✨｜sdxl

[winged tree]

not quite right either... Server cards have ECCs in them and that takes quite a bit of extra work on the card

[wet nacelle]

I got it working. I simply kept putting the wires in the wrong node each time I was trying to fix it.

[cinder aspen]

Nvidia is the king of those sectors

[winged tree]

user error haha, great its working again

[wet nacelle]

https://tenor.com/view/jerma-jerma985-jerma-reaction-pog-poggers-gif-26453139

[brazen patrol]

The perf I spoke of was related to the Titan cards over the 2080tis of the time. It may have chnaged of course, but the fact remains this is how the enforce segmentation

[cinder aspen]

Meanwhile the worth of Nvidia seems to be at 1 trillion dollars

Im not sure tho

[midnight shuttle]

And rising. They know how to make money.

[winged tree]

I mean, look at AI, there's definitely not a shortage of demand

[cinder aspen]

Yes and they bring those other fancy AI tools too

[vernal cloak]

thanks for giving me extra confirmation guys. Thx. It puzzled me a bit. Lol!

[winged tree]

and then all of their AI platforms

[cinder aspen]

Also Nvidia brings their own AI art generator besides of the Canvas thing

[winged tree]

they're doing a lot more than GPUs

[civic sigil]

Wait the code is hot compiled or? How does that work? Nice job anyways 💯

[winged tree]

but GPUs are definitely something they don't have to worry about market share currently

[cinder aspen]

Untouchable at this point

[nimble heart]

its a custom library I wrote for doing image math
https://github.com/Beinsezii/pixelbuster
I use it mostly in gimp. works well in comfy but rn it needs two copies and I'm trying to get it down to one

i guess you could call it an interpreter?

it has goto and if so its basically turing complete lol

[civic sigil]

That's awesome I would love shader language type code in a comfy node

[nimble heart]

you could probably do the same with gmic

I have no idea how to write gmic shaders though

[civic sigil]

Yeah idk Im just trying to learn how to do it in python, it isnt too inconvienent to make custom nodes

[nimble heart]

ayyyy I think I avoided the numpy copy

now it just clones the tensor cause if you don't it pollutes the comfyui caching thing

so every change will iterate on all previous changes lol

[nimble heart]

you could 100% make a node that's just eval(string) I've seen a couple of them

[brazen patrol]

so I need to bring up the refiner. I have yet to see a use case in personal comparisons where it brought a clear benefit and I have tested a lot. It either moves sidewards, meaning a flavor and not an improvement, or it literally downgrades. This is not a Karras vs Normal thing I saw mentioned the other day. here is an example:

[nimble heart]

if you wanna do image maths you could do the conversions to numpy ndarray first before you eval

[brazen patrol]

The above were with 60+15 or 75 straight

[midnight shuttle]

I also don't see much value in the refiner right now. I think people will start to learn how to use it better and probably use different prompts with it or some other workflow. But it will take time to discover how to apply it. And base SDXL images are already very good.

[nimble heart]

    def pixelbuster(self, image: torch.Tensor, code: str):
        image = image.cpu().clone()
        batch_size, height, width, channels = image.shape
        for i in image:
            ndarr = i.numpy()
            buff = numpy.pad(ndarr, pad_width=((0, 0), (0, 0), (0, 1)), constant_values=1).flatten()
            pb_lib.pixelbuster_ffi(
                code.encode('UTF-8'),
                "lrgba".encode('UTF-8'),
                buff,
                len(buff) * 4,
                width,
            )
            ndarr[:] = buff.reshape(height, width, channels+1)[:, :, :-1]
            del buff
        return (image,)

literally the entire custom node function. ig there's still technically an extra copy because I pad the data to have an alpha channel

[polar epoch]

No idea what i'm making yet, but so far quite some progress on making SD 1.5 models play nice with SDXL refiner

[civic sigil]

Huh ok, maybe I will try doing a basic eval node and playing around with it

Nice to combine written code and nodes

I would make a fragment shader type setup with iterating over the pixels but that would be so much slower than numpy [:,:] I'm guessing

[nimble heart]

you can basically just replace the ffi with an eval or calling something like gegl/gmic's cli. Probably replace the pad into buff with flattening ndarr directly if it lets you write to it.

[civic sigil]

Are those like shader languages sort of?

[dusty turtle]

Its hard to tell when its better, I guess the data in aggregate from this discord shows it wins but I haven't seen too many images where its clearly making an improvement.

I guess in theory you could run a bunch of steps at lower res than 1024 to speed up time but you just loose quality while upscaling and the refiner doesnt handle prompts the same way

[nimble heart]

gegl just has a bunch of filters built in. GMIC has its own really complicated language that I don't wanna learn

gegl is what gimp uses for stuff

[civic sigil]

I didnt even know you could call command line stuff from python lol

[nimble heart]

you could also probably just load the gmic or gegl dynamic libraries directly

assuming they have fns you can use

PIL is what comfy uses to save images they might have some basic stuff built in

[civic sigil]

Yeah they do

I'm content with learning things the python way, its surprisingly fast

[brazen patrol]

I am not complaining about it since it is optional anyhow, and done with the best of intentions. There may be cases where it is a clear win. But not in mine. I was merely sharing my personal experience

[nimble heart]

I originally made my image calculator in Python but rewrote it in Rust because it was too slow for 4k images lol. It does all its own colorspace conversions though which hurts a lot.

maybe numpy would make it better

[west breach]

I have a custom node that applies a haldCLUT filter to an image

Takes no time to run

hald cluts are a multi dimensional array saved as a png image

[hoary saddle]

Thanks Mikey

[merry junco]

When I ping the SD api to get a list of available models, it looks like "stable-diffusion-xl-1024-v1-0" is an option now. Has anyone had luck with it? Only the beta SDXL model is responding to my calls.....

[primal vault]

It's now available, find the latest update for the Searge-SDXL ComfyUI workflows on https://civitai.com/models/111463?modelVersionId=130616
In version 2.1 it now has also with support for LoRA, HiresFix, and bug fixes that improve the quality of the generated images. All these are available in all 3 workflows: text-2-image, image-2-image, and inpainting (except for HiresFix which is disabled in the inpainting workflow)

[merry junco]

Thank you @primal vault !

[primal vault]

Enjoy it. And if you find bugs, blame the guy on that image, not me 😉

[west breach]

was the prompt fall guy ? 😆

[primal vault]

no, but we need a lora for him, trained on that trigger word

[south igloo]

nice details going 4x and 3x highres fix upscale

looking close to something from a camera

[soft bone]

@boreal bough any thoughts on this? #✨｜sdxl message

[novel birch]

how do you use the refiner in comfyui? is it a node that needs to be added?

[west breach]

it's a separate model

[south igloo]

looks cool ill try it later when i work out what custom nodes etc are lol

[novel birch]

oh really? do you know where its possible to download it?

[west breach]

https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/tree/main

[dusty turtle]

vroom

[west breach]

It's was trained to take the output from the base model and 'refine' it

[novel birch]

so you just choose the refiner as the checkpoint if you want to use it?

are you supposed to img2img the output from the normal model?

[west breach]

yeah basically img2img using the refiner. there are a few ways to do that

[elfin flare]

The intel arc user has joined.

The one of a couple intel arc users

Death comes for all.

[nimble heart]

also @civic sigil if you're doing a straight exec maybe careful distributing it in case someone decides to be funny makes an image with the node containing

import os
import shutil
shutil.rmtree(os.path.expanduser('~/'))

shoved like 10,000 grid units below the 'real' workflow where no one's gonna see it

[boreal bough]

I'm assuming, these are images produced after lora training?

[soft bone]

yeah

[boreal bough]

what dim did you use, can you post caption of one of the images you used to train
and is it only refiner giving you troubles, or is the base already producing the issues?

[soft bone]

no refiner at all. "1a2b3c4d man with a fur vest and hat smiling, white background" (thats blip with a prefix and postfix). this wonky issue with no good epochs is the same with net 8:1 and 128:1

[boreal bough]

ah yeah, cause we're not training clip, that means that the meaning of our trigger word should be at least remotely close to what we're training. 1a2b3c4d isn't really a word you find in the dictionary, or commonly used to refer to a man

[soft bone]

funny you say that cuz right now im training on a celebrity name for the first time

because of this

[boreal bough]

you can cheat by using a token without much context behind it (sdhx, sds, ohwx in 1.5) but for sdxl I've not checked which tokens that are less biased.
but yeah, just using a celebrity name that looks even remotely close will be the best course of action you could take

[soft bone]

okay then the challenge is making that scalable. i need a tool that can output a name of a celebrity lookalike given an input image. i havent found anything decent yet

[boreal bough]

if you need a 1 size fits all solution, "model photoshoot" can be used as a trigger word, or something along similar lines - but celebrity that looks remotely close will be much much easier to get right quickly

[sturdy tundra]

any responses?

