✨｜sdxl

[pure crystal]

someone on the wing, some...thing

[upbeat summit]

https://tenor.com/view/twilight-zone-strange-shocked-gif-12654145

[pure crystal]

https://en.wikipedia.org/wiki/The_Monsters_Are_Due_on_Maple_Street

[uncut fiber]

tried something peaceful but SDXL understand ... ignite immerse imagination ... Its own way...

[zinc cargo]

is there a way to make auto1111 understand meta info from comfy?

[high skiff]

@noble shoalNew LoRA test proved a substantial success

had a failure in it, unfortunately. Ran out of space mid test, but I was smart enough to save training states

same exact training as my previous new one that was a success, but training with bigger BS, more choppy caption drop out, 2x the epochs, and 1/6th the LR

Results are underbaked due to stopping early, but immensely promising

left is the successful previous one, right is the new one that finished way too soon

skin texture and lighting is massively improved, as well as background fidelity

I used some weight studying in training, so I know only .504/1 of the model has been tuned, so I can resume from where I was and do the other half, hopefully getting to about 0.91, which is right before it starts to deform and over-fit it seems

for example, by epoch 15/30 in the last training, the model was already at 0.907, and by 30 it ended on 0.94 and was way overtrained

While here at epcoh 30/60, its at 0.504, so hoping for about 0.9 by 60

if that proves a success, I will continue and re-train using my full 500 image dataset instead of the current 118

[heavy zinc]

Hi guys, is there a download link for nmkd siax 200_1 upscaler please?

[uncut fiber]

https://openmodeldb.info/models/4x-NMKD-Siax-CX
only siax on that upscaling dedicated page.

[heavy zinc]

Where do I place it?

[high skiff]

if you are in auto, same place you would put SD models, but instead put it in the folder that says upscale

or sorry, no

auto > models > ESRGAN

[heavy zinc]

Okay thanks. There isn’t an upscaler called siax 200k_1?

[high skiff]

I think it might be depricated (outdated)

[heavy zinc]

Okay

I also want the ESRGAN_4x

I couldn’t find it online

[high skiff]

that one is already in Auto

[heavy zinc]

But it tells me couldn’t locate the upscaler

[high skiff]

that one is in the RealESRGAN folder

[uncut fiber]

could be this, it is trained on 200.000 iterations.

[heavy zinc]

I don’t have such folder name

Only ESRGAN

[high skiff]

these should be your folders

Granted, my install is quite old by now

oh wow, you have a lottttt less

I am not sure in that case

[glad grove]

then make a folder and put it there

[heavy zinc]

put what there?

[glad grove]

what model are u asking for?

[uncut fiber]

i think esrgan is as good as realesrgan

[heavy zinc]

I just downlaod the Siax 2000

where do I put it?

and Also I am looking for the ESRGAN_4x

[glad grove]

ersgan folder

[heavy zinc]

Fo Siax?

[glad grove]

yes

[heavy zinc]

I also got formula upscaler

Do I put it in the ERSGAN folder as well?

[glad grove]

idk what is a formula upscaler

[heavy zinc]

based on FormulaXL model

I couldnt find ESRGAN_4x

could you please provide me a download link for it

[uncut fiber]

Siax is esrgan 4x

check page i send you link, choose one you would like.
https://openmodeldb.info

[glad grove]

yea put it in esrgan folder too

[heavy zinc]

In the dropdown list in A1111 UI there are list of upscalers, and ESRGAN_4x is seperate from the Siax

I also downloaded the SDXL refiner

it is called sdXL_v10RefinerVAEFix, where do I put that file?

[high skiff]

VAE folder

or wait

its a refiner?

[heavy zinc]

yes

[high skiff]

just the normal models folder

the first one

[heavy zinc]

where in models folder?

[high skiff]

Stable-Diffusion

[heavy zinc]

what is called

[high skiff]

thats the name

[heavy zinc]

same as where my models are?

[high skiff]

yes

its a model too

[heavy zinc]

It is called VAE in file name

[high skiff]

That means the built in VAE is fixed

[heavy zinc]

https://civitai.com/models/101055?modelVersionId=128080

[high skiff]

all models have a VAE inside of them, a dedicated VAE is if you want to use a different one

[heavy zinc]

got it from here

I think that is the only refiner for sdxl

right?

[high skiff]

yes

put it in the same models folder as the others

its a diffusion model just like base, but its made to run after

[heavy zinc]

Alright thanks

[high skiff]

no problem, hope it helps

[heavy zinc]

NotImplementedError: Cannot copy out of meta tensor; no data!

I got an error

[uncut fiber]

screen of cmd?

[heavy zinc]

it is fixed now

I just restarted the UI

idk what have caused it

AttributeError: 'DiffusionEngine' object has no attribute 'cond_stage_model_empty_prompt'

I am getting this now

[glad grove]

🤷‍♂️

[half cedar]

(Photoreal, POV close:1.3),cosmic evil DMT realm headspace, horned hooved muppet baphomet, Vinton Studios claymation, cinestill cinematic, beautiful dim lighting, night time, contemporary, despair hope, pentagram on forehead, seated cross-legged, wings spread, both male and female traits

[high skiff]

sounds like you were trying to generate Baphomet lol

[stark vortex]

https://tenor.com/bK7lj.gif

[zinc cargo]

?

[upbeat summit]

happy birthday!?

[fierce hollow]

Is there any way to use controlnet tile with sdxl yet? I tried all the stuff like LLLite with blur/ultimate upscalers and such, but it just doesn't look as good

ssd-1b is pretty cool as a side note

[oblique swan]

nope

gotta switch to sd15 for upscaling

[fierce hollow]

big sadge

[oblique swan]

maybe in a year or two sdxl will finally replace 1.5

[fresh path]

no chance even we will have dalle 4 lol

[mellow dome]

It looks so legit 😂

[crisp owl]

sdxl has never looked so good

[noble shoal]

🥹 Thank you

[noble shoal]

I am happy to announce this monstrosity: https://civitai.com/models/183354?modelVersionId=205793

[crisp owl]

https://tenor.com/view/ron-swanson-parks-and-rec-its-so-beautiful-gif-15644547

[noble shoal]

Listen. I need to win this 4090 to keep my quality as high as it is at the moment. 😅

[high skiff]

Kohya might be one of the most annoying tools to use. It's got dozens of settings that are not properly documented, tons of redundant features that interfere with each other, buggy and unreliable bucketing that seem to ruin training, and, resuming from save states doesn't work.

Scam

[noble shoal]

Bro's getting a free ban

[high skiff]

Yessir

[crisp owl]

@smoky patrol scam above

[noble shoal]

If thats with the lora something is clearly off

Way to high quality

[high skiff]

I would have loved to show you guys the results from my latest training for my realism LoRS, but kohya crashed a whopping 3 steps in

[mellow dome]

i 3 times the strength

[noble shoal]

Which model?

[mellow dome]

ssd 1b

[noble shoal]

Oh, no guarantees with that

[mellow dome]

lemme try wif base

yup, theres more like it lmao

with just strength 1

[noble shoal]

Ah, see. Much worse

[smoky patrol]

Feel free to report messages when you see them!

