✨｜sdxl

[trim orbit]

i'll fight it

[autumn forum]

oh nooo the toaster is coming to life!

[true wing]

he just looks happy to be there

[timid sonnet]

GUYS

How do I do the step by step time lapse generation to increase details?

[trim orbit]

https://tenor.com/view/goblin-truck-gif-20619963

[upbeat summit]

yeah - true 😉 let's bring on the singularity
*I'm just having fun - j/k

[nimble heart]

if you're in comfy it's probably easier to just point OBS at the preview window

[timid sonnet]

what is OBS

[nimble heart]

screen recording tool

[autumn forum]

recording / streaming software

[trim orbit]

open broadcast studio

[timid sonnet]

was that a joke?

I was asking a serious question

[trim orbit]

look it up

[autumn forum]

huh

[trim orbit]

it's software people use to catch screen and send it to either a stream or file

[timid sonnet]

He answered my question with POINT OBS at it

[trim orbit]

why so hostile about an answer lol

[timid sonnet]

I asked a serious qeustion about ComfyAI

[nimble heart]

I'm 100% serious. Comfyui doesn't have an XYZ grid thing yet so if that's what you're using you'd have to either manually render an image stopped @ 1 step, then another stopped @ 2 steps, etc.

or you can just enable realtime preview and screen record it

[timid sonnet]

@nimble heart you dont know what I'm talking about please stop

There is a method of increasing details Sytan pointed out when SDXL released.

Something to do with setting it 1 step back or something like that?

[nimble heart]

Oh why didn't you just say so

[true wing]

why don't you search for posts by Sytan and look through them for it then?

[nimble heart]

set the refiner to have 1 less total and end steps than the base

[upbeat summit]

yours looks good! got it here on the right side. sometimes it blends into an object like a wall but it's still there. it could be prompt related. I guess I need to take it apart, but it comes up a lot for me in specific resolutions.

[timid sonnet]

@nimble heart That is what I wanted to know

[trim orbit]

this was exactly what he answered. now you're just changing the question to seem like you're in charge?

mr large i guess. gj

showed up good

[nimble heart]

the way you phrased "increase detail with each step" made it sound like you wanted a timelapse animation

[trim orbit]

asks for help, people help, gets mad and asks for different help, ????, profit

[timid sonnet]

@trim orbit he said point OBS at it, I never asked about OBS.

He answered me correctly the second time around. Quit circle jerking

@nimble heart ty

[trim orbit]

fucking bs

i'm out

[autumn forum]

i mean he just made his account so what do you expect xD

[timid sonnet]

@nimble heart So

If I set the total steps to 80

then set my refiner stop at 79

it should push details?

[autumn forum]

damn how many steps you need????

[timid sonnet]

@autumn forum I don't really need this many steps, but it helps when I generate a specific prompt

[autumn forum]

if i want extra detail i just upscale 1024 to 4096 then downscale back to 2048.

[timid sonnet]

@autumn forum I'm talking specifically about time lapsing the image to cause it to rush details

@autumn forum That increases the quality though?

[autumn forum]

idek what that means lmao

[timid sonnet]

@autumn forum I don't either that's why I'm asking but Sytan was able to produce quality images with this method

I've looked through Sytans chats and I didn't understand what he was doing. But he was clearly posting great results

[autumn forum]

so what hes talking about is what i call the time skip method using his partial diffusion setup where 0-13 steps on base then 14-20 on the refiner it skips the step 13-14 causing extra noise to be left over during the process which the refiner adds as detail. the total steps needs to be 20 still though. so basicallt the base never gets finished with all 20 steps. it stops short at 13 and the refiner takes of the rest. @timid sonnet... is this what your talking about?

[nimble heart]

Point OBS at it 👍

[trim orbit]

oh look a timelapse!

[timid sonnet]

@autumn forum OHHH SO I CHANGE FREAKING 0-30 30-60 TO 0-30 31-60 61-70

[wicked frigate]

that's a very good way of describing it

[nimble heart]

with 80 you might need to stop earlier than 1 to notice it

[autumn forum]

EXACTLY yes man you got it

[spring fulcrum]

and so a new definition of blue balls emerges

[timid sonnet]

thank you @autumn forum

Sorry @nimble heart I thought you were being sarcastic when you said point OBS at it

[autumn forum]

no problemo 👍

[autumn forum]

be careful with that tho. skip to many and itll look like hot garbage

[timid sonnet]

@autumn forum I'll show you my results here in a few, 80 steps takes a bit to generate

[autumn forum]

oh i bet

i just did it and here a before and after @timid sonnet

[timid sonnet]

@autumn forum It definitely enhances the overall detail, but it kind of throws it out of contrast I feel like

[nimble heart]

I wouldn't use it outside of photoreal cause it essentially just interprets the extra noise as small details it should refine into features

[timid sonnet]

hot garbage 100%

@autumn forum

[nimble heart]

yea

[autumn forum]

only skip once lmao

[nimble heart]

though that's probably too strong as well

[autumn forum]

before and after

[true wing]

The refiner is interesting because for some things it helps a lot but I've noticed others where just going 100% base model is better. It really seems dependent on subject and style

[nimble heart]

notice it actually lost details in the flower centers because the extra latent noise is large enough to obscure the pollen

and a lof of the small flowers have less petals

[autumn forum]

i agree. there are only a handful of use cases that this would be needed. but its a 1 number change. so its easy to test to see if its worth it in an image.

[true wing]

sometimes it definitely 'refines' away details you want

[nimble heart]

comfyui caches the latent stages so if you think it needs extra sparkle just reload the latest history item and it should only resample the refiner pass after you decrement the step count

[autumn forum]

yep

[nimble heart]

but saving it into an actual workflow.json is madness

[timid sonnet]

tf

[nimble heart]

okay that ain't right

[upbeat summit]

too much refiner

refiner likes to to do dots when giving it too many steps with textures, water droplets, snow and any kind of particle

that's at least my experience with some prompts

[trim orbit]

the 4096 x 4096 w&h in the text encoder nodes, does changing those give better widescreen generations?

even 4:3 sometimes catches attention problems for me

wider always does

[upbeat summit]

afaik officially 4096x4096 should be the setting for the best quality independent of your resolution

[trim orbit]

i was thinking maybe they needed to be ratio'd too

[nimble heart]

@timid sonnet for the love of the old gods turn on latent previews. it shows exactly where it happened.

[upbeat summit]

yeah I tried that, but it didn't give me any improvement

[autumn forum]

why are you skipping so many steps lmao. only skip like one or it leaves wayyy to much noise

[trim orbit]

okay good to know. i won't waste cycles on it. thanks

[nimble heart]

latent previews should be on by default tbh they cost like 1% render time at the most

[timid sonnet]

how do i turn them on

[nimble heart]

its a flag when launching comfy. --help shows them all, latent is in ther somewhere forgot what it was

[timid sonnet]

okay ty

[upbeat summit]

https://github.com/comfyanonymous/ComfyUI#how-to-show-high-quality-previews

[autumn forum]

god i love making a bunch of people and looking at there tiny little distorted faces

[nimble heart]

also try the bf16 vae while you're at it. improves performance since you're not using the fp16 patched one

[timid sonnet]

@nimble heart Where do I get the bf16 vae?

[nimble heart]

it's a flag

--bf16-vae

[timid sonnet]

Okay cool

I love how chatgpt contradicts itself constantly

[nimble heart]

reject character sheets, embrace anatomical diagrams

[trim orbit]

chatgpt has no concept of what it's saying. it's only predicting the next most possible character

[nimble heart]

i mean when you think about it our brains are just a bunch of neuron-dense meat predicting the next commands to send to our body

[trim orbit]

a much different and less understood structure

[nimble heart]

it's really quite interesting when you think about it.

your friends and family? just meat.

president? congress? just meat.

[trim orbit]

we communicate by flapping meat and squeezing air through it

[glad grove]

is the bot gone? noooooooooooooo

oh wait maybe im blind

[trim orbit]

lol

[spring fulcrum]

Does anyone know how I should set these settings. Every time I enable this I get a light grid on the image... Is there a way to enable this and have it benefit the image. I would imagine its there for a good reason and could help me in some way?

[shy kelp]

is there an nsfw filter on 0.9? I can't get it to worky ?-?

