✨｜sdxl

[upbeat summit]

Most of the time when I want to work on a prompt build I just go through all the pop cultural experiences I had in my life - art, film, music, video games. Or when I want to capture a mood I think of places I've been.

Lighting is for me one of the most important aspects when creating visual media. I mean if you look at a game like Octopath Traveler. Oldskool pixel graphics combined with real time lighting, It's really pretty.

I also use technical terms that describe a certain camera technology, rendering engine or art style. I do research on some topics to know about it - like how a certain look was achieved.

Since the training data includes vast amount of data from the web, it knows many of those terms.

And all the other times I'm just looking at what you guys are doing and getting inspired by your work.

[urban fjord]

I must say I'm really impressed with how well SDXL does inpainting out of the box.
I can't figure out how to continue from the original image instead of colouring it grey as I'm unfamiliar with inpainting in Comfy but despite a bit of inconsistency with seeds it works a lot better than 1.5.

(Click on the images as discord is cropping weirdly)

[visual glade]

set latent noise mask instead of VAEEncode for inpaint

[uneven dove]

lmfao @ that cropping

"great inpainting", my first thought

[winter raptor]

I’ll keep that in mind, thanks!

[uneven dove]

"they're the same picture"

[upbeat summit]

you're welcome. anytime.

[sudden cliff]

deterministic denoising for any seed!

[fresh path]

[fresh path]

easiest job sdxl will ever get

[shy kelp]

bro what? lower resolution noise? this example is not even upscaling, it's just a clear proof that the refiner ONLY works on 100% leftover noise from the base model and does not work on added Gaussian noise from img2img generation. https://imgur.com/a/5VefX0M

[sudden cliff]

That's not how it works on my machine, so 🤷‍♂️

[fresh path]

bro got a quantum pc or something?

[eternal fog]

What we arguing about now?

[sudden cliff]

No it just does denoise added noise, I do wonder if there is a subtle difference but it looks like something is wrong with his pipeline

[high skiff]

Who's, mine?

[sudden cliff]

Whatever beric is using

[shy kelp]

it isnt?

try this. generate an image with lets say 20 sampler steps 100% base model. Then do a simple img2img with the refiner model at 35% denoise.

Generate another image where the the first 0-13 steps is base model and the last 13-20 steps is refiner (basically same amount of noise as 35% afaik)

Compare the two results.

[high skiff]

Not sure if it's mine or not

[sudden cliff]

I've already done this countless times. I agree the nature of the images is different, but I don't end up with blurry images like your example

[upbeat summit]

also a couple of times I have been pleasantly surprised how effective SDXL's CLIP can process your prompt. the robot image is pretty much how I imagined it. storytelling is a lot more fun now.

[shy kelp]

the example i provided there is not that much blurry, it just has way less detail and quality...

anyways i guess the copium in here is too high. ive yet to see a single result of hires fix working as well as SD1.5 despite all your complete denial of it not working well at your particular settings.

[urban fjord]

It seems to mess up the background a bit but is perfectly capable of decent changes I don't think I would have managed without controlnet in 1.5. The user interface isn't the best to play too much around with inpainting but I hope it will manage things I 've failed at in the past.

[trim orbit]

the feedback from bots will likely get used going forward in some way

it's a proven approach to training

[fresh path]

yes probably

but I think they are more interested in what the community achieves with the tools stable diffusion are providing

[eternal fog]

How does this look for a latent upscale?

I'm playing with a few things to see what I can get it to do.

[sudden cliff]

I wish I could automagically copy paste some comfy nodes to Windows clipboard and paste them into a different workflow

[vale eagle]

Save as Template

[sudden cliff]

I'm guessing that's in some extension? I have basic repo installed

[vale eagle]

But the template only stored in browser. I want it could be share to others

[vale eagle]

It is in basic repo

[civic sigil]

How do you do that

[sudden cliff]

Oh I see it, right click

huh!

[shy kelp]

looks like you used very high denoise which kind of works but not great

[eternal fog]

You always had to use high denoise for latent upscaling

[shy kelp]

the issue is moreso with 0.5% denoise and below

[upbeat summit]

slighty overcooked look for me, but the textures look good

[shy kelp]

oh yeah but img2img upscaling/hires fix isnt a latent upscale

[sudden cliff]

well that worked well!

[eternal fog]

You can do it either way. The other way is to pixel upscale and then try and get back from details with img2img

[civic sigil]

So then how do you paste it in another workflow

[sharp robin]

Yo! this was amazing, before this i couldnt gen anything bigger than 4k res, with this i can do up to 6k, maybe even 8k and it is significantly faster, thank you❣️

[civic sigil]

Thank you

[shy kelp]

you can do what either way?

[eternal fog]

What do you think highres fix is? Because people keep using that terminology for clearly different things.

[sharp robin]

6k img

[shy kelp]

generate an image with no leftover denoise->upscale it in pixelspace-> encode it back to latent and do an img2img generation on it.

[eternal fog]

Ok yeah.

I'm meaning you can do highres fix either that way or as a latent upscale. If you are going off the Auto1111 terminology, it had an option for either.

[shy kelp]

you can but hires fix dosent work well for sdxl 🙃

[eternal fog]

I mean it seems to work reasonably for me and I'm trying to make it better.

[shy kelp]

didnt you show an example of a latent upscale?

[eternal fog]

That one was because I'm testing stuff

I can do pixel upscales too, I'm just tweaking some settings

[shy kelp]

well see you in a week when you lost your sanity trying to get the right settings (cause there is none)

[civic sigil]

Yeah I had bad luck with latent upscale

[visual glade]

you want to see a latent upscale with SDXL?

[eternal fog]

I mean, it's not perfect, but left is just a normal pixel upscale. Right is after Img2Img

[mystic cosmos]

What is style in sdxl dreambot? Is this prompt addition or embedding?

[shy kelp]

well latent upscale only really works if you use a high denoise, in which case it generally degrades composition the higher resolution you do

but you can probably do like 1.25x-1.5x and have decent composition

[shy kelp]

sure

[delicate grotto]

mmm
im abit confused, why does it give me a fully purple output?

[eternal fog]

What's wrong with the example I posted above. What sort of thing are you expecting.

[civic sigil]

The first sampler needs to return with leftover noise

[delicate grotto]

i tried that

[shy kelp]

some samplers and schedulers do not work together

[delicate grotto]

same thing

[civic sigil]

I mean why are you doing the full denoising onboth of them

[eternal fog]

I'm going to kill my GPU with a 4x upscale to try prove a point here

[civic sigil]

Make the start at step on the second the same as end on the first

[shy kelp]

you cant start generation at step 0 and have add noise disabled.

[eternal fog]

RIP OOM

[shy kelp]

actually i have no idea what you are even trying to do there 😂

[delicate grotto]

yep, it is the same

[delicate grotto]

what do you mean?

[civic sigil]

Is it between 0 and 10

[eternal fog]

I mean I think the one I posted is fine and no one has replied saying it's not so....

[shy kelp]

he's either gotta return with leftover noise or add noise, otherwise theres no noise to interpret. but he didnt even increase start at step from 0 so its kinda wierd

[eternal fog]

We're into System Memory lmao

[shy kelp]

or you* sorry

[delicate grotto]

im trying to return him to the 2nd with leftover noise, but either way it is like it doesn't care

[eternal fog]

Because then I can't test what I'm doing. I can't use refiner properly with the Ultimate SD Upscale

You have options to start and stop at certain steps?

