✨｜sdxl

[cyan crown]

he has a Mike Helmet 😂

[native knot]

Put my neg into the positive prompt just to see and got this as one of the outcomes. Not terrible on its own. 🤣

[native knot]

https://giphy.com/gifs/morgan-freeman-bruce-almighty-J4JN1fW9HsOHu

[glad grove]

brp tryna gift me nitro 💀

[native knot]

BRO

[crisp owl]

I've always wanted to click those and lose access to my discord account!

[native knot]

Super Saiyan Mickeypool?

[trim vigil]

yea 😄

[ionic gulch]

workflow 10.0 is online - finally i could fix the performance problems by going back to regular reroutes instead of multipipe nodes. so far wires also didn't get lost when saving after using my own 5in1 switches instead of the any-switches - so i hope this works around those two problems that stopped me from releasing the workflow...
https://github.com/JPS-GER/JPS-ComfyUI-Workflows
features:

• txt2img
• img2img
• control net canny
• control net depth
• inpainting (much better results and less artifacts than the example workflow provided with the model)
• 2x ip adapter
• 2x revision
• prompt styler (includes sai presets) wip
• prompt handling (use clip_g + clip_l, only one, autofill, etc. - many options how to handle the two positive prompts)
• better seed handling with option to recycle the previous seed
• 2x upscaler, can be disabled by cutting two wires in the menu section
• 4 versions per generation (100% base, base+refiner + upscaled version) - can be reduced by cutting wires in the menu section
• choose sdxl recommended resolutions through dropdown, imported images get cropped and resized to the nearest fitting sdxl resolution
• 4x lora
• use vae from model or separate vae file
• menu section fitting a single screen (if your screen is big enough)
• same file name for all four versions of the image by including current date and time, instead of just a counter, option to add your own keyword, that will be used for filename and folder
• full workflow with 4 version including upscales takes 60-90 seconds on my rtx4090 depending on the selected options and generation mode
• reduced 3rd party nodes by 50% compared to the last version - so it's less likely to break and easier to setup - comfyui manager is recommended to download the required nodes

https://i.imgur.com/csgvjja.jpg

[dim steeple]

Will updating my version of Torch from 1.2 to 2 make SDXL faster?

using auto1111

[vital ermine]

didn't for me

[steady grove]

you were able to get sdxl working with torch 1.2?

[west merlin]

Have a question - how the heck do I get this model to work with clips in ComfyUI. Just started messing around with SDXL today, and I've no clue what to do with this file. Doesn't work when I just put all the data into the clips folder:

https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k/blob/main/open_clip_pytorch_model.bin

[vital ermine]

https://civitai.com/models/144011?modelVersionId=159938

[dim steeple]

How come when I set img2img denoising strength to 1.0 it still invokes some of the img?

I thought 1.0 meant it completly ignores the image

[hardy cipher]

it's interesting how patterns work on with the ipadapter

those aren't really best efforts on my part. 22 steps through the regular ksampler with no refiner or anything

just trying out different parameter combinations

[ionic gulch]

i'm trying to mix in movie names and directors. the last one was 2x ip adapter + blade runner

[hardy cipher]

these are ipadapter + clipvision + input image for latent

pretty straight forward and it works well sometimes

[ionic gulch]

the same one with "black swan":

[ionic gulch]

this is my planed next addition to my workflow / styler node - a list of 100-200 movies to pick

[hardy cipher]

nice. I was thinking of styler nodes, but hadn't thought of something like that. have some other ideas though

[ionic gulch]

avatar:

this should make it easy to prompt for great images, if you can pick 100-200 movies known for their specific style.

star trek:

pan's labyrinth:

and if you reduce ip adapter weight (still pan's labyrinth):

star trek:

black swan:

mad max:

[hoary saddle]

anyone using wildcard style prompts with success?
using something like this simple example
{male|female}
but getting almost all males, maybe 10:1 ratio, am i missing something?

[heady vale]

probably need the hard brackets. try with [red|blue|yellow]

[dim steeple]

How come when I set img2img denoising strength to 1.0 it still invokes some of the img?

using controlnet depth

[heady vale]

is the prompt different to the image?

with a depth map and same prompt it likely generates the same thing

[crisp owl]

Because if it's from someone's workflow, it's likely just a rough calculation based off the start steps

[dim steeple]

is that a comfyui thing?

I am using auto

[crisp owl]

Yes and no. It's ultimately how it's done regardless

[dim steeple]

I am trying to get different facial expressions, prompt expressive while the reference image is neutral

And the depth is for the head shape

but for some reason the output image seems to be invoking the ref image face

[heady vale]

depth can capture more than just the basic head shape, even if you cant see it

[dim steeple]

ah okay

i didn't know that

[heady vale]

lower the strength of the depth to make more changes

[dim steeple]

this is the depth

You're saying that it has more data than shown?

[heady vale]

as mentioned its hard to see with the eye what is on there

[dim steeple]

Is there no setting to make it more basic?

[heady vale]

lower strength

[crisp owl]

Different depth models

[dim steeple]

does preprocessor resolution do that?

[uncut fiber]

I think it had some issue with randomization. Your brackets are o.k. Would try {man|woman|woman} probably. @hoary saddle

[spring fulcrum]

Does anyone have that workflow for combining two images or more? I think some were getting a bit out there with combining like 4 images or something like that??? It was like image to image on steroids

[hoary saddle]

will give it a shot with diff brackets and then doubling the woman token

[crisp owl]

I've been playing around with ideas on doing that, tried using the image blend option, but unless you've got two pictures that are perfect for each other, my results were meh.
Talking direct image2image flow

[uncut fiber]

@hoary saddle just tested with old brackets red yellow and blue and result is 3 2 4.
I think if SD has not issue with tell this is man and this woman 😄

with [] you get balls with multiple colors

[uncut fiber]

@hoary saddle tryed this {(Man:0.8)|(Woman:1.5)} portrait. and it is working, just some ballance is needed. This gave me 2:4 man vs woman.

[hoary saddle]

will give that a shot, other 2 suggestions still getting a pretty heavy lean on first option

[uncut fiber]

{(Man)|(Woman:1.45)} portrait.
this i got 5:7 seems ballanced

[hoary saddle]

currently doing it via python with fstrings which definately works as it's just injecting a random item from different categories, but sure would be nice to get same type of results from pure prompt
#-------------------------------------- dyncat = { 'head': ['glowing demon eyes', 'sunglasses', 'vrheadset', 'tattoos', 'baseball cap', 'military helmet', 'gasmask', 'spike mohawk', 'bucket hat', 'huge beard', 'long goatee', 'beautiful flowing hair'], 'clothes': ['spacesuit', 'torn tshirt', 'bloody scars', 'puff jacket', 'hawaiian shirt', 'biker jacket', 'business suit', 'gold chains', 'sexy lingerie'], 'background': ['tropical island', 'foggy mountains', 'city', 'lush jungle', 'desert', 'alien planet'] } selected_words = { category: random.choice(words) for category, words in dyncat.items() } dyncathead = selected_words['head'] dyncatclothes = selected_words['clothes'] dyncatbg = selected_words['background']

was using man/woman to see how it fairs as an example before i get deeper into testing

[uncut fiber]

i now get 6:6 man:woman from prompt. () and (:1.45) weights.

