🆕｜sd3

[dusky thistle]

and make sure that actually changes it

[bitter hearth]

that was before we realised SD 3 2B was not gonna last

june or july or something

[dusky thistle]

idk what your prompt is, but try changing it to "an impasto oil painting of a nuclear explosion in a rose garden in september, vibrant radioactive shockwaves in a dark night"

sd35 undrestands stuff well enough that if your prompt is too similar it's not gonna change much even at high noise levels

[short thicket]

Do having custom schedulers and samplers like this actually improve the output of the model, or do they mainly offer extra variations of outputs to choose from? Is there a specific setup that generally gives higher quality outputs?

[dusky thistle]

big improvement in quality

and yeah more variety

my entire goal with this stuff has been to improve quality as much as possible

[short thicket]

Any specific settings other than the ones you've shown so far? Part of me want to play and expiriment, but another part of me wants to know which direction to head in.

[dusky thistle]

ohhhhhhhhhh

sdxl has much higher sigma values

the equiv of 1.0 with SD3.5 is something like 14.26 with SDXL

so try setting it to something like 3 or 4 for starters

when i only wanted to make tiny changes with sdxl i usually set it at 2 or 3

yup

sdxl has a high sigma scale

14.614642 = the equiv of 1.0 with SD3.5 or Flux

[bitter hearth]

red panda too good

from L2 discord https://cdn.discordapp.com/attachments/1148864356906442793/1300205608976388177/RDT_20241027_1709561986655669478107523.jpg?ex=671ffe55&is=671eacd5&hm=ac136cd50058b0206d9385854999a14a75af0fefec1193b0495bc191d2e60e4d&

[dusky thistle]

a few... i'm not sure what the max is for pixart sigma, but it's sometihng crazy like 150 for hunyuan DiT and something like 99.95 for stable cascade

[bitter hearth]

is like 317 for NovelAI Diffusion V3

no wait I misread

20000

https://arxiv.org/abs/2409.15997

that's pretty funny

closed source though

someone on reddit pointed out that the red panda is a chinese animal

maybe this is next pixart

[short thicket]

@dusky thistle if you have SamplerRK hooked up the the Shark Sampler node do Noise Type, Alpha, and K do anything in the Shark Sampler node or do the settings in the SamplerRK node take precedence?

[dusky thistle]

they do separate things

the ones in shark determine the noise in your starting image

the very first blast of noise, so to speak

your starting point

the ones in SmaplerRK determine the kind of noise that is added after each step

[short thicket]

gotcha

[craggy crest]

prompt: professional studio photography in a commercial studio. commercial lighting, dynamic shadows, hard rim lighting. a dashing, debonair, male actor, a twinkle in his eye and a sly grin

[winged seal]

kinda sorta seeing some mild improvements, but I don't think I am gonna be gettig good results out of SD3.5 any time soon with AIT. I might try out another trainer, but I am not sure if I wanna take on that whole project at this moment. Maybe

these results are nowhere near good enough to make SD3.5 viable for me, so I will be just waiting until I can try out LucaTaco's trainer

so far this has been a monumental waste of compute, but hey, I gotta learn some way haha

Its gonna take a massive amount of compute to make SD3.5 really solid, but I think it will be done at some point. I also hope we get a pony version of it so it will be better for niche subjects

[winged seal]

deciced to try in flux, cause why not

I think I am gonna take a break from training for a bit

dumped 20k training steps into SD3.5 and it looks a littele less bad lmao

I will find a better approach some other time

I will say, SD3.5 is a lot faster on small GPU's cause the small 4 bit quants don't have to page out

[dull star]

oh nice

[craggy crest]

:) you've come a long way. breaks are good.

[winged seal]

I'm just holding on to the hope that its the trainer I am using, and that its not gonna be an uphill battle 😅

[craggy crest]

learning curves are always uphill battles. you're doing well, just stick with it

prompt: classic western movie scene inside a saloon. close up on the bartender as he glares at a cowboy. high definition photography. backlit, diffused shadows, dark and moody

[craggy crest]

someone stepped on his biscuits, stole his gravy, an now he's mad

prompt: hdr photo: red to gold to blue to green gradient, oil on canvas, hyperdetailed, intricately detailed, sapphire, volumetric lighting, mist

that man looks seriously about to start trouble

14 karat rose: hdr photo: red to gold to blue to green gradient, oil on canvas, hyperdetailed, intricately detailed, sapphire, volumetric lighting, mist

sd3.5 large

[winged seal]

@craggy crestscrew it

i LIED

I am gonna try to set up Luca's trainer lol

I am kinda doing nothing anyways lol

[craggy crest]

ROFL! at least go eat dinner first

[winged seal]

I just had some food lol

@craggy crestDo you have a link to the LucaTaco trainer? I am finding a bunch of like articles about it, but where's the actual code?

[craggy crest]

i gave it to you already but here https://replicate.com/lucataco/stable-diffusion-3.5-large-lora bookmark that this time

[winged seal]

ohhh, so that WAS the right one, ok

I didn't click on it cause it was replicate. I was on there before 😅

I found the github link on it, thanks

[craggy crest]

welcome

[winged seal]

I thought that was just an online version

@craggy crestWait a minute, am I crazy? Is this a cloud payment version? Can you not run it locally on hardware? I am not seeing anything on running it local

[craggy crest]

replicate isn't free. you do have to pay replicate for the use of their GPUs

it's not very much, however

[winged seal]

Oh yeah, ok, nevermind. I don't do that

thank you for the recommendation, but absolutely not lmao

I will wait for better local options

[craggy crest]

up to you. read through the loras on huggingface, i think a few of them might mention what they've used

[winged seal]

I was looking through some just a little while ago. I have not seen a single one that mentions their LR so far, which is... interesting

haven't seen too many tho

but yeah, I don't do off PC compute or anything like that. I am not gonna be reliant on other peoples services. If I can't run it locally by myself, its not happening. Might as well be worthless in my eyes, personally

Great for people who like that sort of thing, but I have never and will never support that

[patent cape]

guys can u help in the main chat lol

[winged seal]

@patent capetry in #🤝｜tech-support especially if its not about SD

[patent cape]

idk man. i am as confused as it gets

[sacred jewel]

Thanks for this...

[craggy crest]

and alternate link https://replicate.com/lucataco/stable-diffusion-3.5-large-lora-trainer without the large image

[winged seal]

aside from the faces/skin, these look really solid. Nice stuff!

[sacred jewel]

Just a pixel space upscale and second sampling pass. I can't get Flux to give me great skin at upscaled stage ever... been struggling with it ... one day 😄

[sacred jewel]

Thank you... yeah, as I said above, Flux just can;t get me the high res skin I'd like to have. Still looking for a good option.

[winged seal]

Wait, are those flux?

if so, that explains why I was impressed with their coherence lmao

[sacred jewel]

sorry... I know everyone here went back to the true nature of the channel... but I am still struggling with SD3.5L 🤭

Someone has identified Flux output of any human face as having a Butt Chin so they're easy to spot 😛

[craggy crest]

what sort of issues are you haveing with sd3.5 L?

[dusky thistle]

hey that's great to hear 😄

[sacred jewel]

My workflow takes the default sampler output which for me typically is 16:9 1344x768, Euler/beta, 20 steps, 3.5 guidance ... That's around 2 iterations per second.

The image out is put through either a 4x model upscale or a 2x model upscale and then either down sampled to half if 4x or passed through the vae encoder straight if 2x

Depending on the model I select, the model upscale can be slow (SIAX) or fast (ESRGAN)

Then that's fed to ksampler 2 at 20 steps and .20 denoise... Same settings as first pass.
The second pass I get about 2.5 seconds per iteration.

Total time is typically around 100 to 120 seconds.

If using LoRAs, I have seen the generations on second pass go to 3 sec per iteration.

RTX4090 BTW

[sacred jewel]

I kept getting noisy images and img2img just would not work well. So i stopped playing with it until the community got a few days with it and I could steal a few working workflows.