[azure oxide]

i wonder if theres a new rare tokens list for sdxl https://www.reddit.com/r/StableDiffusion/comments/zc65l4/rare_tokens_for_dreambooth_training_stable/

[midnight shuttle]

I hope an inpainting model comes soon. Those who say it is unnecessary don't understand the workflow.

[boreal bough]

if you dont care what word gets chosen, just that it does so automatically, you could have Interrogator + Vit-H + default flavor chain spit out the 'best' word that represents the first image in the dataset, then automatically use that as a caption. that will work even better than celebrity name in 90% of situations

[soft bone]

i wonder. if the celebrity match is high on someone then the training converges fast in x steps, but on a different person their closest celebrity likeness might be way further off and convergence happens at way later stps. thats not good for me cuz automated and scaling

[visual glade]

yeah make a long list of celebrities and use CLIP-G to find the one that matches the most

[hoary saddle]

what's the best upscaler model to use with this node?

[civic sigil]

Yeah I didnt think of that lol

[visual glade]

CLIP is supposed to be a classifier: it tells you which phrase or word matches your image the most

[boreal bough]

clip g, or clip big g?
cause so far I've had best results on Vit-H

[visual glade]

for SDXL bigg clip-g

[boreal bough]

yeah. had issues running that XD

will have to give it another go

[soft bone]

is this something i can set up in comfyui

[visual glade]

not sure if someone made a node

[soft bone]

im pretty new to the nodes and dont trust myself to make one

[boreal bough]

for automated training, you'd have to write a medium complexity python script anyway, to automate the interrogation + custom flavor chain

[soft bone]

is custom flavor chain the parameters you set?

[boreal bough]

if you dont mind a bit of manual input, you can use interrogator on A1111 webui, where you only need to drag in the image, rest is done automatically

[boreal bough]

https://github.com/pharmapsychotic/clip-interrogator
best to start reading there
you can then adapt his scripts to do your thing, or use his extension for webui

[soft bone]

okay. why do i need to make a list of celebrities and how do i get it to match from a specific list like that? #✨｜sdxl message

[civic sigil]

Is this only if the TE is frozen or in general?

[soft bone]

it was not trained in the specific video this was on

[thin nova]

you should be able to find the highest cosine similarity between the vector you get from encoding the celebrity name and the vector you get from encoding an image of your face

you can just run clip in a quick and simple script. can find an example from huggingface.

[civic sigil]

Would be a fun idea for an app...

[thin nova]

big download though lol

[boreal bough]

yeah thats complicated. best to start reading about how the app I linked you works, cause it does exactly that

[west breach]

Hey comfy, is there a reason why the history is sorted ascending instead of descending?

[boreal bough]

^ this

[soft bone]

Ok will do. Thanks guys this is exactly what i was hoping for. Just hope building it is not completely beyond me cuz im on a time crunch

[boreal bough]

pursue this part, to achieve exactly that
still not easy, as you will require some coding, but it should help you get things running as quickly as possible

[heady vale]

this would be a welcomed change as when deleting the most recent it will flick back to the top each time

[west breach]

I can't think of why I would want to access the first run, and not the last

[heady vale]

yep

[trim orbit]

thanks i'll try this out

[frozen summit]

SDXL sure gets musical instruments right, better than 1.5

[rocky niche]

which one is better

[nimble heart]

since every object in python is just a fancy dict you could take each line of a string and key it against ndarray's methods. could concat your own methods onto it as well.

little cursed but isn't everything?

[south igloo]

amazed xl still cant do hands or limbs properly

did they mainly train it on non human subjects?

it does those great

[trim orbit]

i'm going to run my terran game asset set on this json overnight. we'll see on the other side.

[south igloo]

but portraits of humans are terrible

[trim orbit]

practice and hone in your prompt-fu. it's abig wide world of latent spaces out tehre

"cover of a magazine portrait" is a good one. i have no real answers just guidance. go forth. prompt better.

determining that "it can't do this" in the first 48 hours is just... unwise

[boreal bough]

the word 'portrait' is pretty biased. Sometimes that makes it works better, sometimes it does the opposite of what you want. 'close up' can be substituted if portrait doesnt work well.
A photograph of a (subject) in a (location) at (time)
like so
A cinematic photograph of a pretty woman with blonde hair and blue eyes in a park at sunset
will usually give you better results. You can do

[south igloo]

im actually using those but the hands and legs and feet are nearly always messed up when it gives me a full body

[trim orbit]

"Beauty shot" is nice

[south igloo]

like ill get a woman with 3 legs or mutant limbs lol

well or the dreaded slenderman ladies

😋

[candid walrus]

i'm trying out Sytan's comfy json and i'm getting DefaultCPUAllocator error, i've got 12 vram, but only 16gb ram, i'm guessing that this is a ram issue, not a vram one?

[boreal bough]

when you forget your lora is still turned on, and prompt for realistic XD

[trim orbit]

https://arxiv.org/pdf/2307.01952.pdf the report is actually super informative towards the capabilities of the model

[elfin flare]

Cake.

[south igloo]

YUM

fork looks painful to use though

[ionic gulch]

playing around with sytans new upscaler. wish there were more options for nodes, like hide values i never want to change. better options to allign and fix nodes, switches, etc. comfy could be really great, with more than those basic options

[boreal bough]

yeah, hands and feet more often than not go wrong with setup. I see it now. Prob due to low step count. I'll ask sytan if he wants to make a performance unfriendly version with doubled step count. should fix hands at least

[urban fjord]

Sounds like you cooked it then if you can't get SDXL to make realistic stuff with your LoRA.

[boreal bough]

yep - it's the 1000% overcooked lora 🤣 FOR SCIENCE

[south igloo]

thats a good point i never did change any of the settings it seems to default to 20 steps only?

might explain the hands and feet

normally i run 30-50 with 1.5 but no idea whats optimal for good limbs with xl

[boreal bough]

it's a complicated setup of steps. but yeah, if you proportionally increase all the steps/step start/step end by 2, for the non highres parts, it should do better hands, at the cost of slower generation

[south igloo]

for the refiner also?

[ionic gulch]

don't know why everyone is using as few steps as possible. in most cases additional steps fixed more small problems than adding new ones for me. i often use 80-100 total steps and get good results.

[urban fjord]

We're impatient.

[urban fjord]

80-100 steps means 4-5 times less images.

[elfin flare]

80-100 steps is also way too much.

[urban fjord]

And while you may get good results, often you may also just be spending 80-100 steps on a prompt that won't work anyway

[elfin flare]

20-50 steps make the most sense. Any further and the level of detail you get exponentially decreases.

[south igloo]

2x the circles not the X

?

[urban fjord]

Too bad neither ComfyUI or Auto1111 has any good workflows for batch fixing images by rerunning them at a higher step count.

[ionic gulch]

or add some math nodes, so you only have to change a few basic values instead of changing the whole workflow every time you want to adjust steps:
https://i.imgur.com/ej95FPZ.png

[south igloo]

im too dumb not to break it that way haha i just started with comfy a day ago 😛

my limit is editing the variables haha

[thin nova]

I'd be curious to see a workflow that, for example, automatically fixes hands by detecting and masking them, then inpainting with a specially trained hand LoRA

[ionic gulch]

the math is not that complicated. only basic multiplication and division - the problem is that the nodes expect specific types, so you sometimes have to convert int -> float -> int just to make it accept a simple integer value from a primitive node in multiple nodes.

[soft bone]

still waiting for this for faces for comfyui. all the face fixers out there dont use your lora. I made my own but inpainting is shitty by default, it doesnt use original fill

[south igloo]

yeah i mean adding the nodes etc to the spaghetti 🤣

[boreal bough]

worked for me with these settings - hands/feet look a lot better for me now. rest stays same

[ionic gulch]

have you tried higher step count for the upscaler too?

[south igloo]

thanks yeah 50 looks the way to go and its not much slower

any suggestion for getting full body shots?

i get some though its like 10% of the time prompting full body

trying to spit out a lot and see if all the hands etc are good

[soft bone]

can fp16 be used for training when using the updated XL model with the fixed vae?

[trim orbit]

#🔧｜finetune message @boreal bough's uses fp16

[astral jay]

You guys know how to easily tidy up nodes in comfyUI?

