🆕｜sd3

[icy drift]

@unkempt compass Converted the quantized version to .safetensors and now it's 15GB. 😕
This seems like a waste of time. I just don't know what I'm doing at all.

[unkempt compass]

Me neither :/

[icy drift]

A crane is flying over a lake at dusk. Yep. That's a crane.

[dusky thistle]

hahah

it's nothing like my workflows from before i started coding stuff for this

at least

some of those had 500+ nodes

ypeah

all SD35M

[noble coyote]

Flux RF Inversion - Style Transfer

[dusky thistle]

it's partily the resolution, actually

it's not trained super well at 1920x1152

[noble coyote]

Impressionist art "is noise!"

[dusky thistle]

generating my ass off over here 😛

lol at this one

these are all one shot

i generated overnight, i'm picking probably one out of two or three here

here's 9 consecutive

what gpu are you using? that's a good first q

so i know what settings will make you feel like the universe is ending from age

k

i'd try these first, with eta = 0.5, 0.25, and 0 for ODE

for SDE it might be better at 0.25 vs 0.5, not sure

SD3.5 behaves quite differently with noise vs flux. not necessarily better or worse but it absolutely responds differently, it's interesting

SDE will take the same amount of time (or should)

the only thing that makes SDE a SDE not ODE, is that it adds a bit of noise after each step

what i just sent should

but def play with it a bit

it's worth having different configs if it gets you better results, that's "fine tuning" your inference params and is very much worthwhile imo 🙂

ahhh sharksampler doesn't really do a whole lot of crazy stuff, it really just gives you extra noise options

clownsampler/samplerRK (they're the same thing) is where the fancy stuff is

i use clown and shark

clown connects to shark... i named em that cuz ppl were getting confused about which sampler went in which order

yup, SDE just means that eta or eta_var are greater than 0.0

eta and eta_var tell it how much noise to add after each step

no prob

i'd generally start with eta thats less than 1.0 btw

oh, you probably need to update

i added that recently

use "git pull" in the folder and it'll update

definitely

it's really fast, you're on (presumably) a slow card if you have 8gb vram

res_2s will run at the same speed as dpmpp_2s_ancestral and dpmpp_sde

[untold valley]

Surprisingly res_3m was faster than 2m i may have broke something

[dusky thistle]

they should be the same speed

what i would recommend is using one or the other

eta_var = 0.5 or 1.0 as a first value to try for that

with eta = 0.0

and then try eta_var = 0.0 and eta = 0.25 or 0.5

def try a few with a couple prompts and see what you like best

there's no correct value or anything, i'd just keep eta < 1.0 for first tests. with noise mode = hard, eta =1.0 is the breaking point, the math won't work so you'll get a black image

[icy drift]

😮‍💨 OmniGen is broken in my Comfy again and nothing I try fixes it. Why is this node so fragile?

[dusky thistle]

not sure you'll just have to see what you're getting

[toxic bone]

all custom nodes are fragile. i try to play in stock comfyui as much as i can. if you do use custom nodes, freeze your version of comfyui and never update it after you get all the nodes working. only update when you have vetted things. This will be your "Long Term Support" stable install location. The one you update everytime you start should have minimal custom extensions and stay as stock as possible

same rules for modding games

[craggy crest]

@dusky thistle do you have a discord?

[craggy crest]

when clicking on either of the workflow images in the readme

it's what 3 WOULD have been if 1. the community hadn't gotten so toxic about not getting a release yet and 2. the community hadn't started saying that SAI wasn't going to opensource anything, any more. the community, here on discord, was so toxic and demanding that they got handed an unfinished beta test model - and then didn't like being told that's what they got becaue they wouldn't be patient and wait till it was done

[untold valley]

Well, since you are a part of “the community” I thank you for shouldering the blame. I on the other hand won’t be gaslit. and thank you for turning a positive message negative and aggressive helps a lot.

[craggy crest]

i'm not one of the people - most of whom thankfully are no longer here - that were the issue. and i didn't turn a positive message into a negative, nor gaslight you. i stated the same fact that's been stated over and over. it was clearly stated when it was released why it was being released - and the community did the same thing then they had been doing - try to burn this discord down, with some literally trying to see if they could destroy SAI as well.

[spark quail]

ye if anything crystal was pretty much always the one being attacked by the toxic folks simply for pushing back on the negativity lmao

big props

[west herald]

Where can I find inpainting and img2img comfyui workflows for 3.5 large?

[dim geyser]

https://comfyanonymous.github.io/ComfyUI_examples/#other-sources-of-examplesinformation

[craggy crest]

you might see what's been posted in the #🧣｜comfy-ui channel

[noble coyote]

Flux RF Inversion

[cedar axle]

(SD3.5L)

[short thicket]

I got it figured out. I've added the safetensors FP8 up here:

https://huggingface.co/ManglerFTW/Mangled_Merge_Flux_V1_Dedistilled/tree/main

[dusky thistle]

wtf, that' sweird...

don't have my own server, nah, just hang here and on L3

[craggy crest]

you need a discord :)

[short thicket]

Yup. It's this one. The hugging face link it just for extra quants.

https://civitai.com/models/788136

[bitter hearth]

there's a comfy node too

https://github.com/comfyanonymous/ComfyUI

personally I perfer tonemap and/or skimmed CFG

[short thicket]

808

https://github.com/mcmonkeyprojects/sd-dynamic-thresholding

Not a lot actually. I didn't really try to focus it on NSFW but there are some in there. The main thing with the loras was to grab any that looked good and were ok to merge by the creator.

[bitter hearth]

few more thresholding options here too https://github.com/Clybius/ComfyUI-Latent-Modifiers

[short thicket]

These setting work pretty well for me. I usually change the mimic scale between 1 and 3.5 and keep the ksampler CFG at 5.

You can always add them in, this model works with loras.

I actually have to update that list now that you mention it... one moment

ok if you refresh the list all 808 are on there now.

Same. It needed help which is why I merged it in. 🙂

[bitter hearth]

I don't mean to nitpick but you mostly want increasing CFG (like cosine up) rather than decreasing CFG (like cosine down)

[short thicket]

I will play with it. Those settings were originally being used on a non dedistilled model.

[bitter hearth]

for what its worth I've enjoyed both up and down but up is more supported in papers

the reason is CFG does more damage early on, when the cond and uncond disagree more

[smoky gust]

anyone here use comfyui? i need some help troubleshooting an issue regarding rhthree's comfyui nodes

[short thicket]

in some ways, sure. But I merged them in a way to try and limit that. Basically I took Pixelwave and Nyanko7's model and merged them 60PW/40Nyanko, then took MMMagic and MMMatrix 50/50 and merged that 50/40 with fluxbooru and then merged those 2 together 50/50.