[cyan crown]

from 70's movies

[crisp owl]

Is that 343 Guilty Spark?

[sinful falcon]

If anyone is looking for something different to play with, I made a LoRA that creates caricature portraits. 😜 Many thanks to Wizard Whitebeard for sharing his horrible and wonderous Gildenface LoRA with me! It was my inspiration and I used it to create the initial training images for the first round of distorted images for this LoRA. I hope you can have some fun with it! https://civitai.com/models/181092?modelVersionId=203235

[hoary saddle]

You ever have one of those days when you learn something that changes things for the better but still pisses you off because you didn't already know it?

wget --content-disposition ugh...but yay

[half cedar]

Does that set the proper file name?

[hoary saddle]

Yeppers

[heady vale]

Taylor Swift as a muppet on cctv

[half cedar]

Guy on the left "definitely her "

[latent bobcat]

https://tenor.com/bHarN.gif

https://tenor.com/XG5V.gif

[crisp owl]

I see what that is

[hardy cipher]

pretty accurate

is that some sort of pear on the right?

[crisp owl]

"portable pear"

[hardy cipher]

sounds convenient

this guy growing teeth like a shark

[upbeat summit]

welcome back, @hardy cipher

[hardy cipher]

https://tenor.com/view/greetings-chat-better-call-saul-lalo-salamanca-greetings-chat-gif-25797757

https://tenor.com/view/newjeans-haerin-have-a-good-day-newjeans-kang-haerin-newjeans-kang-haerin-gif-27662442

[upbeat summit]

https://tenor.com/view/csr2-gif-27359358

[hardy cipher]

https://tenor.com/view/smile-cat-cheese-saycheese-gif-26155989

[glad grove]

https://tenor.com/view/steven-tsang-gif-25331686

[hardy cipher]

https://tenor.com/view/sad-lonely-alone-chandler-bing-gif-22422379

https://tenor.com/view/monica-monica-ghellar-ghellar-courteney-cox-courtney-cox-gif-21815956

[noble shoal]

Wow, that's such garbage that I had to download it. PoP, you might enjoy my MS Paint LoRA. 😬

[hardy cipher]

did you hand mspaint all the images?

[noble shoal]

Nope, I am lazy af. I used a magical thing called the internet to find reference images.

[hardy cipher]

this lora does look like a gem though

I wonder what it would do if you're a mad lad and make it negative

[noble shoal]

🤔 hmmmmm....

[hardy cipher]

perfect

[latent bobcat]

https://tenor.com/bl7PN.gif

[west breach]

Just made this fun and colourful lora. Check it out https://civitai.com/models/184095/neon-style-xl

[void loom]

happy halloween

[void loom]

(guess who)

[peak dove]

SDXL Tensor.art ... dozens of free generations/day

It has LoRAs, base and refiner, and Controlnet

It can be linked to (or is planned to) ComfyUI

[tulip ledge]

Hi. I'm completely new to all AI stuff and need some pointers on how to get started.

I'm trying to use stable diffusion (Installed the easy diffusion version) to create some standard fantasy art (Knights, witches, monsters and so on) for use in TTRPG's I run. Character art, landscapes, locations etc.

But I have no idea how to start, just typing in prompts ended with some... interesting results.

Any pointers on where and how to start would be greatly appreciated.

[half cedar]

You need good prompts and styles, use https://civitai.com/images to look for images and looks you like and build up some prompt templates

[half cedar]

While you're doing that, take a look at the models of images you like, you'll eventually have a few models around

Then if you have the hardware for it you can install comfyui and a1111 and start learning more technical workflows (start by copying workflows).

[velvet shore]

I am trying to get realistic life like images using SDXL for my website so anyone can create realistic and masterpiece image in few words without need to write long prompt

I tried to fine tune sdxl tried different prompt but at last I get these type of images only

Please can anyone help me 👀👀👀

I used this prompt

an old woman holding lamp, photorealistic

[noble shoal]

Happy Halloween

[velvet shore]

Nice bro

[velvet shore]

Awesome,
can you please tell the model name ?

[crisp owl]

looks like it's base model with a mummy lora and an art lora

[indigo vine]

you can drop it in comfy to see the workflow

[velvet shore]

Thanks

[hoary saddle]

happy halloween you comfy peoples

workflow included

https://github.com/smuzzies

custom text chatbox node

[glad dust]

Happy Halloween! 🎃 👻 🦇

...What is hidden in this image? 👀🌀 https://www.instagram.com/reel/CzEbjOMoEp_/?igshid=MzRlODBiNWFlZA==

ramonartist

Happy Halloween! 🎃 👻 🦇

...What is hidden in this image? 👀🌀

#Halloween #Adobe #ComfyUI, #SDXL #Controlnet

Instagram

[jolly brook]

lol I thought this was Pixelmind at first the design is almost identical

[indigo vine]

||netflix||

[jolly brook]

Pixelmind was a text-to-image service that I worked on for about 4 months before Midjourney's beta started

I'm just commenting on the web design

[velvet shore]

Oh sorry I thought about model

[wheat tulip]

go to civitai and search for some models/loras to help u out

[supple knot]

@myg the more specific you are with the prompt the better the results you will get, also the model you download like "Realistic Vision V5.1" is a good one for real looking people, Matters ,because it make the image from that models database also when you get deeper into it you use LoRa and Controlnet to fine tune your outputs.

[clever verge]

Have this been posted here? System Memory Fallback for Stable Diffusion: https://nvidia.custhelp.com/app/answers/detail/a_id/5490

[icy brook]

Aether Pixel.

Dissolving subjects into pixelation.

[noble shoal]

Did somebody said pixel? 👀

[lusty moss]

Asked in the prompt section but didn't get any response. Does anyone else prompt with Text Concatenation and Cutoffs? Seems to help a lot for specific detail. Wouldn't mind seeing other prompt workflows. This is using ComfyUi for those unfamiliar.

[indigo vine]

red paint redemption

[noble shoal]

Another classic. American Paint

[bright valley]

Okay well apparently my lora can do the Netflix logo, just maybe not spelled right every time

[icy brook]

I gave it a 5/5 on civit 🙂

[noble shoal]

I saw. Thank you 🔥

[bright valley]

You made a pixel thing?

we should combine forces

[noble shoal]

Nah, it's just a MS Paint style LoRA

[bright valley]

I bet it would still be really look if used with mine

[shy kelp]

is it not possible to make embeddings with XL? (trying in automatic)

[hardy cipher]

don't see why it would be impossible

[vital ermine]

almost all training in automatic1111 are flat out not working for SDXL. All we have is really Kohya

[hardy cipher]

not quite as good as yours, but only lora I used here

must be from the covid life days. wearing those masks, people socially distancing

[hardy cipher]

need a lora based on this theme

[icy brook]

what theme is that?

[hardy cipher]

characters from the oregon trails game fighting dudes from fight club in the streets of a large city while 8 bit babies that drink gasoline and spit fire add to the mayhem and smoke cigarettes

[icy brook]

a prompt on base sdxl?

[lusty moss]

West World lol

[bright valley]

where's this fades into pixel thing?