[craggy crest]

i'm doing a sampler/scheduler compare chart for comfy at the moment. what sampler/scheduler were you using?

[dusky thistle]

img2img is incredible

input, two outputs

just yodeling over here

[winged seal]

Currently training flux 8b lol

I am shocked this is working

[craggy crest]

but can you make the clocks melt and look like dali painted them?

[dusky thistle]

i can do this

[craggy crest]

i 'see' what you did there

[winged seal]

@craggy crestAlright, full interest is lost in SD3.5 for me, after trying flux lite. I will come bac in a few weeks when SD3.5 has better experience, but now I am training Flux lite. It looks just as good if not better than flux dev, it runs faster than SD3.5, and it learns EXCEPTIONALLY well

it works fantastic

[winged seal]

some massive changes are gonna have to be made for SD3.5 for me to go from flux lite, since its even faster than SD3.5

[bitter hearth]

I was talking about Flux lite on the Open Model Initiative discord
I'm gonna try and take their methodology further and shrink it some more, since its cheap to do this block skipping method
They also made it with BF16 in mind and I want to do it with FP8 in mind, to exploit the FP8 matmul speed boost

[dusky thistle]

input

[bitter hearth]

They dropped blocks such that the quality only dropped very slightly, I would rather drop a few more blocks to lose slightly more quality and gain a bit more speed in exchange

[dusky thistle]

output

SD35L is fucking awesome

[bitter hearth]

wow yeah that's nice

the same unsampling workflows work on SD35?

[dusky thistle]

i've got new nodes up on res4lyf today

[bitter hearth]

ah nice

[dusky thistle]

most recently updated an hour or so ago

really good results now

[bitter hearth]

ah yeah the old face-feather method

easy fashion

[dusky thistle]

hah

[bitter hearth]

I like how it keeps the colours could probably give it some low frequency rainbow noise as the initial and get colour transfer a bit

[winged seal]

I would love that

[winged seal]

I am finding it makes flux lite train even better thol. I seem to be having even better than the already phenomenal results from dev

[bitter hearth]

not sure if it is better for training but seems possible

that dropping blocks would help

[dusky thistle]

flux dev forgets as fast as it learns

[winged seal]

I've yet to see that, personally

[dusky thistle]

we need a base that doesn't have taht problem

[bitter hearth]

I found a lot of Civit Flux loras were very over-fit

[dusky thistle]

well, you're the only one that hasn't seen it afaik

[winged seal]

man, these results are making me very happy lmao

[bitter hearth]

the brown leaves in the foreground look very real yeah

Civit Flux loras also do the old classic of bringing the background with them

[winged seal]

My dataset is specifically for very dense and detailed backgrounds and professionaly shot photographs.. Its only 30 images.. I have one with 150, 339, 1.7k and 3.1k

[bitter hearth]

I have the same taste as you, I want much more dense and detailed photographic images 👍
so I hope your project goes well

[dusky thistle]

you're probably just giving it that big vector space shove with that stuff

i've seen it too with that

at a fairly high LR you can get it to cough up great photos in just 500 steps or so, far less than it needs to actually learn anything

the real trick is getting ALL of that "flux style" out without the catastrophic forgetting kicking in

which is really, really tough

[bitter hearth]

latentmegamodifier was decent at getting rid of the background blur with pyramid noise injection
it stacks on top of whatever ancestral noise you have

[dusky thistle]

interesting

i've noticed high frequency noise can help

[bitter hearth]

it also has a filter for the noise injection it does
can't remember if its low pass or high pass

[dusky thistle]

what node pack is that? frequency filters are something i'm real interested in implementing

[bitter hearth]

but crucially it has div norm
what you can do is add "too much noise"
and then div norm can save your image

https://github.com/Clybius/ComfyUI-Latent-Modifiers

the Spectral Modulation widget is awesome as well

the Gated tonemaping option is good also

[dusky thistle]

oh, yeah, that's right, clybius

he does cool shit

his implementation of RES is what got me started on all of this

[bitter hearth]

haha yeah

I need to go back to sampler supreme

it had a bunch of options that did good stuff

besides which, I wanted to explore his Spectral Modulation stuff in more detail and change the code around

cos he doesn't have any variations for that technique and I think it has potential

[dusky thistle]

i haven't looked into it at all, tbh

still need to get a multistep node implemented here soon so i can delete like 5000 lines of sampler code lol

[craggy crest]

use the tool that does what you want it to do

[bitter hearth]

oh and his node has a good version of CADS as well

[dusky thistle]

soooo cloooose to being able to clean up bigtime

sweet

[bitter hearth]

would be nice if res4lyf code got rly streamlined yeah

[winged seal]

I don't know of one at the moment, which is why I will give SD3.5 to breathe some haha

[bitter hearth]

I'm still trying SD 1.5 stuff I think I can do better than yesterday

[dusky thistle]

yeah that's in progress

SamplerRK is a huge step in that direction

i was able to add 20 different samplers to that thing just by punching in coefficients tables, with a little bit of dynamic calculation for them in the case of dpmpp and RES, and for those i was able to redo the math to have a single nice clean phi function to call

[bitter hearth]

I think the community would really benefit from a general streamlined release with nice docs
cos you legit have the best sampling library in the world right now lol

[dusky thistle]

so it's incredibly concise

aww thx

yeah that is def the goal now

tinkering with samplers is so fucking addictive

[bitter hearth]

I wanna go and learn sampling math properly at some point too
but my priority is training my first diffusion model at the moment

[dusky thistle]

as in some custom architecture?

i did a lot of screwing around with that with jupyter notebooks a few months ago, that was fun

[bitter hearth]

its gonna be replications at first
gonna do one model for each of DDPM paper, DDIM paper, EDM paper, Peebles and Xie paper

there's not much point in planning beyond that because after doing that my understanding will be completely different lol

[dusky thistle]

sounds like a fun project

yeah no doubt

[bitter hearth]

there's a couple of nice bonuses of doing it yourself
you have all the images and captions that you trained with
so you can fine tune an LLM to prompt it, using the original caption data
and you can train an upscaler like ERSGAN, DAT, ATD etc, using the original image data

[dusky thistle]

and the depth of the understanding you get from working with simple stuff from the ground up... ab initio, as we call it in chemistry... is extremely valuable imo

[bitter hearth]

yeah you have to go back to first principles to learn it properly

[dusky thistle]

pretty mutated but i'm excited about how well it's adhering to the structure of the source image

[bitter hearth]

insanely good contrast

[dusky thistle]

def

[bitter hearth]

I do miss that sort of thing in SD 1.5

[dusky thistle]

not burning or blurring

don't even have cfgpp in here yet, though i'll be able to do it now

took forever but i found a way to separate out the x term completely from the denoised/epsilon prediction and the guide

which i needed to do cfgpp

[bitter hearth]

cfgpp might help loads with unsampling accuracy

[dusky thistle]

now that's a real frog

[bitter hearth]

is this AI or garden photo

I rly need to try SD35 more, the way it is blending fantastical with real photo background is really good

since I mostly do scifi that might be ideal

[dusky thistle]

hahah

whoa just found a major fuck up with my third order for noise inv

whooooops

i actually love finding bugs like that

it's like finding candy on the sidewalk: "yay, it'll be even better now!"