[uncut fiber]

also i think more things you have in bracket result will be better.

[peak dove]

SDXL Bot

[floral island]

hmm... shattered minds, very cool ^^

[floral island]

new shell advertisment?

[crisp owl]

In a world ravaged by the unquenchable thirst for fossil fuels, Ethan was a solitary explorer, wandering through the ruins of what was once a thriving civilization. The air was thick with the acrid scent of decay, and the landscape bore the scars of humanity's relentless pursuit of energy. The pursuit that had led to this dystopian nightmare.
One day, while searching the crumbling remains of a massive skyscraper, Ethan stumbled upon something that sent shivers down his spine—a glowing relic hidden among the rubble. It was the unmistakable logo of the Shell corporation, one of the largest and most infamous fuel conglomerates in the world.
Shell had been at the forefront of the race to extract every last drop of fossil fuel from the Earth, heedless of the environmental consequences. Their greed had contributed to the planet's demise. Now, their emblem, bathed in an eerie, ethereal light, stood as a haunting testament to their legacy.
As Ethan gazed at the emblem, a rush of questions flooded his mind. What did it mean to find this symbol in the midst of the desolation? Had Shell Corporation met its own demise, crushed under the weight of its insatiable ambitions? Or had it, like humanity, adapted and evolved, perhaps moving on to exploit other worlds?
Ethan carefully pocketed the emblem, feeling its warmth against his skin. He knew he couldn't ignore the significance of this discovery. It was a cryptic message from a past consumed by greed. What it meant for the uncertain future, he couldn't say, but he was determined to find out.

With the Shell emblem in hand, he resumed his journey into the unknown, traversing the desolate landscape in search of answers. The world may have been scarred by a history of avarice, but as long as explorers like Ethan sought truth and redemption, there remained a glimmer of hope that the planet might heal and find a new path—a path not driven by the greed for fossil fuels but by the promise of a more sustainable and harmonious existence.

[vocal stream]

are people using the refiner much by now?

[west breach]

I don't use it anymore

[hardy cipher]

I do on occassion

[crisp owl]

I use it for both a precon node and after base node as a "refiner" in a sense.
I use different checkpoints though

[hardy cipher]

but only if I look at something and think, "hey, this should be more refined"

that's about 7th generation ipadapter

ipception

[supple knot]

How do you make it so your reroute pipe has a title?

[west breach]

is that the juggernaut lady?

[supple knot]

Whats in the box?

[uncut steeple]

Yes and no

[supple knot]

Look at it closer its Gwenth 😋

[hardy cipher]

so it is

[supple knot]

https://www.youtube.com/watch?v=07qOm2zp5xk

[heady vale]

If you look closer she lost her freckles

[west breach]

matte gloss coat

[supple knot]

https://www.youtube.com/watch?v=fISNGc5s290

https://www.youtube.com/watch?v=dR4BSb63xxA

[spring fulcrum]

There is another workflow that I have seen someone mention... where you can combine more than one picture together... Its where you use 2 images and its like img2img but it makes a hybrid image of the two images you input and you can weigh them and get cool results. Its different than just img2img

[grizzled dune]

Hello. Does soemeone use --controlnet-dir in their commandline_args ? It doesn't work for me in last version.

[hardy cipher]

ipadapter or clipvision?

[uncut steeple]

Why are yall so obsessed with the Girl in the box

Also why cant I un-bypass multiple nodes in Comfy

I dont want to click everything individually

[heady vale]

hehe I was bored and its a good pic

[noble shoal]

Guys....i just trained a lora on 350 medical photos. I defiantly need some eyebleach now, and maybe a therapy. But in the meanwhile i will enjoy this delicious "Burger"

[vernal spire]

im so glad i tried comfyui, a1111 is no fun with low vram and comfyui is much faster in general

[uncut steeple]

I bypassed a note node within a group and now I can not unbypass it

[heady vale]

what if you muted it, does that overwrite the bypass?

[uncut steeple]

Oh yeah it does, big brain time

[uncut fiber]

Symphony of nature 😄

[noble shoal]

I with i could show you the capabilities of my new gross creation, but i think that would break some rules here 😅 . I will just leave this here instead.

[uncut fiber]

he has 5 fingers!

[noble shoal]

Yeah, there were a few hands in the dataset.... don't even know if they had the right amount of fingers.

[floral island]

is this what they mean, a heart of gold?

[uncut fiber]

Woman bicycle.

[floral island]

let me hear you scream AAAAAAAAA

[willow bane]

Hlo can anyone tell me where I can get the prompts

I am using sdxl for 2 weeks but didn't know how to get prompts

Is there any software or site

[floral island]

click image -> open in browser -> save locally -> import in a1111 / comfyui

[willow bane]

I make deform videos , I need prompts for that

[vagrant reef]

222

[uncut fiber]

what UI @willow bane are you using?
In A1111 you get it in PNG info and in comfyUI draging picture in workflow, it will overwrite to workflow and texts it was created. I think clicking on image is enough to save it, not needed open in browser = i am sure 🙂 @floral island

[floral island]

it used to be, perhaps it changed 🙂

[uncut fiber]

he is missing something 😄

[uncut steeple]

Aw man this relies on chat gpt API and I dont see myself paying more than the 20 bucks per month I am paying already

[tribal moat]

Hey guys, which image resolutions to use while generating images using SDXL? Common & default one is 1024x1024

[native knot]

Just so you know, running about 500 prompts through it cost about $0.30. it's not expensive.

[uncut steeple]

Yeah true, but I still prefer being independent, maybe i can run smth locally

[vocal stream]

I hope stuff like resolutions get standardized and added as metadata to the model one day and UIs can just read them and show them as options

[uncut steeple]

For comfy there is an aspect ratio node with comfyroll

[uncut steeple]

Randomizer seems to be working alright

[crisp owl]

Very close to a bird

[uncut steeple]

[peak dove]

Dynamic Prompts in A1111 is also a Randomizer?

SDXL Bot

[uncut steeple]

Yes kind of, but not really since it's only picking a random prompt

[zinc cargo]

what's that file?

[uncut steeple]

Probably an image

[floral island]

you just need multiple degrees of entropy, so that the random can also get extra variations of random 😄

might i suggest my nouns wildcard file 😛 nearly 7000 words, just to add random flavour ❤️

castleception, spoiler: ketchup warning

[noble shoal]

That's Rose juice

[floral island]

aight. you got it chief. rose juice ❤️

i seriously love that prompt with my randomizer

[floral island]

santa is so effing done

"little shits"

[indigo carbon]

which lava cube?

[floral island]

spicy puzzle cube!

[crisp owl]

top right

[hardy cipher]

2

[floral island]

2

[floral island]

rose juice... lots of rose juice

[floral island]

pure horror.