[hollow halo]

With the advanced ksampler yeah

[eternal fog]

No

With Ultimate SD Upscale

It's not

People keep saying different things....

[eternal fog]

Why do I get different results with the same usage of it then across both

No I mean using the KSampler and KSampler Advanced. Matching up the Denoise with what would be the same values on Starting Step and Ending Step gives different results.

[visual glade]

latent upscaling works fine

[eternal fog]

That's wasn't the issue, there's people bitching saying "Highres Fix" which is pixel upscale and then tidy up with img2img doesn't work.

But it works fine for me.

Latent upscales issue is that you have to use high denoise, so it changes the "Original" image

Not that it really matters as you are generating images from scratch anyway

[visual glade]

I prefer increasing CFG instead

[eternal fog]

That gets rid of all the wobbliness that latent upscale causes without enough denoise?

But anyway with "Highres Fix" What's wrong with this then?

[upbeat summit]

please fix my faces, mrs. refiner

[eternal fog]

Like people are complaining but I don't see any issue with what I've just posted

[visual glade]

the images I posted have workflows if you want to check them out

[eternal fog]

Am I blind or something, how is that face blurry?

[hollow halo]

It really looks fine, not sure what you're on about

[eternal fog]

I mean I can do that too, lets 200% zoom on it

You are expecting way too much

You'd need to generate something like an 8k image for it not to be like that

[fresh path]

why troll

[hollow halo]

It really isn't... 4mp on camera looks like ass zoomed in

Okay well resolution is what it is

Those still look like shit

[fresh path]

u might get a good image at 12mp

[eternal fog]

@brave halo Can you generate an image on SD 1.5 that size and demonstrate what you mean by something that isn't blurry?

Because I'm genuinely struggling to understand

[sharp robin]

zoom without blur like these? 2.5 x upscale

[eternal fog]

I mean that one does have some blur around the eyes

But thats more caused by the colour

[vale eagle]

It just works fine with upscale. From 1360x1024 -> 5440x4096

[eternal fog]

Nah

blurry

(I'm taking the piss, just in case it wasn't clear)

[vale eagle]

because I didn't upload the original

[hollow halo]

@brave halo This one blurry? 4x upscale

[eternal fog]

zoomed in, weird marks on the eyes. Failure

It doesn't need to be the same image. I just want to see something like that with a none blurry face.

I'm trying Ultimate SD Upscale to compare

[pallid path]

holy crap img2img is so good

[eternal fog]

Are you ok? Your not complaining lmao

[pallid path]

wha

[fresh path]

mechanised tyrone

[eternal fog]

I want to test Ultimate SD Upscale properly @brave halo You said that If I run it through the base first and then the refiner after at a low denoise it will work the same way as leaving noise left over and then just doing the last few steps?

So you only use the refiner for the 2nd pass through it then?

When I tried doing that, it did make things like eyes nicer, but it made other things more smooth

[hollow halo]

Well he was supposed to have a cybernetic eyeball, and it did that but I didn't expect him to hold it lol

[eternal fog]

ok, I'll give that a go

[eternal fog]

I just tried the ultimate SD Upscale with just the refiner and it's worse than my method. However it's likey because the refiner doesn't change enough of the image around the face to make it clear

I think this is what people are forgetting when they complain

They've been using inpainting, adetailer, custom models

This is the upscale with just the refiner

Nah it's fine. At least you talk about it and try stuff. And don't just dismiss it and ignore it.

This is just with base

I think it's made the overall composition better, but the eyes are awful

[vale eagle]

Am I doing wrong? The ultimate SD upscale seperate the image as 4 part with the prompt.

[eternal fog]

Denoise is way too high

[uneven dove]

SDXL already does inpainting

[fresh path]

(today)

[vale eagle]

oh

[eternal fog]

@vale eagle You have denoise on 1, so it's making a new image for each tile. Use a lot lower denoise.

[vale eagle]

Thanks. Didn't note that.

[fresh path]

although, actually tomorrow, because you will need to download like 50gb worth of models

[grizzled warren]

PRUNE EM

FP16 FTW

[arctic bloom]

butbutbut fp16 means less details ^_^

[urban fjord]

50 GB of models? What?

[grizzled warren]

FP32 means you need at least a 3090...

[arctic bloom]

https://tenor.com/view/awkward-umm-what-what-gif-14694719

[urban fjord]

The refiner + base for 0.9 was a combined 18.5 GB.

[grizzled warren]

actully, scratch that. BF16 FTW ;D

as a large language model, I can't... emmm, excuse me xD

[grizzled rock]

Hello I’m new, I was curious if anyone would be able to guide me on how this AI works and how to use it?

[fresh path]

gb is gigabits... (nice save)

[vernal cloak]

wtf!? It wont run on 10GB of Vram? Not at all?

[uneven dove]

sure there's text encoders, a VAE, and a unet backbone

[eternal fog]

These are the two results, I prefer my way over Ultimate SD Upscale.

[fresh path]

yes it will

[arctic bloom]

when exSSDD will be released we ll be capable to run on microwave

[winter raptor]

I'm desperately trying to figure out a good upscale method, this didn't work very well

[winter raptor]

Whats your upscaling method?

[eternal fog]

Create an image with the base. Pixel Upscale with 4x-Ultrasharp, downsize it with upscale by 0.5, then run it through the base and refiner again.

I didn't lol. I gave it a batman theme and it did that.

[upbeat summit]

damn!

[hollow halo]

painfully, really isn't ideal in my opinion

[eternal fog]

I've not tried to work it out yet lol

There is an adetailer node pack, but it's confusing as hell to setup.

[hollow halo]

Basically you right click the finished image and do open in mask editor, then copy to clip space, load image, feed it into an inpaint node

[urban fjord]

Pretend it is an NFT and right click to save it.

[eternal fog]

disgusting

lol, I'm joking. If idiots want to pay you money for NFTs all the power to you

[livid basin]

prompt?

[winter raptor]

This is awesome!

[trim orbit]

lora training works well. i took a risk and tried 25 images at a high lr. 0.0002. not perfect but it works great then i just roop the face back into it. works for now. 4080 with 16

[mellow tendon]

This is my current Upscaling and Super Upscaling Workflow (SD Ultimate upscale with ControlNet Tile)
ComfyUI Workflow File: https://files.catbox.moe/708p58.json

[uneven dove]

Scooby Doo

[uneven dove]

got DeepFloyd to use the SDXL refiner as output upscaler now

that's from 64x64 to 1024x1024

so it's pretty good at upscaling

[eternal fog]

that dog has 2 tongues

[uneven dove]

all dogs have two tongues, darn it Arron

[eternal fog]

DEMON

[uneven dove]

now wider res

[upbeat summit]

that's an interesting pipeline 😄

[uneven dove]

i think it has human teeths

it's working pretty well

refiner: "damn it jim, i'm not meant for this"

[hollow halo]

Image should have a work flow attached.
I stepped away from my computer but I can check the prompt when I get back if needed.

[livid basin]

Thanks!