60PW/40Nyanko

[smoky gust]

I am getting this error @silver sluice

[short thicket]

I think it depends on the sampler. I use deis which I set between 20 and 30. But it's also good with dpmpp 2s ancestral / SGM Uniform at 15 steps.

[smoky gust]

genius @silver sluice i just deleted this folder called glob and that seemed to eliminate a bit of the issue, thanks

[short thicket]

yeah, play around with it

I have noticed loras do better with more steps. Sometimes 15 steps isn't nearly enough.

[bitter hearth]

what did you like about flux de distill?

[short thicket]

Different output than the normal flux style.

I merged in 808 loras and still couldn't get rid of the "Flux Style"

Yeah, people and animals look too plastic and 2D is too polished.

Im gonna use this model and finetune on top of it.

It all depends on if creator allows merges.

yup

if it's got this sign, it means no merges.

i blame booru lol

3 of 4 are asian lol

she does in the 60 step

[bitter hearth]

there is also flan

[short thicket]

which one?

[bitter hearth]

I tried Flan today but I didn't do an A/B comparison

it was good though

lol yeah

[short thicket]

Sweet! What's the list?

[winged seal]

Any better supported training tools for SD3.5? I am finetuning Flux right now, cause it gives incredible results. I'd love to train SD3.5 if there are better options to train now

[short thicket]

Nice! I'm working on getting the longClip nodes.

now you tell me lol

[winged seal]

jesus fucking christ they are more baked than snoop dog

[short thicket]

testing now

that redhead with the camel hoof

It's still funny that even with all the merging and training involved in this model. It still gives sleight butt chins.

yes

Yeah, I'm not sure about fluxbooru, but I know nyanko7's model wasn't trained to go past 3.5

Yup it's Fluxbooru's fault lol

OK. My next project it gonna be fine tuning off of this. Working on building my data set. I need to make a tool that allows me to skim through all the captions and make corrections. I don't have a lot of $$$ for compute so I'm going for the smaller but really good dataset option.

[winged seal]

Currently inferencing FP16 flux dev on an 8GB GPU with a messily 1GB/s PCIE connection... Taking 25s/it lmaooo

[short thicket]

https://civitai.com/models/686704/flux-dev-to-schnell-4-step-lora this might help. I got pretty good outputs on 4 steps from it.

[winged seal]

I'm validating a fine-tune, so I can't add any additional things

[short thicket]

gotcha

[winged seal]

Training on a 3090, validating on my 3060ti

[short thicket]

hows the training part working for you? I also have a 3090 and intend to train on that. I was gonna use simple tuner for it.

[lunar canopy]

not the place for this prompt

[winged seal]

I only just started fine-tuning. I'm friends with Mikey, the guy who made pixel wave, and he was giving me some training insights. It's absurdly slow haha

[winged seal]

Thank god

[lunar canopy]

no

[winged seal]

Uhg, simple tuner. I always get a migraine when I see people mention it

I just tested my fientune at 3k, and it actually is fixing some issues in flux, but its sooooooo slow lmaoooo

[untold valley]

wow

[bitter hearth]

really like the last one with the guy with the horns

[severe phoenix]

yaaayy!! goat!

[hallow lion]

so Omni is in comfy now?

[proven pecan]

Much ❤️ to the SD team. It's great you can get these results without any refining or upscaling.

[craggy crest]

@dusky thistle @bitter hearth i finished the sampler/scheduler compare sheet - do you want a copy?

[dusky thistle]

sure, always interesting in more data

[craggy crest]

DM sent

[sacred jewel]

https://tenor.com/view/eric-andre-what-the-sigma-gif-15268419697534459708

[craggy crest]

do you know how many layers SD3.5 medium has?

[dusky thistle]

Nope

Haven't looked yet

[gusty trail]

24

[fiery saffron]

Display an e-commerce interface with Shopify and WordPress logos, emphasizing a smooth checkout experience. Use a cool blue and green gradient in the background. Overlay text reads: “Seamless E-commerce Solutions for Your Business.” Show a shopping cart interface with bright accents in green and teal, and ensure sleek UI elements.

[red haven]

Hi All, I updated my SDXL DaVinci Ink Sketch LoRA for SD3.5 using Ostris' AI-toolkit. I think it came out pretty well.

You can find it on Civitai here if you want to give it a try. https://civitai.com/models/212322/da-vinci-ink-sketch

It's more DaVinci's notebooks with a biomechanical twist.

[untold valley]

depth is amazing but why is it hazy?

[signal shuttle]

Onetrainer pushed out an update that made onetrainer support efficient ram offloading which means you can train SD 3.5M on 1024px images with only 4gbs of vram

[untold valley]

no way 4gb for 3.5m?

[craggy crest]

much appreciated, thank you

[untold valley]

thats all?

[signal shuttle]

Yes

That all

https://www.reddit.com/r/StableDiffusion/comments/1gi2w2e/onetrainer_now_supports_efficient_ram_offloading/

[untold valley]

damn going to have to create a dataset

anyone know how to start?

this is bonkers

[craggy crest]

start by deciding what that lora is for - remember, if you can get what you want with a prompt, don't waist the compute on a lora

[signal shuttle]

Onetrainer is easy to use for beginners, it has a nice user friendly UI, I recommend you to read the instructions on their github to start using it https://github.com/Nerogar/OneTrainer

[untold valley]

thank you

[untold valley]

want to train the whole model on multiple subjects/ styles

not a lora

[craggy crest]

not what onetrainer is for - and that's going to be expensive

that's a fine tuned checkpoint

[untold valley]

you cant finetune with onetrainer?

[craggy crest]

do you honestly need to? the point of a lora is to be able to train a small item that doesn't cost much and specifically updates the model's weights for specific information instead of spending the funds necessary to retrain the entire model

[signal shuttle]

No you can fine tune a full checkpoint on 4gbs of vram one trainer just very very slow

[craggy crest]

see you in three months

i don't want that electric bill

[turbid grotto]

I am having a blast with sd3.5m doras
they are not perfect yet, maybe due to undertrained base but it learns anything, I like it
and the interesting part is I trained dora for 6k steps with lr 0.001 at 512px and it did not overbake at all and works at 1024px, only some details not accurate

[signal shuttle]

Jokes on you, my government pays my electricity bills

[craggy crest]

but do they pay @untold valley 's electric bills?

