut gull]

No, actually, to run SDXL and Llama2 together on my system, I put SDXL on the GPU and Llama2 on the CPU. Not sure how consumer-friendly this would be... But I'll just have to hope for now.

[indigo carbon]

You have no idea how fast OneFlow can be. Stay tuned, it might get implemented very soon

[hardy cipher]

J

[indigo carbon]

I heard OneFlow also decreases VRAM usage, but idk about that

[ionic gulch]

star trek style

[hardy cipher]

Input images, output nightmares

[uncut gull]

4K tall generation (no upscaling / hires).

[uncut fiber]

in pixel sizes?

[uncut gull]

1024*4096

[uncut fiber]

nice

it looks way taller

[uncut fiber]

He has nice necklace, ball bearing 😄

[hardy cipher]

holy cannoli

[uncut gull]

It lacks fine details... I wonder if hires can add them.

[shy kelp]

probably, gonna see

[ionic gulch]

"highly detailed portrait shot of a woman" + 40% control depth + movie styler (nothing else):

[hardy cipher]

I see what you may have done there

[ionic gulch]

the depth map works much better than ip adapter to bring out the movie style, as it doesn't give that much information as ip adapter

so at least for testing it works better

[shy kelp]

no hires on the left

[hardy cipher]

goodness, this stable audio model is legit

have you tried it out?

it's about a million times better than the etension in a1111 I mess with a while back

I realize this isn't the channel to discuss such things. just saying

[noble shoal]

You have my attention.

[hardy cipher]

https://stableaudio.com/generate

you can make free 45 second mp3s

now don't expect it to be creating on point symphony recordings or anything. that won't happen for at least a year or two

[noble shoal]

That's highly interesting. Did already someone asked: "When A1111?" 😅

[hardy cipher]

is it going to be released like that? I actually have no idea. there's a channel specifically for it though

[noble shoal]

"Keep an eye out for upcoming releases from Harmonai, including open-source models based on Stable Audio and training code to allow you to train your audio generation models."

[hardy cipher]

oh nice. well I didn't get that far. but made a "psychonautic masterfully mastered rap beats that goes harder than rocks and hotter than the sun"

[ionic gulch]

those spider-man, star wars, ghostbusters and star trek options need some work. looks nice, but not as the movie:

[hardy cipher]

now we just have to wait for the image to audio models

[noble shoal]

Can't wait to hear Steven Segal mumbling My heart will go on with a deathmetal backing track

[hardy cipher]

very cool. I'd like to hear him sing in a barbershop quartet

just asked that channela bout the image to audio models. gonna be a clueless noob in there

whoa. this is going to be a huge time sink. ugh

[noble shoal]

I saw, well...until open source models are released there isn't a huge community working on such implementations like image to audio

If they release it.....4-5 days

[hardy cipher]

true enough. it's pretty incredible tbh because every other model I've seen was very primitive. not knocking them, just weren't quite there

I tried to train a model on bach midi. and well, it trained. but let me tell you, it did not make beautiful music

[noble shoal]

IPAdaptor for Audio 🤯

[hardy cipher]

dude, multimedia projects

[noble shoal]

Audio2Audio

[hardy cipher]

audio to audio, audio to image, image to audio, audio to neuralink

[noble shoal]

OpenPose Dance Moves to Audio

[hardy cipher]

then we just have to wait for the touch and smell models

[noble shoal]

Prompt: The smell of Steven Segal after running for 10km. Neg: Flowers

[hardy cipher]

tangy

this is moderately terrifying

it sounds like sims rapping

[noble shoal]

Oh no. I already imagined the same text problems as our beloved text 2 image model has.

I guess for now we have to stick with image generation. And if an audio model gets released we all need a 4090

[ionic gulch]

is star trek blocked in sdxl? this is what i get from "star trek borg queen":

[hardy cipher]

that'd be weird if it was. but I have no idea about anything

[noble shoal]

I get an old lady with my mix model

[ionic gulch]

that one gets closer, but doesn't seem that sdxl really knows how a borg queen looks like:

the wires are from my depth model. and that is the closest thing to a borg queen in that image

no wonder that my movie styler has problems, if the model misses everything about star trek

[hardy cipher]

hmm

that's just strange to me

[ionic gulch]

portrait shot of a vulcan woman

[noble shoal]

That's definately not the U.S.S Enterprise, or is it?

[ionic gulch]

woman in an iconic star trek tng uniform

very iconic...