[bitter hearth]

lol yeah

if the unsampling code is working well, might not be that far off getting restart sampling

seems that restart sampler adds the noise in a special way I was wrong to just spam k-samplers

[dusky thistle]

it's def possible to implement

prolly would work really well if you used this math i've worked out here

yuuuuuuuuup

before, after:

[bitter hearth]

I should have saved them, the day I tried 300 restarts the images came out so nice

[bitter hearth]

wow yeah huge improvement in the froggy

[dusky thistle]

no kidding

god i love the look of sd35

[craggy crest]

i want some of these growing in my garden

[craggy crest]

me too - and it's almost effortless to steer in the direction you want it to go, unlike flux which is an uphill battle to get it to do anything other than the very narrow range it's massivly over fit for

[dusky thistle]

for real

[bitter hearth]

hopefully SD35 comes to latent interposer soon https://github.com/city96/SD-Latent-Interposer
can send my SD 15 and SDXL latents there

[craggy crest]

and that frog flower is hysterically funny

[dusky thistle]

sai folks are real quiet on here now, probably some change in corporate policy on community interactions etc, but i'd imagine they're still reading... so yeah, hats off to you all and thank you

[craggy crest]

he's got the ggufs out for it, i'd imagine he's working on it

[bitter hearth]

ah okay nice

[craggy crest]

i promise you, the convos in here are being watched and listened to

[winged seal]

I am gonna keep training Flux Lite to get fantastic photographic results, then use an LLM to guide it to make a big dataset to train SD3.5 on when its more viable for me. I think thats gonna be a fun project

[dusky thistle]

well they deserve a lot of praise

[craggy crest]

100% agree

[dusky thistle]

SD35 is a gem

[craggy crest]

you can't call him a cool cat - is he a hot dog?

[winged seal]

I suppose so?

I'd accept it haha

[bitter hearth]

I wanna do huandit but the control nets are pickles 🙃

[craggy crest]

needs some water, he looks ready for the beach

[winged seal]

peekles

[bitter hearth]

apparently Hunyuan-DiT inpainting control net is the best inpainting method out there

not sure if that claim is correct though

[dusky thistle]

yeah, unfucking my code def made a difference, lol

[craggy crest]

a cube of ice on fire

[bitter hearth]

do you think your method would work for outpainting?
like make the square input image into a 21:9 image with outpainted details?

[dusky thistle]

maybe? it's more something that tracks with a source

so not having anything might not be ideal

but if you could get an early outpaint-ish something there, even if it was some crudely photobashed crap... it might work really well

[craggy crest]

i sent you a friend request yesterday... just in case you didn't see it

[bitter hearth]

oh thanks I haven't been clicking those, will go do so

[craggy crest]

@dusky thistle

[bitter hearth]

oh I see, could outpaint with SD 1.5 powerpaintv2 brushnet and then fix with the SD35 unsampling

there's also this node pack https://github.com/Acly/comfyui-inpaint-nodes it does crude inpaint with small models

[dusky thistle]

yeah something like that could def help

so long as you had at least the lighting... you'd have a good starting point

[bitter hearth]

kolors inpainting model is rly nice also

especially since kolors has the strongest IP adapter currently

[dusky thistle]

what time zone are you two on btw lol just curious

EST here

[bitter hearth]

GMT

[craggy crest]

i'm in Arizona - we don't do daylight savings. it's 12:16am for me right now

[dusky thistle]

so i'm the worst one in the room lol

3:16am

[craggy crest]

programmers come alive at night

we're all vampires or something

[gusty trail]

definitely vampires

[bitter hearth]

I'm quite close to the building in Greenwich which GMT is named after
its the zero point for time zones

[craggy crest]

which - is realy not where the actual line is - it's some feet away from it

[bitter hearth]

haha yeah the line isn't in the building

the problem is that crazy hill completely kills you on the way up

cardio challenge

[dusky thistle]

what a surreal concept that is to me over here

[craggy crest]

they don't have tourbuses or shuttles?

[dusky thistle]

GMT has been this alien thing... almost thought you were joking when you said GMT, like you live in outer space so you might as well use GMT

[bitter hearth]

no you have to climb

[craggy crest]

ewwww - okay, i'll stick to looking at pictures and watching videos

[dusky thistle]

lol

cfg 5.0 vs cfgpp 1.5, first test

[bitter hearth]

there's a college at the bottom of the hill, my friend went there
its cool cos the buildings look like this

[dusky thistle]

don't think i botched the code for this as it's pretty similar... should be fairly similar at those strengths

[craggy crest]

drop cfg down to 3. drop shift to 2

[dusky thistle]

wow, that's gorgeous

i'm not using standard SD3 shift

[craggy crest]

you should be

[dusky thistle]

this lil sucker allows you to pick between flux's and SD3's

exponential is the flux one

[craggy crest]

it's rather necessary

yeah. and doesn't work well with sd3

[dusky thistle]

ehhhh i actually think in a lot of cases it's much better

[craggy crest]

a LOT of what they did for flux, to make it work, is jsut patching

[dusky thistle]

with the unsampling, it's vastly better

[lavish sparrow]

what node is that?

[dusky thistle]

https://github.com/ClownsharkBatwing/RES4LYF

[bitter hearth]

seems a lot better

[craggy crest]

did you get a chance to play with APG at all?

[dusky thistle]

not yet

gotta get the sampler shit done 🤣

long march through a swamp

[craggy crest]

what? you can't code with both your toes on one keyboard, your fingers on another, and split your brain in half?

[dusky thistle]

lol

[lavish sparrow]

more samplers/settings to fiddle with? bueno!

usually i think ipndm_v is pretty bad, but with sd3.5 and high steps beta 0.6/0.9 -> it's pretty damn sweet actually

[dusky thistle]

yeah my collection is a bunch of madness

i've gotten SDE sampling working really well with rectified flow

took a lot of fiddling with the math

[craggy crest]

not madness, rabbit holes

[dusky thistle]

cfg 5 vs cfgpp 1.5

[craggy crest]

i don't like the look of the one on the right. it's blurry and the details are lost

[dusky thistle]

yea it's hard to say how balanced the strengths are there at those numbers too

[craggy crest]

do something very mundane, boring, and that looks like a phone photo

[bitter hearth]

rather than using shift I prefer to babysit the high sigmas over the course of a few ksamplers with settings tweaked to minimise error
and then once the structure is done you can blast the lower sigmas with much less conservative settings

[craggy crest]

there's a guy in tech support that might need help @dusky thistle

[dusky thistle]

cfgpp 1.7

[craggy crest]

still missing all the details in the face except the teeth

and it's starting to look like Stitch without lilo

[bitter hearth]

very cinematic vibes

[dusky thistle]

tbh kinda like the look with the face lol

freaky

[bitter hearth]

I noticed cinematic checkpoint tend to be blurrier

[lavish sparrow]

@dusky thistle dammit, why did it have to make pennywise

aight, new try 😄

they just take a while to make

[dusky thistle]

cfgpp 2

[craggy crest]

because #clown

[lavish sparrow]

yeah, crazed clown = pennywise in most image models

however, with sd3.5 good neg prompt abilities, lets see what happens

[craggy crest]

someone please go talk to the guy in tech support who is looking for an 'sdxl model for extreme violence'

cause i'm done with him

[bitter hearth]

not sure it would be ethical to give them tech support

[craggy crest]

can we feed him to one of clownshark's creatures then?

[bitter hearth]

my viewpoint is that the average person should use Dalle 3 in Chatgpt
if they are not a technical person

or other tools like that, where an LLM "helps"

[dusky thistle]

cfgpp 2.5

[bitter hearth]

texture is coming through

[dusky thistle]

idk if you grabbed those nodes but if you do this is something i'd recommend to start with

[lavish sparrow]

@dusky thistle dis u?

[dusky thistle]

cfg 5 vs cfgpp 2.5

[bitter hearth]

couldn't stop laughing when euler showed up in the code base after the dormand-prince monsters

[craggy crest]

i'm not getting the joke? why's it funny?

[bitter hearth]

this is the biggest dormand-prince for example

[craggy crest]

oky but why is euler showing up funny?

[dusky thistle]

it's like watcing an ant follow a parade of tanks on victory day

[craggy crest]

oh. okay. if i wasn't this tired if would probalby seem funnier

[bitter hearth]

its probably not actually funny I've just been laughing at eulers method for about 30 years at this point

[dusky thistle]

try not going to bed for a day or two, then everything will be hilarious

[craggy crest]

did that when i was in my teens. it didn't make stuff funnier, just lost my temper eaiser

[dusky thistle]

cfg 5 vs cfgpp 2.5

[craggy crest]

CROKERS!