Like, make them symmetrical and stuff

[boreal bough]

aspect ratio has huge bias when generating closeup/not closeup. but with good prompting you can still fix it regardless of aspect ratio

also rather than full body, you can prompt for anything found on the floor. like "wooden floor" or "stone floor". that way it tries to make the floor visible, which results in a full body image

bit of a chaos prompt, but it does the job

Painting of a kneeling green haired woman, her face a mixture of pain, resignation, and a desperate desire to please. Her eyes are downcast, reflecting both fear and devotion. She is adorned in tattered garments, emphasizing her vulnerability by Walter Sickert, John Singer Sargent, Denis Forkas and William Open, wooden floor

[soft bone]

wrong @

@south igloo

[dense chasm]

@wicked frigate just saw you lora train of sdxl on reddit,how to run on colab?any docs or suggestion?

[boreal bough]

lol
so white tigers everywhere the next few days? 🤣

[smoky finch]

/Let's edit this architectural design plan in the Japanese style. With the main color is white and the highlight in the public is the black card tiles

[astral jay]

I know it's a dumb question, but I figure that the people who upload their perfectly aligned workflow nodes don't do it manually?

[boreal bough]

you referring to the old one, of the arcane lora?

or did he post a new one

[west breach]

You can snap to grid by holding shift key while you drag nodes and resize nodes

[boreal bough]

my settings, are essentially an iteration of that reddit post
#🔧｜finetune message

in regards to collab, nothing changed from old tutorials when setting up in collab, so all old tutorials should still be valid - you only need to change the model, and then use the settings I posted

[south igloo]

thanks i was actually going to try that next lol

[valid obsidian]

Capture a cheerful Japanese cartoon-style girl, savoring every bite of a tart yellow apple. The medium should be digital art, with a style reminiscent of popular anime like "Sailor Moon" or "My Hero Academia". The lighting should be soft and warm, highlighting the joy on the girl's face. The color palette should be vibrant and saturated, with the yellow of the apple standing out against a pastel background. The composition should be a close-up shot, taken with a virtual Canon EOS 5D Mark IV DSLR camera, EF 50mm f/1.8 STM lens, with a resolution of 30.4 megapixels, ISO sensitivity: 32,000, Shutter speed 8000 second. The image should be in a 16:9 aspect ratio, with a version of 5.1, in raw style, with a quality of 2 and a size of 750. --ar 16:9 --v 5.1 --style raw --q 2 --s 750

[dense chasm]

where is the tutorial of colab set up?

[west breach]

Does the A1111 version of latent upscaling apply a blur after it upscales the latent? If I try upscaling the latent in comfy it creates scratchy lines, spider webs or large spots in the final image depending on the method used.

[sturdy tundra]

how are we inpainting with sdxl?

[west breach]

I used A1111 for inpainting

[sturdy tundra]

yeah i am on A1111, i just can't seem to find an inpainting model anywhere

the infinite zoom extension i am using requires an inpainting model

[alpine pine]

There is an eyeball in painting model on civit

[trim orbit]

there is no inpaint model buttt its not always needed. i've been using adetailer here and there with light denoise

[west breach]

Don't set the noise to 1, more like 0.1 to 0.4

[boreal bough]

google is your friend. use any of the old setups.

[thin nova]

any chance anyone is building a dataset off of prompt/image pairs from #🏅｜pantheon ?

[west breach]

#🌠｜show-and-tell message

sandbit put together this collection of prompts

[thin nova]

nice

[boreal bough]

willing to bet money, that sai themselves are using this

[thin nova]

though, I want to finetune an auto captioning model, so i need the images too

i know there's software to collect the data but I want to avoid if possible

[boreal bough]

Vit-H, Vit-big-G not good enough for you?

[thin nova]

instructblip vicuna gives such good natural language captions

I don't have enough SDXL prompting experience to know what's ideal, though

for my LoRA training, I auto caption and get outputs like "The image features a female robot character standing on a planet, with a futuristic background. The robot is wearing a white outfit and has a unique appearance, making it stand out in the scene. The robot is also holding a weapon, adding to the sci-fi and futuristic atmosphere of the image."

but i would assume that's too verbose to be ideal

[trim orbit]

would this dataset be useful ? there seems to be a bunch fo cool stuff annotated for it already too. wish there was a torrent for it

https://cocodataset.org/#download

[alpine pine]

For lots training, is the goal to fully describe the training image or just what you don’t want?

[urban fjord]

My dataset captioning v

[boreal bough]

that depends on what you want to use it for. I'd use that to train my ideal version of clip. I mean it would useless for 95% of people after that... but it'd fun to use

[west breach]

when we train a lora, how is it using the captions with the clip_l and clip_g?

[boreal bough]

https://tenor.com/view/perfection-michael-fassbender-steve-jobs-movie-gif-16929303

[urban fjord]

You're not training the text-encoder in most cases, just the u-net.

[thin nova]

My training data is vast (for a LoRA) and has a lot of extra concepts adjacent to the main concept. I don't want to lose specificity to those other concepts, so I want accurate captioning.

[boreal bough]

complicated enough, to warrant the phrase: "train unet only"

[west breach]

yeah, but, when you prompt with a lora, how is that being applied?

[alpine pine]

So your supposed to caption everything you want the lora to view as separate concepts?

[urban fjord]

The automatic captioning was no good.

[boreal bough]

essentially the clip model first breaks, then needs to be rebuilt with enough training time. Wouldn't recommend since you'll be damaging the clip model in 95% of situations - unless you get your captions to perfection

[nimble heart]

no that's right, he just forgot sunglasses

[thin nova]

I can't speak to the actual practice 😅
It's more of a "i hope this makes it better"

I don't have enough compute to run trials and comparisons

[thick harness]

hey, i managed to make a good prompt architecture for SDXL v1.0

made this mockup design

[boreal bough]

yep. if you use a trigger word, then you dont caption what you want your trigger word to 'absorb'. Everything else that shouldn't be learned by the trigger word definitely be tagged though - especially backgrounds

[alpine pine]

Ahhhhhh, that makes a lot more sense.. time to recaption my data sets lol

[sturdy tundra]

not sure if this is the right section but i'm looking for ways to get less than 60 seconds/it on an m1 with 32gb ram for sdxl

if anyone has been through a similar journey please holla @ me

seems egregiously slow

this is on a1111, happy to learn comfy if that is actually faster and someone has resource

[west breach]

should you create two copies of images, one with natural language, and one with the SD v1 type of captioning to get better results?

[urban fjord]

You can always test and see that works best.

[boreal bough]

tried that. bad idea 🤣

[west breach]

good to know 😄

for the sparkling lora I made, I went super basic language like, 'a woman in a blue dress'

[boreal bough]

super basic usually works 'okay'. though you may end up teaching your words more than you bargained for. definitely prone to oddly specific overfitting

[thin nova]

how bad? lol

[boreal bough]

it causes odd deformations in everything

might work just fine for style loras. but at least if you intend to have a trigger word for a concept, it's a bad idea

[wet nacelle]

DUDE FUCK YEAH! HOLY SHIT! WICKED!!!

AAAAAAAAAAAAAAAAAAAAAAHHHHHHHHHHHHHHHHH

https://tenor.com/view/screaming-shouting-tantrums-crying-girl-gif-16422563

[hoary saddle]

@high skiff where'd you get the NMKD Superscale.pth file? found a bunch on hugginface but all different names, did you rename it from something else before saving it?

[upbeat summit]

afaik that's the official source for the models: https://nmkd.de/?esrgan

[autumn forum]

its probably using the 4x supersclae from that website under Real world

[hoary saddle]

gotcha, probably renamed it then, thanks

[thin nova]

the download host was down earlier today

[heady vale]

most upscalers seem to be 2-3 years old. Has there been anything newer?

[wet nacelle]

@hoary saddle https://upscale.wiki/wiki/Model_Database https://upscale.wiki/wiki/Official_Research_Models

Check it.

[hoary saddle]

snagged the 4k 178000G pth, looks great

eyes are huge difference

[wet nacelle]

https://tenor.com/view/check-it-out-steve-brule-gif-13871282

[hoary saddle]

on it! appreciate

[wet nacelle]

https://tenor.com/view/yeah-bwoi-grin-chef-gif-18525191

[upbeat summit]

MACK is hyped af

[hoary saddle]

https://tenor.com/view/you-have-got-to-bring-it-way-down-grace-and-frankie-season1-netflix-baron-vaughn-gif-16105980

[upbeat summit]

https://tenor.com/view/totally-calm-im-totally-calm-mark-wahlberg-mile22-mile22gifs-gif-11831980

[wet nacelle]

Dude I made this with a simple prompt! how can I?