[spring fulcrum]

It seems to do NSFW just fine for me

[nimble heart]

on the bot there is

[shy kelp]

im running locally

hmm im trying but it never works, sometimes it does but it looks really fucked up

[nimble heart]

there's no filter per se but the literal pornography was removed from the training data so it'll need to be tuned back in

[shy kelp]

ohhh okay gotcha

that makes a lot of sense

[nimble heart]

so it's only like tasteful photoshoot nudity

[shy kelp]

mm okay gotcha gotcha

[nimble heart]

might not even need a full tune tbh. probably gonna see 85 nsfw loras pop up on Civit.ai the first week after sdxl 1.0 is out

[trim orbit]

honestly some of those guys might be ready on day one

[nimble heart]

if they prep the dataset beforehand I saw in this thread Sytan and someone else making loras in like 2-3 hours

[dusk mica]

faster than the speed of light 😂

[nimble heart]

i feel like by the time I get the news that 1.0 is out people will already be making porn with it

[timid sonnet]

@trim orbit I don't know you very well, so I won't make an assumption based off of little knowledge, right... but I will ask you.. Do you feel there is a need to always be right? Even when you give a response that is not the same context as something someone else has said, you never admit to be wrong or even admit to understanding the other person's point of view.

Rather so you rejustify whatever you said.

[trim orbit]

this got stupid

[timid sonnet]

Exactly my point

Text book narcissist, moving on

[peak dove]

Say good morning to Scarlett ❤️

Ana de Armas ❤️ ❤️

[peak dove]

Arpillera

[clever verge]

Nice work!

[peak dove]

Thank you

[clever verge]

What style are those?

[peak dove]

The Prompt = beautiful woman in the style of a peruvian arpillera, victo ngai, henri rousseau, vladimir kush

Try it for yourself - I'm using ComfyUI SDXL 0.9

[clever verge]

Thanks. And the pink ones?

[peak dove]

a drawing of a woman in a suit and shirt, in the style of retro feel, charming character illustrations, dark white and pink, editorial illustrations, digitally enhanced, I can't believe how beautiful this is, eye-catching pop art, collage, retro, american film, in the likeness of grace kelly, hitchcock, horror

Swap out Grace Kelly for whoever - it is in fact two prompts - the 2nd starts at 'pop-art'

[clever verge]

Yeah it’s amazing how well sdxl responds to prompts like that.

[peak dove]

Some 3D Paper Art/Papercut

[timid sonnet]

@peak dove

[peak dove]

^..^<

[timid sonnet]

lol

im trying to make a papercut image right now

we will see how it turns out

[peak dove]

Yes, use phrases like paper art, papercut, 3D paper art etc

[peak dove]

Cool

[timid sonnet]

not what I prompted though

[peak dove]

ComfyUI? The Bot? Clipdrop? NightCafé?

[timid sonnet]

comfyui

[peak dove]

That gold\dragon looks a tad flat - papercut tends to work where there's a lot of colour variance - otherwise its more like a silhouette

Papercut and intricate detail are best kept apart imho

I'm getting some of these prompts from the MidJourney Magazine ...

[timid sonnet]

I think its the changes i made to my workflow

[peak dove]

Papercut needs large and obvious shapes - faces, butterflies, flowers, leaves etc etc

[timid sonnet]

\

[peak dove]

Better ...

[rose kelp]

Nice 👍

[spring fulcrum]

Thanks

[clever verge]

I got the frame but could be removed with a negative.

[sand idol]

I've heard that its possible to train a lora on 8gb with the sdxl models. Has anyone succeded in this?

[spark pewter]

yes, using kohya. dim 8 lora training just works with proper parameters. dim 16 is possible with 8bit optimizer.

[sand idol]

thanks. gonna give it a try now

[spark pewter]

these are the settings that worked for me:

accelerate launch --num_cpu_threads_per_process 1 sdxl_train_network.py \
    --pretrained_model_name_or_path=<model path> \
    --vae=<vae path (fp16 fix: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)> \
    --train_data_dir=<training data path> \
    --output_name=<output filename> \
    --output_dir=<output path> \
    --network_dim=8 \
    --network_alpha=0.5 \
    --resolution=1024 \
    --train_batch_size=1 \
    --enable_bucket \
    --save_model_as=safetensors \
    --learning_rate=1e-4 \
    --optimizer_type="AdamW" \
    --mixed_precision="fp16" \
    --cache_latents \
    --save_every_n_epochs=1 \
    --network_module=networks.lora \
    --lr_scheduler="cosine" \
    --vae_batch_size=2 \
    --max_train_steps=3200 \
    --caption_extension=".txt" \
    --cache_text_encoder_outputs \
    --network_train_unet_only \
    --gradient_checkpointing \
    --xformers \
    --min_snr_gamma=5 \

[nimble heart]

return escapes are for the weak. Put everything on one line like an Alpha

You can't make spaghetti if you only have one noodle

[clever verge]

Can do decent unicorn's too.

[clever verge]

Looks like there's slapping going on and they are having a blast. 😄

And some stacked paper art with some depth to it.

[normal geyser]

The simple process of MacOS generating picture use SDXL (CoreML), It's very slow, but it's faster than Pytorch.

[bronze arch]

when is sdxl planned to be released?

[nimble heart]

soon

[ionic dragon]

[nimble heart]

Emad will release it when he finishes Tears of the Kingdom

[spring fulcrum]

Does anyone here have a favorite sampler and recommended amount of steps to go with it?

[spring fulcrum]

lol

Does emad or any of the developers of Stable Diffusion ever come in this channel or on their discord server at all for that matter?

[arctic bloom]

yes

[nimble heart]

sdxl?
base only 20-30 steps of dpmpp karras
base+refiner using noise return 20-30 steps total ddim split about 2:1.

[clever verge]

I get decent results with a simple euler normal with 25 steps base and 5 steps refiner

[nimble heart]

Joe monkey Dustin and comfy are in here a lot. Emad showed up once and immediately a large argument started which was a sight.

[nimble heart]

Euler and ddim normal are basically identical

[clever verge]

but that's not the upscaler

[nimble heart]

ddim + uniform looks slightly better imo

[spring fulcrum]

What was the arguement about? Im sure it was quite a sight

[clever verge]

I've run that too, with 5-10 steps less.

[spring fulcrum]

and yes I run SDXL

Do you know what the settings at the bottom of this node are for I currently have it set to none because when I use it I see a grid over the image Are there settings that it should be where it is enabled and does something beneficial for the image???

[nimble heart]

the moment emad showed up someone started having a meltdown about how SAI was the cause of all his life's problems cause SDXL is worse at making vaginas or something, don't quite remember

[nimble heart]

I don't think that's a normal comfyui node? id check the docs from the repo you got it from

[spring fulcrum]

lmao what a terrible problem...

It is for the ultimate upscaler node

it is a custom node

[nimble heart]

I mean based on what it looks like I assume seam fix adds overlap between tiles areas to blend them together

so width blur and pad are just the sizes for that in pixesl

[nimble heart]

think emad said "hug a tree" or something and end of life guy took it very personally. that was the only spicy part the rest was just kinda sad

[spring fulcrum]

hmmm I'll have to mess around with it until it works

[clever verge]

euler normal vs ddim uniform 25 steps base, 5 steps refiner on both

[spring fulcrum]

I do tend to always go back to ddim

[dusk mica]

I always use the sampler with the longest name because it sounds it works the best

[clever verge]

There's a good rundown on how the samplers work here: https://stable-diffusion-art.com/samplers/

[spring fulcrum]

I just read that whole thing... that was great.. Thank you... very solid info to consider. I am going to try generating some images with dpmpp_2m with the karras scheduler and see how that goes

[vestal flume]

Ancestrals are much better for bounce prompts (like [thing1:thing2:0.2] )

Send it

[peak dove]

Some 3D Paper Art and Papercut

[vestal flume]

Controlnet text?

[jolly creek]

Actually I was surprised. A1111 with prompt: photo of bates motel neon sign, cinematic light, fog, creepy, mystical

[vestal flume]

So 1.5?

[jolly creek]

SDXL

[vestal flume]

Holy sh^*

I wonder if that was just luck or a sign of things to come

I haven’t heard any talk in regards to xl and text

[jolly creek]

Actually I have seen people post here with accurate text generated

[vestal flume]

Off subject but is Janis a nordic european name?

[peak dove]

Cool - love The Text!!!
I live near where Alfred Hitchcock is buried

[jolly creek]

Yessir

[peak dove]

Twice out of 100,000 tries in MidJourney I got text to work - one was a VOGUE magazine cover; the other a black cat at SALEM

[vestal flume]

pats self on the back that was 100% a a guess

[jolly creek]

Haha intuitive guess

[peak dove]

<-------------- heads off to try ControlNet text @AUTO1111

[vestal flume]

I had to title a multi-star civilization for a book im writing, and i looked up some names from that area, i wound up with clestrarieabiorn

[peak dove]

Janis Ian (a songwriter) lived in Alaska

[jolly creek]

quite consistent

[peak dove]

Cool with the text

[vestal flume]

The unystarria of the clestrarieabiorn

[jolly creek]

The name is one of the most popular ones in Latvia as in Greece

[peak dove]

Mostly spelt Iannis in Greek

[jolly creek]

I tried to Google it, no results found 😁

[jolly creek]

Yes, spelling is bit different

[vestal flume]

Send it twri

Scard em off

[clever verge]

Send what? 🙂

[vestal flume]

I just saw yoiu were typing so i decided to be weird

[clever verge]

What other than Ultimate Upscale flows are there to upscale an SDXL image?