[dusky thistle]

now that one is def better on the right

[bitter hearth]

yea I really have to remember to always use cfgpp stuff

the gains are big

[craggy crest]

the ddim_uniform scheduler seems to warm the images up

[lavish sparrow]

@dusky thistle this is missing from your requirements.txt 🙂 just a headsup ❤️

[dusky thistle]

that's really nice

[craggy crest]

course i'm using the recommended settings

[dusky thistle]

got a line about it on my horrible description on the repo

opensimplex... i forget what it was, but putting it in requirements.txt caused something to break

[craggy crest]

i thought it came out really good. workflow's in it if you want it

[bitter hearth]

oh you might want to know
there is a problem in comfy with ddim_uniform scheduler and flux, it may also affect SD35
it doesn't start at the max sigma so it is doing img-to-img on the latent at like 0.9 denoise

[lavish sparrow]

ah, it won't play nice in requirements.txt?

[craggy crest]

it doesn't. i'm doing the sampler/scheduler chart now

[bitter hearth]

ok maybe it was just a flux thing

[dusky thistle]

yeah i can't remember what it was

[bitter hearth]

if you use Display Any from Matteo's pack and connect the scheduler to it
you can check if the first number is 1 or not

[craggy crest]

after what robin did to flux? i'ts a wonder you can change any of the settings and have it work

[dusky thistle]

tbh i'm tempted to just take it out and provide instructions on how to put it back in

[craggy crest]

you should figure out why it's causing problems though

[dusky thistle]

i did, and there wasn't any solution

other than doing it manually

just don't care enough to put much more into it, it's only one noise mode out of many, got much more important things to fiddle with

[winged seal]

@craggy crestTraining flux lite is improving its prompt adherence a lot. I asked for harsh lighting from the right side. I am really serious about automating some image gen using flux lite to generate positional training data for SD3.5

[bitter hearth]

if it only affects the simplex noise, that wasn't a great noise anyway

[craggy crest]

someone else'll figure i8t out and come looking for tech support ;)

[bitter hearth]

simplex noise gives Flux grid in SDXL lol

[craggy crest]

what about sd 1.5?

[dusky thistle]

yeah that's crazy, the grid

[bitter hearth]

haven't tried it yet
sd 1.5 can handle noise a lot better so it might be ok

[craggy crest]

useful for texturing stuff

@bitter hearth

[bitter hearth]

hmm

[craggy crest]

prompted for it: a grid of small dots and crosses in the center of a symmetrical heart

[bitter hearth]

it was a mistake to show clownshark a building that I like LOL 🙃

[craggy crest]

heheheheheheh

especially when he's already lighting stuff on fire

night guys. please don't munch on each other

[lavish sparrow]

@dusky thistle yup, liking your sampler ^^

[noble coyote]

Try this - if you haven't already - RF_Inversion for Flux
https://rf-inversion.github.io/
https://github.com/LituRout/RF-Inversion

[dusky thistle]

which one?

[lavish sparrow]

samplerrk

or that one not yours?

[dusky thistle]

that's mine

[dusky thistle]

i'm using something inspired by this in my code

the img2img stuff

[lavish sparrow]

@dusky thistle

[dusky thistle]

try res_2s too

res is a lot like dpmpp, except some of the math is fixed

it should edge it out a bit overall, i think

[lavish sparrow]

song lyrics + LLM t5 + sampler RK. -> that's some scary coherency

[sage burrow]

By lite do you mean shcnell? I never did like that one. Or do you mean hyper?

[winged seal]

No, I mean Lite

its a pruned version of FLux Dev from 12B down to 8B, and it runs faster than SD3.5, and trains better than base flux dev for me. Its fantastic!

they removed a bunch of blocks that do very little, and it so dope. Flux Dev LoRA's work on it perfectly, and you can train it

[noble coyote]

RFI is totally rad 😄

[winged seal]

so its 8B, it uses less memory, and it needs less steps than SD3.5, so I can get a much higher quality Flux image in less time. Its awesome

I wanna train it to generate super high quality random and diverse photographs to train SD3.5 on later

[dusky thistle]

give it a shot from my repo, got SDE sampling and evertyhing working with it

i totally redid the math

cfgpp is supported now, it's using RES sampling which is top notch for diffusion models

[dusky thistle]

WF should be in this nice building that neoninjastro likes that i casually set on fire

[lavish sparrow]

when a demon says: "you're the apple in my eye" ...

[lavish sparrow]

very crispy

[winged seal]

https://huggingface.co/Freepik/flux.1-lite-8B-alpha/tree/main

[dusky thistle]

sd35 is a joy

[winged seal]

it ROCKS

[dusky thistle]

so much better than being stuck with a distilled model like flux imo

[radiant ledge]

a distillation of a distillation?

[lavish sparrow]

yeah, i'm REALLY liking it too, the longer i playh with it

[turbid grotto]

it is faster due to not having CFG or something else?

[dusky thistle]

yup every day i'm loving it more

i think it blows away flux tbh

[winged seal]

its 8B params instead of 12B

[lavish sparrow]

it's more fun/creative, and i feel it can do more complex prompts

[dusky thistle]

flux is pretty boring

[winged seal]

no, they cut out sections of the model that are barely used. Its not really trained, as much as it is surgically reduced

[turbid grotto]

I menat in comparison to sd35 8b

[dusky thistle]

you can't just cut out parts of a model

you gotta distill it again, so it's redistilled i'd imagine

[winged seal]

oh no its faster cause you can get well refined images in about 25 steps, vs 40 for 3.5

Any less than 40 for me looks super melty and messy

[lavish sparrow]

never got flux to do this...

might have to re-run my old prompts, again. see how much they improved

[turbid grotto]

oh okay, thanks

[dusky thistle]

yeah and we're just using a base here

anyone that thinks this isn't great should go compare zavy v7 to sdxl 1.0

[winged seal]

they cut out blocks of the model after finding out some do very little

[dusky thistle]

oh, i see what you meant

i though tyou were referring to knowledge being ablated/removed

[winged seal]

8B actually does text better than dev by a lot across the board from what others and myself have seen

oh no, they removed sections of the models process

[turbid grotto]

how is it possible

[lavish sparrow]

also, not overloading the clip-l and clip-g seems to work wonders in terms of image cohesion

[winged seal]

and applying my dev LoRA's on it, they express and look even better on lite, and lite trains faster and more robust it seems

[turbid grotto]

does that mean Flux did not reach full potential of 12b params?

[winged seal]

yeah, for sure

just like how SD3.5 is nowhere even remotely close to full potential of 8B

image gen models are exceptionally low density

more params do help with some things, but these models could be trained so so so much more

[lavish sparrow]

same as in LLM world -> you could hardly have a coherent conversation early 7b models, now they are plenty capable

[winged seal]

like, with LLM's, the density of smaller LLM's is insane.. They are reaching the limits of transformers, hitting the loss threshold of what can be encoded in human language

[turbid grotto]

we need Meta to do img model

[winged seal]

Dude, I have been messing with gemma 2 2b, and my god, that model is shockingly coherent

it blows mistral 7b out of the park in terms of knowledge haha

[dusky thistle]

img2img, WF in second (output)

[winged seal]

and gemma 2 27b continues to be the single smartest fine information model I have tested. When asking questions about celebs, shows, or things from fandoms, gemma 2 27b leaves even the best closed models (4O, Gemini ultra, claude 3.5 sonnet) in the dust

but, its all about fine information retrieval, its not nearly as broadly functional

but like, my best friend is an accomplished producer. When I asked Claude 3.5 sonnet and 4O for info for an essay, they averaged 12-16 points with about 30% accuracy between 10 responses each

Gemma 2 27b averaged 14-22, with about an 85% accuracy

[remote holly]

is sd3.5 q5 gguf good ?