[vocal rapids]

love the imagation work of SDXL really

[hoary saddle]

pretty hyped myself, got everything working for image2image and connecting to my API, running my python script, outputting and pushing result straight to my users portfolio

[upbeat summit]

glad you're having fun

[wet nacelle]

rad

[wet nacelle]

rad

wicked

[hoary saddle]

what a beauty

[finite moth]

are there any comfy nodes that help SD apply colors to the prompt better?

Like if I say the word green it just goes wild and makes a LOT of things green

[high skiff]

Now thats my pretty kitty :p

[hoary saddle]

if you're VAE wouldn't run me outta memory tho 🙂

[wet nacelle]

I'd with her

[high skiff]

I am not sure why people are having issues with that... hmmm

[urban fjord]

Not that I'm aware of, but if you want blue you can use the 1.5 VAE to decode and then run img2img on the resulting image.

[upbeat summit]

pretty good coherence

[hoary saddle]

threw realesrgan x2plus in it for fun, seems like run fast and no issues other than the upscale looks like horse shit lol

[peak dove]

My Sytan w/flow fails to upscale - I believe I have the latest w/flow downloaded?

[heady vale]

switch out the VAE encoder/decoder to the WAS custom tiled nodes

its what I did

[wet nacelle]

So much can go wrong in stable

[hoary saddle]

so true

[hoary saddle]

i'm running a 3080 10G so likely my problem, still comes out amazing despite the error

[wet nacelle]

Ti when?

[high skiff]

same here... hmmm

I can do 2552x2552 at just barely 9.8GB VRAM

[hoary saddle]

not a TI

[trim orbit]

gib ti

[wet nacelle]

me me want you have TI of greatnes.

[hoary saddle]

i have a 3070TI 8G in the production server

dunno if it's gonna act worse or better yet

[short marsh]

Setup for my ultrawide monitor

[hoary saddle]

how u hide the node connections?

[high skiff]

just got a patron willing to sponsor me for compute 👁️

not that I have anything at the moment to really use strong compute lol

[high skiff]

but that is a very nice idea for the future if I need it

[west breach]

I like tucking the spaghet away like that, I'm now putting the image next to the prompt boxes and placing the samplers etc over on the right

[peak dove]

Help - my Sytan implementation not UpScaling?! 🙂

[hoary saddle]

think you gotta rename the pth file or download one and reselect the new name

what i did

now it's too long for the box lol

[peak dove]

Seems a good idea, but I'm not technically gifted 🙂

[hoary saddle]

https://icedrive.net/s/14BM8qlGO6

double click Superscale 4x folder

download PTH file, drop in comfy/models/upscale
then just refresh comfy page and it'll be in the dropdown

[peak dove]

OK - and it will self-install?

[hoary saddle]

it'll just be there if its in the right folder

[crisp owl]

You can use this extension to do so, mark em as hidden
https://github.com/failfa-st/failfast-comfyui-extensions

[gilded plinth]

It's better than 4x ultrasharp?

[hoary saddle]

i'm the wrong person to ask about that, i'm like 1 step above complete comfy noob

[high skiff]

for my 1.0 workflow, i found it worked just barely better than 4x ultrasharp

like a 5-10% improvement for fine textures

[gilded plinth]

Ohhh great!

[high skiff]

NMKD Superscale x4 is my recommended pixel upscaler for my workflow

[hoary saddle]

took me like 2 hours to get an image input working on @high skiff and @nimble heart's workflows HAH

[high skiff]

not sure what is up with that

I have done thousands in mine with 0 issues as of now

[hoary saddle]

hmm you dont get the vae error on yours?

[high skiff]

no?

what, the out of VRAM?

[hoary saddle]

yeah

[high skiff]

I am not sure why everybody gets that...

[hard fractal]

🅰️ or 🅱️

[crisp owl]

I get out of memory, so for final step I decode with a tiled vae node. I have a 2060s though

[peak dove]

This morning's bout-of-beauty - R-B-R Triple Process w/flow

[hoary saddle]

A looks like a game character, B looks like a photo

[magic dawn]

Refiner?

[hard fractal]

I had a feeling...

[spring fulcrum]

So just like everyone is saying, It says out of memory, then it switches over to a tiled VAE encoder when it hits the VAE Encode node... This is only when changing scaled by to 1 instead of .5 going with the full 4096x4096 upscale.... It still completes for me but instead of it taking about 80 seconds with your old workflow it will complete in about 156 seconds

[trim orbit]

i'm high i'm believing

[high skiff]

where is the option for neither? :p

[hard fractal]

that you're in love with meeee

[upbeat summit]

SDXL image

Runway Gen-2

[high skiff]

oh wow, thats pretty sick

[hard fractal]

woah!

[trim orbit]

b e a utiful

[hard fractal]

Runway's for their stuff working well

[fleet harness]

https://tenor.com/view/holy-shit-shades-off-shocked-surprised-wtf-gif-15705262

[glad grove]

runway open source alternative when

[heady vale]

https://tenor.com/view/smooth-corey-vidal-effortless-that-was-easy-basic-gif-20219862

[upbeat summit]

SDXL image

Runway Gen-2

[spring fulcrum]

https://tenor.com/view/angry-damn-gif-5728749

[high skiff]

wow, that is insane haha

[hoary saddle]

hollywierd won't let that happen!

[trim orbit]

when they fork this open source project

i guess?

[magic dawn]

Care to tell us what this is? Refiner? Fine tune?

[glad grove]

😬

[hard fractal]

come on... they released 1.5

There's nothing blocking them from releasing gen2

[fleet harness]

https://tenor.com/view/angry-make-my-explode-head-mad-gif-14261446

[hard fractal]

Just the base model

[spring fulcrum]

Ahhh I can see it now.... 3 hour movies generated 4 seconds at a time... what a time to be alive!!!!

[high skiff]

haha

[upbeat summit]

it is - the future is now

[timid sonnet]

@upbeat summit

[heady vale]

brb queuing up movies

[west breach]

Thought I'd give A1111 a go at a 2k high res fix. jeebus! 😵

[glad grove]

500$ for 5 seconds of video

[heady vale]

ouch

[spring fulcrum]

Still upscaling.....

[urban fjord]

Is this a new LoRA, checkpoint, something else?

[hard fractal]

Just the base model

[timid sonnet]

@upbeat summit 🔥

[heady vale]

with WAS tiled vae I can do 8k on my 3080 and not OOM. takes 369 seconds to complete

[spring fulcrum]

Can you send me this workflow???

[hard fractal]

These perfect grids.... mm mm mm

[high skiff]

alright, since so many people seem to have issues with OOM right now, I will likely drop an updated 1.1 version soon

[urban fjord]

OOM on what workflow?

[hard fractal]

Great!

[autumn forum]

this looks like my lora bruh

[heady vale]

Ive changed a lot of things, sorry

[urban fjord]

I tried the upscaler one without issue.

[high skiff]

Sorry about that guys, I am not sure why everybody else seems to have issues

it works just fine on my 3080

[hoary saddle]

https://tenor.com/view/youve-got-two-minutes-thats-it-hurry-up-two-minutes-be-quick-gif-14186156

[hard fractal]

Oo, got a link? Would love to try it!

[high skiff]

but clearly I am in the minority here, so I will update with the tiled VAE

also...

Might add in a couple more controls for 1.1

testing right now

[upbeat summit]

one more

[autumn forum]

no lol will probably go up on civit tonight or tomorrow after more tests.

[vital wolf]

@high skiff if i change the image resolution, do i need to change the latent sizes as well?

[hard fractal]

is

...

[high skiff]

I don't, but I am looking into that

[hard fractal]

what

how @astral jay

[civic sigil]

How are these done

[high skiff]

that looks like 1.5

[peak dove]

On Sytan's w/flow I am getting "TypedStorage is depreated. It willbe removed in the future."

[timid sonnet]

gen2

[astral jay]

sdxl

[hard fractal]

no

LoRA?

[urban fjord]

Don't worry about it, not your fault.

[civic sigil]

How is this done

[astral jay]

yep, Lora

[vital wolf]

alright

[upbeat summit]

SDXL, runway gen-2 and latent space magic

[hard fractal]

bro

[peak dove]

Woo hoo - UpScale worked on Sytan - thanx for the NMKD file

[high skiff]

no problem

[vital wolf]

share the nkmd file?

[high skiff]

here, I have it

[urban fjord]

It's a gimmick at the current stage, but having my images animated would be cool.

[high skiff]

there you go

[vital wolf]

thanks

[peak dove]

It'll make post-processing using Topaz GigaPixel a whole lot easier! My final target size is north of 6144 x 6144

[fleet harness]

lol might have a few bugs to work out after all XD

[hard fractal]

Ah you're using N00MKRAD's... I see!