[vestal flume]

It seems to be the conclusion i land on a lot (be weird)

Probably all the usual upscales? They should work on any image. Once sdxl 1.0 hits and it is implemented into the normal workflows of the big webuis

[clever verge]

What works best upscale image or upscale image using model?

[vestal flume]

What ui are you using?

[clever verge]

comfy

[peak dove]

I'm finding it extraordinarily hard even using ControlNet to get anything like BATES MOTEL - LATES MOBEST?

[vestal flume]

Really? I find controlnet text almost impossible to not work

[clever verge]

I get some weird faces in the semi-solid background when upscaling. 😉

[vestal flume]

What is your prompt-controlnet balance like? And which mechanism are you using? Try canny

Or lineart

[vestal flume]

Shes a maaannn baby

I like the silver outfit

[peak dove]

Will try again ... 🙂

[jolly creek]

Have not really tried ControlNet myself

[vestal flume]

Its an insane power

Any picture you’ve ever drawn (basically) can now be turned full-color, 3D, it’s just insane

[jolly creek]

its crazy right

Works with other words as well

[vestal flume]

It’s almost overwhleming, like i havent taken the time to really sit down and do anything huge with it just because there is sooooo much i could do

[jolly creek]

Right, we are coming to a point where the only limit is indeed your imagination

[vestal flume]

I wrote a whole book (well, most of it) when i started trying to find a way to illustrate it (beyond my own skills) and then stumbled upon all this. The book has been completely set aside as far as writing goes, but the the illustrations and 100000000% better than what i set out for. I knew there was ai imagery out there i just didn’t know anything about it.

[peak dove]

TaDaaaah! I got somewhere ... 🙂

"I am a Prompt Wrangler; and my horizons are set to Beyond Imagination ..."

[peak dove]

A bit staid and a bit fusty wouldn't ypu say?! 😄

[jolly creek]

Adventures take its tolls 😄

[peak dove]

I'm bringing him into the 21st Century - I am a Promptfuturist Wrangler; and my horizons are Beyond Imagination ...

... but my Vlad AUTO1111 has just slipped from 1it/s to 10 due to High VRAM Usage ... ?!

[vestal flume]

You the master bates motel prompter

[peak dove]

Sweep up after yourself!!!

[vestal flume]

10s/it you mean?

I now controlnet uses some vram but not THAT much.

Plus vlad is really good on vram

[peak dove]

How to stop "GPU High Memeory Utilization" warning in Vlad AUTO1111?

[vestal flume]

Though for some reason doesnt output as good as invoke does (probably due to the cost of optimizations)

(At least such is the case in my laptop)

[peak dove]

1s/it then drops down to 10s/t with High VRAM Usage warning ...

I might try Tiling ...

[vestal flume]

Hmmm. Thats odd, tortilla

[peak dove]

One annoying thing about Agent Scheduler in VLAD - if you interrupt the process and re-boot - the cmd window opens and the Agent Scheduler is processing away in the background without a Localhost window being open!!! 🙂

[vestal flume]

Its the same thing in any webui

The localhost is juet frontend ui, all instructions are being carried out in the backend, the frontend or localhost is just the controlboard

[peak dove]

Trying to stop Agent Scheduler is like trying to stop a bull with his nads on fire 😄

[vestal flume]

So a set of instructions wont require the controlboard to continue

You can just just go into agent scheduler and hit stop on all your schedules

Should work

[peak dove]

Nope - one last pesky job will persist and persist (and persist) and ... did I say per .... !!!!! 😄

My version of Vlad - Stop and Skip do not function (obviously there is something adrift ...?)

[vestal flume]

Is it an older version?

Try updating the extension

[peak dove]

Most recent

[vestal flume]

Most recent..l oh i dont use that one, not sure

[peak dove]

Updating Agent Scheduler - OK, willgive it a go

It could be that my 4 year old RTXb2070 8Gb VRAM has finally aged itself out of use ... ?

Although with ComfyUI it works fine

[vestal flume]

Idk i have an 8gb 3070 mobile edition (so not nearly as powerful as the desktop version) works just fine, albeit not blazing fast—it’s 100% good enough

[peak dove]

SD Automatic1111 and ComfyUI work good - just SD.Next Vlad AUTO1111 SDXL does not (but it used to ...) !

[vestal flume]

Hm. Maybe it’s the update.

I notice some webui’s put out shitty images

I compare them all using the same exact parameters

Seed, prompt, etc

I also always do it comparitive images in that same sense any time theres an update, to make sure everything is up and up, and no BS is occuring. Will do it again when XL hits

[peak dove]

ComfyUI is now having issues - dropped from 1.5s/it to 35s/it ... goto re-boot 🙂

[vestal flume]

Odd

[peak dove]

Yes, possibly hardware issues 🙂

L8r

[vestal flume]

If reboot doesnt work try it offline

L8

[peak dove]

OK, re-booted - nothing in the Agent Scheduler ...

1.38 it/s

26 seconds to SDXL image after a reboot

2nd image 2.62 it/s ...

Looks cool, but took 60 seconds to generate ...

54 seconds

[hearty ginkgo]

Are you using a 3070?

[noble shoal]

20 sec. on a 4070 1024x1024

[spring fulcrum]

The Old Man and the Sea

[peak dove]

Great movie with Spencer Tracey

[spring fulcrum]

yessir

[noble shoal]

Sorry i meant 1024x1224

[peak dove]

RTX 2070 8Gb VRAM

1024x1024 yes

[spring fulcrum]

Is there a place on this discord server where we can put images we created locally that people can vote on?

[west breach]

you can participate in the COW

[spring fulcrum]

What does COW mean... I know its not the animal.... or is it?

[west breach]

challenge of the weekend

[spring fulcrum]

So there is a specific image to attempt?

[peak dove]

I've seen enough "milkers" to believe its the dairy version 😄

[west breach]

yeah, check out the #1087493421209485393 channel for info on this weekend's challenge

[peak dove]

Yawn - tiny worlds aren't my thing,sadly! 🙂

[clever verge]

Text pretty good

[peak dove]

ComfyUI's back and running ... here's some Venice Carnival

Spooky Carnival

[spring fulcrum]

Challenge accepted

@peak dove Those are pretty cool

[noble shoal]

Here is another try of the Fisherman. This time i have used Realistic Vision V4 as Refiner.

[spring fulcrum]

I like the lighting and the colors look really accurate

[noble shoal]

Feel free to steal the workflow from the image 🤗 . Comfy btw

[clever verge]

I'll have a go too.

[noble shoal]

You probably need to change the base model since i use a pruned fp16 version

[spring fulcrum]

all right it looks like we will have some challengers.

Sent into outer space to find another race

[past cedar]

What was the prompt? Image is sick! btw

[spring fulcrum]

thanks

[spring fulcrum]

As for the prompt that I used:

Linguistic Positive:

A breathtaking image of an astronaut suspended in the ethereal beauty of a nebula, her suit's reflective visor capturing a universe ablaze with color, a cosmic dance between human exploration and the infinite wonder of the cosmos.

Supporting Terms:

Sony Alpha 1, Sony FE 16-35mm f/2.8 GM, sharp focus, highly detailed, rich colors, vibrant colors, trending on Artstation, 4k

Fundamental Negatives:

logo, words, worst quality, low quality, blurry, cropped, lowres, jpeg artifacts, signature, watermark, username, artist name, trademark, watermark, title, multiple view, Reference sheet, Out of Frame, cartoon, drawing, illustration, 3d render, plastic, blurry, grainy, low-resolution, shallow depth of field, bokeh, text, signature

[noble shoal]

Interesting

[peak dove]

Thank you

[peak dove]

How do you change/setup the refiner at all?

[noble shoal]

Do you use Comfy?

I basically use the output of the base model and make an image 2 image pass with another model.

[peak dove]

Here is a Google Drive Link to the prompts which were in MJ Magazine ISSUE 3 - all MJ Prompts are © MJ, but freely usable https://drive.google.com/file/d/1B_SYuQFAg1SrKmjhuXacymxHGyfsDna1/view?usp=drive_link

[spring fulcrum]

Interesting story:

My wife got a gerbil once, It was a male gerbil, Its nuts accounted for like 40% of his total body size.... We used to call him Captain Big Nutz... Not sure if anyone needed to wake up to that imagery but I thought it was a funny story. 😆

[peak dove]

Looks like a Cappybarra?

[noble shoal]

It is 😄

[spring fulcrum]

It reminded me of my gerbil though

[eternal fog]

Hamsters are like this as well.

[noble shoal]

But yeah, here are your nuts

[whole latch]

That looks great!

[noble shoal]

@indigo carbon What the hell is "Maybe_a_good_model2" ? At least this shows up if i drop the image into comfy.