[uneven dove]

the original img

you can tell the refiner really rekt that lettering

[civic sigil]

Oh so deepfloyd can handle super low res

[uneven dove]

that's what it is, natively

96x64 there lol

[sudden cliff]

DF IF presents sub-pixel detail that latent anything will wreck sadly

[vast narwhal]

i've tried to upscale deepfloyd at the launch

[uneven dove]

that has the IF watermarker on it indicating you went from stage 2 to SDXL

i went from DF Stage 1 to SDXL

[livid basin]

Wow, SDXL training data must be super recent

[vast narwhal]

deepfloyd's launch

[livid basin]

As in within the last year

[uneven dove]

oh

[livid basin]

Because I just saw people post images with "rebecca cyberpunk anime", and that anime came out within the last year

So the training data for SDXL must have been recently collected, post Sept 2022 at the very least

[sudden cliff]

did this as a joke in SDXL... first set had this

refiner output if you want a really nice DeepFloyd text image by SDXL

[shy kelp]

could be trained on promo material aswell

[livid basin]

Possible, but I suspect it was on real artwork

[uneven dove]

RIP arm

[trim orbit]

marketing for edge runners was out long before it was. they did start making the dataset for the next models around the beginning of this year though.

[livid basin]

Ah fair point, I didn't really pay attention

[inner ruin]

I am once again asking if anyone got the diffusers LoRA XL script to work for them 😄

[livid basin]

Are you using the diffusers library?

[inner ruin]

yes, but the script gives really weird, noisy loras

[livid basin]

I'm going to take a look at the XL pipeline tonight, will have to generate like 20,000 images using the XL 0.9 checkpoint, and again using the XL 1.0 checkpoint (assuming they release it tomorrow)

[inner ruin]

nice! are you trying LoRAs as well>

[livid basin]

Ah its just for a machine learning dataset I'm making, it depends on if SDXL 1.0 release comes with the style weights (I assume they are LoRA? You know, the cinematic/photography/anime styles)

[uneven dove]

using base model works slightly better than refiner does

[static cradle]

sdxl 1.0 releasing tomorrow?

[livid basin]

If they don't release the style weights with SDXL 1.0 tomorrow, I'll take a look at training my own style LoRAs

[livid basin]

But offhand, that looks like a precision issue

[inner ruin]

how do you mean?

[livid basin]

I suspect it has to do with somewhere in the LoRA pipeline, you are using too low precision

[thin nova]

I think it's just the other clip model. Someone has a comfy workflow that lets you enter overarching style descriptions

[livid basin]

Ah, sorry I don't use AUTO or comfy, just the diffusers library

[livid basin]

How would the style descriptions work in practice?

[thin nova]

diffusers library has a pooled_prompts parameter for sdxl

that's all i really know

[elfin cobalt]

SDXL has two CLIP inputs. Two separate encoders.

[thin nova]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI
example given: "cinematic, fujifilm, national geographic, nature, snowy mountains, flowers, colorful, overcast sky, sunset"

[uneven dove]

😩

[inner ruin]

yeah I'm looking through dtypes in the codebase

[uneven dove]

SAI isn't doing it that way

the 'styles' are just prompt keywords

[livid basin]

SAI?

[uneven dove]

stability AI

[livid basin]

Ah, any reason you think they are keywords only?

[uneven dove]

they explained it many times

[fair crow]

Shoudn't it work on full fp16? Is there a part of the pipline that needs higher precision?

[livid basin]

Sorry, I haven't followed their messages

[uneven dove]

they're not even designed for SDXL, they're from an old experiment

[livid basin]

Would be curious if you can point me to a PDF or documentation on their prompt for different styles

[uneven dove]

ohh okay so DeepFloyd -> SDXL Base -> SDXL Refiner

@livid basin they do not have one

[livid basin]

Fair enough

[livid basin]

Guess I'll have to make some of my own "style" prompts that make sense

[uneven dove]

1.25hrs til 'tomorrow'

[craggy oak]

What is the copyright on these images used? The photos with and without the watermark

[uneven dove]

it's YOLO copyright

[livid basin]

The wild west

lol, you put it better

[uneven dove]

copywrong, more like

[livid basin]

copyright free until you get sued

I kid

but not really

[uneven dove]

freedom isn't free, it costs AIs like you and me

cutting edge technology

[urban fjord]

You can train on copyrighted images without infringement and currently you cannot copyright raw output from SD and similar.

[lament rune]

SDXL doesn't understand vampire fangs 😦

[trim orbit]

prompting issue

[livid basin]

That doesn't seem to be a legally decided issue yet

But I guess we'll get there legally when we get there

[uneven dove]

[craggy oak]

Thanks

[uneven dove]

my deepfloyd pipeline put it through the SDXL refiner before the base

[primal nimbus]

its the 18th in germany

[uneven dove]

utc

[eternal fog]

It will be the end of the day on the 18th anyway most likely

Probably the end of the working day

[urban fjord]

Which part?

[primal nimbus]

can someone explain to me what the refiner does?

[livid basin]

The training on copyrighted images

[primal nimbus]

and how it works, what the options are

[uneven dove]

jesus christ lmao people are really not willing to do basic reading anymore, eh

it's all explained on the refiner's model card on huggingface hub

[livid basin]

Isn't the only important parameter for the refiner is how much noise you add? 1.0 = do not preserve any original content; 0.0 = preserve original image exactly, no upscaling

[uneven dove]

no

[livid basin]

"refinement"

[urban fjord]

The refiner does no upscaling, if you want upscaling you need to upscale yourself.

[uneven dove]

you can simply test it and see for yourself

[livid basin]

Misspoke

"refinement"

[uneven dove]

oh the refiner can refine a blurry upscaled image

[urban fjord]

Yes, it can fix upscaled images but it isn't an upscaler.

[uneven dove]

you kinda need a higher strength than this but that image started out as 96x64

[nocturne dove]

Every image (natural photos by humans, art, etc) is a copyrighted image by default. I don't think there is any such thing as an image generator AI trained on images that are public domain or have some kind of explicit permissive license. There wouldn't be enough training data to get models of the quality we have now.

[uneven dove]

please let's not argue that here

[nocturne dove]

I have no intention of aruging. I was just commenting that up til now, ai training has just been done however the research group felt like.

[vernal cloak]

3.2.1.

[uneven dove]

when the deepfloyd kicks in

[trim orbit]

vampire fangs

[uneven dove]

DF -> Refiner -> Base. now i'll try base -> refiner

[vernal cloak]

You have't tried base XL with the refiner yet? xD

[trim orbit]

i'm all about that base

[uneven dove]

?

[vernal cloak]

What GPU are you using?

[uneven dove]

A100

[vernal cloak]

Wow. Nice. A Quadro is that?

[uneven dove]

wat the heck is happening

[vernal cloak]

Hmm?

[shy kelp]

so they never dropped this bad boy early huh

[vernal cloak]

No. Probably wont until everyone is awake. UK time

[winter raptor]

I’m hyped

[eternal fog]

My guess is like between 3pm and 5pm

[urban fjord]

This also means WaifuXL release I guess.

[trim orbit]

12am in london! OMG WHERE THE FUCK IS IT

🔥

[eternal fog]

Doubt

[trim orbit]

sorry for the language

[shy kelp]

I wonder if itll run on auto tomorrow.

[eternal fog]

I've got the WDXL Beta and it's nowhere near ready

[shy kelp]

anyone know if Vlad ever got img2img to work in .9?