[hardy cipher]

messing with ipadapter and clipvision. other than some pre-clipvision conditioning tweaks these are all identical seed number and parameters.

2nd and 5th are pretty similar. normalized 1x vs normalized 0.5x

first image is without any conditioninmods, then normalized and multiplied by 1, 3, 2, 0.5

with pooled output

3x pooled output vs normalized 3x pooled output

[floral island]

weird turtle

[uncut steeple]

This looks familiar

[floral island]

"yeah, tell me about it"

[hardy cipher]

sweet long pick cropping, discord

[hardy cipher]

anyway, just find these sorts of things interesting

[hardy cipher]

her leg

[stone fossil]

Shes just getting up.

[hoary saddle]

anyone know if the person that created the WAS NSuite is in this chat?

trying to figure out how to increment with 1.jpg 2.jpg 3.jpg
instead of 1.jpg 10.jpg 100.jpg 2.jpg 20.jpg 200.jpg

could batch rename input images first and add leading zero's but hoping there's an easier way to do it from in comfy

[vital ermine]

Hmmm, seems to be trying to say something.

[noble shoal]

Jeez, no matter what he is trying to tell you, he should calm down with the use of pronouns.

[leaden fable]

Hey is there some inpaint models for SDXL?

[eternal fog]

There is 1 that diffusers made

[upbeat summit]

nice hybrid!

[noble shoal]

Nice desert Turtraffe

[cyan crown]

I have one turtaraffe here

[noble shoal]

He looks like he is totally convinced that he found the place from the picture, but boy is he wrong.

[eternal fog]

lol

[upbeat summit]

dystopian battle fairies? nice

[floral island]

"it's an older prompt, but it checks out" 🧚‍♀️

[floral island]

she's so fucking done fighting this shit

[eternal fog]

Not sure whats going on here lol

[crisp owl]

Heh

[leaden fable]

Can you download it and use it like any other models?

[eternal fog]

With Comfyui yes, not sure about Auto1111

Some details on this blog - https://comfyanonymous.github.io/ComfyUI_Blog/comfyui/update/2023/09/02/Weekly-update.html

[leaden fable]

Thanks :D

[indigo carbon]

space burger

[lusty wolf]

Congealed twins..

[lusty wolf]

is the knitwear a lora? or just prompting?

[hardy cipher]

not sure why it went on a golf ball tangent

[crisp owl]

Still sounds similar to what I've been playing around with. It does have potential, but I don't feel I've really "locked" it down yet.

[indigo carbon]

actually made something like that lol

also made a multimodel version where it's an image input

[shy kelp]

anyone knows how to do eye masks in comfyui?

[indigo carbon]

it like gets an image and the prompting is like [IMAGE] holding a.../doing..../as.../etc

[cyan crown]

I have a strange question. Is there any way to create images that have the same impact of MJ ones without using Loras ? I've tried any kind of prompts, but images comes out good but haven't impact. With Lora everything change. Any hint? See these 2 images

[indigo carbon]

specify what do you mean by "impact"

[cyan crown]

see 2 images

[crisp owl]

Seeing some of the images you've been creating Andreac, I'd say you do a fine job at creating impactful images

[cyan crown]

thanks, but lot of them uses Loras

whenm it comes to painting

they have less impact

1st image is from MJ and 2nd from SDXL

only way to achieve that "style" is using a Lora ora an Ipadapter

[crisp owl]

That comes to prompting, or finding an artists styling
Or utilizing other methods tha thave been released like ClipVision stuff, etc.

[crisp owl]

can prompt for an artists style etc

[cyan crown]

Artist is Alena Aenami in the sample

for both

[indigo carbon]

first off, SDXL is automatically better in many ways IMO. just the parameter count itself beats MJ, also the fact it's open source..

[cyan crown]

I'm not making a comparison

I prefer SDXL

[crisp owl]

Just two, super quickly prompted, zero lora's, single ksampler 20 step

[cyan crown]

maybe they apply a shap mask and some contrast trick after generation

[crisp owl]

They for sure do stuff behind the scenes

[indigo carbon]

idk what y'all want, this is way better imo

[cyan crown]

I can reach that kind of images with PP ora lora, but with prompt, no matter what i write (for painting) they comes out less sharper and with less "wow" effect

[indigo carbon]

this is without a lora or anything, I just wrote like 5 words and it made this

[crisp owl]

Yeah this was the prompt I used, nothing special. Could for sure be refined
Alena Aenami art style of a man sitting on a lake shore at sunset, heavy brushstrokes and vivid contrast

I tried MJ in the past and was nothing but unimpressed on the whole.

[indigo carbon]

yeah, the only way I can see MJ being superior to SDXL is maybe accessibility, but when it comes to the quality it's no match

[vital ermine]

closer now but the problem with MJ is that it has the MJ look/style. That is bad imo

[indigo carbon]

comparing MJ to SDXL quality wise is like comparing an inflatable boat to the damn titanic; just the sheer size of SDXL is probably a similar ratio

[cyan crown]

I prefer SDXL

btw I would like to obtain some kind of images without training a Lora

[indigo carbon]

didn't need to use LoRAs or anything here

[cyan crown]

color are different, but blacks are very good

also style is different

[indigo carbon]

I didn't intend of replicating that style, just made an "impactful" image

[cyan crown]

yes it is, but it works only with certain styles

maybe I can0t explain well

[crisp owl]

Well, styles only exist as long as there was data trained off that specific style

[indigo carbon]

nah, SDXL doesn't struggle with styles, you can even create your own style templates without even touching LoRAs

[vital ermine]

I couldn't find info out about that

Is there one for that?

no idea how to even use it

[indigo carbon]

this is how my workflow looks like, the next time I'll update in on CivitAI I'll include that feature, also after I figure out AIT on IPA I'll release the multimodel workflow I finished working on

[vital ermine]

Welp I was with you up to the point of updating on civitai then, ngl, I zoned out.

my eyes glazed over

[crisp owl]

CNET, CNET+IPA, CNET+REVISION, CNET+IPA+REVISION

[cyan crown]

In one phrase. Images with lora comes out with deeper blacks and somehow sharper

[crisp owl]

Could be because they are usually more specifically trained on whatever subject you are attempting to create. So those weights in the model are going to be much more heavily focused on that specific thing

[vital ermine]

When I used to do ti embeddings I saw LoRAs and switched due to that very thing. deep blacks and more vibrant colours for the exact same trainings.

TI is more general so solesbeedude is probably right and that is why TI embeddings are lacking the vibrancy LoRAs have

iow more specialized is better

[indigo carbon]

one disadvantage with LoRAs is they are not general use like normal checkpoints, so imo if there would be like a general purpose LoRA it would be widely used

[crisp owl]

To my understanding, a Lora is basically a checkpoint modifier, which is why they can be merged into a checkpoint. So when added, it's directly affecting the weights of what it was trained on.