[indigo carbon]

I tuned SDXL0.9 on A1111, it's slightly better than base

[peak dove]

My SDXL approach has mainly been artistic - hanging with you guys I'm sure I'm gonna learn some real good photographic technique ... ?

[noble shoal]

Was "artistic lightning" intentional? 😅 I was wondering why the outputs looked to weird.

[indigo carbon]

wdym weird? the images are incredible

[noble shoal]

There is a small difference between lighting and lightning. At least it influenced my outputs

[indigo carbon]

your outputs are probably weird because your using that weird ComfyUI workflow

[peak dove]

Closest I got so far was this - breathtaking night landscape, universe in a bottle

[peak dove]

Mmm pecans

[noble shoal]

Don't know if that was supposed to be an insult 🤔 , but your images are fine. Mine had lightnings in it. I never tried clip skip in the SDXL model, i might give that a try.

[indigo carbon]

the last intension I have is to insult someone. sorry if I was a little too aggressive.

[noble shoal]

It's ok. Lets not judge us by our workflows. The output is what really matters.

[clever verge]

Don't know what this creature is called.

[peak dove]

I like the lightning!

[noble shoal]

The lightning really does wonders

[noble shoal]

Since we are still doing the glass jar universe thing. Here is a sticker

[prisma wyvern]

Is it just me or voting system of only 2 options of a or b is not sufficient to get accurate judgment of images produced 🤔 There should be 4 choices A, B, Both Suck, Neither Match Text Prompt 👍👎😭😱

[indigo carbon]

man, I remember when SDXL beta couldn't do shit like this

[indigo carbon]

this is now by far the best model

[indigo carbon]

Also, I just made a comparison of the base SDXL0.9 vs my altered version

[amber fulcrum]

are 1.5 compatible with SDXL?

[soft zealot]

Have added my SDXL workflow package to Civitai (and credited anyone such as @high skiff ) . My Main input has been to Add automatic daily folder creation (like Automatic 1111) and to add "Jumpers" to enable/disbale parts of the flow

https://civitai.com/models/114733

[indigo carbon]

https://tenor.com/view/tight-tuco-breaking-bad-gif-7215689

[eternal fog]

How and what have you altered?

[indigo carbon]

Math, using the merge feature.

[peak dove]

I think the real problem is that voting cannot truly separate each model one from another that StabilityAI can back as the surest model forthe future. Users are equally wowed (almost) by all three models

[indigo carbon]

I did dreambooth a little bit, then I merged it with base with a complex math equation that altered the base to be slightly better.

[eternal fog]

Ok, I mean those changes you could probably get via prompt tuning or different seeds. But nice to see people trying different things.

[indigo carbon]

well, it sure does do better than base. Like seen in the comparison I sent earlier. the only difference between the two images is the checkpoint, and well, it did get better

[eternal fog]

I mean that's just a few images. If you wanted to properly gauge if it was better you'd have to probably generate thousands of comparisons.

It's also subjective

It's nice though for that image

[indigo carbon]

idk man, I find myself using that model I made all the time, it's just slightly better- like I said. also many people agreed

[peak dove]

I'm thru with messing with A1111 - I will wait until 1.0 🙂 ComfyUI is more than enough 😄

[indigo carbon]

Isn't it supposed to release today? I could have sworn something happens on the 23rd

or was it 26th?

something like that

[tribal gale]

26th

[indigo carbon]

release?

or sdxlA sdxlB sdxlC vote?

[tribal gale]

that's when the super stage is hapenning, tune in

[wicked dagger]

comfyui has a very basic implementation of wildcards - the dynamic prompts library is easy to extend to create a custom node though

[vale eagle]

I have a question about prompting. Since we are using clip g, we use sentense for the prompt. But does the prompt on the left side of the sentense would be more "important" to the model and the words which on the right side of the sentense would be less "important"? So, we should write the reversed sentense to get the important words first?

[sharp robin]

For reals this time? No fake-sies?

[sharp robin]

U can use g and l

[vale eagle]

I knew. I mean "in low poly a photo of banana" would better than "a photo of banana in low poly" to emphasize the low poly style?

[sharp robin]

The clip understands both. There’s no right or wrong. Using common sentence structure would probably work best. A low poly photo of a banana is what I would type. Try both see what u like best. This is a tool for creativity, art, no right or wrong.

[vale eagle]

thanks for sharing.

[ionic dragon]

trained on 0.?

[ionic dragon]

sdxl on 26th?

[vale eagle]

a low poly photo of banana, a photo of banana in low poly, in low poly a photo of banana

[trim orbit]

last i checked there was no official date this time, this was emad's last message bout a release if i'm not mistaken https://www.reddit.com/r/StableDiffusion/comments/153907z/sdxl_will_be_out_in_a_week_or_so_phew/

[autumn forum]

Yes

[trim orbit]

there's no official word on the 26th so i'm going to assume it's meant to be delayed once more

[eternal fog]

I mean the guy who just said it is SAI Staff

[vale eagle]

Never trust anyone again for any release date.

[autumn forum]

I’m getting my hopes up for this one cuz no sense in being negative about it being delayed again. I wanna be excited for its release not down!

[trim orbit]

yeah and?

that's not official word. that's just a dude with a role chatting

[eternal fog]

I mean you could say that about anyone in here lol

[trim orbit]

lots of official staff said the 18th was legit too

[eternal fog]

So did Emad

¯_(ツ)_/¯

[trim orbit]

all the pinned posts for emad say "a week or so"

[peak dove]

Some interesting prompts - MJ, Blue Willow, Wombo, Imagaine, Dawn, Mage, Leonardo, A1111, ComfyUI, fusionbrain, ClipDrop, NightCafé https://drive.google.com/file/d/1B_SYuQFAg1SrKmjhuXacymxHGyfsDna1/view?usp=drive_link

[trim orbit]

i'm expecting an august release at this point

[eternal fog]

The event for SDXL on the 26th has been up there since Emad last came in here after the delay

They will look silly (again) if they don't have something on the 26th

[vale eagle]

Does it require access right on purpose?

[peak dove]

If you've been voting your As and Bs - then you probably already have a good idea of what to expect of SDXL 1.0

[peak dove]

I thought I was just supplying a link - says anybody with link can share

[trim orbit]

i think people who are hyped for 26th expecting a release look the silliest. there's no official word about 26th being the actual release date. why build hype when we're not even a week out of the hype train blowing up in people's faces last time?

[eternal fog]

Apparently not, it requires you to request access to it.

[peak dove]

OK, then I will grant access ... 🙂

[indigo carbon]

there are 3 SDXL1.0 models tho

[trim orbit]

expect the event to announce an august release

[eternal fog]

I have no interest in hype. People get all worked up lol. They've got 0.9 already.

[peak dove]

I know, and its just me, but they look so much like the exciting stuff I'm already doing in 0.9 - this is not going to be an earth-shattering upgrade 🙂

[trim orbit]

well you're here arguing about speculation with me. saying it's certain to release on 26th. feels like hype ot me.

[eternal fog]

Never said it was certain, was mearly pointing out that the people mentioning 26th was soley because someone who works for SAI said it.

This is probably why they've not said anything prior

[trim orbit]

they said the 18th and built the hype

told a secret exclusive channel about the expected delay though

[eternal fog]

Emad's a huge hype man anyway, they do this every time there's something new. They know what happens.

¯_(ツ)_/¯

[trim orbit]

so that the hype can continue

yeah emad tends to lie a lot it seems. he's gotten better.

[indigo carbon]

idk, I manage to get results way better than the bots with 0.9, so idk. I'm most excited for the efficient finetuning tools they promised to release

[peak dove]

What earth-shattering upgrade over 0.9 can 1.0 be? I mean, realitically - at least in terms of the finished product?

[trim orbit]

the bots use 0.9 right. the UI's typically don't. there's a lot of user error on the 0.9 side. that's not such a bad thing though. it's understandable given we have zero documentation about how ot use it right

[shy kelp]

it was delayed again?

[peak dove]

Yes, I think there is a greater level of backroom excitement - new architecture, implementation techniques etc etc - than what 1.0 will actually produce - which I believe will be every bit as good as 0.9

[autumn forum]

Your a party pooper. Let me be excited.

[eternal fog]

The bots are on variants of 1.0 release candidates at the moment, with various settings.