[trim orbit]

i'm running 0.9 on auto /dev branch today

[vernal cloak]

apparently, the dev branch supports XL already

[uneven dove]

takes weeks to do it on a budget tbh

[winter raptor]

I honestly can’t see moving back to auto personally, it’s just so much rougher on my low spec computer

[shy kelp]

wheres the dev branch? is it easy to access?

[urban fjord]

Oh, this message on Huggingface made me think it was already ready.

[vernal cloak]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/tree/dev

[eternal fog]

Unless they have something extra they aren't sharing. The current model is only good at very specific things and it's not even that good at them.

[livid basin]

So if SDXL 1.0 is released, where will it be announced?

[trim orbit]

the WD beta is just an obvious patreon crowdfunding cash grab. theres no way it would be ready soon but they're going full on the hype train

[trim orbit]

#📣｜announcements

[pallid path]

it's not 18th in america + it's not guaranteed

[eternal fog]

They've not asked for any money. They've not even given it to most people.

[livid basin]

Ah I feel stupid

[shy kelp]

what am i looking for there

[vernal cloak]

same. Lol!

[craggy oak]

How many photos can you create per day? With and without the subscription?

[eternal fog]

Honestly for anime most people are pretty happy with what SD 1.5 can do. So no one is going to be rushing to SDXL for that.

[pallid path]

for clipdrop, infinite if you have an account (not paid)

[vernal cloak]

That is the whole dev branch m8. You just need the git repo from that, i believe.

[urban fjord]

Is automatic going to support the refiner or not?

[vernal cloak]

should do.

[civic sigil]

I dont think refiner should be default behavior for txt2img imo

[vernal cloak]

yh. SAI said it wasn't a few days ago when people started using it for that purpose.

[urban fjord]

I've gotten the best result using the refiner over just the base.

[vernal cloak]

I think you might be suprised. SDXL has incredibly good understanding of context in prompts. That alone will fish people in from the anime community.

[primal nimbus]

🎢

[eternal fog]

I mean not straight away. The base model can't do good anime consistently and there's obviously no custom models yet.

[boreal bough]

yeah - while it can do anime, it's a genuinely odd workflow

possible finetuning only the text encoder, and leaving the model alone will already significantly help

[vernal cloak]

Oh. You just mean the raw base model. Yh.. Idk tho. I thought it does pretty well. I also think it does better with anime if you don't use the style prompt for some reason.

[trim orbit]

it doesn't know booru tags as well as the usual novel ai derived anime models

[civic sigil]

There are some booru tags in Laion but not a ton

Idk if they used other sources

[vernal cloak]

Yh. Hopefully we don't need to wait too long before we get a tuned anime model.

[inner ruin]

I'm running it at float32 now on an 80gb a100 to validate that hypothesis

[livid basin]

Are you merging the weights in fp32 precision?

[misty sequoia]

Maybe ppl will use WDXL as a base for mixes and merges similarly to NAI?

[vernal cloak]

That could very well be the case. Yh.

[misty sequoia]

Although, I feel it's lack of understanding of booru tags might be a turn off for ppl that are used to prompting using booru tags

[inner ruin]

i made everything in train_dreambooth_lora.py fp32 precision

[livid basin]

Ah that's nice, going to be a memory hog I bet

[inner ruin]

and then ran it with

  --pretrained_model_name_or_path=$MODEL_NAME  \
  --instance_data_dir=$INSTANCE_DIR \
  --output_dir=$OUTPUT_DIR \
  --instance_prompt="photo of gbrlql male person" \
  --resolution=1024 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 \
  --learning_rate=5e-5 \
  --report_to="wandb" \
  --lr_scheduler="cosine" \
  --lr_warmup_steps=0 \
  --max_train_steps=2000 \
  --validation_prompt="photo of gbrlql male person running on a lavender field, summer" \
  --validation_epochs=5 \
  --seed="0" \```

[vernal cloak]

I personally thought the tags were quite restrictive. So it might end up being better without them. In the long run

[inner ruin]

yeah it's just to validate a hypothesis. We need to fix that script.

[urban fjord]

Booru tags is a bad way of prompting so not much lost there.

[inner ruin]

lmk if you actually get the script to work, though. Kohya doesn't have those artifacts but it doesn't really train faces well

[livid basin]

My guess is the limited precision during weight merge, not the training

But it is just my guess

I'll probably pull a few hundred V100s tonight, assuming they can get the 1.0 weights released

[boreal bough]

That's why no lora training methods are published yet - we're all hoping that on release tomorrow, faces wont be significantly undertrained

since that would solve the lora issue

[inner ruin]

shouldn't the checkpoint/sample prompt inference use fp32 as well?

[inner ruin]

what is your source for that info, if I may ask?

[boreal bough]

"hope"

why do I need a source for hope XD

[livid basin]

It doesn't have to, plenty of acceleration "methods" perform fp32 accumulation, but use fp16 for inputs to the accumulation

[urban fjord]

ComfyUI still have performance issues with LoRA so that's why I've not experimented too much more with it.

[livid basin]

Even Nvidia advertises it in their 4090 white paper, fp16 addition with fp32 accumulate

You can add A + B=C, where A & B are fp16, but accumulate into C, where C is fp32 for better precision

[boreal bough]

but as proven by multiple people here - the faces arent actually undertrained - just uneven datasets T.T
so some faces are overfit, while others are undertrained

[livid basin]

You can even accumulate in fp32, then cast to fp16 for inference

My guess is that would solve your problem @inner ruin , but that is just a guess

[boreal bough]

or run in bf16 - like we have to with kohya, to solve the 'number big' issue

[livid basin]

Or run in bf16 yeah, for better stability

I don't think bf16 will solve the precision issues thou

[boreal bough]

just gets rid of the NAN situations

[livid basin]

yup

[urban fjord]

Yeah, LoRA might run a bit smoother now, thought can't test ComfyUI with LoRA at an A100 atm.

[boreal bough]

didnt he fix lora for low vram pcs?

you should be able to test local now

[urban fjord]

There was issues on all GPUs with LoRA processing happening on CPU and not GPU

[inner ruin]

I'm not even downcasting it to fp16 which is wild. We're directly passing the unet with the trained lora in fp32 from the training script to generate validation images and they still kind of lose quality.

[urban fjord]

I've not tested on an A100 after the latest update, but there was still some issues after the first LoRA fix on an A100.

[inner ruin]

it's a good guess though

[boreal bough]

are you also training text encoder? or unet only

[inner ruin]

unet only

[boreal bough]

sad 😦

[inner ruin]

tried text encoder too though and it didn't help

[livid basin]

Ah wow, well it was worth a shot

[boreal bough]

I got the washed out issue while doing TE

[inner ruin]

oh so you experienced this too?

[urban fjord]

The LoRA performance is atm quite inconsistent it is tolerably fast on my 3060. Even with a model that has both unet and TE trained.

[boreal bough]

on standard kohya - in the beginning before TE was rewritten to acknowledge both clip models

[livid basin]

Frankly I don't quite understand why XL uses two clip models, I read the white paper and I still don't fully understand

[boreal bough]

also kohya produces those exact same ultra washed out images, when using diffusers to create sample images
-but the lora works without issues in comfy

so I dunno

[inner ruin]

what could be the diff there?

[boreal bough]

been wondering the same. Am even tempted to just add a hook and use an external diffuser library to gen the images, since it's just odd

[inner ruin]

kohya fwiw generates smooth images, it just sucks at faces

[uneven dove]

because they can

that's as much as i've been able to gather

they really don't do much experimentation with text encoders. they could have fine-tuned OpenCLIP instead of involving two of them, but they'll seemingly never do that

[urban fjord]

Hopefully we'll get a proper paper soon after release.