[turbid grotto]

but you can do it in OT

[untold valley]

why is it expensive/time consuming, wasnt 1.5 trainings done in like 2 hrs or so?

shouldnt take more than like 3-4 days for sd3.5?

[turbid grotto]

I remember finetunning takes only 2 time longer than lora but I may be wrong

[signal shuttle]

3.5M should be in Theory faster to train then SDXL since 3.5M is smaller then SDXL in parameter size

[dusky thistle]

the power of implicit sampling:

lots of stuf fcleaned up

[turbid grotto]

sd3.5m - 2.5
sdxl - 2.6

not really smaller)

[craggy crest]

1.5 isn't 3.5

[dusky thistle]

not all params cost the same to train

attention can be hard on the gpu

[craggy crest]

and 3.5 doesn't use the same architecture as 1.5

@dusky thistle prompt: portrait of a cougar in the moonlit winter snow, euler_ancestral+linear_quadratic, layer 4 only vrs all layers

[dusky thistle]

here kitty

[craggy crest]

pretty kitty

just amazed at how much of the final result is present in that one layer

[dusky thistle]

yeah that's interesting

[noble coyote]

Flux RF Inversion Style Transfer

[hasty robin]

i have two gpu's one with vram 12 gb on with 16 gb. is there any possibility to run stable diffusion using these two. its would be a great help. i am new learner .

[noble coyote]

Flux RF Inversion Style Transfer

[bitter hearth]

yes please, that would be great

[noble coyote]

Flux RF Inversion Style Transfer

[untold valley]

This doesn’t even look ai. Best ive seen.

[noble coyote]

E x c e l l e n t T e x t on this ReCraft!

[short thicket]

I'm currently exploring the Detail Daemon. That looks like it's gonna be next on my list of things to try out.

[short thicket]

I wonder if that's 2 gpus through SLI or through a network. Can 2 different GPUs connect through SLI?

[errant dust]

I can't speak for this, but in chess it is possible to leverage more than one GPU without SLI. Rigs with 4-8 GPUs exist

Naturally, the software must be written to allow this, but SLI need not be involved.

[bitter hearth]

Detail Daemon is legit

I do it by other means, using SDE sampling with scheduled S_noise, and also adding noise using latentmegamodifier node, but its similar idea

its all just noise injection at the end of the day

[errant dust]

So after a bunch of testing with SD3.5L, I will say this much: for artwork and whatnot, I like dpmpp_2s_ancestral a lot. Very different output, but often the prettier.

[bitter hearth]

in Comfy dpmpp_2s_ancestral is the best out of the default samplers for Flux and SD3.5 yeah

[errant dust]

SOme things SD3.5 just sucks at, but the same is true of other models, and possibly LoRAs can improve this. So not a general slam of it by any means as I really really like SD3.5 for some things.

[dusky thistle]

Def give my sampler nodes a shot

It blows dpmpp 2s ancestral out of the water for versatility and max quality

And also has faster options

[gusty trail]

After many testing on sd3.5m, it is very hard to learn the detail accurately.

[errant dust]

I am open to trying. You have a WF I can use or link for the node?

[dusky thistle]

Yep give me a sec here and I'll get you one

[errant dust]

Today for example is my uncle's birthday. He is 63 and is a known surfer and taekwondo blackbelt (gold in senior division of World Cup). So I asked both to produce images with same prompt. SD3.5 gave me this (and 3 more tries were no better). A third also by Flux shows a typo but nice creativity in text style.

And yes, I asked for a cartoon style

with candles in the shape of 63

[limpid thunder]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[dusky thistle]

https://github.com/ClownsharkBatwing/RES4LYF here's the repo... getting really close to ready to throw up some documentation and get added to the manager

here's an img2img workflow with two different methods

[errant dust]

Many thanks. I will try in Flux first and then in SD

[dusky thistle]

here's txt2img

the node should have tooltips that give some basic idea of what the options do though i'll be adding more detailed documentaiton later

the first two options to play with are the sampler type:

for starters, i'd play with res_2m, res_2s, res_3s

res_2m runs as fast as euler, 2s runs as fast as dpmpp_2s_ancestral, res_3s is somewhat slower but higher quality

the other param to play with at first is eta. set it to 0.0 and it'll be an ODE, anything greater than 0.0 is an SDE (adds noise after each step). values like 0.25 or 0.5 are good starters, if noise mode is hard (default) then eta = 1.0 is by definition the point where the math blows up so it has to be less than 1.0 for that

[bitter hearth]

I could only get noise mode soft working for SDXL

and only eta no eta var

this may have been due to that weird bug though where my sampler was still called RKsampler instead of Clown though

[dusky thistle]

i fixed hard, it should work now

i had done something profoundly stupid with the noise scaling for sdxl lol

[bitter hearth]

lol ah okay nice

I tried the Res_3m in ODE form and it converged really, really fast

much faster than UniPC

so I'm pretty happy with that

for Res_3s I had to use 120 steps for it to finish improving, but the image was probably the best SDXL image I have made

so the repo seems good

as a side note, unipc seems to not work with sd3.5m 🤔

[errant dust]

How do I connect the VAE? I am using SD3.5 GGUF, so VAE is not embedded AFAIK. I added the Load VAE node but it refuses to connect to the VAE output

nvm. It was minimized

[short thicket]

Sounds good. I got a spare 3060Ti. I might look into leveraging it. Might need a beefier power supply though.

[dusky thistle]

awesome 🙂

should be even better with that hard noise, i think

just a lil bit of it

i'm gonna add nodes for scheduling params soon, so you'll be able to do stuff like tail off noise at the end, etc

[bitter hearth]

ah yeah that would be good

in terms of noise flavours to add as ancestral noise I liked Brownian, Uniform or high frequency power noise

[dusky thistle]

perlin is real interesting too

and the pyramid ones... especial pyramid bilinear, and hires pyramid bicubic

[bitter hearth]

ah okay thanks

I do remember some of the pyramid ones were nicely unhinged
with galaxy bottle prompt there wasn't even a bottle but there was a weird Japanese pagoda

[dusky thistle]

hahah wow

yeah what i'm usually looking out for is a change in the visual style... as in, toward something more or less saturated, or toward a painting/illustration style or back toward photographic

that, and the complexity of the composition: subject front and center, or in the background off to the side

[bitter hearth]

ah my goal with noise is always to make crazy stuff happen

well also add detail

really high strengths of Vector Sculptor node are great for creativity

it adds or subtracts nearby tokens to your prompt automatically

probably the worst possible thing for precise work though

[noble coyote]

Try ReCraft - it has superlative text capabilities

[bitter hearth]

I noticed ideogram has good text also

I suspect the private models are using large text encoders, they may as well

[dusky thistle]

i don't think i ever got that to do anything

but i don't think i put much effort into it

[bitter hearth]

the node's limits are way too low

you have to crank it right up

also I was leaving the negative stationary and normalising both by "mean"

this example was good

[errant dust]

Ideogram has long had the best text of all. Now they have serious competition by Flux for less Graphic Design work.