[autumn forum]

No

[indigo carbon]

idk man, whatever I'm using it's better than the bots =\

[trim orbit]

it's a foundational model so a lot of what they're targeting is outside of your scope of use. it's meant to be a model to cover all use cases.

lora compatibility is a big one here.

i hope they release another report on these final stages of 1.0

[peak dove]

Anybody like me watch "Line of Duty" on the BBC? Remember how hyped that ending was - deflation 😦

[peak dove]

Yes, but the images will be just as beautiful - from an artist's POV (i.e. me) the excitement seems to lie in the workings - so its a techfest! Way to go I say

[trim orbit]

we're not even a week out of the last failure to launch. temper your expectations for your own sanity. i'm not forcing you to unhype yourself at all. i'm just giving advice and maybe it's ringing true with you

[indigo carbon]

like, I never saw the bots make stuff like this when they used 0.9, yet, I'm using it on A1111 and it does this incredibly good

[eternal fog]

You don't know what the bots are doing behind the scenes

They are using settings you don't know about, they could even be changing the prompts

[peak dove]

If you've been working the Bots (A and B) then you already know what is possible in any forthcoming version of 1.0

[trim orbit]

i think the bots even randomize settings like step count and cfg. so that they can get better voting data.

the bots are meant to be a candidate search and they're farming as unbiased results as they can with them

[autumn forum]

The bots are using 1.0 plus refiner of the 3 candidates which will be decided based on voting data on the 26th then they will release it(hopefully). The bots are randomized sliders like step count, cfg, ascore, etc. so outputs may vary.

[trim orbit]

cherry picking one good image from your personal workflow is just selection bias

which is precisely why the results are so randomized for the bots

[eternal fog]

Yes, that's what I was saying, you don't know what each output has for the settings.

People having been moaning that they are "Worse" sometimes, but that's kind of the point.

[peak dove]

The selection process fell into an almost Gaussian Distribution

[autumn forum]

Correct sorry I didn’t know that’s what you meant! My bad

[trim orbit]

i think too the bot gives intentionally worse images, to see if someone selects those. they'd obviously disqualify those people's votes as they're clearly trying to hurt the training data

[regal heath]

which other GPU from NVIDIA has 12+GB VRAM without being a 90 xd

might have to stick with the 3060

[trim orbit]

if you're buying a new gpu, get 16

at least

[indigo carbon]

hell no, never by old gen gpu

[eternal fog]

Pretty sure he means 16GB VRAM

[indigo carbon]

way less efficient

[trim orbit]

lol i didn't mean a gtx 1660

god no

[eternal fog]

I keep eyeing up the 4090, but the price is so silly

[trim orbit]

same. they went up in canada, but i'm still like "but i could"

[eternal fog]

And then I'd be tempted to upgrade my whole system to get the most performance

[trim orbit]

i'm at a 4080 right now. might sell it and flip to 4090. was considering dual 3090s

[eternal fog]

I'm fine with my 3080 at the moment

If they drop the 4090 price a bit more I might bite

But it's still over £1500

[indigo carbon]

yeah, whoever buys an old gen gpu when the new gen has better cores and is incredibly more efficient is a fucking idiot

[trim orbit]

yeah. workign with a 16gb 80 series is plenty of leg room i think. i haven't hit many roadblocks.

learning a lot within these limitations. wish i had 24 or 48 gb though

[eternal fog]

I mean depends how much money you have.

[trim orbit]

30 series are still super viable even though they have old cores. i wouldn't get all ego fantastic on someone because thye bought a gpu i wasn't a fanboy of

[eternal fog]

I still see loads of "cheap" pre-builds with 1660's in

[trim orbit]

seriously. anyone getting into this hobby now is one smart fucker. no matter what hardware they're limited by. enthusiasm is key

why be alll ego centric about it

[indigo carbon]

still, a 40xx gpu that costs X is better than a 30xx gpu that costs X. unless your getting ripped off

[ionic dragon]

ok

[trim orbit]

4060ti vs 3060ti

[eternal fog]

Not if you can't afford a 40 series. Also that's not strictly true. Some of the 30 series cards are better than the 40 series because they've gimped the memory bus

[trim orbit]

hmmmm

your logic fails hard in this case @indigo carbon the 4060ti is just absolute balls and nobody should support it

[indigo carbon]

it sure is better than a 3060ti, and they got the same MSRP price.

[eternal fog]

Has anyone found a consistent way yet, to either latent upscale, or pixel upscale through img2img where it doesn't wash out all the details.

[sudden cliff]

If you need the VRAM, say if you're running LLMs, you're better off with a 3090 than any 40x0 card with <24GB VRAM

[eternal fog]

This argument doesn't work at all when 1 is new and another is years old

[trim orbit]

it literally has a slower clock rate and last i checked it generated images slower

msrp is meaningless in today's markets

[indigo carbon]

WHAT, now that's stupid

[indigo carbon]

that's my exact point =\

[trim orbit]

you didn't know? the 4060ti's launch is a total catastrophe. nvidia shenanigans

[indigo carbon]

it's really slower than a 3060ti? man, I'm dumbfounded rn

[eternal fog]

From what I've seen every card in the 40 series stack is overpriced and the performance is not a good upgrade from a 30 series from the price.

[dusk mica]

just wait for the 5000 sries

[eternal fog]

Except the 4090 where the performance is extremely good

[sudden cliff]

Yeah 4090 seemed like the one good card if you have a good reason for it

[eternal fog]

They've said probably not for at least 2-3 years. So you'll be waiting a while.

[trim orbit]

i've got a 4080 with 16gb and generate images and train faster than a 3090. i'm fine.

[soft zealot]

thing is all the testing has been done on In Game Performance.

Has anyone tested how it works with Stable Diffusion etc?

[sudden cliff]

Maybe a bit, but then you can't run llama-3xB locally on GPU

[eternal fog]

It has lower memory bandwidth, so it will be worse

[indigo carbon]

makes sense. newer cores usually means better performance. I'm still skeptical on what you said about that a 3060ti is slower than a 4060ti

[trim orbit]

you should be skeptical. it sounds ridiculous. that's how nvidia built it though

[eternal fog]

Go look at all the youtube reviews saying not to buy it

[indigo carbon]

wtf did nvidia think????

[trim orbit]

$$$$

[sudden cliff]

dreaming of that 48GB Nvidia x0x0

[trim orbit]

titan ada when?

[eternal fog]

They are skimping out of the memory bus to reduce cost hoping that eventually DLSS3 saves them

[trim orbit]

when the titan ada was rumored in december, i started saving then.

[indigo carbon]

the xx90TI is supposed to be the new name for a titan.

[soft zealot]

depends on what you're measuring against surely?

Has anyone empirically tested fow the 4060ti 16GB compares for AI performance against a 3060ti 8gb ?

I doubt it.

Yes I agree, on papaer due to the memory bandwidth reduction it is theoretically worse but has that been proven>?

[trim orbit]

the ti i saw a spec sheet released and it was going to have 24. the titan was due to have 48

i'll buy the first 48gb card available even if it's an amd tbh

[eternal fog]

I'm not talking about just AI Performance I'm talking about in general. The new Nvidia cards are not good value for money.

[trim orbit]

(affordable card)

[spring fulcrum]

Did they cancel the Titan?

[indigo carbon]

idk when they made the first 90TI cards they said it's their new name for a Titan GPU

[trim orbit]

it was never officially announced. just rumored. might've been canned as they develped it

[soft zealot]

Depends on Use Case.

Eg I'm not a gamer but I am interested in AI perfomance

[sudden cliff]

what about A6000 or similar?

[trim orbit]

They'e made a Titan card since then. RTX titan came out after 90ti

[sudden cliff]

oh

[trim orbit]

heh

[indigo carbon]

was it faster?

[sudden cliff]

I feel like the first 48GB Nvidia will be >half the price of a A6000

consumer card

[spring fulcrum]

into the pit springbonnie

[dark jasper]

trained on wdxl

[trim orbit]

perhaps. if they delay it longer i'll have more money saved

[eternal fog]

I don't see Nvidia ever releasing a consumer card like that. There's no money in it. Keep it in the enterprise sector where they can sell them for £15k+ a card

[sudden cliff]

Well 48 sells for way less than 15K punds, they sell much cheaper configurations at 48

[trim orbit]

games are going to start launching using large language models for NPCS. there will be a consumer need

[eternal fog]

That was just a ball park number, they aren't sold at prices a regular consumer can afford.

[indigo carbon]

mean while nvidia finds new ways to scam the shit out of people with their 4060ti shenanigans China is trying to make chipsets that work on photons to multiply the speed of circuits by ~6000x

[trim orbit]

there's also jim keller's company, tenstorrent. They may be doing something interesting too. a second generation lineup is likely to be announced from them

[sudden cliff]

I don't see any reason over long term Nvidia doesn't start selling larger VRAM enterprise cards. Releasing a 48GB consumer card might fear your Meta and OpenAI into buying ultra high end 160GB cards etc

[trim orbit]

https://tenstorrent.com/grayskull/ these are a little out dated now, but he's absolutely a contender to watch out for

[ionic dragon]

do you know it personally?