[inner ruin]

Like I ran the dumb dog example in kohya and it worked without noise

[uneven dove]

it might be your training images then.

[inner ruin]

nope, ran with same images on both and got noise in diffusers but not in kohya*

[livid basin]

I was hoping they would have used the giant state of the art EVA-CLIP02: https://github.com/baaivision/EVA/tree/master/EVA-CLIP
Or one of those CLIP models fine tuned on compositional reasoning (to improve visual object relationships)

Perhaps their use of Open-CLIP was for compatibility with older SD weights

[urban fjord]

They still want the models to run on consumer hardware so their clip-models cant' be too big.

[uneven dove]

@urban fjord according to those groups their text encoders perform better with fewer params

[livid basin]

Ah that's a fair point, but EVA-02-CLIP shows that they are better than Open-CLIP at the same model size

[urban fjord]

Yeah, not a big difference between 0.4B and 5.0B.

[inner ruin]

how many params is the SDXL openclip?

[livid basin]

I don't know for the language branch off the top of my head

[trim orbit]

vit-bigG

[boreal bough]

cries

what was your sample prompt?

maybe I'm just doing it wrong x_x

girl sitting on a chair --w 1024 --h 1024 --d 2 --l 8 --s 30

[boreal bough]

or was that inference? -cause that works! It's only during training that it goes south like that

[inner ruin]

maybe 0.9 is like SD 2.0 – they fucked something up and we are slowly finding out smh

[inner ruin]

that was inference on each checkpoint

this is the best LoRA I got so far

I think the script is fucked up, actually.

I'm talking to replicate people tomorrow to get some feedback. will report back

[livid basin]

weird @inner ruin can you point me to the exact script

[inner ruin]

https://github.com/Linaqruf/kohya-trainer/blob/main/kohya-LoRA-trainer-XL.ipynb just use this notebook

if you need compute juice, just DM me

but it should run on 16

nothing special about the params either

1e-4 is ok. 5e-5 is ok too

AdaFactor or AdamW are both ok

constant or cosine are both ok

I would set grad_accum_steps to 4 instead of the default 1

beyond that it's vanilla af

[eternal fog]

I'm going to try some training tomorrow on SDXL 1.0 and see if it's any better

The issue I logged for the Training Scripts got dismissed saying that it was a "bug" that it worked

lmao

[boreal bough]

does it work differently on diffusers?

cause 1e-3, rank8/alpha1 should be good enough for 95% of usecases

[inner ruin]

these guys should create some hype and set a release time/counter

[eternal fog]

No because then if it's delayed people freak out

[inner ruin]

1e-3 was pretty noisy for me on kohya

[eternal fog]

You'd have people in every channel 1 second after the time

WHERE IS IT

[inner ruin]

don't delay? haha be a competent company

[eternal fog]

WHERE IS IT

HOW DO I USE IT

[inner ruin]

like for real, you raised 100M, stick to your deadline

[eternal fog]

Can my potato laptop use it

[inner ruin]

midjourney is eating your lunch

not to be a dick lol

[boreal bough]

no noise at all, even once with these settings
#✨｜sdxl message

[fair crow]

Maybe I'm missunderstaning something but bf16 makes no sense for sdxl right? Like it ends up beeing lower precision than fp16?

Or any diffusion model

[eternal fog]

It's faster

And it seems to work ok

[boreal bough]

typical output of my loras trained on these settings

2 of those images are base only, no refiner - just to make a point of training faces

[fair crow]

Interesting, didn't know it's faster 🤷‍♂️

[boreal bough]

@inner ruin posted some examples of my lora output

[inner ruin]

those are sexy

[fair crow]

Those look really good! Trying to repro rn

[inner ruin]

alright I'm trying to reproduce your settings in the kohya notebook

[inner ruin]

lol samesies

[boreal bough]

bucket resolution is very important - as I found out last night. Default buckets arent scaled properly min/max needs to be manually set to 512/2048 (if you work with buckets)

[inner ruin]

even if the images are 1024x1024?

[boreal bough]

also, low res images will degrade the whole model

sadly

only poses /styles can be trained on lower res 😦

[eternal fog]

Who needs 2b cosplayers

[boreal bough]

then its fine - only relevant if you actually use buckets

[inner ruin]

I'm going for a 100 epochs and giving it a go

fingers crossed

[boreal bough]

40 max should be enough - unless you have a huge dataset

[inner ruin]

22 images

[boreal bough]

yeah, then it will probably 'finish' around 20 epochs, give or take 5. I wouldn't train past 40

[inner ruin]

ok, gonna go for it

50 steps warmup

[shy kelp]

i should just be able to use the safetensors .9 then right? i get an error when i swtich to it

[boreal bough]

good luck!

[uneven dove]

is there something wrong with the 1.0 weights?

[eternal fog]

Extremely bad seed?

Or they are doing weird shit in the background with settings

[upbeat summit]

looks like you need to reload the level

[uneven dove]

running locally

[eternal fog]

How would you be running 1.0?

[upbeat summit]

yes, how ?

[uneven dove]

they're out, it's 00:22 UTC

[boreal bough]

wtf

[eternal fog]

stop it

[upbeat summit]

lol

"keeps refreshing https://huggingface.co/stabilityai"

[eternal fog]

I'm telling you, it will be like 3-5pm

[upbeat summit]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 404

[urban fjord]

They never said which 18th of July.

[eternal fog]

You've only got another 13 hours

UK Time where they are based

Yes lol

Yes it is

London I think

[urban fjord]

Can we get the beta model too...

[livid basin]

This is a joke right

Is 1.0 really out

[eternal fog]

no

[livid basin]

don't scare me like that

[eternal fog]

@brave halo They are registered in Notting Hill Gate in London https://find-and-update.company-information.service.gov.uk/company/12295325

[nocturne dove]

I trained a lora for 0.9 for a likeness. Well lokr or whatever. It worked about as well as loras I've trained for 1.5. Maybe my settings could be better because skin detail seemed a little lower than the base model. But basically it was fine.

[livid basin]

I swear to god, if they've already frozen XL 1.0 and have just been stringing us along

[livid basin]

Convincing photoshop

Very nice

[uneven dove]

it's not a photoshop, i'm on linux

[sudden cliff]

didn't even bother to update the Model number

[nocturne dove]

Well things have to be finished before release. You know, due to causality and the arrow of time. And I don't think the plan was ever "press the button to stop training and then press the button to start the livestream"

[uneven dove]

what model number

[inner ruin]

link?

[uneven dove]

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

[nocturne dove]

So I would hope that files were already uploaded to the places where they were going before now.

[final veldt]

It's out I promise

[upbeat summit]

but it's not there: https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0 404

[uneven dove]

here's when i disable dark reader

[upbeat summit]

do you have early access? 😄

[sudden cliff]

zomg

[livid basin]

Surely he is trolling us

lol

[uneven dove]

i would never

[eternal fog]

Yes he's just changed the text on the website

[upbeat summit]

pseudo? he would never

[uneven dove]

i don't have that ability, Arron

like i'm some kind of fuckin wizard? changing text on a website? what does that even mean

[upbeat summit]

https://tenor.com/view/i-dont-believe-you-whatever-lies-gif-3602502

[eternal fog]

You forgot to change the description @brave halo

[uneven dove]

text.. websites don't.. it's an IMAGE model for generating images

there's no text involved

[west breach]

the number of models is still 21 in your screenshot

[final veldt]

July 18th 2026

please look forward to it

[west breach]

every time someone asks if 1.0 is out, they delay the release by 18 minutes

[final veldt]

Is 1.0 out

[eternal fog]

?