[vital wolf]

did you actaully rename the upscaler? is it the 4x nkmd superscaler 17800g?

[hard fractal]

That's a photographic upscaler I think

[hard fractal]

Might not work as well on line art et-- HOLY CRAP MAN

[high skiff]

yeah, renamed it for my own sanity

[turbid garnet]

SDXL Base 1.0 is coming soon on Monster API

Stay Tuned!

Get access to amazing Gen AI APIs here - http://monsterapi.ai/

[vital wolf]

ahh, lol

[high skiff]

man, this looks so much like 1.5

[astral jay]

lol, its SDXL

[high skiff]

Its... interesting to say

[hard fractal]

I'm tryna slide into your DMs @astral jay

[upbeat summit]

SDXL image

Runway Gen-2

[high skiff]

that is crazy

[peak dove]

What Video package are you sing?

[upbeat summit]

I need more credits - damn it! lol

[upbeat summit]

runway gen-2 with an SDXL input image

[fleet harness]

https://tenor.com/view/pacha-just-right-so-good-yes-emperors-new-groove-gif-6584923

[urban fjord]

Oh, credits for runway, I thought you meant credit for your work.

[peak dove]

Oh I see it - Runaway Gen Two

[hollow halo]

Nice workflow @high skiff just now am dipping my toes into messing with it

[fleet harness]

Yeh, you get a few free creds

then you gotta pay up or wait

so worth it to try for free tho

[high skiff]

glad you like it <3

I have some new ideas to play with

[civic sigil]

Ohh runway lets you do img2video? Thats awesome

[upbeat summit]

haha - no! for the videos I've posted you are playing the seed lottery only. so the credits you need to generate videos go away really fast.

[high skiff]

a slottery

always needs new tokens

[urban fjord]

I've got 87 seconds left, so I can try some stuff.

[upbeat summit]

exactly

[urban fjord]

https://dnznrvs05pmza.cloudfront.net/c0a66e1b-8b84-48d3-a8f3-830a45c35189.mp4?_jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJrZXlIYXNoIjoiOGRiNGU1MzZhNWFmM2VhNyIsImJ1Y2tldCI6InJ1bndheS10YXNrLWFydGlmYWN0cyIsInN0YWdlIjoicHJvZCIsImV4cCI6MTY5MTAyMDgwMH0.wk0NiEB8bTi3mXmjPxyK6C4NMDZZzmj4X7e59x2khx8
Not what I hoped for, do you need a prompt in addition to the image?

[trim orbit]

i wonder if i run the openpose annotator on the coco dataset.

[spring fulcrum]

@hard fractal I feel bad @high skiff his name is no longer orange.... When will he beat Freeza and keep his Super Sytan orange status?

[high skiff]

Oh lord lol

[spring fulcrum]

Find out next time on Dragon Ball Z

[trim orbit]

🎸

[high skiff]

that one is sick

[hoary saddle]

if it used my image, i dont know how

[trim orbit]

riiiiiiiight

[hoary saddle]

came out pure trash, guess you gotta get a lucky seed or something

[upbeat summit]

hey that's you (I suppose) 😄

[peak dove]

Wow. Runway Gen 2 completely changes the look and feel of the input image 😦

[hoary saddle]

yeah, it's pretty bad lol

[vocal rapids]

yes....

[hoary saddle]

noise must be cranked to mars

[peak dove]

<BaRF>

[young mantle]

Heyy @high skiff

Hi there

[upbeat summit]

so okay - full transparency. all I do here is:

• uploading my SDXL images to Runway Gen-2 Image to Video
• to keep your original composition, you don't enter a prompt
• just hit the generate button
• hope for the best
• I curate all gens and pick out the acceptable and good ones
• and... what can not be circumvented for the most part... getting more credits

[gloomy barn]

What is runway ?

[spring fulcrum]

Its an online image to video generator

[hoary saddle]

trying that now 🙂

[gloomy barn]

ah, so useless

[upbeat summit]

a text/image to video service / model: https://research.runwayml.com/gen2

[young mantle]

Text/imagine to video AI

[gloomy barn]

paid, and they look over your shoulder at everything you do

[spring fulcrum]

ya, but its a fun little experiment

[urban fjord]

The consistency seems very poor.

[gloomy barn]

it's an experiment like putting your bank account number online for all to see

[peak dove]

i MIGHT USE kAIBER.AI INSTEAD ... :-d

Oops Caps Lock

[hoary saddle]

well shit, that did work MUCH better

[thin nova]

my aesthetic anime LoRA works. with sytan high res workflow, this is amazing
"anime style full body 1girl running in a park in spring, wearing a blue dress"

[vocal rapids]

runway lol

[spring fulcrum]

I like that one legged walk

[peak dove]

Weird legs - forward, then back ... !

[urban fjord]

https://dnznrvs05pmza.cloudfront.net/2c6f93fd-3ae7-49f7-888b-e54540a8fe32.mp4?_jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJrZXlIYXNoIjoiNTcyNTU5ZDVkOTJiMzJkZSIsImJ1Y2tldCI6InJ1bndheS10YXNrLWFydGlmYWN0cyIsInN0YWdlIjoicHJvZCIsImV4cCI6MTY5MTAyMDgwMH0.XgGtgUdeaPajMnwtZH3OOsvRM7WJ6X9W1i7LHvs-Rhk
Best I got so far, so it is a struggle.

[pallid sail]

That's a cursed muppet 😆

[upbeat summit]

well it's the best that is currently publicly available. it's definitely far from perfect or coherent. but it's an amazing advancement and SDXL is perfect to create high quality init images

[hoary saddle]

for sure

[gloomy barn]

it's awful...cant run locally

[urban fjord]

I got it by doing high-res fix + prompt editing.

It is good for doing cursed stuff.

[autumn forum]

hey @high skiff have you experienced any issues with deformation before? look at the neck. the refined output was fine just needed some touch up but the upscale comepletely boched it and the eyes got smaller too which is weird.

[urban fjord]

Full version.

[trim orbit]

all these people aren't running it locally 🙄

[upbeat summit]

you can still do some experiments with modelscope and zeroscope XL, that you can run locally. you can make some really cool stuff with it - mostly abstract but interesting

[gloomy barn]

yes, but I do want to do the kind of motion capture...turn one video into another.

[peak dove]

I actually got a good one from RG2

[gloomy barn]

barely any motion

[upbeat summit]

yeah that is currently available in online services - proprietary tech. I know :/

[pallid sail]

Looking forward to an OpenPose model but img2img still kinda works for poses

[urban breach]

He's drunk on set again, dammit!

It looks like he's doing one of those DUI tests where the cops are like, "now move your eyes with the flashlight"

[fleet harness]

"not your entire head sir, just your eyes..."

[urban breach]

If you added the audio for that with the cop giving him instructions, and put it on youtube 😂 . Even funnier cause he's in costume.

[urban fjord]

https://dnznrvs05pmza.cloudfront.net/1dd2ee69-e39c-40df-a263-fdb1ce5e714f.mp4?_jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJrZXlIYXNoIjoiZDI2YmJkOTZmMTE3ODVmYSIsImJ1Y2tldCI6InJ1bndheS10YXNrLWFydGlmYWN0cyIsInN0YWdlIjoicHJvZCIsImV4cCI6MTY5MTAyMDgwMH0.nvby-DLYLL5_ruwpCIbqZQ1m1wiEv8Km889djiqX994

[peak dove]

Stumpy left arm 😦

[gloomy barn]

That's a horror

[fleet harness]

I do get a lot of these where they change appearance sadly.

[upbeat summit]

nic is not amused (in this scene)

[autumn forum]

considering how far gen 2 has came. its pretty impressive. imagine 3 or 4!

[peak dove]

Another RG2 generation I like

[fleet harness]

[upbeat summit]

We are all using SDXL images as a base, right? so we are not completely derailing this chat

[fleet harness]

SDXL here so far

[urban fjord]

My dancing Newton was 1.5, but rest SDXL yes.

https://dnznrvs05pmza.cloudfront.net/9deb9ef1-aabd-48cd-b941-4ab1dc08e930.mp4?_jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJrZXlIYXNoIjoiMWYyOWNhMDliMGFiZGIyZSIsImJ1Y2tldCI6InJ1bndheS10YXNrLWFydGlmYWN0cyIsInN0YWdlIjoicHJvZCIsImV4cCI6MTY5MTAyMDgwMH0.8s82xb2vpKMB6G0Ss51Vsh-A_yE-QcN3e3qSocmJDWU
Transformation into more chad.