[sudden cliff]

If the wee little babby company I work at can afford to buy H100s for ML work without fear then I don't see why Nvidia wouldn't make bigger better cards going forward

[indigo carbon]

also, yes, they are really trying LOL

[trim orbit]

i've seen them in action in a lab i was getting a tour of. ML research out of UVic. near the end of the pandemic situation. I don't have any real world experience with those cards though. i don't think they're meant for this realm. they're a product of a different time.

Jim Keller is the man who devised AMD64 instruction set that all modern x86 cpu's use

[trim orbit]

he also consulted with apple to devise their specialized silicon architecture and their machine learning neural engines.

[trim orbit]

also tesla, to design all their custom silicon architectures

also intel to help them redevelop their tooling process in their new factories

also amd too again, because he laid foundational work for the ryzen architecture

[indigo carbon]

yeah companies are trying to make circuits work in a similar way fiber optics work. photons move at literally light speed. this is the current goal of so many companies

[trim orbit]

totally unrelated and i don't think optical chips will be a consumer product for many years yet. we'll likely see different physical substrates for electronics before optical circuits.

electrons also propegate pretty quick

[indigo carbon]

nintendo should take notes

[trim orbit]

/sues

[indigo carbon]

oh yeah, no one is allowed to play or talk about nintendo LOL

[trim orbit]

i dont write the rules heh

[indigo carbon]

nintendo can go fuck themselves

[trim orbit]

/lawyers intensify

[indigo carbon]

i will make as much pokeballs made of fire and there is nothing you can do about it

[spring fulcrum]

What was your prompt for that last pokeball?

[trim orbit]

that pokeball contains the legendary flashichu

[spring fulcrum]

What did you use for a prompt for that last pokeball you posted

[indigo carbon]

same prompt as the first one

it wont be as detailed with base SDXL0.9 though

[noble shoal]

Couldn't get the pokeball working, so i went with a stuffed animal. 🤷

[autumn forum]

idk if yall know comfyui well or not but can anyone tell me why when i save a photo in comfyui it goes to my downloads and not the outputs folder?

[noble shoal]

Do you use the "Save Image" node?

[autumn forum]

yes.

using the save image node doesnt actually save it automatically for me. i have to right click save image.

[soft zealot]

I prefer the WAS Node Suite Save Image Node

Has more options including file type to save as

[autumn forum]

yeah i may try and use that

[soft zealot]

that sounds like you have "Preview Image" rather than "Save Image"

[autumn forum]

nope its a save image node lol

[ionic dragon]

do you know anyone in their team, i wanna talk with them

[autumn forum]

i used the image save node from was and it acted like it saved it but its nowhere to be found.

[trim orbit]

not really no. it was just a tour i got the opportunity to go on through a friend

their ML stuff was for really mundane stuff like systems controls. one project he explained a little about was a greenhouse control system

[soft zealot]

what does it say in output_path?

[autumn forum]

"[time(%Y-%m-%d)]"

[soft zealot]

preficx that with ./ComfyUI/ouput/

you ve probaly got a folder in your comfy root with todays date

[autumn forum]

nvm... i found the folder.

[autumn forum]

im soo dumb haha

[trim orbit]

we're all tabula rasa to begin

just remember, babies are always dumber than you. if you need that confidence boost.

[autumn forum]

theres 5000+ images in that folder i never knew i had for some reason. jeeez

[soft zealot]

lol

[lusty raptor]

since we're on the topic, the default location for image preview node is in comfyui/temp

[trim orbit]

wait it saves the previews as files? does automatic do that too??

[soft zealot]

everyone quickly checks temp fles for pr0n

[lusty raptor]

yes and yes. webui saves to appdata\local\temp, but windows should be clearing those as it hits quota i think

[soft zealot]

yes ,A1111 saves them to a temp file somewhere in your user profile that it allegedly deletes from on exit

[trim orbit]

lol i was mostly concerned with space as that would be 3x my output folder possibly. but yeah, if you're wanting to hide your pr0n digital trail that's a huge concern too. great point

oh yeah i sure can rely on windows to maintain itself properly without my intervention 🙄

[eternal fog]

If you use the load image node, it also saves a copy of that img into your comfyui folder as well

puts them in the input folder

[trim orbit]

good to know

[eternal fog]

temp it clears when you close it, or next time you open it

But it doesn't clear the input folder

[trim orbit]

likely just once per image? instead of a copy every gen

[eternal fog]

once per image

[lusty raptor]

never noticed that, thanks

[trim orbit]

okay yeah i just checked the local temp folder. its not so bad. got spooked for a second

[eternal fog]

Auto1111 saves all the images as well in your Appdata temp folder

And it didn't used to clear them up either

[trim orbit]

thats where i'm at. appdata/local/temp

[eternal fog]

I think it finally cleans them now

Comfyui has it's own temp folder in it's directory

Although if you've been using Auto1111 for a while and haven't cleared your temp directory you'll likely have thousands of empty folders like this

[trim orbit]

i need to migrate my os system drive to the new faster ssd still. my data is getting sloppy and all over the place. i have my diffusion software spread across 3 drives at this point. oh wait, discs? what are SSDs? they're not a disc, no spinning disc, or a drive, a motor spinning a disc.. device i guess. yeah. storage device.

/rant

[trim orbit]

few hundred but yeah you called it

[soft zealot]

guessing CivitAI needs updating to take into account that with SDXL (0.9 at least) there are 3 prompts 🙂

[indigo carbon]

ah, you see, that's what you think. I did something far more interesting here

[soft zealot]

generated in A1111 rather than ComfyUI or just added the details in manually ?

[indigo carbon]

I tweaked A1111 a little. still no refiner though. but yeah, that's probably the best way to use SDXL rn, well, at least in my opinion

[soft zealot]

I just find Comfy faster with my 1080ti then an optimised A1111 which I was using. I also really like the way COmfy clears out of VRAM and swaps to sytem RAM as soon as its done unlike A1111 which holds everything in vram until you shut it down

but thats my view 🙂

[trim orbit]

It seems a lot of the builders in the community aren't taking the 2 positive prompt approach seriously. comfyui is the only system i've seen that allows multiple text encoders to be accessed in the front end. I feel like most UI's have entrenched their UI's and aren't willing to make any significant changes anymore. Only cosmetic.

[indigo carbon]

I also tried that method. I find SDXL to work the best on the new A1111 fork, by a lot. I actually did a benchmark, I got 7it/s on images I got 5it/s on comfy

[soft zealot]

In fairness to them I can understand that approach fiven SDXL isn't finalised yet and it appears uncertain whether the 2 model approach will be retained in the final release

[trim orbit]

2 prompts is towards the dual text encoders. not the 2 stages

the refiner has 2 clips as well

[indigo carbon]

not true, the version of A1111 I tweaked has default 2nd universal positive for added detailing.

it IS 2 prompts, just one of them gives it an idea for details

[eternal fog]

There's not really such thing as a "Universal Positive"

[indigo carbon]

almost it improves most stuff

[eternal fog]

That second prompt can cause huge changes to images, locking it into 1 prompt is just bad.

[trim orbit]

what branch is that on? i was using release candidate a bit last night and dind't see it

[soft zealot]

regardless of any specific SDXL related things it cannot be denied that Comfy is far better with its's memory management compared to A1111

[indigo carbon]

@soft zealot

A1111 is a little faster

at least on 40xx cards

[soft zealot]

I did say I found it faster for me on my 1080ti, I didn't generalise 🙂

[indigo carbon]

ah, sorry.

[soft zealot]

YMMV

[trim orbit]

comfy is faster on the 4080 i use

[indigo carbon]

weird. probably because A1111 uses cudNN and you don't have the corresponding version. if that's not the case, IDK.

[trim orbit]

for python 2 that doesn't matter. and i've got the right corresponding sdk involved. 11.8 and 12. the difference is 2 seconds ~

[soft zealot]

I will admit though that f you have less than 32gb of system Ram then Comfy may cause you issues (thats based on me having 64gb and regularly seeing 30gb in use when Comfy is running (likely because it switches the models to System Ram from VRAM)

[indigo carbon]

that's your problem then. the optimal Python version is 3.10.9

[eternal fog]

I've got 16GB and Comfy is fine

Sometimes it stutters when loading large images through VAE, but other than that it's fine

[autumn forum]

i have found that programs use as much ram as possible to increase performance. more ram you have the more itll use

[trim orbit]

i meant pytorch 2 mb. i've got 3.10.9 installed but any 3.10.x works. newer is more optimal and i should install the newest 3.10 build

[indigo carbon]

still, pytorch 2.0.1/cu118 exists

[soft zealot]

I appear to be using Python 3.10.10

[trim orbit]

the speed difference between comfy ui and automatic isn't because of the abstraction layers. it's because comfy is architecturally more optimized

[indigo carbon]

A1111 said 3.10.9 is the fastest, but unperceivable difference

[trim orbit]

the cudnn dlls were only needed when automatic used pytorch 1 builds, since it didn't distribute the newest dlls with it

[trim orbit]

if it's not perceptible it's a margin of error. you likely misunderstood his reasoning

[indigo carbon]

it does, idk what's wrong with your A1111 inference then. I'm getting 7it/s on a1111 where I get 5 on comfy.