[winged seal]

yeah, should be fine

I use the 4, and it looks great

[lavish sparrow]

i'm using the hybrid fp16 which just fits in 24gb, 3090 XD

[remote holly]

i use fp8 sd3.5 but is slow , if i use a quant model it should run faster ?

[lavish sparrow]

quants are slower, actually

[remote holly]

or quant is just for low vram ?

[lavish sparrow]

quant is for lower vram, correct

[remote holly]

ha

[winged seal]

Yeah, ggufs get slower the lower accuracy they are

[remote holly]

i should stay in fp8

[bitter hearth]

the top models like Gemini, Claude and GPT are designed with the idea that you add the subject knowledge yourself via RAG or fine tuning
so its correct that they have less subject knowledge out of the box

[winged seal]

I hope we can get exl2 support for image gen models, cause that would be insane. Its a little more size efficient, its faster for inference, and it also gets faster the lower the precision

like llama 3 8b is about 70t/s for me at FP16 but at 4 bit exl2 its closer to 140t/s

[turbid grotto]

could it work with omnigen as it is based on llm?

it is crazy long rn

btw, sd35m will be about 5gb fp16, so it is ~2.5gb q8 and ~1.25gb q4 😂

[sage burrow]

By trainable, do you mean anything which can be done online? Small gpu here.

[winged seal]

I would assume so, I would love to see that

[noble coyote]

Do these PNGs contain your w/f? 🙂

[dusky thistle]

yep

[winged seal]

anything that can train flux can train flux lite, so any cloud service you use

Also it works in Kohya for FFT, which can run on 6GB for dev. Its slow-ish, but it works!

[dusky thistle]

you'll need my res4lyf repo to use em

[winged seal]

flux dev LoRA's also wor great on lite. I am training on lite and it trains faster, and seems to be gathering info faster as well too

[lavish sparrow]

just milk.

[sage burrow]

Ohhh! That's what all those nodes are!

[sage burrow]

My flux nsfw loras never worked that well (neither dud anyone elses), so it's tempting to try.

The sfw ones worked well, with a lot of work.

[dusky thistle]

all kinds of stuff, but it's mostly sampling and sigmas nodes

i have a few image manipulation ones like stuff for frequency separation

[winged seal]

I saw tons that worked for NSFW dudes, but thats all I look for. 90% of LoRA's are poorly made straight NSFW get off slop LoRA's lol

[noble coyote]

CSBW Nodes Missing - Manager Cannot Locate

🙃

[dusky thistle]

this was the source image

[dusky thistle]

gotta git pull res4lyf

https://github.com/ClownsharkBatwing/RES4LYF

[sage burrow]

There's a few that show images by the lora creator, but only a couple which produce good results for other people 😦

[noble coyote]

OK, update RES4LYF

[dusky thistle]

If you are using a venv, you will need to first run from within your ComfyUI folder (that contains your "venv" folder):

Linux:

source venv/bin/activate

Windows:

venv\Scripts\activate

Then, "cd" into your "custom_nodes" folder and run the following commands:

git clone https://github.com/ClownsharkBatwing/RES4LYF/

cd RES4LYF

If you are using a venv, run these commands:

pip install -r requirements.txt

pip install opensimplex --no-deps

Alternatively, if you are using the portable version of ComfyUI you will need to replace "pip" with the path to your embedded pip executable. For example, on Windows:

X:\path\to\your\comfy_portable_folder\python_embedded\Scripts\pip.exe install -r requirements.txt

X:\path\to\your\comfy_portable_folder\python_embedded\Scripts\pip.exe install opensimplex --no-deps```

for anyone that hasn't used it

[winged seal]

I would gladly train my own, but only for men 😅

[dusky thistle]

yep, new stuff there

[lavish sparrow]

love the coherency in the image

[sage burrow]

There's one that works, it's on civitai.

[turbid grotto]

sd35m still tomorrow?

[dusky thistle]

afaik

tuesday

[turbid grotto]

curious how they'll make it work

[sage burrow]

2weeks

😉

[turbid grotto]

[winged seal]

I saw a few really good ones. No idea how the straight crap is. Its always looked horribly trained for me. I guess cause there is a lower skill floor to get "popular" for it lol

[noble coyote]

Updated RES4LYF, added OpenSimplex - still Missing the Nodes

[dusky thistle]

is there an error at the console

did it actually load the node pack

[dusky thistle]

check the console

[noble coyote]

Console?

[dusky thistle]

yeah

[noble coyote]

Guide me towards the console - do you mean the terminal? Ah, OK

[sage burrow]

I did futa ones (twice as difficult lol), now wanting to use my training data for sd 3.5 loras. Trying to figure out replicate...

[dusky thistle]

most often you just need to reload your browser window

[noble coyote]

Says RES4LYF failed to load

[dusky thistle]

comfyui is pretty annoying about it

k can you paste the error?

[noble coyote]

Gotta go disable a few nodes ...

[winged seal]

oh god, my bane

[sage burrow]

Does anyone on here use replicate for lora training? I can't tell if it's my HF setup that's messing every one up, or?

[turbid grotto]

replicate moderators:

[dusky thistle]

you'll want to go farther up

that only says it failed but not why

[sage burrow]

Think that is why it keeps failing? Apparently nsfw is allowed on the replicate api, but that's currently beyond my skill level

[noble coyote]

Here's the Console

[bitter hearth]

I think Fal might be a better choice than replicate

in general

[sage burrow]

I have $10 in credits on replicate tho! Will go look at fal lol

[dusky thistle]

Cannot import X:\ComfyUI_windows_portable\ComfyUI\custom_nodes\PyramidFlow-ComfyUI module for custom nodes: No module named 'tensorboardX'
Traceback (most recent call last):
File "X:\ComfyUI_windows_portable\ComfyUI\nodes.py", line 2001, in load_custom_node
module_spec.loader.exec_module(module)
File "<frozen importlib._bootstrap_external>", line 940, in exec_module
File "<frozen importlib._bootstrap>", line 241, in call_with_frames_removed
File "X:\ComfyUI_windows_portable\ComfyUI\custom_nodes\RES4LYF_init.py", line 1, in <module>
from . import extra_samplers
File "X:\ComfyUI_windows_portable\ComfyUI\custom_nodes\RES4LYF\extra_samplers.py", line 11, in <module>
from .refined_exp_solver import _refined_exp_sosu_step_RF
File "X:\ComfyUI_windows_portable\ComfyUI\custom_nodes\RES4LYF\refined_exp_solver.py", line 12, in <module>
from .noise_classes import *
File "X:\ComfyUI_windows_portable\ComfyUI\custom_nodes\RES4LYF\noise_classes.py", line 12, in <module>
from opensimplex import OpenSimplex
ModuleNotFoundError: No module named 'opensimplex'

there's your culprit

[turbid grotto]

idk, maybe flux too overbaked to learn something new that easily?

[dusky thistle]

in your venv you gotta run this

pip install opensimplex --no-deps

[noble coyote]

pip install tensorboardx

[dusky thistle]

If you are using a venv, you will need to first run from within your ComfyUI folder (that contains your "venv" folder):

Linux:

source venv/bin/activate

Windows:

venv\Scripts\activate

Then, "cd" into your "custom_nodes" folder and run the following commands:

git clone https://github.com/ClownsharkBatwing/RES4LYF/

**cd RES4LYF

If you are using a venv, run these commands:

pip install -r requirements.txt

pip install opensimplex --no-deps

Alternatively, if you are using the portable version of ComfyUI you will need to replace "pip" with the path to your embedded pip executable. For example, on Windows:

X:\path\to\your\comfy_portable_folder\python_embedded\Scripts\pip.exe install -r requirements.txt

X:\path\to\your\comfy_portable_folder\python_embedded\Scripts\pip.exe install opensimplex --no-deps``**`

look at the escond half there

depends on if you're using a portable or not

[noble coyote]

Says I already have OpenSimplex ... mebbe I didn't install inside venv

[sage burrow]

Trying with sd3.5 but it keeps failing. I'm half sure it's just something like wrong HF setup

[bitter hearth]

look on fal discord there is $10 coupon

[dusky thistle]

so you either activate your venv and run:

pip install opensimplex --no-deps

or you just run this

X:\path\to\your\comfy_portable_folder\python_embedded\Scripts\pip.exe install opensimplex --no-deps

and then you should be good to go

[turbid grotto]

I am still waiting tools

[bitter hearth]

fal have a rly fast endpoint for SDXL and Flux Dev
is why I was recommending them

[bitter hearth]

bird has come to defend the naval college that got destroyed

[noble coyote]

Doesn't like me updating inside Scripts

Where is that '"D:\a\ComfyUI\python_embeded\python.exe" coming from? I don't have anything on my D:\ drive?!