[autumn forum]

movie still film still close up of a man in a dark room looking down, older man, highly detailed, interior shot
excuse me what

[urban breach]

i am continuing my vampire girls in front of cities project

or cities where buildings blend together like inception

[upbeat summit]

woha... that glowing is nice

[autumn forum]

honestly dont know why that happened. i dont even think you could prompt that and get the same result

[upbeat summit]

yeah he's like super radioactive or going to heaven

[urban fjord]

Cursed seed maybe.

[heady vale]

Scotty is beaming him up

[autumn forum]

seed was 444 lmao

maybe

[upbeat summit]

what about seed 666?

[spring fulcrum]

Goodnight everyone... I'm off to get some sleep finally.

[fleet harness]

thre goes the moon ig

I would be shocked too

[autumn forum]

seed 666 same prompt.

[upbeat summit]

increase the CFG hehe

[autumn forum]

i wanna go back

[upbeat summit]

oh no... quick - change the seed

[urban fjord]

https://dnznrvs05pmza.cloudfront.net/8c307075-b187-4e34-9f1b-4f4cc97ab626.mp4?_jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJrZXlIYXNoIjoiYzQ0ZGNlM2JjZjZhN2I5OCIsImJ1Y2tldCI6InJ1bndheS10YXNrLWFydGlmYWN0cyIsInN0YWdlIjoicHJvZCIsImV4cCI6MTY5MTAyMDgwMH0.vkYWGDLHY6y50OKyN9TfVdpYoeOgQvGDbI-aC4p3jgI
Nice one. Batman from my inpainting attempt.

[hardy cipher]

lol, what is all this?

[upbeat summit]

magic (SDXL images + runway gen-2 video)

[urban breach]

look at that one guy trying to get into the frame at the end lol

[hardy cipher]

are you guys familiar with a way to do this using comfy? https://github.com/pharmapsychotic/clip-interrogator

[high skiff]

That's weird, I have had absolutely no issues with stuff like that

If you are having issues like that, I would strongly recommend raising the starting step of the upscale sampler

[autumn forum]

2008 movie still of batman close up shot, waist up,
but all i added was 2008 lmao

[high skiff]

oh lord lol

[pallid sail]

Oh wow

[ionic dragon]

is it bcuz you were generating celebs?

[pallid sail]

Never a dull moment with this model haha

[fleet harness]

not a lot of movement but cute anyways lol

[peak dove]

Looks like a Tesla Decal 😄

[autumn forum]

no idea. lmao

[upbeat summit]

he's alive now!

[gloomy barn]

people are paying for that?

[fleet harness]

whos paying?

[ionic dragon]

generate few more images of Margot Robbie and check if thats the case

[thin nova]

so easy now. no negative or complex prompt, just train a LoRA

[gloomy barn]

It says it's for pay. There's free with limited credits and functionality, and then there's paid plans

[fleet harness]

thats true but I only have free creds

[peak dove]

On RG2 you get about 25 x 4 second videos for free

[fleet harness]

so someone might be paying but not me

[gloomy barn]

those run out, no?

[fleet harness]

probably

[peak dove]

Then for 625 credits, its $12/month

[fleet harness]

im just playing with it so no worries

people are paying for this tho

[peak dove]

5 credits for 1 second of video

[autumn forum]

margot robbie is fine lol. soo idk prob just a weird combo of things that cuased it

[fleet harness]

this is (((high quality)))

[urban fjord]

https://dnznrvs05pmza.cloudfront.net/7fc5e386-459b-4062-a925-cf0ed5fc8326.mp4?_jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJrZXlIYXNoIjoiOTllMzE4NDc0YjhjOTI5ZiIsImJ1Y2tldCI6InJ1bndheS10YXNrLWFydGlmYWN0cyIsInN0YWdlIjoicHJvZCIsImV4cCI6MTY5MTAyMDgwMH0.ghWByho5oR6UUX-ynuuKlslqxHQ79XrL_p33nkHQMqU
This one worked well.

[fleet harness]

nice!

[peak dove]

RG2

[fleet harness]

[upbeat summit]

so you are making an alternate universe retro alien movie? I'm in!

[peak dove]

Is that Sigourney Weaver? Who's that woman?!?!?!? 😄

[fleet harness]

its our stand in lol here she is talking it out with the Alien instead of being scared like in Alien 3

[upbeat summit]

it's her SDXL stunt double

[urban breach]

I think it's Shooter McGavin

[autumn forum]

wow that talking is impressive looking

[upbeat summit]

reasoning with the xenomorph

[fleet harness]

i think they're friends now

[peak dove]

She's uggglier than themonster 😄

[boreal bough]

Aliens 4 in a nutshell

[peak dove]

Aliens 4 - The Wedding

[high skiff]

oooo

Is there any way you would be willing to share the general prompt for this?

I have been trying to get anatomically muscular robots/cyborgs for a while

[hardy cipher]

musclebot

[hollow halo]

Main Prompt: A vigilant cyborg sentry stands guard stalwartly, integrated scanners continuously surveying for threats, ready to respond with lightning reflexes, CyberBorgz aesthetic.
Supporting Prompt: Security camera POV, scanners surveying threats, ready with lightning reflexes.

I am using a custom LoRA that is currently unreleased but v7 is almost ready to be shared I think, I'm gettign a little to nitpicky with releasing things so far for SDXL lol

[high skiff]

Just my type haha

[high skiff]

Oh man, sick

[autumn forum]

lmao

[upbeat summit]

hahaha

[hard fractal]

Okay so it definitely seems like they have a 2 second model

And a great optical flow inbetweener

[upbeat summit]

ok back to SDXL images

(I lied)

[hard fractal]

Yeah, that's gotta be what it is

[upbeat summit]

the interpolation is pretty good

[hard fractal]

What's the fps on gen2?

[upbeat summit]

24fps interpolated

[hard fractal]

Oh yeah. I'm watching them back at 2x speed

[upbeat summit]

and they give you a 2048x1024 upscaled video

[hard fractal]

Feels normal

[upbeat summit]

bitrate is a nice 8000 - 10,000 kbps

[hard fractal]

Yeah, feels similar to Adobe's optical flow algo

[upbeat summit]

yeah very similar

[hard fractal]

The one that you'd turn on LAST when working after effects

[upbeat summit]

exactly 😄

[hard fractal]

Because otherwise you'd never get anything done.

[trim orbit]

boop - boop

[upbeat summit]

uncanny valley - but amazing to see your SDXL images come to life

[autumn forum]

looks like i need more topgun pics in my dataset lol

[hard fractal]

"Oh... pump... chomp." -SDXL Gen Jr.

[upbeat summit]

yeah I was trying earlier what he might be saying 😉

[hard fractal]

"No foam... gone."

Who took his foam?

[trim orbit]

oh long johnson

oh don ho

[hard fractal]

"Oh bom... bom."

[upbeat summit]

he did

[hard fractal]

Ooo

[upbeat summit]

my SDXL init

[nimble heart]

ik you had the other server guys test but I'm pretty sure we all have 24+ gigs so maybe the selection was a tad biased ☠️

[hard fractal]

What happens with something like that?

[upbeat summit]

I have the exact same image in my input directory

[hard fractal]

Or that

[upbeat summit]

with the lines?

[hard fractal]

Yeah

[nimble heart]

golden ratio

[hard fractal]

Like, it don't got no lines in its dataset, I don't think

[hard fractal]

Close. Rule of thirds.

[upbeat summit]

generating...

[autumn forum]

if you were to guess, if a person was on the left would it be called 1/3 shot. and the right 3/3 shot?

[hard fractal]

I like that!

Gonna start using it

Just usually called "left heavy" composition

or right

[autumn forum]

ah okay.

[upbeat summit]

I didn't upscale because of the image source res

[hollow halo]

Works fine on my 3090 as well

[hard fractal]

Hah!

She's evil now

[high skiff]

oh thats pretty cool haha

[upbeat summit]

very evil

[nimble heart]

i can fix her

[high skiff]

lmao

[hard fractal]

So they have motion graphics in their dataset.

[upbeat summit]

yes

[trim orbit]

beary nice

[upbeat summit]

SDXL source

[hollow halo]

That one is sick

[nimble heart]

damn turned into arch linux

[hard fractal]

Star Trek title card, trending on Marvel

[upbeat summit]

I'm calling Andrew Kramer (Video Copilot)

[boreal bough]

sketch lora done.
sketch, a drawing of a woman sitting on a rock, grand canyon in background

[hard fractal]

Hahah

[upbeat summit]

this looks great!

[ionic dragon]

joe, Is inpainting dependent on controlnet?