[soft zealot]

TBH I've stopped using A1111 since I made the initial switch to COmfy

[trim orbit]

it/s isn't really meaningful. speed to image output is.

[visual glade]

are you sure you don't have one of the a1111 options like the one that removes the uncond for lower steps?

[trim orbit]

different cfg, steps, samplers, resolutions can all affect it/s

[soft zealot]

I just look at the total number of seconds to generate an image

that's my metric

[indigo carbon]

I know. I'm saying I get better and faster results with identical settings with A1111

[trim orbit]

it/s is good for benchmarking and comparing values against identical parameters

[ionic dragon]

Oh ok, do you know how I can reach them apart from emails, as they don't workout

[soft zealot]

i would love to get it/s

Sometimes I get s/it lol

[trim orbit]

haha yeah. sometimes you turn settings up and see some maniac step times

[visual glade]

I'm pretty sure your 7it/s is wrong somehow because even people with stronger cards are not getting that on either UI

[trim orbit]

i feel like it's using default a1111 w/h values like 512x512

[eternal fog]

If it was it would look like shit

It just generates blobs at that resolution

[indigo carbon]

I'm getting those speeds with images like this

[indigo carbon]

👆

[eternal fog]

However I'm not convinced A1111 calculates it/s properly

It also has a bunch of shortcuts enabled by default

[visual glade]

you are doing a hiresfix right?

[trim orbit]

yeah. i fee like information is being made up and woven a bit. like the 7it/s claim came from 512 gens, but they pick out a quality generation that saw 2s/it

[visual glade]

where is your it/s coming from, the first pass or the second pass or the whole thing?

[trim orbit]

there seems to be a bit of bovine scat weaving. skibainindiboodbodapdbop i'm a scat man

[indigo carbon]

first pass, seccond pass gets 3

[trim orbit]

comfy brings up a good point that 7it/s is strange claims

[eternal fog]

Speaking of highres fix, have you worked out a workflow yet for doing a highres fix that's consistent and doesn't kill background details.

[visual glade]

and your first pass is 1024x736...

[trim orbit]

"i got shoes with rockets in them and i went to the moon last night" kind of claim

[indigo carbon]

I literally sent proof. whatever man

[trim orbit]

of 7it/s?

[indigo carbon]

yeah. that's what I said. first pass on ComfyUI is 5it/s

[trim orbit]

sure it wasn't 5s/it and 7s/it?

[visual glade]

on comfy at 1024x736 I get: 6.15it/s on my 3090 TI

[indigo carbon]

yes, I am sure. if it was 7s/it it will get VRAM errors

[visual glade]

but for benchmarking when we talk about it/s we usually refer to 1024x1024

at 1024x1024 my 3090 TI gives 4.5it/s

[trim orbit]

that doesn't make sense. when i was generating on my vega 64 i was seeing 10s/it

might be a different reason you're getting vram errors

[soft zealot]

admittedly this is a SD1.5 model at 512x512 and used the same sampler/cfg etc in both
A1111 reports 2.69-2.93 it/s
Comfy reports around 3.6 it/s
1080ti

[visual glade]

yeah that's the expected behavior because I optimized my unet like diffusers did

[soft zealot]

I'll be honest I havent compared SDXL between Comfy & A1111 but I see no reason (unles sI'm told otherwise) I wouldnt see similar behaviour

[indigo carbon]

you probably won't see any difference. most optimization are ment for newer GPUs

[soft zealot]

don't disagree but for poor old me surreing with my sub £200 watercooled 1080ri Comfy fits me better 🙂

[trim orbit]

thinking back to the 1080 release, i honestly thought the next one ws going to be 1180 and the 1280 and 1380

so naive

shit. that means my 4080 is actually a 1380.. sppoooky

[soft zealot]

lol

I still say the 1080ti was a mistake from Nvidia, still hold s up well even after 6 years and 3 more generations

[jade hill]

(thanks for report)

[indigo carbon]

I didn't even notice haha. I saw it on the side of my eye and thought nothing of it LOL

[jade hill]

it was in multiple channels, slowly trickiling in

[autumn forum]

yeah. not good lol. thanks for taking it down so quickly 🙂

[soft zealot]

hmm when the 108ti was launched in 12017 it was an MSRP of $699 which adjusted for inflation is $870 today.........

[indigo carbon]

also, hey man. I miss those times when you made a model for every POW lol

[soft zealot]

The launch price fo the 4080 (non ti) was $1199

[jade hill]

yeah, sorry I took a break from the online world, and PoW isn't a thing currently. if/when I get back on the training model Horse, I'll sure update that complete edition I did

[indigo carbon]

also, is model datasets from now on will just be top AI generations?

[clever verge]

Looks mean!

[regal heath]

does anyone here have a tesla P40?

[fresh path]

why?

oh they are quite cheap now

[shy kelp]

What is sdxl?

[delicate grotto]

@jade hill did something happen?

[elfin cobalt]

This speaks for itself, I figure~

Bot works quite well now.

(And has nothing to do with a GAN, but historical raisins.)

[delicate grotto]

What is the aesthetic 20 exactly?
New property in sd?

[elfin cobalt]

It's a parameter to the refiner, yeah.

[eternal fog]

Does anyone use the tiled ksampler custom nodes in comfy?

It does a much better job when used for upscaling, except it leaves patches of noise for some reason.

[delicate grotto]

Mmm odd, didnt see that yet in comfy, how can i access that?

[eternal fog]

like this is a normal attempt at an upscale. It's ok, but not brilliant

But with the tiled custom node it looks better overall, especially the character, but it leaves bits of noise

[molten gull]

hi guys 🙂 whats new ? i've missed the last 4 days

[clever verge]

Where do you get that custom node? What model are you using for the upscaling?

[eternal fog]

https://github.com/BlenderNeko/ComfyUI_TiledKSampler

4x-Ultrasharp

[clever verge]

Had it in my favorites already it seems. Something that probably was on my todo-list. Thanks for reminding me!

[eternal fog]

Ah looks like this is a known bug, I'll try a different sampler https://github.com/BlenderNeko/ComfyUI_TiledKSampler/issues/9

Of corse that sampler gives by far the best quality result

[lilac wren]

@indigo carbon
wow you are in good shape today

[lilac wren]

I have 32 Gigs, and sometimes I still have a problem like this. Comfy_UI keeps the 2 models in the RAM, then after a few generations, or when I interrupt one and then generate another, I don't know what it does but I have more than 20 GB of RAM used, and the generation becomes very slow, where normally it only takes a few minutes, I think @visual glade that there is a problem on this side.

[soft zealot]

Personally I like this behaviour. Its better than the way A1111 does it which is to keep everything in VRAM.

At th eend of the say system RAM is relatively cheap and upgradeable, VRAM isn't.

[lilac wren]

I didn't say it was bad, I'm just saying that sometimes I think Comfy keeps the same model twice in the Vram, and that strongly degrades its performance

Personally, I often interrupt a generation, because I don't want it upscaling or refining on a bad basis. And often, when I launch another generation after that: it takes a lot longer, probably Comfy doesn't always properly purge the RAM/Vram

Dear @visual glade, if you ear my voice 🙂

[clever verge]

Something is cooking up

[bronze arch]

does sdxl need more vram

[autumn forum]

this guy is posting this in all channels..

not really a bad thing but idk why lol

[autumn forum]

not necessarily

[upbeat summit]

I've seen this device a couple of times 😉

[eternal fog]

Must be a magical circle

[eternal fog]

The refiner does virtually nothing for anime style images does it

[upbeat summit]

you can remove details with it

[eternal fog]

Yeah that's not helpful lol

[eternal fog]

That's the main thing that frustrates me most about this setup

When you try img2img upscale the base model will kind of work but has a tendency to wash out background, then the refiner does a fantastic job of fixing eyes and hands, but it washes out all the texture

Most of the time your better off not even doing a 2nd pass

But then you have dodgy eyes 😢

It does work on stuff like this though

fixed the eyes, nose and mouth

It removed some buttons from the shirt, but in this case they look weird anyway

[visual glade]

so it's the interrupting that causes it?

[eternal fog]

Ok this isn't working perfectly, but it seems acceptable for most things now

[shy kelp]

When fine tuned sdxl models start releasing, will they have to bundle their own “refiner” or is it recommended to just use the refiner that comes with the base model instead? Or no refiner at all?

[eternal fog]

No idea yet, I guess it depends on what they do with the finetune

[shy kelp]

Got it

[eternal fog]

You can make decent things without it

[shy kelp]

I was just wondering if it was like a “one size fits all” kinda deal

[eternal fog]

It just helps a lot with human details like eyes, mouths, teeth and fingers.