However their new model in terms of imagery impresses me a lot less

[bitter hearth]

Ideogram probably has best text yeah, I agree

I was hoping for a new Pixart model but they made Sana instead this year

[bitter hearth]

the situation is kinda dire
there are only 3 modern models with MIT or Apache 2.0 licenses
Lumina, Auraflow and Schnell

I've been preparing for a big fine tune but I am not sure which one of the three to pick

[errant dust]

is it hard to hooks this up to Flux?

[dusky thistle]

nope, you can use the same WF

just switch to loading the flux checkpoints, vae, and text encoders 🙂

dual clip loader instead of tri clip loader, etc

[noble coyote]

img2img Flux RF Inversion

[errant dust]

Ok, I did that. I disabled the Negative prompt and the SD35 model somethingsomething

[dusky thistle]

should work great with flux, i actually originally wrote this code to work with that and was then very happy to see SD3.5 came out and worked fantastic with it (absolutely love SD3.5 tbh)

oh, yea, and then you need the guidance node too

[errant dust]

oh, guidance node. wopps. Should be fiun to see what mess I get

lol

[bitter hearth]

with enough whacky noise, inpainting and refiner passes
you can kinda make Schnell look good

its still not the level of Dev though but not as bad as the Schnell from launch day

[errant dust]

What about the Conditioning Zero node? Leave it or remove it?

[dusky thistle]

doesn't matter too much now since you should set cfg = 1.0 with flux, unless you don't want to use the distilled guidance and want to use cfg... which 95% of ppl don't want to do, myself inculded usually

takes too damn long, and the quality degrades quite a bit, i only turn off flux guidance and turn on cfg if i'm desperate to get a certain look that the distilled flux guidance is forcing to happen

[errant dust]

the question I had was this:

[bitter hearth]

I quite liked flux with CFG

its extremely dependent on the nodes you used to fight the CFG burn

[noble coyote]

ReCraft better at long strings of text

[bitter hearth]

not bad yeah

[craggy crest]

that's a nod to one of my players in an old GURPS game - he made a superhero that was normally just a coat, a hat, and a pair of sunglasses. but if you hung them all on the same hanger, then they activated and turned into a superhero

[rapid pivot]

crystal

others

[turbid grotto]

I am yet in testing fun shit but so far had good results (ignoring anatomy and consistency as it is not lora's fault), all loras been detailed and flexible. likeness was fine. And great results with 512px training which scaled to 1024 greatly, even to fullhd with little padding artifacts

however I am not sure about lr yet, it takes more steps than sdxl and I was not able to overcook model yet, even with lr 0.001 at 512px

[gusty trail]

I tried to overfit one image to the model with prodigy. The lr raised to 0.011 end up with 0.007 step loss but the image still couldn't reconstruct perfectly. Some small detail messed up. Overall is reconstructed.

[turbid grotto]

maybe it overfits by messing up structure instead of looking "overbaked" and reconstructing image? I had one run where after certain steps proportions became more and more distorted

[gusty trail]

The structure is reconstructed, including hand, anatomy, etc, but some small part of image doesn't reconstruct correctly.

[turbid grotto]

gonna test more...

[craggy crest]

https://gradientflow.com/lora-or-full-fine-tuning/

[gusty trail]

One image learned around 12,000 time to get 0.007 using prodigy. Many repeats and epoch.

[turbid grotto]

we need Furkan to test all parameters 😆

[tough oriole]

For comfyui is there a way to use an older version of the frontend? the updated one is kinda jacky for me.

[gusty trail]

You always need different parameters with different dataset.

[craggy crest]

he'd probably find the project interesting, and he does a good job with that sort of stuff - you should suggest it to him

[turbid grotto]

He already planned to do this, just waiting for tools to mature

[errant dust]

I cannot say whether it is better or not, need more testing no doubt, but overall good stuff, and I will say that the birthday banner was clearly much better using your WF

[dusky thistle]

it's gonna be more accurate overall in the end

with the right settings you can replicate dpmpp_2s_ancestral exactly

RES is basically dPMPP with some fixes to problems with the math

[errant dust]

my birthday banners with the standard WF in SD3.5 were crap. These are not

I used res3 each time as you said it would be a bit slower but better

I would only switch for potential better

I have no clue about the various math options. I put all in Brownian, but it was a mental lottery

and switched linear Quadratic for SGM

what is noise mode Hard, Soft, and so on?

and do you have a suggested choice?

[noble coyote]

Go to Settings and disable the new gui

[runic tusk]

What isn't working or don't you like?

[craggy crest]

is the spectre young? or is the girl young?

which is wearing an eye patch?

[errant dust]

I'd assume the girl is not 'it'

[brittle moat]

the spectre is young

[brittle moat]

the spectre

i'm ale_256

[craggy crest]

if you're trying to generate this, you need to use the Artisan channels, and you need to read the information in this channel #artisan-faq first

[brittle moat]

it's paid only?

[errant dust]

No, but this channel is for pics of cats and chess pieces.

[noble coyote]

Or 'neon hedgehogs'!!!

[rapid pivot]

Cursed

Stuff @sage burrow would make

[rapid pivot]

[rapid pivot]

Did not expect an ambulance on the second one lmao

[cunning lintel]

neither did sd3.5, did 4 gens in 3 the ambulance turned out too realistic 🤡

[rapid pivot]

Lmao

[runic tusk]

A confusing timeline, to be sure:

[untold valley]

I changed my mind SLG is a must have

[bitter hearth]

yeah the idea of SLG is sound

its not gonna be on the same level as PAG on the Unets but its similar

[untold valley]

I think i missed the class on PAG, but compare the images above.

[craggy crest]

if you're looking for a place to generate with 3.5 that has free accounts, you should take a look at mage.space

[craggy crest]

rendering only layer 7, scale 1 vrs skipping layer 7, scale 1

[craggy crest]

try switching your sampler to euler_ancestral, and your scheduler to linear_quadratic

[untold valley]

im alternating between that and res 3, i think im liking eulera^2 for people and res 3 for everything else, euler ads more details works great but for like plain stuff res gets accurate/sharp but no thrills/extra details.