No you need to get access

I asked Hakurei access for it

It's not really worth using yet though in my opinion

[uneven dove]

ouch

lol

[urban fjord]

What's wrong with it?

[eternal fog]

Just doens't do anime very well yet

It's only 6 epochs

[uneven dove]

it's not haru's fault

[eternal fog]

Training still doesn't work on less than 13GB VRAM 😢

No that just makes it take so long there's no point

I hope they sort it out, I'd like to make some LoRAs

Give me something to put in WDXL and I'll show you how it is for now

[remote bronze]

where is stability gonna release 1.0 cuz theres fake leaks everywhere lmao

[urban fjord]

You can train LoRA on 12 GB VRAM.

But it's faster to just use a speedy cloud GPU

[eternal fog]

I have 10

And they keep changing stuff on the scripts and it takes different amounts

[livid basin]

if you were evil, you would take the 0.9 checkpoint, add a bit of noise to the weights, and release it as 1.0

[eternal fog]

This is what you get currently unless you use heavy negatives. But I've not been able to find any consistant ones

I think they are planning on re-training on 1.0 anyway though

[boreal bough]

the irony is I'm fine with 0.9 - just tell me I'm not getting an update T.T

[eternal fog]

For some reason the negative "watercolor" cleans up some of the mess

[livid basin]

I thought cheap anime usually have lines that bleed through

fair

[eternal fog]

It can still do 3D stuff okish for now

[livid basin]

I don't typically watch anime, but when I do watch the seasonal stuff, I notice often times lines overlap while ignoring occlusion, or colors are slightly offset from the edges

What's up with the overlapping edges in anime? Eg a line should be blocked by hair, but still appears

Is that just a thing?

[west breach]

isn't ghibli hand painted like watercolor?

[livid basin]

Or are the animators just too cheap to properly hide the edges

[west breach]

unless you're from the early 2000s with those frosted tips

[high skiff]

IT ARRIVED

[eternal fog]

Does it work

[high skiff]

I don't know, I literally just unbox it lol

[livid basin]

Ah I got excited and thought he meant SDXL 1.0

[upbeat summit]

hi @west breach 🙂

[west breach]

nice hands

that explains it 😄

[west breach]

hey 🙂

[eternal fog]

https://tenor.com/view/zotac-3080-gif-18549684

It's so tempting lmao

[west breach]

I got the Galax 4090, so far been rock solid and barely goes over 50 degrees (celcius)

[eternal fog]

It's just so much money though lmao

it's over half my monthly pay

Doesn't seem like a smart decision lol

Why would I get a 3090 when a 4090 exists

https://tenor.com/view/champoy-el-risitas-kek-issou-etu-gif-17837830

It won't have any more VRAM and will be a tiny bit faster for £500 more

[livid basin]

why get 4090, when there will be 48gb 5090 next year?

[upbeat summit]

why 4090ti when jensen has a 5090 in the oven

[eternal fog]

They've said no 50 series for like 3 years

Because I play games

[livid basin]

h100 is slower than 4090 in pure fp32 CUDA cores

[winter raptor]

I’ve been desperately trying to create something in this style with SDXL, anyone have any ideas?

[eternal fog]

As in a poster like that, or just that art style

[livid basin]

the H100 should only be faster than 4090 when it comes to tensor core or fp16

[winter raptor]

The art style

[eternal fog]

Wait for a Style T2I or Reference ControlNet XD

I can't sleep so I'll try something,

No idea, they've not said

[sudden cliff]

If you can get h100 at 15K...

I should have access to 1-2 H100 soon

[winter raptor]

So probably no way to accomplish without external plugins/loras/embeds? And if you figure something out I’ll be super impressed and grateful, don’t wanna keep you up tho lol

[high skiff]

That is insane lmao

[west breach]

akira x moebius?

[high skiff]

My 3080 is considerably bigger than this 3090

[uneven dove]

i've been running 1.0 for an hour now

[winter raptor]

How is it?

[uneven dove]

it's definitely their best model

[winter raptor]

Any workflow difference from 0.9?

[eternal fog]

First attempt

[high skiff]

Did they release it while I was out?

[eternal fog]

No he's being an arse

[winter raptor]

This is pretty sick

[thorn lion]

Lol I open discord to this. Slight arrhythmia nbd

[lusty raptor]

yeah we all got it, man

[winter raptor]

I’m starting to think it’s impossible to force SDXL to reproduce that grainy, semi-silhouetted style. It seems to want to go all or nothing with the silhouetting

[thorn lion]

Which means he doesn't think he's a toy but he acknowledges the difference between Andy and himself. Implications?

[fossil thorn]

I have been seeing some good dark images posted here. Is offset noise no longer needed with SDXL?

[winter raptor]

This is awesome, how did you get the text like that?

[upbeat summit]

@tender timber I've send you a message

[sudden cliff]

that's a completely unassisted output. It can do something like that easily

[winter raptor]

I LOVE the second one

[autumn forum]

meow

[winter raptor]

This is my fav so far, that’s awesome

[eternal fog]

Yeah how do you do text

I've never been able to get it to work

Yeah I do

[sudden cliff]

something like 'with the word "SOON"'

[eternal fog]

Doesn't work

[high skiff]

[upbeat summit]

award winning photography, a cute monster holding up a sign saying SDXL, by pixar...

[thorn lion]

I was just asking myself if it came in yet. Woot!

[upbeat summit]

now generate images and make sure all is used

[winter raptor]

What so far have you guys been finding to be the best method of upscaling with SDXL?

[sudden cliff]

imagine if the shadow spelled out 'never ever'

[high skiff]

Just got here haha

[upbeat summit]

I mean we have been through this phase but... it's here 🔜
Still from 1972 british comedy film, a smiling very british man holding up a sign that says SDXL is here!

[high skiff]

Driver installing

[winter raptor]

My mind would actually explode if someone got SDXL to do that lol

[high skiff]

DONE

[boreal bough]

this one, right?

[sudden cliff]

next level: modify the downloads graph

[high skiff]

its workinggggg

mmmm

[boreal bough]

GPT4: "Write me a safetensor file for what you think SDXL 1.0 looks like"

[uneven dove]

sorry was busy genning 1.0 imgs

[high skiff]

har har har

[boreal bough]

good enough? xD

number go up!

[high skiff]

just tested it

[eternal fog]

BING HAS SPOKEN

I don’t have control over the release of SDXL 1.0, but if I did, I would probably release it at a time when most people are awake and online, so they can try it out and share their feedback. Maybe around 10 am GMT, which is a reasonable time for most regions in the world. Of course, this is just my hypothetical answer, and it does not reflect the actual plans of Stability AI. Please don’t take it as a confirmation or a promise. 😅

10am boys

[high skiff]

16x 1024x images in 115 seconds, or just 7.18 seconds an image

[boreal bough]

or else, the AI must buuuuuuuurn

retorches his pitchfork

[high skiff]

[sudden cliff]

this made me notice there is that shadow layer

oh crap 25 1.0 spaces I failed

there's always something

[boreal bough]

but you missed the spaaaaces

yep XD

i made sure to only give access to time travel repo

[high skiff]

@visual glade Question for you good sir

I have 24GB VRAM, and yet the base model is still unloading before the refiner starts?