[hardy cipher]

well someone was posting spock earlier so there must be either a workaround or a lora or something

[ionic gulch]

lol

captain picard:

captain picard captured by the borg - sorry, but that is not just leaving those images out - they fed in some stupid stuff for those key words as well

[hardy cipher]

remixing some of my old sd1.5 wins

now that my flow has made prompts obsolete I can just plug new images in and go

[ionic gulch]

will remove star trek from my movie styler - not worth it that way

[hardy cipher]

yeah, that sucks. and also pretty weird

[ionic gulch]

star wars also just gives some fan art.
darth vader in front of his stormtrooper army:

[hardy cipher]

interesting

[ionic gulch]

still "movie scene from Darth Vader in front of his stormtrooper army"

[ionic gulch]

movie scene of a star wars x-wing fighting the death star
those random results of fan art or off topic stuff doesn't work for a movie styler. star wars will be removed too

[wide crane]

https://tenor.com/view/bornskywalker-dap-me-up-woody-woody-handshake-woody-toy-story-gif-26021440

[supple knot]

I cant get Chewbaka to play drums.

[ionic gulch]

do you think this captures the style of spider-man into the spider-verse?

this is the prompt - how could i improve it?
"a woman in style of 'Spider-Man Into the Spider-Verse', mix of realism and stylization, dynamic camera angles, comic book-style, vibrant color palette, exaggerated character expressions, urban landscapes, graffiti-inspired visuals, onomatopoeic comic book sound effects, motion blur"

[supple knot]

Smiling, Jumping on a building, NYC

[hardy cipher]

whoa

[kindred plinth]

whats the easiest to use SDXL workflow?

when it comes to prompting

[uncut gull]

I keep getting a "blocked" message for my post, and I don't know why. 😕

[kindred plinth]

do a binary search and figure out what word it is

[uncut gull]

Meaning what? Is there a list of disallowed content somewhere?

[kindred plinth]

just test half at a time till you figure out the word thats banned lol

[uncut gull]

Wouldn't it be way faster to just not post???

[kindred plinth]

bro if you wanna figure out why your message is blocked thats how ur gonna figure it out

[indigo carbon]

https://github.com/NExT-GPT/NExT-GPT just found source. I can't figure out if it's finished or a webui yet

[zinc cargo]

what did you prompt for?

[noble shoal]

photo of a clear (transparent:1.4) (translucent:1.3) (evil:1.2) worm alien creature in a dark organic cave, evil face, sting, intricate veins, wet, horror, inner biological cell
But my model is something.....special, i guess

[zinc cargo]

i love it

[noble shoal]

Thank you

[vital ermine]

wubba lubba

WOW, even has a "bonus hole"

[noble shoal]

Its for the plumbus

[vital ermine]

LOL

[ionic gulch]

you have to wash your hair even in the matrix trinity:

[noble shoal]

Or it's just ultra greasy

[ionic gulch]

spider braids:

guess i could sell that idea to some hair stylist in nyc.

[noble shoal]

Noice

[zinc cargo]

yea, i like those ugly bastarads alot

[noble shoal]

Is this still my prompt?

[hardy cipher]

you should try "if dolphins could sing like birds"

lovely sounds

[zinc cargo]

slight change

this one with a lora

[hardy cipher]

need to set up ipadapter with the song model

[zinc cargo]

these are with Protovision

[noble shoal]

The man she tells you not to worry about

[noble shoal]

Binpba Poulbes looks like its located in a cooling tower. What a lovley place.

[hardy cipher]

yeah, I noticed the pretty background as well

[zinc cargo]

to counter @noble shoal 🙂 (maybe to compliment)

[noble shoal]

Very nice eyebleach. Very nice

[noble shoal]

What did you expect? 😅

[hardy cipher]

olive heads

[indigo carbon]

REAL BITCOIN POG

[solemn tide]

What up SDXL fam. I finally took some time to figure out how to get my xformers updated correctly to have sdxl work.. I took a random prompt from civitai for the juggernaut model and the step preview looked on par with what it should be

but the output turns out like this

[indigo carbon]

looks like VAE error, also could be an A1111 issue, most people use ComfyUI as of now

[solemn tide]

got it will start with VAE

does comfyui do Deforum?

[indigo carbon]

I'm sure you can make a workflow for it, if it exists in A1111 there is no reason you won't be able to make some sort of workflow for it

[indigo carbon]

ComfyUI also allows for modern optimizations like AIT, but that's way too hardcore if you just installed Comfy

[ionic gulch]

der alte fritz:

[solemn tide]

I'm just trying to run Deforum videos with sdxl lol. Either comfy has that ability or it doesn't. Deforum way to complex to "build it out" in comfy if it isnt there

I'll stick to SD V1.5 for now

[ionic gulch]

and the oscar goes to...