[lusty moss]

https://tenor.com/view/the-sword-in-the-stone-madam-mim-im-an-ugly-old-creep-jump-jumping-gif-5366013

[icy brook]

Wdym?

[bright valley]

is that a lora or something? it looks awesome!

[lusty wolf]

Good Evening... who's neck is available...

[icy brook]

Yes. Aether Pixel. Will release it soon.

[bright valley]

Ohhh okay it's not out yet. It looks really great so far, can't wait to try it out!

[bright valley]

My lora generates custom text, and I think that effect would be so sick with it

[icy brook]

Cool!

[bright valley]

just made this for somebody on another server

I feel like that particle effect would look so sick with these

How do you control what pixelates/from where?

[strong copper]

Happy Hellowin )

[native knot]

Cutie Audrey Hepburn if she was alive today. 😉

[vital ermine]

I always had a huge crush on her even old and shit

Btw, that ain't Kathryn that is Audrey

[native knot]

Fixed. This is why I shouldn't do 3 things at once.

Do one and do it right.

[vital ermine]

Katharine never did it for me but she would be good as your tom boy friend to just mess around with. Later in Life she sort of went "I am Kate, and I don't give two shits anymore" on everyone.

[heady vale]

new feature added to NV drivers for stable diffusion. mainly for users with 6-8gb cards https://videocardz.com/newz/nvidia-introduces-system-memory-fallback-feature-for-stable-diffusion

[pallid path]

allows you to make it perform like it did a bunch of updates before

[rustic garnet]

I thought they had that on Windows since long time oO

the question is: when do we get that on Linux, too

[heady vale]

Ive not seen it before and they claim its new

[rustic garnet]

weird. I always thought windows is using the RAM as fallback when vram is full 🤷

if you google that you get so many explanations about that. But I have no clue, I don't use windows for that

[heady vale]

the new feature allows you to decide to allow that or not

previously you had no choice

[crisp owl]

Well this is interesting, implemented, I go from about 3 minutes upscale process, to 35 seconds upscale process.
But then I OOM if I try and push to my FaceDetailer node

oh yall chatting about it lol

[heady vale]

yeah thats the down side. it may crash if it cant handle it

[crisp owl]

For straight up simple gens though, I get an absolutely monstrous boost in speed though,

[heady vale]

vram is blazing fast compared to system ram

[rustic garnet]

ram ist fast, too, but the transfer between ram and vram is slow I guess

[crisp owl]

Yeah wonder where exactly that bottleneck is.
Imagine if that could get resolved and RAM could be utilized efficiently.

Could be a pipedream though haha

dreams in 64gb RAM

[tribal lantern]

This one is so creative and well executed, care to share the prompt that managed to create it?

[glad grove]

https://tenor.com/view/booba-popping-eyes-eyes-popping-out-pepe-eyes-booba-pepe-gif-23805447

[mellow tendon]

Needs bigger fangs, like this. I added "like sabre tooth tiger teeth" etc.

[slender coral]

Could someone explain the difference for me:

[glad grove]

yes

[heady vale]

post the link

[slender coral]

https://civitai.com/models/91807/biopunkai-2-sdxl-sd15

[heady vale]

it was SD 1.5 and now thy updated a new version for SDXL

[slender coral]

Okay wouldn't it make sense for it to be 1.5 and 1.0?

[hoary saddle]

It's the way people do their lora versioning. It's version 1.0 of their model based on SD1.5, normally you would see version 2 based on SDXL and eventually version 3.0 of the lora based on whatever version of SD is out. You can see Sergeant's workflow on civitai is version 1, 2,3,4,4.1 etc...just depends on the user and how they want to number it

[crisp owl]

Ooh glossy detail, nice

[slender coral]

Spawn? 🤌🏿

[velvet shore]

✨

[velvet shore]

What the F

[glad grove]

he's massive and ballistic 🫡

[velvet shore]

Bro has balls of steels 🤣

[glad grove]

ball-istic 🪩

[velvet shore]

🥶

[vital ermine]

Let's Dance.

[indigo carbon]

how tf, I literally just made these

that's some hive mind shit

[heady vale]

nice one

[native knot]

A Spawn/Venom mix.

[noble shoal]

Shameless self-promotion:
https://civitai.com/models/185138/sdxl-sanddrawing

[icy brook]

looks good! 🙂

[icy brook]

nice!

[heady vale]

thanks!

[hasty smelt]

hey guys, I'm happy with this result, after a long round of adjustments, what do you think?

[smoky patrol]

Love the lighting; nice!

[noble shoal]

I like the overall image quality and the background details, but i can't unsee this neck

[stone fossil]

😄

[static prawn]

is it a general problem in sdxl when u have something like sunlight in the bg ,the person in front stands out a lot looks overexposured? Like there is too much contrast

[surreal siren]

hey guys can i train sdxl with 512 images?

or do the images need to be 1024

[noble shoal]

You can do that, although i can't 100% recommend it. Problem will be double features. Like SD 1.5 on 1024.

My smallest dataset was 85x111. Just for fun

[surreal siren]

ok cool thanks

[surreal siren]

have you tested loras vs ckpt for training people? Like do you think a lora produces better results of the subject over training a whole ckpt

[noble shoal]

I have not tested it since i have only a 12gb card on my hands. But feel free to ask in #🔧｜finetune

[peak dove]

This IPAdapter is way cool!

[crisp owl]

I've never wanted to eat a cat so much

[pure crystal]

sushity

[stone fossil]

Ya man. 😄

[pure crystal]

a very serious wizard made a small error in his transmogrification spell, he used the wrong form of you

[upbeat summit]

cool style! alice in wonderpunk land

[heavy zinc]

hey guys, can I upscale an image without using the Hires setting at the time I generate?

[half ivy]

i am here
what did i miss

[half ivy]

noice 🔥

[half ivy]

bro

[native knot]

No time to say, "Hello", goodbye!

[thorny glacier]

Yes, but you will need extensions like Ultimate SD Upscale, ControlNET Tiles, and perhaps Multidiffusion—perhaps even all together—to get good results. But you WILL get much better quality from such methods

Here’s an example; it starts off with a crappy painting I did, then I put it into the AI at a low resolution, and then up scale it to 2.1K:

[peak dove]

Cool

[thorny glacier]

And also, if you decide to use them, you will need to read the GitHub documentation as it can be a bit confusing to use. In these images, I used ControlNET Tiles and Multidiffusion to create a high quality painting that doesn’t have the “Upscaling” texture that you get when you only use ControlNET Tiles + Ultimate SD Upscale

[peak dove]

Good to know, thank you!

Only for A1111?

[thorny glacier]

I personally used in A1111, but I believe it is possible to use these in ComfyUI

[full idol]

why does CR Latent Batch size increase s/it so much and increase the time it takes to generate by 1000x?

I put it at 16 since I want to generate more images at once

[barren birch]

#1072220168534642768

[heady vale]

you need more vram for larger batch sizes

[peak dove]

IPAdapter can go really slow the more Photo Inputs you use. I found that 3 inputs and Photoshop really slowed things; so closing Photoshop made things a lot faster! But Bingo - there is Online PS now - which makes no difference to the IPAdaper at all 🙂

An SDXL ComfyUI original ... using Generative Fill in Online PS

[full idol]

in auto1111 you can generate 100 images np

[heady vale]

that is batch count. generating them one by one which any gpu can do

batch size generates x number of images at the same time

[full idol]

any node for batch count in comfy?