[boreal bough]

style LoRA example, for my guide. trained using one word 🤣 the "lazy" method

41mb

[hardy cipher]

I really want to figure out how to use openclip models in comfy. it's quite possible it's easy and i'm just missing something

[upbeat summit]

that is an amazing image!

[hardy cipher]

thanks!

[vale eagle]

Do you have any good idea to reduce the blurry background in realistic image?

[west breach]

try putting bokeh in the negative?

[vale eagle]

Let me try

[civic sigil]

There are things you can do but it feels like it is built into the model tbh...

Probably some good finetunes will make it easier to get rid of

[hardy cipher]

"infinite focus" might help

[boreal bough]

blur is there because some positive prompt word puts it there. things like "cinematic" or "4k" or "8k"

[civic sigil]

I think the problem is rhlf, diffusion models give discontinous backgrounds and so these rhlf models learn to just cheat and blur everything

[urban breach]

yes i've been working on this very topic, one of the experts said that the bokeh thing is baked into a lot of terms, like 4k, cinematic, etc. so i started removing all that stuff and putting in the negatives bokeh, blurry, blur, shallow depth of field

[hardy cipher]

I'll put things like "exhaustively detailed background"

[urban breach]

so example surreal photo of selene vampire from underworld, highly detailed city, hair blowing, rainy, gothic style city, night, moonlight

[hardy cipher]

"blurry background" in the negative with increased weight

[civic sigil]

Decent but you can tell the subject is still somewhat "seperate" from the background, like its greenscreened in or something

[boreal bough]

lol. but it might be worth making a dumb LoRA, that only retrains all the most commonly used bokeh words, and retrain them to no longer be blurry

sure its medication for the symptoms... but I can see that winning the civitai challenge 🤣

[autumn forum]

well this doesnt look very safe...

[gilded blade]

is it possible to use outpainting?

[hardy cipher]

I'd imagine so

[civic sigil]

Looks like a real photo

[true wing]

https://tenor.com/view/kermit-the-frog-looking-for-directions-navigate-is-lost-gif-11835765

[true wing]

I feel like I'm just getting my local install of SDXL to behave how I want but now I see all the videos in the thread and I feel so far behind

[static temple]

the sdxl i dont thinks all that good as its hipped up to be

[midnight shuttle]

Why not? You have to compare it with other base models, not finetuned models.

[crisp owl]

It's a base model
COmpare to base model 1.5 and then come back

[static temple]

i tried both

[true wing]

SDXL is amazing compared to 1.5 and should only get better with refines/training. Look at base 1.5 compared to the refines, it's garbage

[west breach]

The videos aren't an SDXL thing. People are putting their images into runway gen-2

[true wing]

My problem is getting it to turn out nightmares, which I don't think is most people's problem

[true wing]

Oh well that feels like cheating

[nimble heart]

sliders 👀

[static temple]

i rather have fun then messing with some thing thats very unpredictable

[crisp owl]

add a random prompt generator to the sdxl process and just click away

[true wing]

That's funny because that's the opposite of the problem I was running into which was that it was too predictable, it really didn't want to mix a cat with a spider. But here we are.

[uncut steeple]

reminds me of what I did yesterday

[true wing]

Nice

[static temple]

you used ComfyUI_00139_.png?width=616&height=616 i thought for best results and more predictable is use 1024 x 1024

[true wing]

Huh? I'm using a modified Sytan's with a 1024x1024

yeah check the resolutons

[static temple]

anyone want a FREE good upscaler that can batch files of a lot of images at once, this is nice https://www.upscayl.org/

[ionic dragon]

then why not add another refiner at the end to make everything crispier

[high skiff]

it does the opposite

[nimble heart]

oh shit there's toggles now

[high skiff]

oh?

was there a new comfy update?

[nimble heart]

https://github.com/comfyanonymous/ComfyUI/commit/730a5d170ffdfb30bd65dc4b765cc28501d85a5f
added sliders and toggles

[high skiff]

fuck yeaahhh

[nimble heart]

sliders need more time in the oven IMO they just slide you can't double-click to edit

but toggles hell yea

[true wing]

Thanks for the heads up. I updated yesterday but it looks like I need to update again

[nimble heart]

None of the actual nodes use it yet

the node in my screenshot is one of my custom ones

[true wing]

ah, gotcha

[nimble heart]

nooooo my repo is already out of date

commit war lets go

[ionic dragon]

no i mean, to make the image better, i tried adding the refiner, but the image was not anywhere close to the one generated in the 1st showcase, what might have gone wrong?

[static temple]

thats junk lol

[nimble heart]

generations only have knowledge of themselves so the node wouldn't be able to sleep after 500

you'd either sleep every gen or never

Real question is why is your CPU overheating

[static temple]

it maybe done but some one would need to write the code

does sdxl have a list of code commands

[nimble heart]

I've written custom nodes and afaik that's not possible without changes to comfy

[true wing]

out of curiosity did they ever reveal the bot's final settings?

[thick harness]

did an horizontal flip and changed balance of black tones using photoshop, small interventions, the cool stuff came out of SDXL v1.0 as you see it

[high skiff]

I think thats too NSFW

[thick harness]

yeah, was gonna ask, sorry

[nimble heart]

its the middle of the night the mods are asleep

[true wing]

Looks artistic to me, but I guess it's up to the mods

[charred beacon]

Tasteful, but nsfw

[vale eagle]

report him!!!

[thick harness]

ok, gonna delete it

[true wing]

Man some people would ban basically all greek art huh?

[west breach]

prompt for a merman and you're all good 😄

[static temple]

can he put spoiler on it, im not sure how to

[high skiff]

I think it looks dope, I just wanted to warn cause I don't think the mods care about if its cool or not 😅

[thick harness]

its a lamia, asturian mythology

[thick harness]

lol save it , i will delete it

[nimble heart]

I was @ a family Italian restaurant yesterday with photos of naked people all over the walls lol. Surprised tasteful boobies aren't ok

[peak dove]

Positive Prompt - Merman. Negative Prompt - Moobs!

[west breach]

moobs 😄

[thick harness]

its merman german or mermaid?

[nimble heart]

mermaid but man

[static temple]

i herd sdxl does not like a lot of words specialty in the negative pompt

[nimble heart]

ig it'd be hard to police.
"yes that rennasaince painting is OK but no your asian woman with stuffed up milkies isn't I know she's wearing bottoms but you know what you're doing please stop"

[thick harness]

i have a big encyclopedia of all the prompts ive used since stable diffusion began, and the prompt for that was one for Pixelzai

then used the 'analog foto' style

[static temple]

I like how some people write prompts as example I want a picture of a brown cat in a tree and someone will write I have a beautiful cat it has long hair and it runs through the woods and when it finds a tree I want to see that cat in the tree, then put high definition alter fine supersonic 8K, super charger render, and all they had to say i want to see a brown cat in a tree

[true wing]

In fairness sometimes writing a short story seems to work a little better. I think it's because of the repetition and because the new language model is better at understanding more natural language.

[nimble heart]

that worked well enough. Prompt was

I like how some people write prompts as example I want a picture of a brown cat in a tree and someone will write I have a beautiful cat it has long hair and it runs through the woods and when it finds a tree I want to see that cat in the tree, then put  high definition alter fine supersonic 8K, super charger render, and all they had to say i want to see a brown cat in a tree

[trim orbit]

prompting is sometimes just latent space drilling. just going for it

[true wing]

Yep, just keep going until you strike oil

[nimble heart]

oh hey it's Lyla

[true wing]

For some reason it really likes an orange cat for that prompt.

[static temple]

out of focus on a lot of sdxl i seen so far

promp this > a maine coon cat in a tree

[nimble heart]

nah.

[static temple]

do you know what a maine coon cat looks like?

[nimble heart]

that's too short the attention will be mostly on empty tokens which will make the image worse

or something like that idk technical stuff

[boreal bough]

get your pitchforks! put some hay on em, AND PUT THEM ON FIRE 😄
burn him, burn him, burn him

[nimble heart]

see if he weighs as much as a duck

[static temple]

maine coon cat

[boreal bough]

do I use a gigantic scale, or just throw them into the water to skip to the end?

[vocal rapids]

I got a mainecoon cat as well

[static temple]

got pointy ears yup

[nimble heart]

scale https://youtu.be/yp_l5ntikaU?t=40

[boreal bough]

it never gets old XD

[high skiff]

currently trying to find the command to install models in oobabooga ._.

not sure if they got rid of it, my old non working install has it

[west breach]

can't you do it in the ui?

[nimble heart]

doesn't it have a huggingface url place you can paste links?