[indigo carbon]

that's not how ComfyUI works. it doesn't have "tabs" like A1111 has, you are the one that makes the workflow, making the possibilities pretty much infinite

and yes there are nodes for deforum

[solemn tide]

oh then definitely setting up comfyui. Thanks for the help friend!

[ionic gulch]

what movies should i add next?

[indigo carbon]

the only disadvantage Comfy has imo is it doesn't look fancy after you finished making a workflow.. we need some kind of fancy frontend made that comes with nodes you can add to comfy to allow it to modify certain parts of the workflow while maintaining a simple look

[noble shoal]

Saving Private Ryan

[glad grove]

Fun with Dick and Jane

[floral island]

i've been mixing. a lot

[noble shoal]

Amélie, or Le fabuleux destin d'Amélie Poulain

[ionic gulch]

does this look like the correct style?

[noble shoal]

The colors and style come close, yes. Well i guess the realism can be improved.

[floral island]

the REAL teenage mutant ninja turtle

[noble shoal]

Intense

[ionic gulch]

my goal is to get the style, so i can apply it to different content. like in that example (last one is the source):

[noble shoal]

Are you doing it only with prompting?

[solemn tide]

100% VAE issue. Had it set to vae8000 and needed to switch to automatic

[ionic gulch]

the source is canny, depth or ip adapter, the style is just prompting

[solemn tide]

its go time

[ionic gulch]

this is the same one (the youngest) with matrix:

[ionic gulch]

or ex machina:

[ionic gulch]

mad max witch in paris:

[noble shoal]

Nice work so far 👍. Are you releasing it on your GitHub?

[ionic gulch]

yes, it's already on there, but not with all movies and the latest small prompt fixes. it's in my node pack and in my workflow.

[noble shoal]

Cool. Will check it out tomorrow. I am going to bed. Gute Nacht 😉

[ionic gulch]

gute nacht :)

[hardy cipher]

that bottom pick has me a bit flustered

[upbeat summit]

Spock / Leonard Nimoy lora?

[upbeat summit]

that shirt design - haha nice

[vital ermine]

He still has to exercise, lol

I have been fighting with this one A LOT

Never have issues with styles but people seem to not work for me

[upbeat summit]

the base SDXL Leonard Nimoy doesn't look like Leonard Nimoy at all heh

[vital ermine]

supposed to be the other way around

no, looks funky

[vital ermine]

What made this sort of work, not happy with it though, is I did 1 epoch and all repeats

never tried that before

[upbeat summit]

cool 🙂 the likeness has definitely improved by a lot

[vital ermine]

why does it sometimes ignore me?

[vital ermine]

I used this from my 1.5 success to learn about chars in lora and not really having any luck

back then it was DB so I tried DB now and success but I wantr this as a lora not a db turned into a lora

An issue I seem to have in comfy is this in the lower right I can't neg away without a HUGE value, and sometimes not at all.

[floral island]

who did it best?!

[vital ermine]

3

[hardy cipher]

2 is a bit of a smoothskin

[vital ermine]

yep

[floral island]

none of which i'd want to encounter tho 😮

2 is probably the best model out of the 3, it's a lot more creative 🙂

[vital ermine]

I am analytical in nature I don't see things like that

[floral island]

this is what i meant with creativity -> i feel protovision is lacking in creativity and just made "generic evil thing"

[floral island]

it has one of the best if not simply the best image quality right now, so i'm just trying to get some creativity back in there

fck. my cursed lyrics. i HAVE to do this, and scar my eyes probably

[floral island]

my mission to add creativity to protovision is probably a succes. To celebrate, have a 3 image comparison of the cursed lyrics. warning, you cannot unsee.

[hardy cipher]

I need 3 separate images to celebrate

[floral island]

one more serving of cursed lyrics, coming up ❤️

[hardy cipher]

I need your help in creating a vision. 3 images that have impacted your life

like this, for example

[floral island]

more of the cursed lyrics, who did it best, #1 #2 or #3

[ionic gulch]

protovision doesn't like my upscaler:

[floral island]

well that's weird

which upscaler and renoise are you using?

[ionic gulch]

my own solution in my workflow.

doesn't happen with default sdxl model

[floral island]

that sucks 😦

this model might be pretty damn nice for me ❤️

[ionic gulch]

dynavision doesn't break the image, but also turns it green or blue for some reason

[floral island]

self-made upscaler?