[sudden cliff]

^ did you do high-vram command line flag?

[eternal fog]

--high-vram

[high skiff]

oh, I didn't realize that was a thing

let me test then

[civic sigil]

You should start all your messages with "I have 24GB VRAM, and yet" today

[visual glade]

--gpu-only if you want max speed

[boreal bough]

if anyone has a downloader script they wanna let retest every hour, this link will prob work upon release

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors

just noticed that all the links are super simple

[west breach]

does it crash if you run out of vram?

[eternal fog]

It should just OOM

[high skiff]

ah, its --highvram

[visual glade]

optimizations that auto detect if it should unload the model or not is something I need to add

[eternal fog]

Ah ok

[sudden cliff]

I do batches of 8 and it stays around 20GB utilized at gpu-only

[visual glade]

so for now it uses the "safe" method of always unloading the model by default

[civic sigil]

BTW is it possible to load high bit depth images or even EXR in comfyui? I have heard EXR is difficult in python

[high skiff]

HOLY CRAP

--gpu-only

makes it WAY faster

[dim steeple]

Is 1.0 still coming out tomorrow?

[high skiff]

[boreal bough]

only the gods know, for the devs hath not spoken

[eternal fog]

We made a rule that every time someone asks it's delayed by 18 minutes

[high skiff]

6.82 seconds for a 1024x image, waow

[eternal fog]

How much for an 8k image

[west breach]

how many steps?

[boreal bough]

20 steps - assuming his regular setup

[high skiff]

20 steps

[eternal fog]

What do you actually gain from more. most of the time it's well done at 10 steps

[high skiff]

this GPU is pure stock

I bet I can get at least another 10% perf out of this card

[boreal bough]

my face = your profile picture

10 steps?!

[tender timber]

@quasi remnant thank you for posting the latest JSON you posted in #🌠｜show-and-tell . I will give it a try

[eternal fog]

Yeah

[high skiff]

I believe 10 steps works with uniPC

[eternal fog]

works with dpm++2m SDE as well

I shall demonstrate, one moment

[uneven dove]

1.0 is so much better at text

[eternal fog]

12 steps, no refiner

[high skiff]

I hate how this made me snort lol

[eternal fog]

For some reason on 10 steps she has no top on, so I couldn't post it

[civic sigil]

Lets see 30

[eternal fog]

I changed the seed hang on I'll do the same seed on 10, 20 and 30

[boreal bough]

vs exact same everything, but on 25 & 60 steps?

[eternal fog]

10

20

30 the image changes completely

[boreal bough]

25 is usually the first 'ideal' I find, and then around 60 for the second ideal

[eternal fog]

She comes back at 40

[boreal bough]

but the eyes are good now - which is also what I experience

[eternal fog]

50

[high skiff]

interestinngggg

on this card, its faster to do single image gens than batch image gens

[civic sigil]

Yeah 10 does not look like enough based off that

[eternal fog]

60

[analog wren]

/dream, girl with bloody hand

[high skiff]

at 1 image it takes about 6.4 seconds for my GPU, but 4x takes about 32, which comes down to about 8.16 seconds per image

[boreal bough]

try batch 2 & batch 8 - since those work fastest for me on 4090, which has same vram

[eternal fog]

dpmpp_sde at step 5

[high skiff]

sure thing

[boreal bough]

(or batch 9 if you have all the vram available - unlike me)

[uneven dove]

tried to get tony honk's video game art

[sudden cliff]

SDXL is out, gonna crank out some images on this bad boy

[boreal bough]

only makes sense with bf16 support turned on though, else you get stuck on vae decoding

[high skiff]

ok, so for me, anything over BS 1 maxes out the VRAM, which seems to be the speed limit

[high skiff]

oh?

how do I turn that on?

[high skiff]

Tony HONKNOLK

[boreal bough]

--bf16-vae

after gpu only

you need to be on nightly torch though

[eternal fog]

uni_pc at 10 steps

[boreal bough]

either via comfy prebuilt - or running your own venv

[uneven dove]

a tony honkn

[eternal fog]

better with a different noise scheduler

[boreal bough]

Download Link with unstable nightly pytorch
https://github.com/comfyanonymous/ComfyUI/releases/download/latest/ComfyUI_windows_portable_nvidia_or_cpu_nightly_pytorch.7z

[high skiff]

oh oop

yeah, i assume so haha

[uneven dove]

TONK 7

[boreal bough]

at least on 4090, vae decoding is now non existent. it passes right through

[high skiff]

ok, so how would I run this? Is it its own special portable verion I just boot up?

[boreal bough]

exact same as default comfy - you just extract it, update once, and then ready to run

no need for any config

[high skiff]

delicious

[nimble heart]

what res is that?

[uneven dove]

don't wipe your SSD while you instaaaaalll that

[boreal bough]

any res that fits in vram

[nimble heart]

interesting

[boreal bough]

at tiled, basically any size
untiled, around 2~3 batch, 1024x1024, or 1 1920x1024

[nimble heart]

1k^2 decodes instantaneously while 2k^2 takes > 1 minute for me

[boreal bough]

i usually switch to tiled vae if I'm not looking for 'final' pictures, since the difference is small enough to not be worth the waiting for me

[boreal bough]

then I can batch 4 or 8

[eternal fog]

SONC

[nimble heart]

3060 still plenty for inference

[eternal fog]

lmao video game box art gives some interesting results

[sharp robin]

@west breach how do you make a github thing for nodes for comfy? like is there a guide somewhere? .gitignore, i dun get it.

[eternal fog]

lmao, I put in Boris Johnson Video Game boxart and it made this

[sharp robin]

github to distribute i guess

idk im dumb dumb but wanna learn

[west breach]

create a repo and choose python as the template, select a license

[sharp robin]

oh so it does .gitignore license etc all on its own?

tf I thought there was some dark arts i had to know

[livid basin]

@sharp robin you can use github basically like it is google drive

[livid basin]

Not that I recommend it, but you can

[sharp robin]

ohhh okey! thans, so then all i have to do is upload the init.py and hte nodes.py and that is it?

[livid basin]

yeah you can even drag and drop into github, exactly like google drive

[sharp robin]

ohhhh

wow

[sharp robin]

im going to assume pychace is done autmaticly by github too right?

[west breach]

pretty much, i'm using vscode to sync with github

[livid basin]

Why do people use vscode over pycharm? Genuine question?

[sudden cliff]

W I D E N B I D E N

[west breach]

because it doesn't take 600 years to index my library

[livid basin]

Ah fair enough

But I like my IDE to look over my shoulder and assume I'm an idiot

[west breach]

mainly use at work with jupyter notebook files

[livid basin]

That's why I like pycharm and not VScode

pycharm = ide for dummies, perfect for me

If I wanted to use a text editor with a python plugin then vscode is fair game I guess

[eternal fog]

He's so orange

[livid basin]

did sdxl fail to converge, or did it converge perfectly

[sharp robin]

what should I select for how many team members will be working with me?