[fleet meteor]

What model is that? 👀

[craggy crest]

@dusky thistle scale 1 vrs scale 2 vrs scale 3

rendering just layer 8

[tough oriole]

I figured it out had to run a argument first.
The newer versions are very bad when i run comfy to my phone.

[dusky thistle]

they change how much noise is used in the beginning or end

[untold valley]

SD3.5Medium.

[dusky thistle]

soft and softer start strong and drop off fast, and faster

hard is a constant amount based on a fraction of the current noise level

generalyl, best has been with hard and eta = 0.25 up to 0.5

[untold valley]

Woah this is amazing to look at. Scales apear to be exponential. Less is more maybe perhaps.

At scale 3 cannot tell there’s even a leopard there.

1 and 2 still can make out the image

[craggy crest]

i'm sure that you can refine it farther, scales can have decimal points. i'm just doing a rough compare sheet on this

some layers appear to have a much greater effect on the image, than others

[untold valley]

Was told 4 is compositional layer and 7,8,9 is finer details like hands feet.

[craggy crest]

not what i'm finding. i'm going through this, right now, one layer at a time

when i get this sheet finished, and it will take me a few days, i'll post the link

[untold valley]

That’s will be great to look at. I do wonder how many total layers there are.

[craggy crest]

24

[untold valley]

Model Description: This model generates images based on text prompts. It is a Multimodal Diffusion Transformer (https://arxiv.org/abs/2403.03206) with improvements that use three fixed, pretrained text encoders, with QK-normalization to improve training stability, and dual attention blocks in the first 12 transformer layers.

Maybe first 12 are the most important

[craggy crest]

they are definitely important, and qk-norm is extremely important

https://arxiv.org/abs/2010.04245

[untold valley]

The pain in the but is that many layers affect other layers. Learned that trying to big brain model merging by layers a while ago. And when you get humans good sometimes that messes up animals or shapes or landscapes. It’s not as easy as bad hands? Adjust this lever. That sheet should be really handy.

[craggy crest]

layer 9 - scale 3 vrs scale 2 vrs scale 1

would suggest keeping scale between 1 and 2

[untold valley]

really neat to see it visualized i will say

[runic tusk]

I finally rest and watch the sun rise on a grateful universe.

[craggy crest]

:) yeah. i have to have a visual - i can't just look at the math and see what it's doing

[cedar axle]

Been having fun exploring SD3.5L for the last few days. These are some of my favorite gens so far.

[craggy crest]

@untold valley rendering just layer 11. scale 31 vrs scale 2 vrs scale 1

skipping just layer 11

[untold valley]

this one is interesting, seems to be affecting color in a big way

[craggy crest]

it is. if you look at the three with only 11 skipped you can see a definate red shift by scale 3

[untold valley]

some details too, the whiskers now all look normal.

[craggy crest]

look at the trees and background

[untold valley]

thsi is very cool, so far layer 11 seems to improve the details over the whole image better than all others

[craggy crest]

this is goign to be a fairly indepth sheet. right now i'm doing the single layer compares - with this layer skipped, with this layer rendered. i'll do that for at least 3 different prompts. then i'll do compares with 2 layers skipped and rendered. and with three - we'll see if i go farther than that

[untold valley]

man hope you have at least a 3090

preferable a 4090

[craggy crest]

i have this https://www.newegg.com/abs-aqa14700kf4060ti16g-stratos-aqua/p/N82E16883360436

layer 12 - scale 3 vrs scale 2 vrs scale 1

no background, at all. amost no foreground. just the subject

[untold valley]

hey that one looks exactly like the image

[craggy crest]

skipping layer 12

[untold valley]

the higher ur going the weirder noise is getting in the sense noise is stronger?

[craggy crest]

yeah. as scale increases, the noise increase. that doesn't mean it's bad, because injected noise can help refine the details

[untold valley]

this one made them go on a a diet

they shrunk, background changed some

[craggy crest]

look at the slope of the back, too

[untold valley]

foreground became 2d

[craggy crest]

so far, every single layer does do things, there aren't any layers that don't - it's learning exctly what effect each has, and how you want to set the values in order to tweak.

and that's the first 12. so now i work on the second 12 and see what's going on with them

[untold valley]

hopefully its not diminishing results from here on out

[idle matrix]

This is a picture of a perfume, add a background to it

[craggy crest]

layer 13 - scale 3 vrs scale 2 vrs scale 1

[craggy crest]

if you want to generate in this discord, you have to do it in the Artisan channels, and you need to start by reading the information in #artisan-faq that channel

[untold valley]

this one is confusing

[low inlet]

Hello guys

[craggy crest]

why so?

[low inlet]

[untold valley]

from nothing to scale 3 bam full image without background

[craggy crest]

second set of 12 layers - i can make assumptions what they are doing but i'd rather not

[low inlet]

Errrm I have a question for SD3.5 runners

[craggy crest]

what's the question?

[low inlet]

wh sd3.5 is not much better than sd3?

I tried the sd3.5 large i expected it to be much better but yet my hopes are down for sai

Flux is still better specially the new 1.1 and also red panda is here

[craggy crest]

a couple of things, jack. 1. better is a compareison word - it means nothing if you dont' compare stuff - so what are you expecting

[low inlet]

what is the big difference between sd3 and sd3.5

[craggy crest]

and 2. do not come in here and try to start a battle over sd3 and flux

[low inlet]

Now that's not what i'm saying i just want to know what are the changes or improvements over sd3 ?

I'm talking about 3.5 large vs 3 large 8b

it still doing the anatomy wrong

[craggy crest]

did you have questions about comfy?

[low inlet]

I'm not saying sd3 is bad but it's still good with creative styles and text rendering and realism

[low inlet]

But this is in the comfy channel i'm talking about sd3 vs sd3.5 large

[craggy crest]

yeah, but you started out by saying you had questions about comfy

[untold valley]

you need to try SD3.5 Medium, its works exceptionally well and is more atristically apt than 3.0 or 3.5large. its got styles and does text, follows prompts better than flux.

[low inlet]

Errm yes

But this is not the right channel for it

[low inlet]

I didn't try 3.5 medium tbh but is it better in anatomy ?

because most of the images i could get out of sd3.5 are not usable anywhere

🥲 💔.

[craggy crest]

so here's the deal - sd3.x is a base model. if you are having issues with people in cerain poses it is trainable. and you should train a lora for those poses

[untold valley]

yes, and you can push it very far in that area i think u may be leaning.