[sage burrow]

Where? 😄 (I joined)

[bitter hearth]

they mentioned it at some point the the last month or so
I hope the offer didn't expire

[noble coyote]

🥳

[turbid grotto]

for some reason flux8b about the same as sd35l in generation speed while not having slowdown from negative unlike sd35l

but not as slow as flux12b

[raw shadow]

Are there any authors of sd3 or sd3.5 here?

[sage burrow]

How come? BTW do they have sd 3.5 lora training yet?

[bitter hearth]

fastest SDXL and Flux Dev endpoints

they also make auraflow, which is nice

[sage burrow]

lol

[bitter hearth]

lol auraflow is nice its just not done yet

[rapid pivot]

strawberry juice flower

[lavish sparrow]

dragoncat sleeping in cozy hot spot

[noble coyote]

RF Inversion for Flux - Prompt = Angry Boy Cartoon Style. Input Image on left

[turbid grotto]

i love auraflow

[noble coyote]

I need a TensorRT for AuraFlow ...

[winged seal]

that is impressively MJ

[gusty trail]

sd3.5 with lora

[winged seal]

oh yeah, I know, just impressively MJ aesthetically aligned

@gusty trailwhat trainer are you using? I don't remember if you have said

[gusty trail]

I used my t2itrainer

https://github.com/lrzjason/T2ITrainer

[bitter hearth]

lower sigma_noise (s-noise) and possibly also lower eta

[dawn spindle]

hey yall, what's the most powerful model with the best speed to quality ratio to use on say an 8gb vram card? are there any nf4 bnb models like with Flux? Thanks

[noble coyote]

Anybody want Corel Painter 2023 for $28 (legit on HumbleBundle)

[noble coyote]

My 8Gb VRAM works well with flux-dev-fp8, flux.dev, Flux_GGUF_Q8

[sacred jewel]

Any plans to include them in Manager?

[lavish sparrow]

i love sd3.5 more and more each day

concidentally 🙂

the OG prompt was "an abstract creature made of shadows and mist" -> gets fed into LLM, which then makes 3 prompts for clipl, clipg, and t5

for flux, it didn't matter that much tbh, but i notice that sd3.5 REALLY likes if you do that

[pseudo owl]

Yeah flux doesn’t even need the clip text encoder really, it has very very slight impact on the image.

[lavish sparrow]

overloading the clip gives the images the SD3.5 instability, so keeping that one short and concise helps a lot

[noble coyote]

SD3.5L Turbo 25 seconds, Flux.GGUF_Q8 90 seconds

[lavish sparrow]

SD3.5 FP16 50 steps + Res 2s (so 100 steps) + LLM inference: 130sec

[noble coyote]

Extraordinary!

[noble coyote]

Is the w/f in the image at all?

[lavish sparrow]

of course, i have nothing to hide ^^

now trying res 3s, but that's gonna slow it down even more

[noble coyote]

450 seconds

[lavish sparrow]

but, it does get results ❤️

now experimenting with noise samplers

[dusky thistle]

what settings were you using

[dusky thistle]

yea... just got more cleanup to do

[hallow lion]

I can;t even imagine what the promt for these is...

[noble coyote]

Ten Hag - Ten Gone! S a c k e d at last!!!

https://discord.com/channels/1002292111942635562/1237460360634830939

"Generate marzipan and custard"

[proven pecan]

SD3.5T ❤️

[noble coyote]

Hello

SD3.5L Turbo LLM

[signal shuttle]

Oh right, SD 3.5M is releasing tomorrow, shame i can't test it out tomorrow because of work 😦

[craggy crest]

:(

[craggy crest]

@dusky thistle in case you missed this #🆕｜sd3 message

[dusky thistle]

should be fun testing out all these sampling nodes with that... SDE works fantastic with flux and SD35L

[craggy crest]

:) i can not wait to see what you post once you get your hands on it

[turbid grotto]

but how is it for t2i?

[noble coyote]

How will it cope with marzipan hegemony?!

Or herringbone custard?

Will we ever know?!

[lavish osprey]

Very detailed but of course Large is better overall.

It's good for specific use cases (a small model with very low vram requirements) or as an upscaler.

[mortal mesa]

if you guys trained a cnet also 🚀

[somber briar]

Hello everyone. Have any plans to realese 64ch vae for SD3 L?

[bitter hearth]

oh nice its the SD3.5 2B\

looks great, this level of high frequency detail would be good for tiled upscale

upscale pass doesn't need as strong a model

[craggy crest]

@lavish osprey hasn't said anything that i've seen

[bitter hearth]

the original SD3 paper ablated VAE size and found 16 a good compromise

if I remember rightly

[lavish osprey]

64ch VAE 👀

holy hell, even 16ch is already overkill

and it's basically underused most of the time

also how slow would that be uhm 🤔

[bitter hearth]

Sana is trying to go the other way and make VAEs more efficient

at high res and low step count (I mean like 1-4 steps) VAE decode is an issue currently with existing models

[sacred jewel]

Oh yeah. 🥰

[somber briar]

That may be the case. I tested with sdxl when training a model for anime, the vae was a bottleneck and really broke the characters eyes if they were not up close. I haven't tested the new 16-ch vae yet, but on reddit I think it was written that the sd3 8B model will have 64channels vae for the sd3 8B model

[lavish osprey]

sd3.5 Large (8b) is already out and has the same 16ch vae as sd3m

also sd3.5 Medium (2.6b) is made to reach 2k resolution and still uses the same VAE

[bitter hearth]

as far as I know there is a cost to beefier VAEs beyond just their size and execution speed, because they make the diffusion model harder to train
so the juice has to be worth the squeeze

[craggy crest]

i'm going to assume the reddit poster might have misunderstood something

[pseudo owl]

yeah sdxl's vae isn't that great compared to the 16ch vaes and I also did experience that even realistic human faces far away were broken. That shouldn't happen with 16ch vaes.

[bitter hearth]

if you get the final resolution of the image high enough then VAE issues mostly go away
even for SDXL or SD 1.5 VAE

16ch VAEs will allow better quality at 1024 though

[craggy crest]

come back next week and we'll see if it's ready?

commercial photography studio shoot. soft, diffused lighting. spotlight on a table. on the table is a crystal perfume bottle full of pale pink liquid. a gold label with the words "Rose Touch" adorns the bottle

[pseudo owl]

mochi-1 with your prompt, removed the "rose touch" part tho since it can't do text.

basic but ok

[craggy crest]

oh pretty :) nice start to a commercial, that would be

[lavish osprey]

cool effect, prompt?

[lunar canopy]

i was just abt to ask the same thing lol

[cunning lintel]

There in lies the problem :p SD3.5 is lovely chaotic, but in this case doesn't really follow the prompt, the effect doesn;t repicate at all for other prompts

A textured, tempera painting style on white ricepaper, the textured tempera painting is digital anime themed and depicts a close-up shot of a woman's hand, holding a paintball gun, the gun is small and compact, with a bright yellow paint splattered on its side, a vibrant splash of color against the muted tones of the background, the woman's face is out of focus, a soft blur in the background, but her eyes are determined, her gaze fixed intently on some unseen target, her lips are set in a firm line, a resolute expression that speaks of focus and resolve, the background is blurry, but the faint outlines of trees and a road are visible, creating a sense of movement and action, as if the woman is poised on the brink of something, ready to take the leap.