[heady vale]

its on the toolbar. tick extra options box

[lusty wolf]

Preserved...

[crisp owl]

Harry Frogger and the Galactic Jedi Shrooms

[pure crystal]

tron mechanic

[fringe pawn]

guys i have to get a new video card i guess, my pics suck

[crisp owl]

gpu doesn't change quality of image

[fringe pawn]

well i cant run xl and regular crashes alot from no vram

[crisp owl]

that's a separate issue lol

[indigo carbon]

I have yet to see ComfyUI give VRAM errors

[fringe pawn]

wish i could do something with this one im on. oh and the one i have is not nvidea

i have lots of issues eh? or comfyUI may solve most?

[crisp owl]

Cheaper to try ComfyUI than buying a new GPU.
Unless you're dead set on purchasing a new one

[fringe pawn]

Does it let you allocate some regular ram to it by chance?

[pure crystal]

Boba Tron

[crisp owl]

not exactly.
How much vram do you have?

[fringe pawn]

says 4 but i think 3.6

64 mobo ram, gpu is only part i didnt buy when i built this box

[crisp owl]

you may squeeze by with 4 in Comfy. It'll be slower as it'll likely use tiled vae for everything

[fringe pawn]

cause i didnt game

this is mine

[crisp owl]

If you use SDXL at 512 you'll get ugly images

stick to these

[indigo carbon]

yep, that's exactly what I was about to say

[fringe pawn]

Do you think a $400 or less video card would work for SDXL

[crisp owl]

I use a 2060 Super just fine

[fringe pawn]

i have 6400 XT

[crisp owl]

Wish it was faster, wish I had more VRAM, but I can get by with nearly everything

[fringe pawn]

no 6400 RX sry

[crisp owl]

Probably generated over 10k images at this point, I couldn't even count

[fringe pawn]

yea but 2060 has cuda and prob more stream threads

[crisp owl]

Yup, it's a good card for what it is

[fringe pawn]

ohhh i see what your saying

[indigo carbon]

old AMD cards are a bad idea for AI imo

[fringe pawn]

yes i checked its good price and quadruple ram

[indigo carbon]

I think used 3090s/ti from ebay or something are the best budget rn, right?

[fringe pawn]

u have the 12gb or 6gb?

[crisp owl]

Not sure who you're asking, my 2060s is 8GB. Not sure if there are other options for that

[fringe pawn]

yea diff manufactures put different fans and stuff on that chipset i think, they range from 6-12gb from what i seen.

those 24gb topshelf or are there higher?

[crisp owl]

Consumer grade, I believe 24 is about the max

[fringe pawn]

Might try to get a 8 or 16 cyber monday, will try comfyui this weekend

thanks

[crisp owl]

If you're gonna continue with this route, you won't regret getting more vram

[stone fossil]

Yes.

[indigo carbon]

yeah, that must change in the future.. if we would want models that use something like T5 we would need a little more than that

[fringe pawn]

what are the cards that ship with m2 slots these days?

or soon

[crisp owl]

I'm really curious what nvidia does with the 5k series, since between the 4k release and now, there hasn't really been anything released, and AI has exploded since then

[indigo carbon]

I'm thinking 32gb cards maybe? it probably won't cost them too much to bump to that, then SAI could consider having a bigger text encoder on future models, and using something like LLaMa with SDXL would be a possibility

[crisp owl]

If they don't tread the right way with it, it'll open the doors for something else to really swoop in

nvidia

[fringe pawn]

it has a 2tb ssd but i havnt read it yet, looked it up again to sharehttps://videocardz.com/newz/asus-shows-off-rtx-4060-ti-graphics-card-with-m-2-ssd-slot

[crisp owl]

Interesting if it's comporable speeds as built in (it may specify in the actual article)

[indigo carbon]

bumping the VRAM to 32 would defiantly be smart from NVIDIA, considering they are actively making driver updates SPECIFICALLY for Stable Diffusion

so we could expect the 5k series to have some capabilities that allow SAI to start to consider making models with a fat text encoder

I personally don't expect too much from NVIDIA considering how much they fucked up the 4000 series, but judging by their recent moves; there is a chance that they will actually make the GPUs capable of such a thing

[velvet shore]

My fine tuned sdxl model running on nvidia A100 large

Model : fine tuned sdxl
Prompt : A mystical house in Forset, concept art

Model playground : https://ionium.web.app/sdxl

[fringe pawn]

its nothing, just an additional slot to be efficient. nothing more. weak.

[fringe pawn]

$15k gpu and u got that? which one u braggin about lol

[velvet shore]

Lmao bro it's SDXL you can try generation time is just 20 seconds for 2 images

[indigo carbon]

I'm on a 4070ti and for me it's 30s for batch 4

[cyan crown]

steps?

[indigo carbon]

takes me a little bit over 30 seconds to make stuff like this

[fringe pawn]

why do you have that card? I dont know anyone that would have that... @velvet shore

[cyan crown]

I've 4070

[indigo carbon]

I use AITemplate so the sampling is near instant

[velvet shore]

Yes, but I have some users also

[velvet shore]

Hehe 😁

[fringe pawn]

I liek that pink

[cyan crown]

how many steps?

[indigo carbon]

I did 60 there

[cyan crown]

only 30s good

AITemplate is ?

[indigo carbon]

like TensorRT, but a little faster and entirely flexible

just acceleration

[fringe pawn]

im for real.. did you write it off? what kinda work?

[cyan crown]

I'll try Nvidia accelaration on A1111 as soon as possible

[velvet shore]

I just wanted to take reviews as I am making a complete fast stable diffusion web playground with many models

[fringe pawn]

im looking at specs.. thats super insane. like having a 1995 ferrari f50 or something

[indigo carbon]

I'd personally recommend ComfyUI, A1111 is outdated

[fringe pawn]

world’s fastest memory bandwidth at over 2 terabytes per second (TB/s)

When u upgrade please dm me

[velvet shore]

I think it must be supercomputer
2Tb/s

[cyan crown]

Nvidia new drivers extension is for A1111

[indigo carbon]

it's still not as fast as AITemplate. NVIDIA finished making that extension a few months ago, they just now released it because licensing or whatever..

also you'd need to compile EACH checkpoint individually to turn into TRT engines, AIT is architecture specific; not checkpoint specific.