[shy kelp]

Yeah I saw quite a difference in interior design with and without the refiner

[eternal fog]

I'm not sure how much like the Add Detail loras it is

Maybe we won't need it at all once someone makes a good one of those

[shy kelp]

I did this generation before

[shy kelp]

Left is without refiner, right is with refiner

[viral aurora]

i know this doesnt belong here, but i couldnt get answers anywhere so far. so please someone help me out
i am getting this error when using "finetuning" method in kohya_ss: fine_tune.py: error: unrecognized arguments: --no_half_vae. It worked for me in the past just fine, but after the update i simply cannot figure out why i am getting this error, this only applies for "finetune". Dreambooth and loras are working just fine.Hope someone can help me out.
i cannot figure out where --no_half_vae, args are in fine_tune.py and if removing it would even help.
also i have no extra arguments specified in the gui itself, its for 1.5

[shy kelp]

The flowers look a lot better in the refiner version

[eternal fog]

Is this on the SDXL Branch?

[viral aurora]

its just after the last kohya update, but i didnt pull a specific branch

[eternal fog]

Not sure then, I've only ever used the SDXL Branch and that command works fine for me

[viral aurora]

but its exacly the same if i try to fine tune sdxl or 1.5 no matter what, i get that --no half vae error

also tried fresh install and nothing seems to work

[eternal fog]

They might have changed it / broken it on the last update. I'd suggest logging an issue on the github page.

[viral aurora]

already done, no asnwer so far. any idea, how could i revert to pre-sdxl update?

[eternal fog]

The SDXL stuff is on a different branch

But you can just get the commit id for a version that worked and do

git checkout commit_id

[viral aurora]

not sure about that, because i did a fresh basic install and all sdxl stuff is included, so have no idea

[shy kelp]

I see the bot is a bit less crunchy today, thats nice

[upbeat summit]

cool style! try wind as a negative for the hair to calm down (if not intended)

[eternal fog]

It had a super sayian prompt in, so the crazy hair was expected lol

[upbeat summit]

haha makes sense than

[lilac wren]

No, sometimes even without interrupting it happens.

[spring fulcrum]

I have a 4090 and I get about 7it/s when the refiner is running with 1024x1024 with 16 tiles... It goes up to about 19it/s if I change that to 512x512 on the refiner with 64 tiles, which goes fast but does take longer because of more tiles... this is on comfyui which I really like.

[inner ruin]

I have trained a bunch of cool face LoRAs but SDXL loses resemblance when prompt is complex. It also doesn't extrapolate the body as well as SD 1.5 or even 2.1 did. Any tips on how to keep resemblance?

[visual glade]

try running the refiner on the whole image without tiles

[vernal cloak]

We will likely get other extremely powerful models built on top of the SDXL 1.0 soon anyway. You will very like have less issues then.

[spring fulcrum]

How do I do that. do I just increase the refiner dimentions to 4096 x 4096?

[visual glade]

do the same upscale but without a tiled upscaler

[spring fulcrum]

It doesn't like when I do that it went oom.

[sharp robin]

Download more vram.

[inner ruin]

lol glhf fixing the text encoder with a fine-tune

[vernal cloak]

Can't be any worse than 1.5. I'm sure it would out perform 1.5 with a finetuned model.

[spring fulcrum]

lol.. ya ill do that right after I download more hard drive space and upload my document on downloading a faster internet connection

[sharp robin]

Glad I could help

[sullen bison]

It's getting delayed again, isn't it?

[prisma wyvern]

Does anyone know if SDXL 1.0 will fix hand issues? ie thumbs coming out of bottom of palms, palms facing inside out, holding objects wrong way in reverse grip, pointing downward instead of up, objects going through arms instead of holding in fists 👎😱

[sullen bison]

Knowing how they are neutering each new model they produce, hands might be worse

[boreal heart]

+1 for this. Hands is the biggest pain point for me too, wasting so much time to go through all the images trying to find the ones without hands problem. None of the lora or ti on the market can fix them

[urban fjord]

I feel hands has mostly become a non-issue. If 1.0 isn't fixing it then I'm sure fine-tunes will do. I just hope people do as good of a job with SDXL 1.0 as SD 1.5.

[prisma wyvern]

Have you tried to fix hands with ADetailer extension?

It supposed to fix faces and hands automatically

And give better results than inpaiting

[solar tusk]

any updates on release date?

[nimble heart]

mount a google drive folder as swap space and you actually can download more ram

[urban fjord]

Given that they have a stage event the 26th that seems like a likely release date.

[half thicket]

Does someone try to parallel inference with SD models?

[visual glade]

you could do parallel inference by sending the cond and uncond to different GPUs

but it's difficult to split it up more than that

[boreal heart]

I tried it along with many other loras/ti. They worked only in some occasions but also change the whole style vastly the whole time. It really depends

[tall lion]

needs access rights

[tall lion]

I totally agree. It would be least incremental and some additonal tuning maybe. Do not expect the output to be earth shatterlingly different

[boreal heart]

I only wish the hands would be less shattering 😆

[spring fulcrum]

I could also just assign space to paging file

[hearty ginkgo]

Lol I got like 64gb min and 128gb max in my paging space thingy

It ain't do shi

[spring fulcrum]

thats my point.. just get more RAM

[hearty ginkgo]

I got 64gb of ram

U mean vram?

[spring fulcrum]

either way. there just isnt a way to make it effectively more than what it is in a substantial meaningful way

[hearty ginkgo]

Ye

[spring fulcrum]

I Just shaved off my beard tonight.... Bad idea... I haven't seen my fat face in like 2 years... the beard is going to stay next time

[nimble heart]

Boring. How are you going to flex on plebians by telling them that yes, you did in fact download more ram

[spring fulcrum]

lol

[ancient roost]

Was the model updated?

[nimble heart]

No. Emad is still playing Tears of the Kingdom

[ancient roost]

Ah lol.

[nimble heart]

also low key don't bother if you have plentiful ram. just unnecessary SSD wear, and it might even be faster since some things like file caches get offloaded to swap/page before you even start to run low, so removing that will keep them in actual RAM instead

[chrome flicker]

this week sdxl 1.0 release?

[spring fulcrum]

Doubtful

[chrome flicker]

[ornate elk]

anyone around who has set up sdxl with automatic111?

im following this guide here to set it up as an extension https://stable-diffusion-art.com/sdxl-model/

but it ran for several hours, then I stopped it. is this normal?

[spring fulcrum]

this might as well be skim milk

[ornate elk]

sorry?

[nimble heart]

it'll come out when emad finishes Tears of the Kingdom, so about 2-3 days

[ornate elk]

I dont understand

[spring fulcrum]

Until the full release... there is no refiner for it...

its like getting half the experience

[vale eagle]

Do we know how did SAI caption the images? Or the dataset already have the caption?

[ornate elk]

it refers to a refiner on that page

[spring fulcrum]

oh... in that case i take back what I said while putting my foot in my mouth

[ornate elk]

...

[spring fulcrum]

nvm

[ornate elk]

check the section "Using the SDXL demo extension"

[spring fulcrum]

If you can run SDXL in A1111 with the refiner then it should be a pretty good experience

I really do hope that SDXL 1.0 drops this week but I'm not holding my breath

[ornate elk]

what changes are expected in 1.0

[autumn forum]

My cat told me it’ll come out this week but idk if I trust him…. Hmmm

[fossil patio]

well the longer it's in development the better it gets, I personally never understood the 'sooner the better' outlook. i'm more 'the better the better'

especially now that it's being refined by us

[spring fulcrum]

You have a good point

[autumn forum]

It’s all done. Sense there were 3 models they decided to put em in the bots for us to vote on cuz we didn’t have a clear winner last time. And maybe they did a little extra ✨refining✨.

[fossil patio]

i'm voting based on three points 1) prompt consistency--does the prompt match the image 2) vibe--does it make me feel good, 3) is it aesthetically appealing (in that order)

well narrowing 3 down to 1 is refining

technically

[spring fulcrum]

Those are good measures to take into consideration

[fossil patio]

indeed!

what ui are you planning on using XL on ?

probably a silly question--like you can only choose one lol

[spring fulcrum]

I'm using ComfyUI... not sure if you are asking me.

[autumn forum]

Comfyui! I was a a1111 guy but comfy is just too fun lol

[ornate elk]

making progress

[spring fulcrum]

Whats your connection speed?

[ornate elk]

it's fibre, can't test rn as it's maxed out obvs

[spring fulcrum]

Possible COW submissions... What do you guys think?

[tawdry knoll]

is it just me, or images are getting over sharpened?

[spring fulcrum]

They might be

[molten gull]

morning guys 🙂

say, is it possible yet to train embeddings in sdxl ? and is 1.0 out by now ?

[molten gull]

nobody knows ?

is it possible to make sdxl embeddings now ?