[livid basin]

0

[high skiff]

@boreal boughoh my godddd

[livid basin]

Your repo is going to be public right? You don't plan on hiding your files?

[high skiff]

this is even faster by a lot lmfao

[sharp robin]

dont think so

[high skiff]

wow, no joke, it just flies through VAE

[livid basin]

oh lawd he's coming

[high skiff]

on BS 8, it took maybe .5 seconds

[eternal fog]

These game covers are hilarious

[boreal bough]

He'll get the girl and he'll win your heart. The love story of the North you weren't expecting, but needed all along.

[livid basin]

Yeah, then just make it public, set the number of collabs to 0 or whatever

[eternal fog]

Praise Be

[high skiff]

down to 5.4 seconds an image this way

officially more than 2x the speed of my 3080 for the same power

[livid basin]

I'd like to report an assault, Kim stole my heart

[strong bramble]

it's almost like people dont now Discord has a search feature....

[livid basin]

Can I see Kim and Trump hugging it out

I dream of world peace

[tender timber]

Cool... first test. The brown box sure changes the output quite a bit

[livid basin]

That's the world peace I dream about

But why are Trump's hands so smol

Does SDXL think our god emperor Trump has smol hands?

Because he does lmao

[uneven dove]

looks like an HBO show

[high skiff]

this is incredible, this means my image gens use less than half the power and less than half the time compared to my 3080

[eternal fog]

Free Propaganda tool

[boreal bough]

Kim 2, in the sequel, he must beat his fathers golf record, of finishing the 18-hole tally in 18 shots, by somehow using only 17. Will he do it? Will a new star appear in the skies to immortalize this achievement? Find out in Kim 2: The legend of the 17 strikes

I was imagining it flying back out and into the next hole - but your solution is a lot better XD

[eternal fog]

lmao, what is this

CREAM

[boreal bough]

Drop'n Cream: DONT EAT THE PROPAGANDA

[boreal bough]

Cream. Serve. Delicious: The Road Trip that saved the USA

[eternal fog]

This looks awful, but the text lmao

Now this is an interesting perspective

[uneven dove]

@eternal fog

[high skiff]

some people are so out of touch with reality lmao

[eternal fog]

lmao

[high skiff]

just had a guy tell me that my repadded/repasted 3090 shipped for $600 was "not a good deal, should have been $500 max"

bruh

[high skiff]

I love the energy here haha

[livid basin]

well the 3090 is a much worse card than the 3090 Ti

[high skiff]

how so?

the 3090ti is like 5% faster lol

[uneven dove]

it doesn't have a "Ti" on its name, Sytan

[high skiff]

true

[soft bone]

mine was 500 but used

[high skiff]

[uneven dove]

trying to improve it

[livid basin]

Not the speed difference

But 3090 had heat issues with the memory, 3090 Ti changed the design

[high skiff]

mine was used as well, but that was $600 with shipping and him having full reworked the cooling system to run insanely cool

[uneven dove]

my 4 image gens take 5.9 seconds

JUST FYI

whew, ego saved

[high skiff]

this card came with specially repadded VRAM from the seller, my max VRAM temp has been 84C so far, with a 900mhz OC on it

[livid basin]

I think $600 is a good price

[soft bone]

ngl the 3090 is by FAR the coldest part in my pc even during endless inference. And that's with the ftw3 which pulls like 450watts.

[high skiff]

the thermals on this card have been next level

[uneven dove]

sytan abuses his gpu

[livid basin]

But 3090 had double sided memory, they fixed that in the 3090 Ti with single sided memory. The 3090 had heat issues that were fixed in 3090 Ti

[high skiff]

I do not

I love all my silicon unconditionally

[uneven dove]

that poor silicon

jinx

[high skiff]

can it be a jinx if it was not read at the same time? lol

[livid basin]

So despite the small performance difference, sometimes the 3090 Ti is a lot more than the 3090

Because thermals on the 3090 were fixed in the Ti

[uneven dove]

in my boomer-ass brain wrinkles they reached at the same time

[high skiff]

yeah, i could see that

This card was known for VRAM temp issues (like 110-115 C)

but now I can't even get it to 85C haha

[livid basin]

But the guy was still an ass, $600 is a great price

[uneven dove]

he was jealous

[autumn forum]

my 3090 doesnt get above 67 lol

[uneven dove]

i mean it SHOULD be 500 bucks because that'd be fucking awesome but that's not how things go

[high skiff]

he also said that the 4090 will be selling for $500 a few weeks after 5000 series drops lmao

[high skiff]

I didn't mean core, I meant VRAM TJ

[autumn forum]

ohhhhh yeah mine is about the same lol

[uneven dove]

lol

$500 is his magic number

[high skiff]

I guess so lol

[uneven dove]

that must be how much money he got for his birthday

[high skiff]

headass lol

[uneven dove]

if he just says it enough times, it's certain to become true

idk i saw his name and stopped there

[high skiff]

2080ti was NOT $500 when 3000 dropped lmfaooo

it was like 3k cause of shortages lmao

[eternal fog]

He's from Florida, figures he's a bit slow

[high skiff]

[uneven dove]

my suggestion to you is to just enjoy your hardware and not post about it on reddit lmao

[high skiff]

floribruh

[sudden cliff]

wait it's really 'florida bois holla at me'?

[uneven dove]

the mating call of GPU sellers

[high skiff]

I was posting a review for the seller in the server (which is requested) as it was his first high value sale

which means my review gets him the certified high value seller tag now

[uneven dove]

oh cool does that dipshit actually have a $500 3090 for sale? because i will buy it lol

[high skiff]

not likely lol

[uneven dove]

didn't think so

weird how he can dictate price of stock he doesn't have

[autumn forum]

hey man

[high skiff]

he also said he sold a 2070S for $650 back in covid, yet he thinks the 2080ti was just sliding out of peoples hands for $500 lol

[uneven dove]

everything costs $500

[high skiff]

its all $500

that was a certified spedrun lol

[livid basin]

lol a 4090 will not be $500 when the 5090 launches

[high skiff]

lmao

[uneven dove]

the guy you're arguing with

[high skiff]

I didn't even argue, I just flat out walked away lol

[livid basin]

I bet the 4090 will be at least $1000 when the 5090 is launched

You are a good man Sytan

Now, what about a $500 3090 for me?

[high skiff]

yeah, likely considering the 5090 will probably be like $2k MSRP

[placid edge]

5090 will be mcm, will be great

[uneven dove]

5090 is going to cost 15k

jensen needs to pay his bookie

[high skiff]

I just hope its a 5090 and not a 509-eh

[uneven dove]

https://giphy.com/gifs/images-monkey-drums-Xvz7fb8rEWZtS

[high skiff]

lmao

https://tenor.com/view/gladiator-russell-crowe-maximus-are-you-not-entertained-gif-3477451

[uneven dove]

ohhh i love shia laboeuf

is that him?

ahhh that's tim robbins

[winter raptor]

Oppenheimer leak?

[winter raptor]

This is awesome, love the fidelity

[uneven dove]

tim wobbins

[high skiff]

have anotha

[uneven dove]

lul

the russells

[upbeat summit]

I mean a good caricatures for sure

[uneven dove]

chris farleyXL

yay it 'knows' "Eric Andre"

[upbeat summit]

how are you getting these faces so good and coherent? can you give a hint? prompt, processing? magnets?

[eternal fog]

It depends on the person and the prompts really