[indigo carbon]

I remember there being a LoRA called "offset noise", I think that has that potential

[vital ermine]

checkpoints are highly specialized only holds more subjects. At least with a lora, or embedding, I can use it on most checkpoints but checkpoints are what you see is all you get with just it

I still have that

one

[crisp owl]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_offset_example-lora_1.0.safetensors

[vital ermine]

I am about to retire all my 1.5 and 2.x checkpoints and embeddings/loras

[cyan crown]

Don't look a the image itself, just focus on blacks and clearness. Not upscaled or PP

[crisp owl]

I've purged a ton of my 1.5 stuff recently, mostly stuck with my merge anyways

[vital ermine]

yeah, next month I get a 2tb NVME to replace the 512gb and put the HDD to rest. with it goes the 1.5/2.1 stuff

[indigo carbon]

yeah that guy, I think if we would make something like that but an even better version, it could be a general prepose thing unlike all the other LoRAs

[vital ermine]

Next year I move on to AM5 or 6 (if it comes out) and can have 4 nvme drives

[crisp owl]

Yeah I love th eoffest noise lora, I'm sure there will be another one created by the community.

[vital ermine]

1.0 it is baked in

I hate how they baked the vae in

[crisp owl]

1.0 vae was broken anyways

[indigo carbon]

does offset noise reduce coherency anatomy and text ability though?

[vital ermine]

yeah, but it is baked in. for training that is wasted memory when I have the fp16 version I use

[crisp owl]

ooh yeah, that's true

[vital ermine]

every drop of vram matters now which sucks

[crisp owl]

Not that I've really tested, but I use it in nearly 100% of my images

[vital ermine]

Dreambooth I feel the lack of ram the most with only 24GB. Sad that ONLY is used with 24GB now

[indigo carbon]

I'm pretty sure you can use the save checkpoint feature in ComfyUI to save a checkpoint with or without certain parts

[vital ermine]

I have not see it stripped out, or the ability to do it. PLUS they trained 1.0 with a noise offset of .0357

[indigo carbon]

you can make a version of base that just has UNET or whatever part you wish to train

[vital ermine]

DUDE, gimme the ability, seriously

[indigo carbon]

hold up, I'll screen shot it, I remember doing that in the past

[vital ermine]

I do train the unet and tes but the vae I use the fp16 less problematic one

seems stupid I have the baked in vae AND the fp16 one all shoved into vram

[vital ermine]

Oh, LOL, good one joe

Pikachu

[indigo carbon]

you can do this

heck, even save it with a different VAE if you want

[vital ermine]

Does that actually shrink it?

[indigo carbon]

yeah, I used that many times before, haven't tried that on XL though

[vital ermine]

I am training or I would

just finished

[cyan crown]

Left is PP right is original. Training a Lora will give first result. I must try to train a general "impact "lora

[vital ermine]

I am now using BS12 and can't go more with only 24GB

[hasty smelt]

Hello everyone,

I have been having a lot of blue screen issues when playing SDXL on my 16GB RAM system, as the memory usage sometimes reaches 99% during the VAE pass in ComfyUI.

I want to ask if upgrading to 32GB of RAM will resolve this ComfyUI issue? Here in Brazil, 64GB of RAM is very expensive and I only have a budget for the 32GB upgrade.

My system > RTX 2070 Super FTW3 + 16Gb ram DDR4 3200Hz + AMD 5700X

[vital ermine]

you only have 16 now?

[hasty smelt]

yes

[vital ermine]

I think comfy has a lowvram flag

I know it has a med one

[indigo carbon]

if you have 24gb, you might be able to compile AIT for that batch size; the reason why the node only supports up to 4 is because when supplying the precompiled modules the highest we could streach it is 4

[crisp owl]

Does increasing swapfile/pagefile work for SD?

[vital ermine]

I wish I could train faster but AIT is for training too?

[vital ermine]

BSOD for me in auto when it went to 128GB and kept growing. I told it 16GB and rebooted and BOOM, auto BSOD me in a min or two

[indigo carbon]

it can be implemented into anything when done correctly; but there is something called DeepSpeed which is already implemented for that and it does the same thing AIT does

[crisp owl]

ah gotcha, so doens't like it. didn't think it would

[vital ermine]

used up all 48gb and still 128gb swapfile

[hasty smelt]

I used this, it solves it in many cases, but even so, at times the RAM consumption goes up to 99% and I get blue screens.

[indigo carbon]

idk if DeepSpeed works on windows though

[vital ermine]

32 will help, yes. It will off load more if you tell it

[vital ermine]

nope

funny cause it is Microsoft too, lol

[indigo carbon]

WSL?

[vital ermine]

I dual boot

Ubuntu 22.04

I normally train in it since it is almost twice the speed over w10

[hasty smelt]

I'm going to upgrade to 32GB of RAM then, thanks for the support.

[vital ermine]

You are welcome and investigate --medvram or w/e comfy has

[crisp owl]

COmfy defaults to Medvram

[hasty smelt]

thanks buddy

[indigo carbon]

than use that lol, DeepSpeed works flawlessly on Linux; the reason why it's not being used for running SD is because it only supports training and running only LLMs, but it can EASILY be used for training SD

[vital ermine]

Ahhh, then wubba will be golden

[vital ermine]

tbh, I am unsure how. I saw it in the accel config but said no (the default) cause Windows tore me a new asshole one Saturday. I gave up on it.

[indigo carbon]

if you use KohyaSS for training, you just do pip install deepspeed then start Kohya with DeepSpeed selected

[vital ermine]

wait, I gotta tell accelerate yes, though, right?

when I start kohya_ss I select nothing it just gui.sh

[indigo carbon]

I think you add a --FLAG for that

[vital ermine]

accelerate config it on and a flag. Hmmm, let me check his flags

[indigo carbon]

idk what are the flags for kohyaSS, do start gui.sh with a --help flag, it will explain itself

[indigo carbon]

there should be one for DeepSpeed though

[cyan crown]

this is the best I can reach only with prompts

[crisp owl]

I tell ya, the things that you can simply prompt for in SDXL is pretty awesome

[vital ermine]

I can tell I am going on an adventure

That's me and the guy touching my lifeless body is Deepspeed

[floral island]

my attempt

[indigo carbon]

Why? Was it not simple to set up?