[lavish osprey]

I was thinking it could have been just very low cfg

you get crazy stuff like that

thanks!

[cunning lintel]

is 3.5 low?

[lavish osprey]

4.5 cfg if you want consistency, lower if you want craxy looking art

[cunning lintel]

Cool, hadn't noticed a whole lot of consistency loss going below 4.5, but i'm biased towards crazy looking :p

[pseudo owl]

Prompt?

[lavish osprey]

https://civitai.com/images/36533271

[craggy crest]

jedi musk

[pseudo owl]

lol what’s on pikachus leg? But still pretty amazing for open source.

[patent acorn]

its the dayy

[lavish osprey]

https://x.com/Lykon4072/status/1851012939538710738

Lykon (@Lykon4072) on X

SD3.5L

Twitter · Oct 28, 2024, 9:27 PM

[patent acorn]

wait then, can you try 3 woman in 3 different artstyles (split screen)? i remember doing this prior sd3 medium

[dusky thistle]

using any SDE samplers on SD3.5?

[lavish osprey]

it works better if subjects are different, 3 women will just get mixed up. I already di in that case

[dusky thistle]

dormand-prince SDE with SD3.5L

[lavish osprey]

nope

[dusky thistle]

they're pretty great

[craggy crest]

the superheros gather for a groupshot under a tree at the park. they are: iron man,spider-man,deadpool,batman,wonder woman

[lavish osprey]

oh cool, which one?

[dusky thistle]

i had to rework the math but really happy with the results

[craggy crest]

oh you missed his flaming shark images last night as he was playing with that

[dusky thistle]

https://github.com/ClownsharkBatwing/RES4LYF/

[patent acorn]

a women at the left in photographic, a women at the middle in anime style, a women at the right in watercolor style
so this doesnt work?

[dusky thistle]

currently building this node that will have as many explicit samplers as possible implemented as a butcher tableau:

[lavish osprey]

oh interesting

[dusky thistle]

i'm working toward a much more general release, once i implement a multistep version of that, should be able to delete like 80% of the code bloat from screwing around in my repo

def cuts down on the mutation/artifact rate, having just the right amount of noise

SD3.5L is real consistent once that's dialed in

[dusky thistle]

fantastic model btw!

can't say enough how glad i am 8b was released

[craggy crest]

the term watercolor has such a lot of weight it tends to overpower everything else and turn the image into a watercolor

[dusky thistle]

img2img

[patent acorn]

oh that sucks

[craggy crest]

it does, unless you do want a watercolor image, then it's really nice

[lavish osprey]

nobody is happier than me

[craggy crest]

create an image of three women. draw one as a sketch, create a photo of one, and do a watercolor painting of one

[craggy crest]

that's debateable. he's been dancing

[dusky thistle]

it's such a huge step up over flux

first model i've been truly excited about since cascade

img2img

flux is really limited and boring, just no soul, 8b has life

[craggy crest]

and that shark has teeth

[lavish osprey]

I'd prompt that differently.
Something like "an image that's a blend of 3 different styles, from left to right, three women are side by side each with a different style. The first woman is photo, the second woman is watercolor, the third woman is a sketch"

but if they were 3 different "names" that would be easier, like a woman, a dog, a plant, or a woman, a man, a cat

[dusky thistle]

i've had my eyes on 8b since day one

[craggy crest]

an image that's a blend of 3 different styles, from left to right, three beings are side by side each with a different style. The first woman is photo, the second dog is watercolor, the third flower is a sketch

i think i work on this one for a bit

[lavish osprey]

uhm I'll try to find the one from my image

anyway maybe remove first, second and third and don't use 3 beings

I'd rephrase it completely with 3 different entities

[craggy crest]

artist at work

Vibrant, iridescent, swirling scales-inspired mosaic, shifting colors in response to light, with shimmering turquoise, emerald, and sapphire hues, evoking the luminous beauty of opalescent glass, intricately arranged in a mesmerizing, fractal-like pattern.

[lavish osprey]

https://x.com/Lykon4072/status/1851047697370783966

Lykon (@Lykon4072) on X

SD3.5L + M Upscaling

Twitter · Oct 28, 2024, 11:45 PM

prompt Portrait of a female rocker, expressive eyes, tattoos, attitude, sultry, pictorialism, dystopian, grim dark, goth, chiaroscuro, contrast, surrealism, gritty background, , grunge, graffiti, photorealistic, professional studio lighting, extreme photorealism, epic insanely beautiful, on the beach, (ultra-wide angle:1.3), pale skin, shadow, underexposed, polaroid color sx-70 instant film, faded, blurry, motion blur, ((grainy)),nicolas delort, light silver and turquoise Fujifilm X-T4, Sony FE 85mm f/1. 4 GM --ar 51:91 --style raw --s 750 --v 5. 1

[bitter hearth]

they look really sharp, but not over-sharpened, its a good balance
very natural skin texture also

[bitter hearth]

sdxl

[gritty steeple]

A tiled upscale for SD3.5

[cursive frigate]

very demure

[bitter hearth]

LOL I heard about that meme

[patent acorn]

atleast better than dalle 3

[bitter hearth]

I think we have finally caught up to dalle 3 yeah

and midjourney for that matter

at the moment open source is essentially ahead, feels a bit weird

[craggy crest]

lots of movement in the AI video sector

[bitter hearth]

oh yeah I forgot about video completely this summer/autumn

[craggy crest]

an impressionistic painting of people having lunch beside a lake

[errant dust]

It isn't a crippling issue or anything, but SD 3.5 L has issues with such high res. 1280 x 1280 or 1920 x 896, etc.

[winged seal]

the dreamshaper is exceptionally strong in this one

[cursive frigate]

Does anyone know of a way or a node that can either schedule to run a queue or pause a queue at a set time say right around midnight so when the date changes it will start the queue again on the new day for folder naming purposes?

[craggy crest]

@dusky thistle might write you one if you bribe him

[craggy crest]

how's the taking a break going?

[winged seal]

I have just been training flux lite with extremely good results, but its not the most hands on, so I consider it a break haha

[craggy crest]

somehow i didn't think you were gonna really be able to take a break

[winged seal]

That read is very accurate for me 😅

I have spent an absolutely absurd amount of time and energy on AI. Probably one of the most dedicated people in the scene, in my own way. I wasn;t kidding when I said I have trained over 1k LoRA's for SDXL for research purposes

[craggy crest]

let's see - starting the week after sd1.4 was released: 12 to 16 hour days, every day, put in a one word prompt or short phrase, generate. change the seed by 10 numbers. again and again and again. go another direction, explore the next small space, same schedule. lots of extremely deep dives. till now. is that your schedule?

[craggy crest]

what makes you think a lora was used for that?

[winged seal]

I started 6 days after SD1.0 was released, and I was there for release of MJ V1

I have bee using local LLM's since GPT2 was serious beans, back in the start of 2020 with AI dungeon and novel reader, or whatever its called.

I used all of the 1.X models and had over 1M generated images just stored locally on my PC, generating for friends and family, as well as massive grids haha

[winged seal]

oh no, I just meant its overbearingly "dreamshaper aesthetic", which makes sense case... Lykon

no idea if a LoRA was used or not

[craggy crest]

wasn't lykon that posted that i don't think

[winged seal]

I know, I know, but Lykon trained 3.x

SD 3.0 Medium was basically a dreamshaper model lol

[craggy crest]

i predict the need for new drivespace soon

[craggy crest]

i'd argue with you but i'm tired of beating a dead horse

[winged seal]

3.5 seems to be much more neutral, but its dreamshaper aesthetic roots pop out very intensel sometimes

[craggy crest]

pictures of your new training?