[crisp owl]

I've noticed that with some checkpoints, upscaling tossed extra lips on, or with NSFW images adds extra nipples or belly buttons and such

[ionic gulch]

colors with sdxl default model:
colors are the same even before upscaling

protovision and dynavision turn it green. prompt includes vintage grading, so the default model is closer to the prompt

[floral island]

perhaps the vae to re-encode the image

i know socalguitarist has a built-in vae, that might be different

[ionic gulch]

yes, maybe vae if it's different from v0.9
juggernaut does normal colors:

[shy kelp]

speaking of vae... uhhh how do i force tiled vae usage in comfy?
my hires usually breaks saying out of memory and then switches to tiled vae but why can't i just use tiled vae from the start

[ionic gulch]

there is a special tiled vae node afaik

don't know if it is a default node or in some of the popular packs

[crisp owl]

Believe it's default. It now has the option also to select the tile size

[hardy cipher]

it's default and it's hardwired into the code that you default to it if your vram gets maxed

[shy kelp]

512 sounds like a lot

[hardy cipher]

well you could change it to 2 or 4

lots o tiles

[shy kelp]

hmm lowest i can input is 320

[hardy cipher]

4 seems ideal

256^2 tiles

not a big deal

[vital ermine]

I don't like lora for this

[hardy cipher]

more nimoy than nimoy

[vital ermine]

either his hands are huge or his ears or it just isn't right. DB it just works so might witch to that and see if extraction will work

[hardy cipher]

have you messed with the weights at all?

[vital ermine]

yep

[hardy cipher]

ever go nuts and put the weight into the negative for a lora?

[vital ermine]

yes, of course

[hardy cipher]

whoa

[slender coral]

Does anyone know if there's currently a mnode in comphy that will allow to easily read wildcards from a text file?

I want to do something like thing = [cat, dog, man] and replace those in something like A {thing} in a scene with xyz

[hardy cipher]

you know, that terrible fan art lora seems like it'd be great for that

that's gandalf's evil inbred cousin

[vital ermine]

LOL

[hardy cipher]

that image is moderately disturbing tbh

[vital ermine]

Damn vampires.

[hardy cipher]

I like it when pictures are on the wall in these weird renders

[vital ermine]

yeah, not even asked for

[hardy cipher]

I'm sure someone will come up with a method for blowing those images up and rendering them on their own

[vital ermine]

hmmm, these are a bit disturbing in a freaky way

[hardy cipher]

remember "this person isn't real" from a couple years ago? that was some mind blowing stuff. and there'd so often be that secondary person/demon all deformed on the side

soon the ai is going to get too good to gift us with such things

[slender coral]

Could someone advise where to hook up my lora?

[vital ermine]

I like this one

[slender coral]

Am I going insane or missing somerthing I thought text was in the core?

[willow eagle]

9 suns in the sky

[crisp owl]

looking for #1100170312106127410 ?

[hardy cipher]

I love bots

sent the image through an interrogator and used the result to make a lovely theme song for it

[rigid bobcat]

Any reason to use Hires upscaling instead of SD Upscale in Img to Img?

It's bugging or something I'm not sure...it's taking a ridiculous amount of time to produce even one image

[raw cedar]

Sorry do you mind explaining more in detail, the lower the tile size the less VRAM but Longer the generation? Am I correct?

[hardy cipher]

Yeah, that's essentially how it works. But I was being sarcastic earlier. Smaller tiles would very likely degrade the image significantly

[raw cedar]

Lol okay thank you so much, I bow to your superior skill! hahaha (seriously tho sick generations my man). Another question, is there a correlation with the tiled VAE and SDXL being trained on 1024 or can you use 512 tiled VAE

[crisp owl]

If you have VRAM limitations, tiled VAE of 512 will help you as it's split your image processing to 512x chunks.
Slower, but allows to continue without OOM.

If you don't use a tiled VAE node, Comfy will automatically detect if you're gonna OOM and switch itself to tiled processing. But it takes additional time to detect that

[hardy cipher]

hmm

yeah, it sends me to that regularly, lol

[crisp owl]

just put a tiled vae node where that happens to give yourself back a few seconds

no sense in waiting for it to detect that

[hardy cipher]

yeah, I made my own vae loaders to try and better understand how that stuff works

but it only ever happens at the tail end of a long run of things

always at the end, a stupid upscale or something

[crisp owl]

Yeah my last vae node, after upscale, I have swapped to a tiled vae

[hardy cipher]

if I make the mistake of leaving in a regular vae I'll usually just close the program and restart because better than sitting there for what seems like several minutes

[hoary saddle]

thank you @visual glade for adding the badge with ID's, i just found it and very appreciative! really helpful with all the python scripting, just wish'd i'd found it earlier LOL

[visual glade]

the badge with ids?