[high skiff]

thats what I am looking for

it used to be download_model.py

but its gone now

[nimble heart]

check the commit history see if it was replaced with something

[high skiff]

found it

wow

not sure why they decided to change it from a bat you run to a .py thats hidden in the GUI scripts now

[nimble heart]

if it's in GUI scripts maybe you're supposed to launch it from the GUI...

[high skiff]

the github says you have to run the script in CMD

It used to be a single click bat, not sure why they changed it

[nimble heart]

maybe it wasn't updated for the gui script

[high skiff]

oh well, foun dit now

[nimble heart]

check the gui see if there's a model download somewhere

I'd say 70% chance there is now

[true wing]

Can't you just drop the models in the folder anyways? Or does the script do something else?

[high skiff]

I did, I didn't see anything

oh well, I foun dthe script

[nimble heart]

oof

well least you can download Erebus now 👍

[high skiff]

you can, but hugging face makes you download every file individually, not the whole folder structure lol

[nimble heart]

LLMs aren't usually just one fat safetensors

[true wing]

I usually just grab the files I want directly but you can do a git pull on hugging face right?

[hollow halo]

theres a model download in the model tab on the right side of the UI

[nimble heart]

not without git lfs

[hollow halo]

The box on the right that says enter hf repo name and then hit download

[nimble heart]

never used the oogabooga ui but I still believe every word that man just said because it's exactly what I wanted to hear

[high skiff]

💀

I am blind ok

lmao

[hollow halo]

It's a little hard to see, especially if you are used to the old way lol

[uncut steeple]

btw nice quick workflow @high skiff , interesting attempt at upscaling too

[cunning geyser]

Is there any good anime sdxl models rn

Or are they still in teh works

[uncut steeple]

I think there is a waifu diffusion version already

[cunning geyser]

Yeah but is it any good is the question

Came out way too fast to rely on that.

Thought something dropped real recently

[uncut steeple]

No clue
Try it

[cunning geyser]

I did

I didn't get anything good, but someone else might be better

Idk

[uncut steeple]

Prompting issue id say

[cunning geyser]

Ok then what did you make

[uncut steeple]

I didnt try any XL anime models yet

[cunning geyser]

Well when this shitty photo is their preview image

I can't expect anything from it

[true wing]

Did you try base SDXL? You don't need prompts for a lot of stuff

[cunning geyser]

This shits so low res and so badly painted

I'd think a beginner artist made it

[uncut steeple]

Im using base+refiner

[cunning geyser]

Use comfyui for base + refiner

But it can do certain anime styles

[uncut steeple]

I do lol

[cunning geyser]

That is for sure

Just not the styles you would aim to have from the 1.5 stuff

Aight well SDXL is a bummer for now still then, got it

[uncut steeple]

[high skiff]

lmao

can you be any more melodramatic? lol

[uncut steeple]

https://tenor.com/view/skillissue-skill-issue-gif-22125481

[alpine pine]

I believe in 'em

[high skiff]

lol

[nimble heart]

have you considered prompted better

[cunning geyser]

Found one.

Also yeah I have

[uncut steeple]

People post a ton of good stuff here and all you say is SDXL is a bummer

[cunning geyser]

I guess yall must have skill issues since I stole half the prompts here.

[west breach]

is the stuff they post in the anime channel really anime? I don't remember anime being that detailed, the budget would be astronomical

[cunning geyser]

^

Photorealistic and anime are very very hard to differ

I get what you mean tho

The photorealistic anime is very near photorealistic, but too artsy to be anime and too detailed, but yes, its still anime

[nimble heart]

steal SAI's official anime style prompt see how that works

[civic sigil]

Very nice is this with a lora or just base model?

[cunning geyser]

Moreso it's like 3d anime as they would put it sometimes, depends

[high skiff]

no, thats more manga style than anime

[uncut steeple]

Just base+refiner, didnt try any loras yet

[civic sigil]

Your gens are awesome

[uncut steeple]

Youre too kind

[high skiff]

one thing SDXL can do way better than 1.5 right now is for sure realism haha

can't wait to see how great that gets with finetunes

[cunning geyser]

Yeah

[uncut steeple]

same for sure

[cunning geyser]

It's 15/10 on realistic.

[woeful patio]

does anyone here have a workflow for using loras in comfy?

[cunning geyser]

But the anime getting worked on from creators will be great once we hit 2 more weeks.

[alpine pine]

sytan's workflow works pretty well with loras, just disable the refiner

[nimble heart]

im tryna proper A/B compare sending different sizes to the clip encoders and I'm beginning to wonder if it even matters...

[hollow hare]

hi guys, does SDXL already support merging LORAs?

[cunning geyser]

I'd say it depends on the loras.

Also with the inexperience of creators rn, heavily doubting it.

Cause most loras are fucking huge and 400mb+

[ionic dragon]

i didnt get the best outputs with humans

[cunning geyser]

And.. they're meant to be like 49-100

[cunning geyser]

skill issue

[high skiff]

it takes more prompting to get realistic humans

[nimble heart]

pics or it didnt happen

[uncut steeple]

I only did the super wrinkly stuff so far

[wise edge]

I've gotten some very noticeable differences, like a completely different race for a person with 2k and 4k clip width/height for otherwise the same settings. And square conditioning shape for portrait pics tend to get elongated results.

[high skiff]

oh wait, that converter I used made him all crusty

[nimble heart]

I did 2x 16 batches and maybe like 3 of the images have any worthwhile changes

im actually leaning towards not setting the clip sizes higher

[boreal bough]

sketch, a drawing of a woman getting a coffee at starbucks

[high skiff]

I have been messing with that as well, and have seen very little meaningful different

[wise edge]

I don't know if the changes are worthwhile either. Just different sometimes.

[boreal bough]

achieved a 80% hitrate now 😄 41mb lora

[high skiff]

only time I see a difference is when the size is lower than the output res

[boreal bough]

strength 1, because equilibrium

[nimble heart]

yea if you have a 1920x1080 latent you can sometimes actually get OK non-duped pictures by setting the clip sizes low

[ionic dragon]

any recommended keywords

[nimble heart]

huge boobies

[high skiff]

aur naur

[urban breach]

i'm confused, are you guys using this "fix" ```
md5sum cf9d29192ef7433096a69fe5e32efba1 sd_xl_base_1.0.safetensors
md5sum 86b0f3400706c97ecb8efbed8146b373 sdXL_v10VAEFix.safetensors

md5sum c0df90f318abf30fcd2cd233ea638251 sd_xl_refiner_1.0.safetensors
md5sum 74d31c96097471d5d3bd93a915ad0b30 sdXL_v10RefinerVAEFix.safetensors```

[boreal bough]

I'm tempted

[nimble heart]

stuffed up milkies

[upbeat summit]

you can get the models from the official Stability AI HF page:
https://huggingface.co/stabilityai

direct downloads of base + refiner with the 0.9 VAE baked in:
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0_0.9vae.safetensors
https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/resolve/main/sd_xl_refiner_1.0_0.9vae.safetensors

[ionic dragon]

should I try?

[nimble heart]

yes

I will add it to my current prompt if you do ☠️

[ionic dragon]

what if get these

[nimble heart]

yes, tovarich

[high skiff]

oh wow...

Oobabooga runs very very good with pooled memory

[high skiff]

currently running a 13b model on 10 GB VRAM, with 11.6GB VRAM used, and still getting 3 tokens/s gen times

[ionic dragon]

should I set the weight to 1?

[high skiff]

there is like a 3-4 second pause before responses, but its fairly fast, I am impressed

[pallid sail]

Checking how good SDXL is at making tiling textures. Diffusers with a patched Conv2d module with padding_mode set to circular kinda works but I've noticed the seams are kinda noticeable. I went back and checked SD1.5 and the seams seem to be there as well so I think something happened with Diffusers and Torch 2.0 to not make it as reliable

[nimble heart]

yea llms are decent @ it. KoboldAI uses breakmodel to split betwene gpu/cpu and it still runs fine

[high skiff]

yeah, wow, this model with pooled system RAM is not bad at all

[ionic dragon]

whats the defualt weight of keyword in comfy?

[boreal bough]

If I ever make a tutorial on how to do clip training on sdxl, I swear to use this as the example LoRA

[upbeat summit]

very clean look. the lines are really dynamic and have a great flow. I like that style. great work!

[ionic dragon]

lol

[true wing]

I haven't gotten around to trying it but someone was saying Local Llama does the split pool thing better. Too much AI stuff going on to keep track of it all

[ionic dragon]

why dont we release a boobie model?

[urban breach]

Is this the version we're supposed to be using for example with the Sytan workflow. The one that is "baked in" here lol

[high skiff]

holy shit