[low inlet]

It's not i'm leaning it's just like not everything good is good until the end some other models are better but they might be slower , harder to train , have different style of art

[untold valley]

3.5m is good trust

[low inlet]

I will try it out <3 thanks

[craggy crest]

3.5 large and 3.5 medium are 2 of the most easily trained models out there. they're also incredibly easy to steer in the diredction you want them to go with just prompts. they're almost effortless to use.

the first lora for 3.5 large came out within a few hours, literally, of it's release

[low inlet]

Yes that's what I'm saying but i don't have much experience in training models I'm not very into this specific area that much since i'm new to get deep into AI further than just running the model itself

[craggy crest]

there are quite a few on civitAI now, and https://huggingface.co/models?other=base_model:adapter:stabilityai/stable-diffusion-3.5-large a large number on huggingface

[untold valley]

sdxl had full trained models on release. that is a whole other story tho.

[craggy crest]

and if you want to try your hand at training, there are a lot of people in this discord that can walk you through the steps

[low inlet]

I'll wait for someone in the comfyui section I've H100 so I'm trying to play with it a little bit

But i noticed that sdxl models are lightning fast to run than other models for example flux

[craggy crest]

sd3.5 has a turbo version.

[low inlet]

Yeah i noticed that one but the sd3.5 runs fast i just had to get that clip_g thing

[craggy crest]

that's one of the encoders. you should be using 3 encoders with sd3.5. flux only uses 2

[low inlet]

yeah clip_l clip_g and the t5xxl

flux uses clip_l and t5xxl

but there is another version of flux called nf4 it's all vae and encoders included you don't add anything

I really love what we have as a technology right now

I mean who could imagined that in two years we would reach this point

Ai era

[craggy crest]

for sd3.X - clip_G is the workhorse encoder. it drives the entire process

[untold valley]

@craggy crest one of these is res3 the other is eulerA lin-quad

[craggy crest]

not about to guess which is which

[untold valley]

euler left res 3 right

[craggy crest]

you so need to go animate that

[untold valley]

after we finish messing around with "optimal" settings lol

[craggy crest]

so in a month or two?

[untold valley]

yeah sounds ab right, then thers a new lates and greates model

[craggy crest]

the push is really on the video side of stuff right now.

[untold valley]

ill save video when i get a 5090 rofl im genning on a 1080ti

[craggy crest]

prompt: world of warcraft elven druid

[low inlet]

what is the seed ?

[craggy crest]

not a clue.

[low inlet]

I will try with sd3.5 medium

[craggy crest]

you should probably download the workflows that were released for it from the SAI page, and play around with the one that has SLG (skip layer guidance) - ask @untold valley to tell you about that

[craggy crest]

@sage burrow would like those

[untold valley]

@low inlet so was sd3.5M better or nah?

[dusky thistle]

got some more samplers added to ClownSampler... all the deis ones

28 samplers running with the same code

[untold valley]

any that are better than euler a, res 3?

i know better is subjective but those two stick out for sd3.5m

[dusky thistle]

euler_a is pretty weak tbh

euler in general isn't all that great

[untold valley]

why is it that out of most samplers it almost always puts out great work?

[dusky thistle]

it's the equivalent of... you need to get to a distant mountain you can see, so you start walking directly at it without looking to see if there's any cliffs coming up

[untold valley]

i dont understand

[dusky thistle]

it results in lower quality outputs

lmaybe you like that lol

[untold valley]

as in grain or distortions?

[dusky thistle]

both

if you want euler speed, use res_2m

if you want euler_a speed with the ancestral part, use eta = 0.25 or 0.5 with res_2m

it's a significantly more accurate sampler

once you go to stuff like res_2s or 3s it starts getting waaaay more accurate

you'll notice too euler tends to have a dusty look pretty often cuz it's not developing the details as well as a more accurate sampler

you'll see more hazy images more often

things with incoherent small details that don't make as much sense

i also got masks working with the img2img stuff, really good results with that

with a mask over the clock face

[untold valley]

my brain is hurting, been doing comparisons and while euler a doesn't often follow prompt as well as res 3, it makes it better quality, res 3 leaves me with like a smudged look, though it does understand prompt way better.

[dusky thistle]

are you using res_3m?

the multistep one will have more issues with sd35M

cuz what the 3m one is doing, is using the previous two steps to improve the guess for the next step

with low step counts, and for whatever reason, sd35M has some issues with that

i'm only using those when i'm going for that look

2m is more stable

[untold valley]

yes res_3m on sd3.5m

ok im dumb, guess i need to go back to 2m

[dusky thistle]

hahah no worries

if you want the crazy quality, it's 2S and 3S

what's cool is SD35M is pretty stable with the outputs

[untold valley]

euler a, then res_3m, and 3s yikes lol

[dusky thistle]

so what i've been doing is scoping out seeds with res_2m, and if i find something i reallllllly like, i set it to res_3S, with implicit_steps = 1 and go get a snack

whoa your card must be really slow dang

[untold valley]

1080ti

[dusky thistle]

sell a passenger door from your car if you have one, the stereo, some seats, get yourself a 4090

[untold valley]

ill wat for a 3090 when the 5090 releases

should come down to 450-500$

[dusky thistle]

might go up

they stopped making the 4090s i think

and the supply of 5090s will be scarce for a while i'm sure

that might put pressure on the 3090 market

[untold valley]

That did not occur to me that could happen

[sterile pendant]

Yeah because it's doing 2x and 3x as much work per step. Like 30 steps with a _3s sampler will take as long as 90steps with euler. But they definitely do usually come out better than if you just ran something with ruler at 90 steps.

[dusky thistle]

yup

been exploring that stuff in depth

got lots of SDE modes working too

with RF

big gains in quality and coherence with that for sure

[sterile pendant]

But the tradeoff is far longer inference times

I prefer to be able to experiment at a faster rate and then try to switch to a higher quality sampler in the same family of solvers to get roughly the same image, but higher quality if that makes sense

[dusky thistle]

yea that's pretty much what i do

fortunately RF gives stable enough outputs you can usually just swap samplers like that and not get a totally different output

espec since i've got everything implemented under the same framework here, there's no weirdness with implementations changing from one sampler to the next

ultimately even better obv is just to go nuts with the hardware and get the best of both

hoping the 5090 fits well enough to do a dual rig

it'd be nice to have one for pos and one for neg conditioning... get 45k cuda cores ripping away at a single latent lol

[untold valley]

ive got 3.5k cuda cores take it or leave it?