[fringe pawn]

in 10 years we should have specs of that a100. its happened time and time again for almost 60 years

[indigo carbon]

H100 is even stronger

[velvet shore]

Price 💸💹

Wow! So much rule breaking message

[crisp owl]

bro watch the language

[pure crystal]

just a heads up, typo on your main page "cinematic" vs "cinemantic"

[pure crystal]

world war 2 soldier werewolf, flamethrower expert, propaganda photo, Bernie Wrightson (American, 1948 – 2017) photography, 1986 aliens cinematic lighting by Denis Villeneuve

[floral island]

if i see stuff like that, i feel like i'm looking brightprotonuke v1.0 (my first model had so much of those twirly artifacts)

[nimble heart]

if they do it'll be something like
5090 32GB: $2200
5090 24GB: $1600
5080 16GB: $1100

[crisp owl]

There's gonna be price pressure for sure given the state of the economy, as well as maintaining market share

[indigo carbon]

nah, it doesn't hurt too much to bump the VRAM up

[nimble heart]

in manufacturing costs no but nvidia will 100% charge a shittone extra for the priviledge of having 32

is my point

[heavy zinc]

I just used the extras tab which is upscales pic using gfpgan or code former

[indigo carbon]

that's dumb, more people will buy the 5000 series if it will have a normal price and 32gb

it would be more profitable and better for both sides

[heady vale]

NV have limited caacity purchased with TSMC, so its more profitable to use that time making server chips

[nimble heart]

if 1000 people buy a $2200 card with a 30% margin it's more profitable than 5,000 people buying a $1600 card with a 5% margin

though arguably the Titan class cards failed so it hopefully wont come to that

[indigo carbon]

that's not considering the likeliness of SAI making bigger AI models due to the consumer cards being stronger. SAI brings NVIDIA plenty of attraction

the last few drivers are made SPECIFICALLY for SD

[nimble heart]

i dont think its a question of whether or not nvidia wants to make the best Stable Diffusion cards I think it's more whether they feel threatened by AMD's current cards

there's definitely been some movement with a lot of new ROCm support PRs being brought up in major repos, more specifically for the XTX

whether its enough to dampen the nvidia gouging is the question

[indigo carbon]

speaking of; I managed to quantize T5 to 6bits and it was almost identical to FP32

[nimble heart]

what's the test?

quantizing can work well to your meat brain but might not work well when it's part of a diffusion pipeline is what im wondering

like the seemingly subtle losses could be magnified

[indigo carbon]

just asked some general questions, responses are almost identical

I think we could have a quantized T5 as a text encoder? or will that need adaptations to the other components of the model before training

[nimble heart]

no idea. Pixart still hasnt released their weights so our only hands-on demo is DFIF which is certainly one of the models

[indigo carbon]

there's no way to do that, the only way to test an encoder as a diffusion pipeline is if a compatible UNET would've been created. this means SAI must be 100% sure about each component before they start training it

[indigo carbon]

DFIF's T5 isn't quantized I think?

[nimble heart]

dont think so

i dont have it installed anymore

havnet run it since i got my xtx

could probably add it to my diffusers CLI script ig

[indigo carbon]

I can try to quantize whatever version of T5 DFIF uses then replace it with the included one I think?

[deft forge]

Etc

Hello

Anime boy with dark brown hair and blue eyes blue flame effect

[nimble heart]

https://discord.com/channels/1002292111942635562/1101178506530013284

[deft forge]

Пр

Пр

[nimble heart]

i dont speak in runes

[hasty smelt]

To give you an idea, these 2200 USD become the value of a motorcycle here in Brazil, it's always like this, waiting for the 50 series to be launched before being able to buy the 3070. 🥺

[nimble heart]

buy like three 3060 12 GBs and string them together using deepspeed or something

what's the worst that could happen

[indigo carbon]

DFIF loads T5 in 8bit

it fits in 10gb cards

[nimble heart]

render some images with and without 8bit

see if it combusts

[indigo carbon]

I just read the repo and people talking about the repo, will give it a try in some time

"Text encoder needs a bit above 8GB of VRAM loaded in 8bit, so the whole thing should be able to run on 10GB cards." this means the UNET takes 2gb

it should take less if they load it in 6bit though

[nimble heart]

oh you mean dfif uses 8bit by default?

[indigo carbon]

yep, according to people on the repo

[nimble heart]

fucker OOM's my 24GB xtx so I think 10GB is bullshit

unless diffusers has made memory improvements since

[indigo carbon]

I have 12gb, it should OOM for me if it did for you as well. I'll test it out

[nimble heart]

dfif was like the first thing I tried on my XTX and I had to reduce the res below base to render cause the 2nd upscale would OOM

[indigo carbon]

I'll also change the code to use 6bit on the text encoder to see what it changes as a diffusion pipeline

[nimble heart]

maybe the script I used was just trash idk. i made my own diffusers script for SD models and it works almost as well as comfy for the most part

[static prawn]

i really think i gave up on upscaling

trying like 12h now and still having the same issue

no upscaled

upscaled , RIP Details

[nimble heart]

for SDXL run either ultrasharp for art or swinir for photos then re-run XL at the final res after the upscale with like 0.3 denoise to clean up the artifacts

[static prawn]

its the same for me, kinda no matter if sdxl or normal sd...

it always deletes any details, i prefer using extras tab now

left img2img, right extras tab

[indigo carbon]

I usually use something else and it works insanely well for all stuff

[static prawn]

every detail always get smoothed out, i hate it

[nimble heart]

do both together

[static prawn]

i dont understand why others dont have this problem

[nimble heart]

or at the very least preprocess your img2img input with an unsharp mask or something

[nimble heart]

they do you just dont see all the failed results lol

[static prawn]

have no idea what u mean 😮

[indigo carbon]

oh they are on A1111, that must suck

[static prawn]

the only solution for me is running the upscale with 0.01 denoise

then it looks "ok"

[nimble heart]

sharpen the image in GIMP/Photoshop/something then feed the sharpened image into img2img so SD can pick up on the details easier

sometimes helps

[static prawn]

mh... i mean i could just use the extras upscale, seems to work better for me at all

[nimble heart]

if you're talking auto, what I do in other uis equivalent is extras upscale first then img2img after

[static prawn]

mh might try that

or maybe its just i like the grain of the pixelation more and its matter of taste

[nimble heart]

"grain of pixelation" is certainly some words

[static prawn]

the more grain of the low res, kinda brings more textures

dunno how to describe it

[nimble heart]

your meat brain fills in the missing details at low res better than your computer's sand brain

[static prawn]

right just feels more detailed for me

because of that texture on the skin

[nimble heart]

you should look at SD.Next's sampler configs lmao. They're all hardcoded at like the opposite settings they should be using.

like ddim is set to "linspace" I think for the timestep spacing which I dont think any SD models use???

and the set alpha thing forgot the exact name is wrong along with step offset

[indigo carbon]

yeah, ComfyUI is easily the best UI; there is no function it can't do that others can

simply because you create the functions

[nimble heart]

i feel like a proper diffusers frontend could be competitive but right now the only major one I know of is SD.Next which sucks ass

comfy is omega specialized for SD so it generally outperforms diffusers but diffusers is super flexible. Train an XL model with velocity prediction instead of epsilon and comfy cant sample it correctly

• the non-sd pipelines in diffusers are fun

[visual glade]

anything based on diffusers can't really be competitive because the library isn't organized correctly

[nimble heart]

What makes you say that?