It should be really easy on Linux

[vital ermine]

Nah, I will try after 7pm my time but there are no commands to kick it on and when asked bmalt basically shrugged

I think it works via the accelerate command but installation is dead easy

[vital ermine]

going to force me to bake it in

I'll bake in the fp16

[indigo carbon]

Damn, I remember that working on 1.5 models, there must be some kind of node that allows you to save certain components, I don't see any reason for it not to be possible

[shy kelp]

How do I add a sampler to ComfyUI? The only one available is the default Ksampler

I want to add K_DPMPP_2M

[floral island]

with this, imma go to bed 👋

[nimble heart]

the fp16 converted vae from hf has precision problems. it can nan out on non-standard resolutions or when denoising < 1.0

[heady vale]

I tried that vae and didnt like the results. It didnt nan out but wasnt better

[nimble heart]

the fp16 wont be better just faster on some hardware

[vital ermine]

YET, trainers, and even articles, say then baked in one is problematic

btw, where did it save this as I don't see it

[nimble heart]

the baked 1.0 vae is

bake the 0.9 vae

[vital ermine]

I don't have that one

can't even find it

[nimble heart]

its a download right on the same place you got the 1.0

[vital ermine]

not when I went it reverted to 1.0 with a link

they had just removed 0.9

[vital ermine]

well, Ican't find it but it looks like it saved it so must not work

[nimble heart]

https://huggingface.co/stabilityai/sdxl-vae/tree/main
opposite. they removed the 1.0

so its just the 0.9 vae now

[vital ermine]

Yeah, that is the one I have

335mb

[nimble heart]

they'll be the same size

look at the checksum

[vital ermine]

I can't check that for vae

[nimble heart]

sure you can

just open the terminal or PowerShell or whatever and run the checksum command for sha256

[vital ermine]

no biggie

I am trying to figure out where comfy saved this at

[nimble heart]

find my-model.safetensors

in PS

at least I think PS has find like that. it might be a different name

[vital ermine]

nothing in checkpoints

[nimble heart]

maybe it didnt save

[nimble heart]

idk just run a find on your entire %USERPROFILE%

drives are so fast nowadays it only takes a few seconds

[vital ermine]

no, because we have no idea what it called it and I don't want to find 300+ checkpoints across 5 drives

[nimble heart]

looks like you named it ComfyUI?

[vital ermine]

I suspect it didn't save it

no, comfyui named it comfyui I didn't changeit

[nimble heart]

checkpoints/ComfyUI is the filename prefix so

file probably starts with ComfyUI

[crisp owl]

Check where your images save for a file named "checkpoints"

[vital ermine]

every place I have my safetensors it is not there so it either did not save or went into limbo

[crisp owl]

odd

[nimble heart]

lol

its in your output folder

like where the images are saved

[crisp owl]

that's where I said to check lol

[vital ermine]

I already checked there right off

[nimble heart]

it makes a folder called checkpoints

[vital ermine]

Oh, I didn't check new folders

[nimble heart]

so it'll be outpout/checkpoints/ComfyUI_00001_.safetensors

[vital ermine]

wtf comfy?

[crisp owl]

that's what checkpoints/ does

[nimble heart]

yea

[crisp owl]

makes a folder called checkpoints

[vital ermine]

nope, did not work

[nimble heart]

skill issue

[vital ermine]

never saved it, or created a new folder

must bomb out on xl

[nimble heart]

works fine for me

lemme try XL i guess

[vital ermine]

you are in linux so might work there

[nimble heart]

works fine. ran in 7 seconds

xl 1.0 base with the 0.9 vae

I mean you know what the files are named now

find ComfyUI_00001_.safetensors

[vital ermine]

oh, lord that is on the slow HDD.

[nimble heart]

noooooo hdd

flash storage is so cheap now

[vital ermine]

if it is where it saves the images. what he should have done is to save it in the same damn folder where the checkpoints are currently

[nimble heart]

searching is literally the worst thing an HDD can do because its just a billion tiny ops

[nimble heart]

nah

I like it like that

means I can save checkpoints to my ramdisk

[vital ermine]

then give me the abilty to tell it where to save don't just decide for me

[nimble heart]

lol

[crisp owl]

worked fine for me on windows

[nimble heart]

lol

output/checkpoints

just like you'd expect

[vital ermine]

well, I am not saying it didn't work just I have no idea where the fucker put it. Frustrating as it should have asked me where to save it.

[nimble heart]

where does it put your images?

just like

look there

[vital ermine]

my downloads folder and there is nothing new in there

[nimble heart]

to clarify its in a subfolder of the images. so the folder is mixed in with the output images

could also check the default output folder in your comfy install dir

just in case

[vital ermine]

having it search now on the hdd for ComfyUI_00001_.safetensors

[nimble heart]

nice

[vital ermine]

I did, as I said

[nimble heart]

nah

up one

[crisp owl]

check your outputs folder

[nimble heart]

ComfyUI/output/checkpoints

[vital ermine]

thank you

first one ran out of vram it said

[nimble heart]

so your images render to downloads but checkpoints are still output to /output/?

[nimble heart]

???

[vital ermine]

yes

[nimble heart]

how

[vital ermine]

¯_(ツ)_/¯

[crisp owl]

didn't you say you ahve a 4090?

[vital ermine]

yes

[crisp owl]

wtf lol

[vital ermine]

yep

[crisp owl]

How you run out of vram and I have 8gb vram and i don't lol

[nimble heart]

it only used like 9 gigs to save the XL + vae on mine

and im pretty sure thats high cause AMD

executed in 4.8 seconds

[vital ermine]

because it loads all into vram in comfy so it doesn't have enough room for it all over again and is too stupid to know it already is in vram

[nimble heart]

?

[vital ermine]

once I shut it down and restarted it worked fine

[nimble heart]

it unloads stuff when you're high vram

[vital ermine]

tell that to comfy cause it didn't I got an OOM at first. I don't care why, or how, just know I did

[nimble heart]

you can also use the old allocation system with more aggressive unloading with a cli flag

[vital ermine]

????

[nimble heart]

but a couple weeks ago they added a new allocation system that only unloads things when the mem is high

[vital ermine]

Oh, remember I am on 531.79 so easy to OOM now

[vital ermine]

any driver past that is ungodly slow as in 3s/it to 200+

[nimble heart]

I think that was the flag

[nimble heart]

3s/it doing what

also might be the shared memory thing

someone was here earlier getting like 300s/it with a 1650 on XL cause new drivers would load the model half in ram half in vram instead of OOMing

[vital ermine]

Trainers already know, as does Nvidia who acknowledge this, that past 531.79 it has smart memory built in and make ML/AI training next to impossible

saves gamers from OOM but at the cost of ML training

[nimble heart]

nvidia saw AMD with their driver issues and thought "hey I wanna try that"

[vital ermine]

LOL, yeah

[nimble heart]

also still waiting on ML libraries to catch up for my XTX. Don't think rocm 5.7 has anything new support-wise

[vital ermine]

Sad.