[low inlet]

uhmm not really i didn't test much with it as i moved to the animate section now i'm trying moochi

[winged seal]

@craggy crest Flux is exceptionally good at learning really varried and creative art styles en masse, so I wanna see if SD3.5 is even more flexible and capable. I know its gonna take way more compute to get it to the same level, but we have access to a monumental amount of compute this time around, so 😅

[craggy crest]

you won't listen to me, so i'm not going to bother with much of a response other than to say 'learn to prompt'

[untold valley]

[winged seal]

I am not even sure why that response is there. Learning to prompt has nothing to do with wanting to drop millions of steps worth of training onto SD3.5

[untold valley]

not even Sytan is immune to crystalwizard

[craggy crest]

because what you're setting off to do is 100% unnecessary

[winged seal]

why, because SD3.5 is perfect out of the box with no deformtations or gaps in knowledge?

every model can be better

[craggy crest]

no, but you don't need toi make a huge massive thing when you could make a few specific small loras for just the data that you actually need

and then you can sell each of them. - 6 or 7 products instead of just one

[winged seal]

The goal is to make a much more stable base that is more robust, like Pony, but less... janky at the start 😅

[craggy crest]

pony is garbage. if that's what you want, have fun

[winged seal]

a 3.5 tune that can make this level of coherent information out of the box, like flux can

[craggy crest]

3.5, with the right prompt, can do that with the base model

[winged seal]

coherently? I would seriously love to be proven wrong if you wanna try to generate it. I have low expectations and experiences with SD3.5, which is why we are willing to put time and money into fixing it up considerably

[craggy crest]

yeah, coherently. and yeah, i've done plenty just like it and no, i'm not going to run off and create stuff for you to then pick apart and come up with stuff you don't like even though it might match yours exactly

go waste your compute power doing unnessary training and jumping through hoops

[winged seal]

uh, ok then? I guess you're not really serious about changing my perception of SD3.5 then, which is fine, cause I don't need to be tricked into liking things

[mortal mesa]

ya you do

[craggy crest]

i gave up on that way back there when it became obvious you weren't actually interested

[winged seal]

I have high hopes for what we will be able to do with SD3.5, which is all I need

[craggy crest]

you won't. because you're a flux person and you will only be happy if you turn 3.5 into flux

[winged seal]

oh my god lmao

ok bud

[mortal mesa]

fluxperson LOL

i bet you vote a certain way

[winged seal]

for future reference. I hate flux's aesthetics and looks out of the box with a flaming passion. I only like flux because its been exceptionally easy and reliable to train. Thats all I like about it

[winged seal]

I think its too big, too slow, overbloated, I held off from using it for months cause I thought it was a failure to the community. I don't like flux. I tolerate it

[craggy crest]

that statement there ' flux is exceptionally easy and reliable to train' has got to be the silliest thing i've ever heard anyone say

[winged seal]

I am still very interested in jumping to SD3.5 when I seem more accessible training tools, strictly just because medium is so much smaller

@dusky thistleWould yo say you have had an easy time training Flux?

[craggy crest]

i wonder why all the experineced devs have had to fight so hard to get flux to train at all when you can jsut breeze along with it

[turbid grotto]

it is very trainable in onetrainer rn

[dusky thistle]

nah, i wouldn't say so

i think it's really easy to get tantalizing results

[winged seal]

oh really? Thats dope and good to know. Finally something thats not simple tuner

[dusky thistle]

it picks up on character likeness very easily, and it's easy to shake some stuff loose with just a couple thousand steps from the model

[winged seal]

yeah, thats fair enough honestly, I can see that much

[dusky thistle]

but it's very difficult to teach it a lot of diverse concepts without it losing a bunch of stuff too

[craggy crest]

you can also use luca taco's trainer for 3.5 large and medium

[winged seal]

is that available for local yet, or still no?

[craggy crest]

he only puts stuff out on replicate

[winged seal]

yeah, thats fair enough honestly. I started doing very low LR training on it like my friend did, and found it was rapidly improving, and prompt adherence got way better in no time

I was having issues with using higher LR's and getting good results which would then hit a wall and prompt adherence would fall apart. Turns out that just cause flux CAN stay coherent for a while at very high LR's, doesn't mean the damage doesn't add up lmao

[craggy crest]

#🆕｜sd3 message

way back there, @gusty trail told you to use prodigy

[untold valley]

[craggy crest]

this needs to be a book cover

[untold valley]

quick q what scheduler for res_2m

[dusky thistle]

depends on the model, and what you're doing to some degree, but i gotta say, quadratic has been pretty damn good

next up has been beta scheduler with alpha = 0.5, beta = 0.7

[craggy crest]

cougar moon

[dusky thistle]

these are all sd35L

Acrylic illustration depicting a vast landscape with a sprawling pink-blossomed tree, intricate texture of bark, lone figure with sketchbook, delicate waterfalls cascading over rocky cliffs, distant cityscape of towering spires, afternoon glow enhancing warm tones, crisp horizon with cumulus clouds, moon faintly visible, gentle wind hinted by drifting petals, vibrant greenery patches, heightened contrast adding depth and dimension, invoking inspiration.

[winged seal]

I love the diverse styles my friends Flux tune is able to do, which is why I wonder if the same amount of time and training put into SD3.5 would yield even better results

[untold valley]

thank you, i learned my lesson on euler a and what you meant about the mountain. it was just not obvious to me

[craggy crest]

every single one of those can be done with 3.5L and 3.5M without fine tuning

[dusky thistle]

it's not exactly easy to figure out wtf any of this stuff means, tbh

there's no nice textbook on it that's easy to read or anything like that

[winged seal]

You can have the prompts if you wanna try, but we all know you're not gonna "waste your compute" on it lol

[dusky thistle]

and all the information online is polluted by huge amounts of misinformation from authorative sounding sources that don't know wtf they're talking about

[craggy crest]

did you look at any of the images clownshark jsut made you?

[untold valley]

here is 1000 million settings and variables go figure it out

[dusky thistle]

and then the sources that do know what they're talking about... tend to only share their thoughts in papers, where you gotta get past all the notation and terminology, so there is def a barrier

[winged seal]

oh, i didn't realize thats what he was doing

[winged seal]

the general aesthetic is kinda there, but man the coherence is not

which is why, again, I think longer training on it will be very beneficial with how diverse it already is

[dusky thistle]

left hand looks like a foot

[untold valley]

i say go for it Sytan then share ur model with me

[craggy crest]

;) look at the workflow in mine

[winged seal]

Our goal is for it to be all available to the public if things go well so, yeah!