[visual glade]

https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/stable_diffusion

it's a mess basically

[indigo carbon]

heh, yeah, I can easily fit DFIF

I used a custom node, works like a charm

[nimble heart]

could maybe use a touch of deduping lol

[visual glade]

the whole point of a stable diffusion library is to be able to use as much stuff as possible at the same time

[visual glade]

diffusers just isn't good for that

[nimble heart]

maybe i should look back into making a library with libtorch directly instead of through pytorch lol. probably a ways beyond my experience though

[nimble heart]

is that with both scalers for the full (1024?) output? that's the part that ate like 28 gigs for me

[visual glade]

and sd.next isn't the only frontend that uses diffusers, invokeai also uses it

[indigo carbon]

invokeai was actually faster than A1111, but not as fast as ComfyUI

[visual glade]

it's not hard to be faster than a1111

[nimble heart]

from what I recall invokeai doesnt support other models outside of SD which is a big part of diffusers.

right now I just have my own cli script so I can do hacky stuff like fix the DDIM timesteps in 2 lines of code instead of trying to make a patch for a whole ass UI

[nimble heart]

yea that's diffusers. Usualy a few % behind comfy and a large chunk ahead of A1111

diffusers xl 1024 for me is like 2.95 it/s on diffusers with split cross, 3.17 on comfy with subquad, and i think like 2.6 or something on auto last I checked.

with a triton compiled unet I can get like 3.39 on diffusers

[visual glade]

I recently improved subquad so it should OOM much less on AMD now

[nimble heart]

Oh nice. Guess I'll try re-enabling smart mem for a bit

[indigo carbon]

I also heard AIT is about to be officially built in UI?

[nimble heart]

i compiled a flash attention 2 wheel but when I tried it on a llama model I guess it doesnt support the shuffled forward v2 fn needed for inference so big sad

[visual glade]

with my recent comfyui refactors it's very simple to change the scheduling and v prediction, etc...

I need to write a node to expose that functionality

[nimble heart]

lol i saw that. I made a cursed node that just

model.model.model_type = comfy.model_base.ModelType.V_PREDICTION
return (model,)

and it actually worked

[indigo carbon]

alright, so T5 loaded in 4bit takes about 4gb VRAM, easily enough for most cards

[nimble heart]

4bit is super rough

it always falls off hard for llama models in my experience

[cyan crown]

Good night

[indigo carbon]

when used as a component of the model the precision of the encoder barely does anything to the quality/coherency

[visual glade]

SD at least is very sensitive to imprecision in the text encoder outputs

if you inference it in fp16 vs the fp32 (with weights in fp16) you will notice lower quality images

[indigo carbon]

I tested that on DFIF, to see if SAI has a reason not to use a bigger text encoder on future models

[nimble heart]

so the natural solution is a new brainfloat8 type. All exponent no precision

god damnit i just pulled master and now the vpred node doesnt work. 0/10 literally unusable

[visual glade]

https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/model_base.py#L117

you just have to set that model_sampling to the right class

[indigo carbon]

8bit T5

4bit

[nimble heart]

i feel like you'd enjoy Rust's trait system from the way you make a shell class that exists only to inherit from two others lol

[nimble heart]

you stole his god damn legs

ayyy it works now

[indigo carbon]

Still, the degradation isn't too bad, I think a bigger text encoder would be a good idea for future SD versions

CLiP's language model is probably too small to have a coherent language understanding, so that at least certainly need to be switched

[lusty moss]

Using After Detailer in ComfyUI to adapt the face and it actually yields pretty good results.

[pure crystal]

all great

[lusty moss]

What does the envelope reaction mean?

[shy kelp]

what sampler do you guys prefer?

[rare rapids]

Good Models to generate Air-Planes like B-24 Liberator?

[lusty moss]

I haven't extensively looked into it, I usually go Euler as of now but have no benefit why.

[lusty moss]

https://stable-diffusion-art.com/samplers/

[shy kelp]

thanks. theres so many options its overwhelming lol

[zinc cargo]

can i get some help running sdxl on auto1111? on comfy everything works fine

but auto gives me black screen

[lusty moss]

Briefly looking at what I just posted i will stick with Euler or UniPC. Obviously take your time to look through and see what works best towards your goal.

[high skiff]

Anybody know where you can apply for money grants for LoRA's and stuff?

Cause I am spending so much time per day on this realism LoRA, and while its current handily beating results from realism models, I have a ton of work to go, and this is turning into a full time job, and I just can't support myself as it is

as it is now, this is my new workflow I have to do for every single image in my dataset:

find the image, favorite the image, pixel peep the image, download the image, hand crop the image, down sample the image into a custom bucket res, rename the image, and tag it all by hand following a strict formula all by hand

And I have to do this 500 times for my next test, which is likely gonna have at least a week just for the collection

If I deviate at all, the quality falls off significantly

if anybody knows where I could apply for a grant or funding or something, I could really use it cause I am an unemployed 19 year old doing something I am passionate about, but I can't keep doing this if I am not at least making some income off of this.

I can't just dump hundreds of hours down the drain for nothing, its not sustainable

[bright valley]

Civitai

But you'd get paid in "buzz", which unfortunately you can't exchange for money atm

[high skiff]

Thats not exactly what I meant, but I do appreciate the suggestion. I was meaning like actual funding, not offering some buzz for their limited rewards 😅

[bright valley]

So just... somebody to give you money to make your own models...?

[high skiff]

Yes. Its more common that one would think, and I have just put in some applications from some other people who reached out in DM's. However any OTHER suggestions would be extremely appreciated

Hoping the results I am having can interest LAION enough

[lusty moss]

Indeed or Only Fans like the rest of the world. Sorry to say it but seems like everyone is doing this for fun.

Client requests could help.

[bright valley]

Yeah I have no idea then. I've done commissions before but they always want something specific. Or want something.

[high skiff]

Yeah, this is more of a very powerful LoRA I am working on that is honestly handing fine tunes their ass, but the scale at which I wish to do the next iteration is just too big for me to commit to without some support from a benefactor. I have been told about a few I am trying out, crossing my fingers

[bright valley]

I just made my first about a week ago

It solves text generation for stable diffusion XL

[glad grove]

try with kofi/patreon so ppl can support u each month

[high skiff]

I have been making them for a long time, but this new one I am working on is by far the most promising I have ever seen, and EASILY the most time consuming one I have ever seen as well haha

[glad grove]

if ppl like your models/loras they will support you but not always sometimes they just dont care 😔

[high skiff]

I am not looking for the support of individuals at the moment, but the support of groups/orgs that see what I am doing at the moment. I am holding out on a few options right now

[bright valley]

Yeah I mean obviously I'm new the to model training side of things, but it's the highest rated XL model on civit this week, and I don't see how I'd really make a dime off of it

But if you want to just turn a profit, there's many ways to do that.

[high skiff]

We are working on very different types of LoRA's. All of my previous ones have been like that as well, and I felt the same way

but I am hundreds of hours deep into this one, and I likely have hundreds more to go

but my small scale tests are so promising, I think it could be attractive to training groups

[bright valley]

I was a programmer for many years but I fell in love with creating art when AI tools came around, and now I do it full time

Both as my job and side hustle

[high skiff]

with less than 1/25th of my full dataset, only 1/4th trained, I am already crushing dedicated realism finetunes

[bright valley]

Aren't there like 100 of those already though?