[nimble heart]

saw someone make a custom 5.7/torch2.1.0 build on github and it was no faster than 5.6 apparently

[vital ermine]

That is why, at the last minute I switched to 4090

[nimble heart]

amd has their own forks for FlashAttention and AIT and everything else, most of which have navi3x branches

but idk they're all stagnant

[vital ermine]

Hardware for AMD has so much potential but software always has sucked. Better for gamers now than it used to be but still, damn 😦

[nimble heart]

so I guess they've just got a million things on the todo list with like 3 people working on it

[nimble heart]

yea been amazing for games. I run battlebit @ like 8k so I can snipe people > 1KM away using the low magnification scope

it does pretty good with AutoGPTq too

stable diffusion is mostly the rough spot

[vital ermine]

Yeah. I almost bought the 7900xtx but too much headache so 700 more USD hurt immensely. All because AMD can't get their software side shit straight but I firmly believe they are in the same rut as Nvidia and afraid they will hurt their MI300 sales.

last min I swapped it out but if I had pressed the button I would have had the 7900XTX

[nimble heart]

ah yes cause the cards that people get from best buy are totally gonna ruin the sales of special-order 192GB cards with no display outputs

[vital ermine]

yep

[nimble heart]

like kneecapping the vram and fp64 perf is enough for most people to not use them outside of hobbies

may as well add the extra support to get people in the ecosystem

[vital ermine]

Point blank (and either option is bad sign for AMD) is that they are 100% inept, or they drag their heels due to the MI lineup

[nimble heart]

(apparently the MI lineup is having software problems too)

[vital ermine]

ahhh, shit 😦

MI has been around for almost a decade now too

MI25 was far ahead of its time

[nimble heart]

rocm stack specifically. ig if you buy a server with 50 of them you just write your own software to run sims or whatever

[nimble heart]

I heard it mostly affected MI 100/200/300

[vital ermine]

MI25 is no longer supported

[nimble heart]

ah

[vital ermine]

People were using it for 100USD to SD with now XL not so much

[nimble heart]

according to google its basically a low profile vega 56 with extra vram so cool but not setting any speed records

[vital ermine]

yeah, but in its day it was state of the art

plus for only 100usd it was a great buy. sometimes going to 80usd

[nimble heart]

their pytorch builder fork already has an rocm 6.0 branch so maybe that's the "super amazing" release the ceo keeps talking about that'll officially support the XTX and 6900 XT

[vital ermine]

If it doesn't then I have lost all hope for AMD which sucks

why did my save make a bigger file than the base?

[nimble heart]

?

mine's the same size

[nimble heart]

oh by like one megabyte probably just metadata or something. Or maybe the old one had the broken vae baked in

[vital ermine]

it did

old vs new vae

Nothing changed

maybe does for more realistic

I am not sure what this style is but I like it

simple

[nimble heart]

you can literally see the bad artifacts in the flowers

on the old

got lines

eyes as well

[vital ermine]

I didn't zoom in

[nimble heart]

compared to the new vae

or old I guess since it came first

[high skiff]

I feel like I can finally breathe

It took over 3 months, but I finally got my fucking money back from pay pal

[nimble heart]

breathe in that sweet sweet polluted california air lol

dannng

[high skiff]

Air quality is excellent here, luckily haha

[nimble heart]

let me guess guy still has his $600 too

[high skiff]

Likely

[nimble heart]

everyone gets free money

[high skiff]

I don't give a fuck anymore

[nimble heart]

paypal communists confirmed

[vital ermine]

LOL

[high skiff]

It took 24 phone calls, over 15 hours on the phone, several dozen messages, at least a dozen emails, and over 3 months to finally get my fucking money back

I can finally pay my mom back, because she floated me the money while I bought my new 3090

[nimble heart]

hey congrats you made $40/h

[high skiff]

$40 an hour of money that I don't get to see, and never had lol

That's just a little over a monthly electric bill here, which is fucking insane

[nimble heart]

WHAT

[high skiff]

What?

[nimble heart]

ours is $400/m and we're in the top like 15%

you must be top 1%

[high skiff]

Nope

We just have criminally high electric rates

[nimble heart]

Ah. Our city has its own grid so we're not on PG&E

immune to brownouts too

[high skiff]

We have SCE, which is worse

[nimble heart]

never heard of them but somehow I feel like I don't want to

[high skiff]

Southern California Edison is the worst electric provider in the United States

[nimble heart]

still doing the solar panels?

[high skiff]

They have gone bankrupt three times now because they are responsible for all of the horrific fires that happen in California.

Just a couple years ago they got sued for like 8 billion or something

Because like 1.8 million acres of wildfires in 2019 or something

[nimble heart]

PG&E got sued for that too didnt they

like in 2021

[high skiff]

I don't know about the solar panels, I'm not too educated when it comes to electricity, and I'm not sure I want to be handling the wires for system that can output 9,000 watts of electricity

[nimble heart]

just like

wear gloves

hide your electrically conductive human juices

[high skiff]

I also don't wanna risk starting a fire or something. We don't have insurance for accidents. I wanna start on a smaller solar project first.

[nimble heart]

if a fire starts just tell it "no". It legally cannot damage your property without your consent

[west breach]

rooftop solar systems are pretty cheap here

[nimble heart]

the cost to put it in the roof is considerably more expensive than the actual panels

[west breach]

also helps to upgrade to an energy efficient heatpump AC

[nimble heart]

in cali heatpumps are standard but idk maybe not on older houses

[west breach]

I've seen posts on hackernews about heatpumps, like they're some new fandangle invention or something

[nimble heart]

?

its literally just an AC running in a reverse cycle

[west breach]

maybe a lot of americans haven't bothered with heatpump systems since electricy used to be cheap?

[nimble heart]

so the coolant decompresses and compresses in opposite ends to reverse the effect of AC

[vital ermine]

https://civitai.com/models/144527?modelVersionId=160636

[nimble heart]

yea the average cost of electricity in the us is like 15¢/KWh or something

combine that with most states not being literal deserts and power efficient cooling isnt top priority like it is in california and texas right now

[nimble heart]

https://youtu.be/7J52mDjZzto?feature=shared

good video

[west breach]

what worked well in AU is government offered feed in tariff for solar systems over a decade ago which kicked off the industry here

[nimble heart]

good yt channel in general

[west breach]

thanks. I know about them 😄 I was just saying it seemed like a new thing to a lot of americans. I have ducted reverse cycle AC and a heatpump dryer

[nimble heart]

but yea propane is also pretty cheap in most places so propane heating is more common than heatpumps

[west breach]

and rooftop solar

[ruby ibex]

can someone help me? my stable diffusion for the past 2 days is only generating this kind of images

[west breach]

grey images is interesting. I've seen black, but not grey

[ruby ibex]

it worked fine until a few days ago..

[west breach]

evaporative AC is a good option too in dry areas

[nimble heart]

update anything?

[nimble heart]

idk what that is but it sounds like some Tatooine shit

[ruby ibex]

windows

[nimble heart]

what model?

[ruby ibex]

all of them are doing it

[west breach]

https://www.seeleyinternational.com/products/cooling/evaporative-air-conditioning/how-evaporative-cooling-works/

[nimble heart]

using the 84k vae or the anime one?

anime one borks images sometimes

[ruby ibex]

ok i just deleted all the models

[ruby ibex]

and vaes

and got them again..

and now its working

[nimble heart]

now why'd you go and do such a thing?

wait what

[ruby ibex]

10gb fiber optics XD

[ruby ibex]

i download them instantly

but its working now

[nimble heart]

probably still limited by the website's rate cap

[nimble heart]

I dont even get 1 gig on anything but Steam

[hardy cipher]

Nice

[native knot]

Emcee Escher.

[nimble heart]

if it works again with the same models run chkdsk cause that might be a drive issue or something

[hardy cipher]

Dude, I made several mc escher images quite a while back

[native knot]

Yes, but were they Emcee Escher?