[hoary saddle]

or whoever is responsible (maybe creator of the manager addon)

[crisp owl]

manager thing

[crisp owl]

ltdrdata's work

[hoary saddle]

ah, well kudos to that person and comfy anyway cause he's aweesome 😉

[supple knot]

My MiDaS Node stopped working ,any idea what I can do?

[crisp owl]

new midas node I think

There's this default Comfy one

or these WAS pack ones

[wide crane]

guys

how do i unfuck faces

i like this imag

[hardy cipher]

face fixer

[wide crane]

but the face is

irredeemable

[wide crane]

pls how

[floral island]

found new waifu, don't think she's happy tho...

[crisp owl]

This is my face fix process I use

[hardy cipher]

I don't know. I like goofed faces

[floral island]

isn't there an adetailer alternative for comfy?

[crisp owl]

probably, but for the rare times I need it, this does what I need

[hardy cipher]

there are at least 2 different face fixer things I believe

[wide crane]

does this actually work well

would it be able to fix this

[wide crane]

fucking

thing

[crisp owl]

dufuq 49 lora launcher?!

[hardy cipher]

this is the perfect face in my opinion

[wide crane]

?

whar?

[wide crane]

that is not the case

[crisp owl]

hahaha weird, loads like that in mine 😆

was gonna say, who the hell uses 49 at a time 😅

[hardy cipher]

what's up with these tools not having regular inputs and outputs?

[wide crane]

talking about mine?

[hardy cipher]

yeah. I'm not trying to knock it. I just don't understand the why part

cause seems like it'd be simpler to stay consistent with other lora loaders

[wide crane]

im not sure what you mean

[hardy cipher]

nevermind. anyway, fixed her face

[wide crane]

so real

[crisp owl]

In my opinion, I think it's in two part. One for the creators own purposes, such as just simple ease of attaching a single point, but you have to stay within your own nodes. So your flow consists of purely "efficiency node" nodes.
And two, to reduce the amount of noodles everywhere, but that's now mitigated with the ability to hide the noodles

[wide crane]

like

these images are so cool

but the faces mna

idk what to do about them

[hardy cipher]

I do get that, but let's consider lora loaders for a second. you can stack them vertically, line them up horizontally. I can think of few reasons why they'd need to be separated. so no real need to run some weird pipe that cna only go to one other node etc.

[crisp owl]

unsure on anime type, but probably. Just still goes off your prompts/models/etc.
Works well enough on my realism images when I need it

[hardy cipher]

just make 3 or 5 stacks

and have regular inputs and outputs

but whatever makes people happy. doesn't matter to me

[wide crane]

yea im not going to even try with these faces

thats a headache im not ready for

[crisp owl]

Yeah I opt for using the Comfyroll lora nodes. Normal connections, plus the on/off switches, so I can keep them inline

[hardy cipher]

I just made my own with a few loaders in one and switches

but similar concept to those

[wide crane]

@crisp owl so am i just out of luck

because inpainting doesnt fix the faces

and i dont know what to do to fix them myself

[crisp owl]

did you try the facefix node I showed?

I screenshot so you can just copy over my settings

worth a try.
Or you can just keep mashing images till you get a good one 🤷🏻‍♂️

[wide crane]

i can barely see the settings

[peak dove]

Is this a custom stacker, or is it available on ComfyUI?

[wide crane]

its a custom module

[peak dove]

Available?

[hardy cipher]

all done

[wide crane]

what do i do with these

[crisp owl]

The top image will be the actual finished image. All the others are the seperate processes going on. like "cropped_refined" would show you a preview of the zoomed in face it touched up. etc

You can just throw a preview on each of the outputs to see what they all show.
But the top is the finalized image

[wide crane]

YO

IT ACTUALLY FIXED THE FACE

[crisp owl]

bueno

[wide crane]

@crisp owl you beautiful man

i could kiss you on the lips

[crisp owl]

https://tenor.com/view/ugly-love-me-dentist-ilvi-no-filters-gif-15353098

[wide crane]

except weird clip idk

what that is

ill see what i can do abt that

oh nvm

thats just hard shadows

that arent good

[crisp owl]

If you wanna keep some minor details from one vs the other, can just toss into a photo edit program and mask out the one vs the other

[wide crane]

yea

thanks again dude

you saved me from a year of headache

[crisp owl]

as I said, when you need it, it works well enough haha
I'm sure there's other methods, but that was one i got setup working well for me

[wide crane]

now the real question

is when will someone make an extension that lets you put clipping on modules

so they look

all nice

and aligned

[hardy cipher]

clipping?

you mean resize them?

[wide crane]

no i mean like

grip snapping

is a better word