Like really good ones

[high skiff]

100's? yes. Good? I have yet to see one ATM

[bright valley]

I guess I don't know what your standards are lol

[high skiff]

I am stretched so thin I don't have many good examples of my newest tests, and my newest yet failed unfortunately, but I have some really old examples, and trust me, the results are much better now

[glad grove]

from my experience ppl give u money for a lora when they want a lora from certain char or celebrity that they cant find anywhere

[high skiff]

here are some tests from my much older model. this is less than 1/25th of my full data set, trained only 50% of the way. top left is mine, top right is realistic vision XL, bottom left is Realism engine XL, and bottom right is RealStockPhotoXL.

[bright valley]

this is an example of what mine does

[high skiff]

Text LoRA's are dope. I have no idea how to train one, so thats all magic to me haha

[bright valley]

I don't think I've ever seen one before tbh

[high skiff]

but yeah, those results from above are less than 1/25th my datasset, and only about 1/4th the full training potential even from there

also, they require no key words

[bright valley]

i don't quite know what that implicates though

like the alpha version of my LoRA was trained on 25 images just to test and see if it worked

https://cdn.discordapp.com/attachments/1037799583784370202/1169720206315237488/image.png?ex=65566dd5&is=6543f8d5&hm=156b35bd0cb288f898a9b9531440f29cda82513df150096d51c5027aabace4ed&

[high skiff]

top left is base SDXL with a 100MB LoRA compared to 3 of the best realism finetunes for SDXL

and that one is already a step behind my hybrid 1.5 version, which is a step behind my 2.3 failed training which is only 40% of the way done

And that new one is less than half the quality I am expecting from my 3.0 demo

and the 3.0 is 1/5th the size of the 4.0

[bright valley]

yeah if there is a direct relationship between the size of the data set and the quality/capability of the LoRA, I'm not sure what it is

[high skiff]

my new results are so promising, but I can't properly share them cause they are not done, and are very volatile

[high skiff]

oh, sorry I left that out

my LoRA allows for control over the following:

lighting luminance, lighting direction, Subject, crop, aperture, color grading, location, time of day, and more

in addition to not needing any of that info to produce the results above

[bright valley]

With no trigger or activation terms?

[lusty moss]

Reminds me of Army Men: Sarge's Heroes the old game for some reason lol..

[high skiff]

if you wanna control those values, you need to prompt them, but if you don't, it works as see above

so its like a DSLR, it shoots good in "auto" or you can dial it in with manual to get better looks

[thorny glacier]

I don’t tend to use the Extras tab for upscaling because utilizing Stable Diffusion itself for upscaling along with those extensions makes for dramatically better results than using dedicated upscalers alone. I only use the Extras tab for upscaling when I want to have a raw “backup” upscale that I can edit and “collage” for inpainting into the real upscale created with the extensions if SD created messed-up parts

[bright valley]

Well if it really is the best realism model for SDXL, then maybe if everyone else agrees they may help fund your following projects

[high skiff]

ok, here we go, I have another example of just how good my new failed LoRA is

Left is base SDXL, 2 is my realism 1.0 (in the results above), 3 is 1.5 hybrid, and 4 is my new crashed training that only got 40% done before it crashed. And you can already see how promising the 4th one is here, even severely under trained

And I have new tricks up my sleeve that are showing potential to increase quality greatly again, then a 5x increase in datset, then another trick a friend is writing from scratch, and then another 5x dataset increase

Also, this LoRA is made to upscale, so its really 2048x base with support for 1024x images

prompt was "Blonde woman in forest"

no negative

you can already see how much more realistic the last one is, barely even 40% trained

[bright valley]

Sure, but like I said, I not convinced that larger data set is a good thing always

[bright valley]

^^^facts

[high skiff]

its better fro the specific variables I am testing, and every dataset increase so far has yielded much better results. Also, the full size dataset will be accompanied by full text encoder training with subject breakdown using nested folders, courtesy of a research partner

my original dataset was 20, then 60, then 90, 120, next is much much more

[bright valley]

I guess we'll see how it stands up when you release it

[lusty moss]

Oregon Trail remaster anyone?

[high skiff]

also, my results beat out the competition wayyy more if I cherry pick them lol

but I chose general prompts from my research partners

[bright valley]

dang that's some sick lookin pixel art

[high skiff]

hell yeah it is

[glad grove]

you have died of dysentery

[lusty moss]

https://tenor.com/view/horror-disgust-ohno-jimcarrey-liarliar-gif-12183274

[high skiff]

@bright valleymy progress so far from base, without the last one even being fully trained

[lusty moss]

The key, ([Pixel | Voxel]:1.1)

[high skiff]

on a subject and location not represented in the dataset

[bright valley]

I trained my lora to do pixel art the best I could, it does 8-bit, 16-bit, and 32-bit

[high skiff]

actually here, let me try something

[lusty moss]

Female Maximo!

[bright valley]

here's 8-bit

[lusty moss]

How do you prompt the text you want?

[bright valley]

the trigger is starting the prompt with
YOURTEXT text logo

[lusty moss]

Drop a link?

[bright valley]

https://civitai.com/models/176555/harrlogos?modelVersionId=201745

[lusty moss]

You use ComfyUI or Auto1111?

[bright valley]

comfy

[lusty moss]

Noice

[bright valley]

i usually use this VERY simple workflow to make shit with it, takes 6s per gen

It doesn't work perfectly everytime, it's only v1.0 and a work in progress, so sometimes you need to go through a few seeds to get the correct spelling

but all things considered, I'd say it's a HUGE step foreward already

[lusty moss]

Is there a picture I can drag in for the workflow on CIVIT?

[high skiff]

@bright valleyhere you go

My LoRA is trained on foliage, the others are not lol

[bright valley]

here i made you one

[lusty moss]

TY

[bright valley]

why do the first two images look like trash tho

[high skiff]

cause one is base SDXL, and the other is realistic vision

[bright valley]

I mean

[high skiff]

realistic vision sucks at animals

[bright valley]

I've used those models lol

i feel like the output wasnt that awful usually

[high skiff]

I am not spamming tiggers and negatives

just "Photograph of a white tiger in a forest at dusk"

its not always as big of a win, but my model is trained specifically for forest images

[bright valley]

i guess if somebody wants to gen animals in the forest, youre the guy

[high skiff]

or people

anything in the forest, my LoRA destroys the finetunes lol

cause I originally trained it to do just that lol

[thorny glacier]

Personally, I tend to front a lot of my prompts with “RAW Photo, taken with Provia” to get higher quality images—though sometimes I also use “stock photography” to get higher quality. I also prefer to keep my prompts relatively short if possible

[bright valley]

ngl the panther in the second one looks pretty diesel

[high skiff]

and thats why I am making such a big dataset, cause the most images in my dataset are in a forest, and it shows cause my LoRA does forest portraits really good

[bright valley]

hopefully that works out for you

[high skiff]

A portrait photograph of a woman wearing a purple dress in a forest

[bright valley]

which is the good one in that

[high skiff]

remember, this one is only 40% trained, which is a shame

[glad grove]

the one that doesnt look like from a stock photo

[lusty moss]

You tried messing with clip -2?