[hardy cipher]

At least 8 months ago, I have them somewhere

No, wasn't cool enough to go with emcee

[native knot]

I think the first two are my favorite, but I like all of these.

[heady vale]

cloning Trump

[native knot]

I get inspired.

[native knot]

https://giphy.com/gifs/everyone-remake-rw-s239QJIh56sRW

[hardy cipher]

I've been going down the ipception rabbit hole. Putting results back in and moving things around

[heady vale]

making pink great again

[hardy cipher]

Also messing with the conditioning data

[native knot]

Nice

[hardy cipher]

Left a bunch to render while I went to do other things. Some of them turned out alright

[native knot]

When you say "messing with the conditioning data", can you tell me more?

[nimble heart]

back when he first announced he was running for president there were countless memes about cloning him so he can become immortal god-emporer that takes humanity into the space age to conquer the galaxy so we can all get alien war brides

[vital ermine]

Welp, time to switch over to Linux and get deepspeed working. I don't have anything to train so will train what I just released

I wonder if I will see any speed increase?

[nimble heart]

maybe the drivers arent shit there

[vital ermine]

just going to linux is 50-100% faster than W10

[nimble heart]

assuming your distro actually has up-to-date ones

[vital ermine]

I train in Linux unless I am doing business (fuck dyslexia) Windows only work too

[nimble heart]

newest is 535 so idk how that compares to windows

[vital ermine]

far better. I even tried the windows version of the drivers released the same time and 200+ s/it

[heady vale]

537.34 just released on windows

[nimble heart]

is that just a game release?

[vital ermine]

unless they make mention of AI/ML I want no part of it

[native knot]

So I put "Donald Trump" plus the text from your reply into my workflow and got this back...lol:

[nimble heart]

linux doesnt get releases for specific games

[heady vale]

the regular game ready driver

[vital ermine]

Linux gets a new one about twice a year

[hardy cipher]

Split it coming from the text prompted. One side normalized and multiplied by 3. The other side I multiplied by 2. Then concatenated the two streams.

[nimble heart]

yea but is that just for Starfield or something? no other changes?

[hardy cipher]

Then it hits clip vision

[vital ermine]

Nvidia did say a future version will fix this ML/AI windows issue just not when. Could be in the version released in the year 2030.

[nimble heart]

you'll have to see what your it/s are on linux before adding deepspeed and other shizzle

[heady vale]

yeah its game optimisations. nothing for SD that I know of

[nimble heart]

yea that wont hit linux. They only release major revisions, not minor game-specific ones

since they dont matter on linux anyways because everything is DXVK translated

[vital ermine]

As I just told you I do most all my training there and it is twice the speed of W10. For giggles I will do a training the install deepspeed then train again just to see side by side

[nimble heart]

Ohhh

w10 == win10

[native knot]

Interesting. I assume you're doing that in comfy? If you have a workflow to share, I'd love to poke around with that at some point.

[vital ermine]

yes

[native knot]

Or is it embedded?

[nimble heart]

I thought w10 was cloud service or some shit

[vital ermine]

screw the cloud, lol

[hardy cipher]

Yeah, it's in comfy. Made a couple nodes so I could do it

[native knot]

In honor of tonight's impending Ahsoka episode.

[nimble heart]

wonder if all the people running just inference would get a good speed boost on linux too

[vital ermine]

yes, we do

[nimble heart]

I never compared when I had my 1070

[vital ermine]

first thing I tried

[native knot]

Would love to mess w/that if you'd be willing to share.

[spring fulcrum]

does anyone have a recommendation for Sampler, Scheduler, Steps, and CFG for img2img?

[nimble heart]

whatever works best for you

[hardy cipher]

For sure. When I'm back on my pc I'll show you the workflow. It might be in the images too

[nimble heart]

no real 'correct' answer on those besides using ddim + normal when using the step swap method

[native knot]

Right on.

[vital ermine]

I just can't use comfy there as his install for windows is shit and doesn't even use, or work with, a virtual env. I tried but it infected my main system so not trying on Linux. Auto ran smooth as butter on Linux but I removed it.

[nimble heart]

since the refiner was trained on that

[spring fulcrum]

And does anyone still have that workflow for combining 2 images together to make a hybrid? I remember that floating around here somewhere?

[hardy cipher]

I don't really know if my approach is optimal, but it's interesting. I just like messing with things

[nimble heart]

it's just the clip vision node

s

[native knot]

same here

[hardy cipher]

I'm trying to figure out what causes the Nan black images with ipadapter, and if they could be fixed

[nimble heart]

precision problems. could be a lot of things. usually vae but with custom nodes its a whole mystery box

I managed to cause it by directly mutating the cache lol

[hardy cipher]

Yeah. I'm just wondering if I could somewhat normalize the data to the point that it'd still render something

[vital ermine]

A lot of stuff being updated on linux

even the base which I suspected

[nimble heart]

like python libraries or for the distro you're on?

[vital ermine]

everything as I update irregularly

[nimble heart]

ah. rolling release distro?

if I dont update arch in a week I come back to 300 package updates

[vital ermine]

Ubuntu 22.04

yeah, arch is like that

mint was sorta like that too

[nimble heart]

didnt like mint

[vital ermine]

I liked it but what I despise about Linux is the quirks

stuff I like from windows it doesn't do cause it is like windows

[nimble heart]

thats why I like arch; the quirks are documented well enough its a fast resolution instead of going through forum posts from 2005

[vital ermine]

well, in Linux I know I saw this but then lost the link, but how do I get the damn windows to minimize to the task bar not that stupid grouped thing in the lower left side?

[nimble heart]

idk depends on your desktop environment. if its ubuntu you're probably on Gnome which I've never used

[vital ermine]

I use cinnamon

so it is a gnome, yeah

[nimble heart]

thought cinnamon just minimized to taskbar like windows

unless they changed it

[vital ermine]

they did

[nimble heart]

damn

well change it back ig. probably a setting somewhere..

you used to be able to right click on the taskbar to customize its behavior

[vital ermine]

I couldn't find it. On a fresh w10 install I see it did the same thing. I was like, damn. Changed that one back fast

[nimble heart]

why I like Plasma lol

[vital ermine]

I tried that and reinstalled everything

[nimble heart]

i use plasma for laptops n stuff where I might need to hand it to someone who's only used windows

[vital ermine]

I knew it was time to update because it was bitching about some MSFT thing as it booted. Gone now.