[craggy crest]

for all you know, that image was created with flux. he didn't say what he used

[winged seal]

was messing with a mix of old trains on flix. Left is Lite, right is full Dev

Not too serious of a test, just a random assortment

[winged seal]

yeah I know, I didn't claim it was anything specifically. I could totally believe it came from something else. It still has an extremely strong dreamshper aesthetic tho

[craggy crest]

not too bad. is that the same black couger/cheeta prompt you used last night?

[winged seal]

yeah, but at a different aspect ratio, so it changed considerably

[craggy crest]

what are you calling a dreamshaper aesthetic? other than digital art

[winged seal]

the huge amounts of artificial texture and lighting. its an extremely "AI" aesthetic. Lots of people like it, I am just not one of those people, personally

its not bad, just very very strong in that image is all

like dreamshaper turbo

like this

just soooo much everything

I get why people like it, its just not for me

the people in my researchgroup refer to it as "Aetheti-slop" lol

[craggy crest]

so - digital image. i can get the same look out of Daz studio - and have in the past

what do you think of this out of curiosity?

[winged seal]

Its a specific type of digital image/processing/aesthetic

Just had dreamshaper get really popular with that aesthetic specifically, so I and a lot of people refer to it as the dreamshaper aesthetic

[winged seal]

looks interesting, definitely not the aesthetic I am talking about

[winged seal]

here like this

Sooooo much texture and "detail" and tonal contrast to make it more visually interesting, but it just passes a threshold, and ends up lookig super cheap and ugly, at least to me

just so visually messy, to the highest degree

[winged seal]

this is just like 2005 RPG vibes haha. I like it

[craggy crest]

i don't see what you see apparently. to me that just looks like one of millions of digital images that have been being created for years in a number of rendering programs

[winged seal]

this doesn't look even remotely like what I am talking about, if thats why you sent it

[craggy crest]

jsut posting stuff

[winged seal]

ah ok

well it looks cool 😉

[craggy crest]

mutant brocolli

the thing in the back of your fridge when the light's off

[dim geyser]

what's the M upscaling? these are really good!

[craggy crest]

he's asleep i think

[dim geyser]

ah maybe he'll see tomorrow

[winged seal]

likely just a pass with medium. He said Medium is supposed ot do 2048x, so it might be better as a refiner

[craggy crest]

he did post one that did use medium as a refiner. the quokka in a suit

[dim geyser]

ahh 3.5M isn't out yet until tomorrow right? he must have access I guess?

[winged seal]

That leads me to believe that medium will be a lot less coherent, but more "detailed", like the original refiner

Curious to see

[craggy crest]

you'll get your chance tomorrow if the world doesnt' blow up before then

[winged seal]

haha, good point

[craggy crest]

he's, um, one of the SAI researchers. so ... yes?

[dim geyser]

gotcha, should've looked at his user roles.

[craggy crest]

the developers and programmers usualy have access to the code

[craggy crest]

not sure what role he has on this discord. a lot of SAI people don't have roles that identify them

[dim geyser]

he does have a role for it

[turbid grotto]

today is the day

[remote holly]

Will sd3 medium have a decent inference time ?

[turbid grotto]

I expect about 20s on rtx3060

it's the size of sdxl

The only thing I am worrying is that it being advertised as refiner

[flat oracle]

I tried on 3060, took about 40-50 secs with no negative promt. (comparable to flux)
Although i did run comfy in reserve-vram mode with 8 gigs,

[turbid grotto]

sd35m did not out yet

[lavish osprey]

He's talking about Medium

[lavish osprey]

Prompt?

[lavish osprey]

So cool

[turbid grotto]

hope bytedance will make hyper lora for sd35l 🙏

[lavish osprey]

Nice!

[dusky thistle]

what's not nice is the mess i made tonight... fuck lol

[craggy crest]

backups are your friend

[dusky thistle]

oh it works great

just going to take a bit to reduce the math back down to something portable

and easier to manage/modify/etc

[craggy crest]

if i didn't know better, i'd think you made a math mess just to have the fun of distilling it

[bitter hearth]

advertising it as refiner is probably optimal TBH
since upscales don't need as strong a model it makes sense to cascade down the model sizes as steps go on

[dusky thistle]

yeah like how turbos are used for upscales a lot now

[bitter hearth]

yeah my upscales are usually TCD or PCM and 2-20 steps

whereas the initial image was often 60-1500 steps

[turbid grotto]

true, but it would be cool to have sdxl replacement for t2i

[turbid grotto]

[bitter hearth]

it would yeah

[dusky thistle]

just get pair of 5090s, one for positive and one for negative conditioning

thats pretty much my plan

[bitter hearth]

I think SDXL/SD1.5 are still strongest models overall, given their amount of libraries

[turbid grotto]

I am always using hyper lora 8steps

[bitter hearth]

I'd consider used A6000 non-ADA as well
after the date of the 5090 launch
will be slower but 40GB at home might become viable with that card's price drops at that time

[dusky thistle]

yeah

[turbid grotto]

they were absorbing data past years, I am not sure they are going away anytime soon

[dusky thistle]

but for inference, dual 5090 will be insane

gonna have to modify the comfy code a bit to get it on separate devices but it'll be wort hit

[bitter hearth]

it would be insane yeah

[dusky thistle]

will be insane, i'm totally gonna do it lol

43520 cuda cores for inference

[bitter hearth]

H100s/H200s aren't even that much faster than 4090s
dual 5090 will likely be faster than a single H100 lol

[dusky thistle]

prolly eah

[turbid grotto]

[bitter hearth]

I thought that not having ZTSNR and 16 channel VAE would make SD1.5 and SDXL not useful any more
but if you have done tiled upscale up to 8k+, the VAE artifacts are minimal with SD1.5 and SDXL

plus SDXL actually has ZTSNR I am just too lazy to set it up
you can merge CosXL with each SDXL model you want to use, and you get your ZTSNR

I don't know if using CosXL would break a ton of SDXL comfy nodes or not

[dusky thistle]

it does i think

it's too bad cosxl came out so late

[bitter hearth]

at some point Emad said that CosXL nearly made it as the original SDXL release 😿

[dusky thistle]

i really wish it had been it

[bitter hearth]

we could have had good blacks and colours since July 2023

[dusky thistle]

yeahhhh

first we got was cascade

[bitter hearth]

wait, cascade is ZTSNR? 🤔

I might have made a mistake by skipping cascade tbh

[dusky thistle]

i think?

it's great with it

well, just like cigarettes, it's never too late to start 😛

[bitter hearth]

is there a best setup that you guys worked out for UltraCascade?
I remember there were some fine tunes and a tiled ksampler or something

[noble coyote]

SD3.5L Turbo

[dusky thistle]

yeah i fine tuned stage B

got some real good WFs in the thread somewhere

the most aesthetically plseaing shit i've ever generated was with ultracascade, zero question

[bitter hearth]

somewhere in the thread someone layed out the folder structure that is needed, I've gotta go digging

[dusky thistle]

it's really simple now

just need everything in models/unet

where it normally goes

[bitter hearth]

ok so I can just take this one Intriguingly, the ultrapixel_t2i.safetensors file available on the original HuggingFace repo is not actually a safetensors file, but rather a pytorch save (a pickle). I converted it to a legitimate safetensors file: https://huggingface.co/ClownsharkBatwing/ultrapixel_convert/blob/main/ultrapixel_t2i.safetensorsand this one I finetuned stage B lite and highly recommend using it, even in place of the full weights. I've found it generally leads to sharper, more coherent details, with a significant reduction in "nasty Cascade noise": https://huggingface.co/ClownsharkBatwing/Cascade_Stage_B_CSBW_Refined/blob/main/stage_b_lite_CSBW_v1.1.safetensors
and then install ClownsharkBatwing/UltraCascade to custom nodes folder?

[remote holly]

Yeah

Same

[livid plover]

medium 3.5 drops today?

[bitter hearth]

not sure

[cobalt moon]

In 2-4 hours ig?

UK just reach midday

[noble coyote]

3.5L dropped at 16.15 UK Time ...

Flux/Florence2 i2i