For the text lora

[bright valley]

not yet no

[high skiff]

I am just showing different examples. The one on the right is mine, and has more accurate lighting on the face and chest. The light is from above, and thus the chest is brighter than the neck and face

My LoRA is made specifically above all else for better lighting

[bright valley]

if you make some shit with it, I'd love to see!

[lusty moss]

Trying one now will post it

[bright valley]

🤘

[high skiff]

A portrait photograph of a black woman from the side by neon lights at night

[bright valley]

yours is looking towards the camera, not a profile shot

FAIL

[high skiff]

if you ahve any you want me to try, let me know

[bright valley]

yeah try my lora with it

[high skiff]

I asked from the side, not a profile. Her body is from the side

I don't care for any of the results

[bright valley]

see if you can make the most realistic letters possible

[high skiff]

I am not testing with other LoRA's, just any prompts you'd like

this LoRA isn't even fully trained, not made to pair with other LoRA's lol

[bright valley]

oh, well I'd say finish the dang thing then

[high skiff]

bruh, thats why I need funding lmao

[bright valley]

mines not made to either but people do it constantly

[high skiff]

the next training is likely gonna be at least 100 more hours of work

cause of all of the improvements I am making

[bright valley]

unfortunately nobody is going to pay you to do that though

so you can either do it or not do it lol

[high skiff]

We've yet to see that

There are a few groups that do it all the time, some of which I have just reached out to

[bright valley]

give people money to make their own LoRA's?

[high skiff]

yeah, its not uncommon

models and LoRA's

[bright valley]

I mean, I work in the industry and I've never heard of it

[high skiff]

RunDiffusion paid a ton for JuggernautXL, a model my LoRA already bitch slaps in realism

[lusty moss]

Clip -2 1024x1024

[bright valley]

fk yeah

[high skiff]

thats dope :p

[lusty moss]

Clip -2 Upscale 2x

[bright valley]

went straight for the pixel art i see

[lusty moss]

Normal Clip 1024x1024

Normal CLip Upscale 2x

[high skiff]

if anybody has any realism prompts you want me to compare, let me know

[crisp owl]

clip 1 and clip 2 are the same exact image

[lusty moss]

Was wondering if the clip would cut back on some of the artifacts, going to dial it back to -4

[high skiff]

I am sure there are plenty I would fail at cause of my small dataset

[lusty moss]

Like always?

[crisp owl]

Pretty sure it does nothing in sdxl

[bright valley]

dude the first version of this one was 25 images

and it kicked ass somehow

good and big are not directly related always

[high skiff]

some concepts are a lot easier to train than others, and it depends a lot on how you captioned it too

[glad grove]

try clip skip 12

[lusty moss]

Goog to know, it has still beem in my workflow.

[crisp owl]

It was a "trick" used in 1.5 for certain models

[high skiff]

just skip ALL of clip lol

[bright valley]

how about achieving custom text generation in stable diffusion

[crisp owl]

maybe 2 also, but was necessary for some in 1.5 I know for sure

[lusty moss]

-12 lol

[high skiff]

I know somebody else who has done it as well, a few weeks ago

[bright valley]

LOL

[high skiff]

I have 0 clue how you would ever tho

[lusty moss]

Still pretty cool though

[high skiff]

I don't train stuff like that

[bright valley]

Link it

[high skiff]

let me find their examples, IDK if they posted it, but it can do sentences

[bright valley]

with no control net or any bullshit right?

and with stable diffsuion>?

[high skiff]

with SD only and no control net. Just a LoRA

he trained it at only 768x res I think, and he uses a second diffusion on top to upscale. I am not super sure, but he had some even better results I didn't save

[bright valley]

looks like dalle tbh

[high skiff]

its not lol

and it looks nothing like dalle3 lol

[glad grove]

looks like stock photos

[bright valley]

yeah its tough for me to glean anything from a bunch of images

[high skiff]

also, you can see the images names lol

[bright valley]

I have no idea how they made them

[high skiff]

I need to remember dudes name

hes in here

[bright valley]

is their model not out anywhere?

[high skiff]

it starts with an A

if I can find him I will ping

found!

@noble shoal Any info on your text LoRA?

more of his examples he has sent here

[bright valley]

Theyre cool, but don't mean much of anything without knowing how they're created

[high skiff]

hmm, it looks like he doesn't have his LoRA out, so IDK actually

[bright valley]

Strange, if it can do all that

[high skiff]

but its for sure not dalle, cause of the content he had it make lmao

[lusty moss]

Pic or it didn't happen.

[high skiff]

I honestly have no idea, I will have to see what he says

he told me he trained it at a low res, so it got the shape of the words, then used img2img to upscale

[bright valley]

and now it can gen full sentences without any control nets or anything?

or thats how he makes the images

[high skiff]

this is when he was showing it off

[high skiff]

not like full full sentences, but what I showed

"NO PINEAPPLE ON PIZZA"

Not like a super eloquent sentence, but a string of words properly spelled in order

[high skiff]

lol

[bright valley]

Yeah I was just wondering if he has a model and the lora and presses generate and gets that

[bright valley]

I get it man

[lusty moss]

Same lol...

[bright valley]

still meaningless withou tknowing the workflow

[high skiff]

its just a LoRA according to him

[lusty moss]

@bright valley

[high skiff]

he trained it at like 768x res, gens with SDXL at the low res, and passes it to normal SDXL to upscale if he wants to

[bright valley]

oshit i missed it lol

sick

[high skiff]

thats what he told me at least

[lusty moss]

Yea got bumped...

[high skiff]

actually wait, when did Dalle 3 even get launched?

[bright valley]

made this for someone in another server

very goosebumps-esque

[high skiff]

yeah, for sure

[pure crystal]

what was their name

[crisp owl]

sarah

[bright valley]

Adamantium

[high skiff]

also, a friend of mine retrained all of SDXL from nothing using just his single GPU, and the results are shockingly good for having destroyed all the weights and rebuilt them lol

[bright valley]

but yeah, they go by sarah

[high skiff]

hes actually the owner of the research group I am in, and he spent so much time on it, OMG

its no finetune, just a completely new base for SDXL to train finetune on with terminal SNR and gamma correction, as well as v-pred

[bright valley]

seems a little redundant, no?

[high skiff]

no, its not even close to normal SDXL

it looks nothing like it

and it trains completely different

[high skiff]

he took base SDXL, nuked it to incoherence, and then trained in Terminal SNR, Gamma, and V-pred from scratch using his own dataset of like 500k images he scraped from all over the place

so he changed the way the model diffuses as a whole

[bright valley]

now thats what I call a Blood moon

It's trained on all my own artwork, so I love how it puts horns on shit like I do

[high skiff]

I don't fully understand what he did, but he basically tore apart all of SDXL and retrained it on different papers with drastically different tools and stuff

Hugging face has been pestering him to release it, and he finally did after several months of work

[bright valley]

yeah seems like everything you talk about isnt done or isnt out

[high skiff]

he also retrained the whole base of SD 2.x to 1080p, and it can rival SDXL sometimes

[high skiff]

it is out lol

just a sec