Successfully installed deepspeed-0.10.3 hjson-3.1.0 ninja-1.11.1 py-cpuinfo-9.0.0 pydantic-1.10.12

now to test with and without

I swear it is so much faster in linux with the exact same everything. Not like 5 or 10% but 50-100%. It is insane just how bad Windows is for this.

stopped it so now to activate deepspeed

This is the missing part

apparently it really does want a json too

needs it

too much stuff I have no idea about

crash and burn

no info on this error either

AttributeError: 'DeepSpeedEngine' object has no attribute 'text_model'

[nimble heart]

opens 'getting started' docs

50 pages long

quits

I feel that

[vital ermine]

LOL

see

for all I know kohya hasn't implemented it

almost works

[2023-09-12 20:07:05,739] [INFO] [config.py:957:print_user_config] json = {
"train_batch_size": 1,
"train_micro_batch_size_per_gpu": 1,
"gradient_accumulation_steps": 1,
"zero_optimization": {
"stage": 2,
"offload_optimizer": {
"device": "cpu"
},
"offload_param": {
"device": "none"
},
"stage3_gather_16bit_weights_on_model_save": false
},
"steps_per_print": inf,
"bf16": {
"enabled": true
},
"fp16": {
"enabled": false
},
"zero_allow_untested_optimizer": true
}

[hardy cipher]

hmm

[vital ermine]

@indigo carbon I cannot get deepspeed to work with Kohya it throws this error - 'DeepSpeedEngine' object has no attribute 'text_model'

Looks like it may not be implemented in Kohya

[hardy cipher]

lame

[vital ermine]

New error tells me the lycoris is missing the forward function.

NotImplementedError: Module [LycorisNetwork] is missing the required "forward" function

Now that is lame

Yep, Kohya

NotImplementedError: Module [LoRANetwork] is missing the required "forward" function

shit

DB it must work only 24GB is not enough mem

iow deepspeed is far too limiting, and just not even implemented to be of any use

looks like I need about 34-40GB of vram

[hardy cipher]

what's going to happen?

well okay

[hardy cipher]

lolwut?

[heady vale]

hehe

does not compute

[hardy cipher]

it's pretty quality though. this workflow has all sorts of weird things going on. but didn't expect that

[heady vale]

medusa housewife at the seaside

[hardy cipher]

ipception has changed things a bit

[spring fulcrum]

@indigo carbon I haven't been on discord in a while... I was wondering if I could get the latest workflow you are willing to release for you AIT... I want to feel fast today. Crank out some volume.

[heady vale]

they have a couple of workflows on civitai

[spring fulcrum]

oh cool they also have the image blend thing I was asking about earlier.

[spring fulcrum]

Thanks 🙂

[untold dawn]

can anyone teach me how to use stable diffusion

[hardy cipher]

there are probably quite a few people that could

[untold dawn]

what tool is this

[spring fulcrum]

first question is what platform do you want to use to utilize SDXL? ComfyUI or A1111? That might help narrow down who helps you or what videos they point you to...

[untold dawn]

i want to able to generate the same quality image you generated lol

[untold dawn]

is there a link to download the platform or what

[hardy cipher]

several

[uncut fiber]

@untold dawn and what is your gpu btw?

[untold dawn]

im on gaming laptop with rx6700s

[spring fulcrum]

Yes but it depends on if you want to use A1111 or ComfyUI?

[untold dawn]

im fine with anything tbh if its easy to use then i dont rly care

[nimble heart]

a1111 is easier but it's a good 30% slower on AMD cards

[untold dawn]

oh ok

is a1111 free?

[nimble heart]

I get something like 2.5 it/s on auto and 3.15 it/s on comfy

both free open source

[untold dawn]

k thx

do u download it somewhere?

[nimble heart]

github

oh

also amd cards dont work on windows for stable diffusion

[hardy cipher]

lol

[untold dawn]

what

bruh

[nimble heart]

yea linux only at the moment

[untold dawn]

what tool can amd cards use

[nimble heart]

ROCm support on windows is coming soon™️

[nimble heart]

for AI? nothing on windows

[untold dawn]

oh what

😭😭😭

[nimble heart]

except maybe SHARK but that's complicated to set up and doesnt work well

[uncut fiber]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs

[nimble heart]

they said they're gonna support windows eventually

but now you either install linux or it runs a compat layer through directml which means it'll take 20 minutes instead of 20 seconds

[untold dawn]

thx

[heady vale]

you can use the #1100170312106127410 on this server though

[nimble heart]

linux is free to install too so if you have the luxury of lots of free drive space you can just shrink your windows volume using Disk Manager then install linux to the empty space

[untold dawn]

thx

[untold dawn]

what tool is that

[hardy cipher]

comfyui with some custom nodes

ip adapter, clip vision, various other things

[untold dawn]

nice

its free?

i thought it needs payment

oh well ill see

[spring fulcrum]

ya and it uses sdxl 0.9

[untold dawn]

thanks for the help tho

[hardy cipher]

just use the bot

wat?

[dense chasm]

pika is awesome https://twitter.com/pika_labs/status/1696198588295184420?s=20

Pika Labs (@pika_labs)

They arrived from distant stars.
Our world was their prize.

Introducing Invasion, a 100% AI-generated movie trailer.
Directed by Aldizm, all scenes animated in #PIKALABS.

The future is here.
One person, one computer, one movie.
Would you watch a movie like this?
#image2video…

Likes

156

▶ Play video

Twitter · Aug 28, 2023, 4:30 PM

[indigo carbon]

if you want a WIP

[spring fulcrum]

That sounds good to me... What are some of the changes/improvements you've made with v5?

[indigo carbon]

added a styles node and made it so you can bypass the refiner node without effecting outputs too much

[spring fulcrum]

Can you add a node to use a LoRA with it? My wife and I are working on some things and we found some good LoRAs we like. I honestly don't really get how to connect that stuff myself.

[indigo carbon]

yeah, it shouldn't be challenging to add that

currently my goal is to figure out IPA multimodel with AIT but LoRAs should be way easier

[hardy cipher]

I can use 10 loras at once no problem

[indigo carbon]

with AIT? I was wondering how LoRAs effect the speed boost from AIT

[spring fulcrum]

I would be happy to test that out for ya with my 4090 🙂

[hardy cipher]

nah, not with ait. sorry

last time I tried ait it kept giving me some error about incorrect data so I uninstalled it. didn't want to

but no idea what was going on

[indigo carbon]

it could be the comfy commit you were using, it seems many people have compatibility issues recently.. me and FizzleDorf were considering making a ComfyUI fork for that reason

[hardy cipher]

yeah, I wasn't sure what it was about. I'll try it again one of these days

really a strange journey with these images

[indigo carbon]

heh, it appears the latest version of the AIT node is compatible with latest ComfyUI

[hardy cipher]

I'll look into it in a bit

this stuff is wild

[indigo carbon]

IPA?

[hardy cipher]

yeah. first image ipa. second is clip vision g. third is the input latent at 70 percent denoising

also modified pooled output a bit prior to the clip vision

[spring fulcrum]

I feel like I have some wrong settings here... The image on the left always looks great then the image on the right always looks like they got hit by a bus... lol

This is for the image blender BTW

[hardy cipher]

seems to have helped with the black box images. or maybe I'm imagining it

but haven't gotten one recently

[indigo carbon]

you can always bypass the second stage and/or change to DPMPP_2M

[hardy cipher]

put "haggard" in the negative prompt

tbh your cfg is too high. especially for that second pass

[spring fulcrum]

How do I bypass the second stage?

[indigo carbon]

oh wait, are you using the one on CivitAI? It's a prototype