[craggy crest]

told you that's what you'd do. they are more coherent than yours was

[winged seal]

to be fair, the SD3.5 ones are a lot lower resolution, so maybe thats where the discrepancy is for me

[craggy crest]

no comment

[dusky thistle]

yup, boost resolution and things resolve better

gotta generate at the same res to make a fair comparison

[winged seal]

My friend did show me these SD3.5 gen's a few days back and I was really impressed with how textured they are

[craggy crest]

add the term cubism into your prompt

[winged seal]

I don't have any way to gen with SD3.5 at the moment, but if/when I do, I will mess around with that

[craggy crest]

@dusky thistle don't know if you saw this yesterday, but there are 24 layers in 3.5medium

[winged seal]

wow, lotta layers for a small model

[craggy crest]

https://www.runcomfy.com/comfyui-web yes you do

[winged seal]

I meant as in I didn't want to deal with it lol

[craggy crest]

hmmm

[winged seal]

I have comfy and the resources to run it, I just don't want to right now

I think I am gonna go mess with SD3.5 training in one trainer though

[winged seal]

oooo, I love the colors on this one. I am a sucker for good color grading haha

that one looks much more coherent than the other one you sent for sure

[craggy crest]

so you like more saturated stuff then

[winged seal]

I like that one much more for sure

[winged seal]

it depends on the style more so, but that one I think looks really good with those colors

my tastes are very very dynamic when it comes to stuff like that

I am assuming these are all large gens?

[craggy crest]

had a discussion one night with @bitter hearth - and what it boiled down to - images only look like photos to him if he can see film grain. even digital photos, without film grain, don't look like photos to him. that's a personal taste, but valid. so maybe you need to identfy what it is you actually are looking for in an image

[winged seal]

You know what... I could set up a direct comparison between my friends model and SD3.5 to see how they fair against each other. I am sure SD3.5 is still stronger in some ways, but I would be really curious to see how far his model has diversified flux

[winged seal]

I don't have one factor things like that in my preferences typically, but there are some specific things I want when I am generating photographs, at least

[untold valley]

medium gens are more "refined" ready to go than large. medium is really surprising. large needs large more training tbh, medium needs a push and a tiny teapot nsfw for the uhh scientist and doctors

[craggy crest]

sweepy time

[craggy crest]

large is exactly what it's supposed to be. its training is fine. medium is deliberately more artsy

[dusky thistle]

yeah they're complementary imo

[craggy crest]

they are, yes

[winged seal]

Thats what I predicted would be the case several months ago, so I am glad to hear that 😅

I never thought large was gonna be very good, my eggs were always in Mediums basket because of density and easy of access for people on lower end hardware to train and run it, meaning much more support and iteration

I shouldn't say very good

I should say viable in the long term compared to medium and its accessibility

[craggy crest]

i'm really sick and tired of hearing you say this

[dusky thistle]

idk how much it helps having a model be small so more ppl can train it

tbh, most ppl just train trash

[craggy crest]

yeah. or train stuff that's unnecessary

[winged seal]

that is true, but there are also people who train good who haven't been able to

[dusky thistle]

the really good finetunes take a lot of prep and usually a lot of hardware

it's not something ppl can make a serious contribution toward by screwing around casually in their freetime for a couple hours a month

[winged seal]

and plus, you can take on way bigger projects with a model 1/4th the size, which is what I really am excited about

[craggy crest]

and someone that knows what they are doing - which 90% of those out there training... don't

so you get civitAI packed with loras that all do the same anime girl in states of undress

[dusky thistle]

i remember how eye opening it was seeing some of the datasets ppl shared

captions were literally shit like "dog"

the end. dog. lol

[winged seal]

yeah, i am tired of that for sure. But there is also good that comes with it as well

the more people training, I mean

[craggy crest]

I did a couple of Sdxl loras for faces - male and female, one each - and genereated all the images for the data sets on thispersondoesnotexist

and then cleaned them all up in photoshop

several hundred each

they came out good

[winged seal]

A more accessible model means that people who didn't previously have the means can contribute, and there are a lot of very smart people who don't have a lot of money or resources

One of my closest friends works with a SOTA Audio training company for AI generated audio and she works with 10+ H100 systems, and she herself only has a RTX 3070 and can't really afford more because of medical issues

[craggy crest]

ordinary looking people

[winged seal]

Yeah, I am all for more ordinary looking people

[craggy crest]

that is what 3.5 IS

[winged seal]

I know, thats what I am saying

[craggy crest]

but that doesn't mean that people wont' still make the same anime girl in various states of undress

[winged seal]

Thats why I am saying I think 3.5 medium will get a lot more support because its easier and faster to run

[winged seal]

oh of course not, that well never not happen

I don't mind waiting a long time for a training or inference if the result is worth it, but lots of people would rather have a much faster result than wait that long

[winged seal]

like on a 6GB card. I can only imagine the speed difference of large vs medium for that. Its gotta be at least like 10x faster

[craggy crest]

the cat is medium. comments?

[craggy crest]

not on my machine. speed's about the same

[winged seal]

I don't see any major issues, just needs a better photographic style tune and thats honestly pretty great

its a base model, its meant to be trained more in specific directions

[craggy crest]

i'm not tuning a model that is perfectly good without it.

[craggy crest]

agreed. but i'm doing some very specific tests on medium right now

[craggy crest]

that cat means business!

[winged seal]

yeah, and if you like it how it is, then thats great man. I personally wanna have more of a photographic/professional shot look, which means I will just train that in myself 😅

[dusky thistle]

that's masked unsampling

i think im' gonna make a node that allows you to interpolate from one mask to another throughout the diffusion process

[winged seal]

I will say, that image looks a hell of a lot better than what flux dev base would do lmfaoooo

[craggy crest]

dude, i was shooting photos professsionally before you were born. you have a specific look you like, that doesn't mean everything you don't like isn't professional quality

[dusky thistle]

maybe make the weight between the two determined by the sigmas

[winged seal]

what type of cat is this meant to be?

puma?

[cunning mesa]

While I don't agree with what crystalwizard is saying at all, if you want your idea of professional photo quality you should just drop the image in Lightroom and twist few settings instead of finetuning a model.

[craggy crest]

prompt: portrait of a cougar in the moonlit winter snow

[winged seal]

its not an editing thing, its a composition/detail thing

output from my flux tune (downsampled cause noise issues 😅)

[craggy crest]

someone was asking about masks the other day. wonder if that would have solved their issue

[craggy crest]

that's really washed out

[winged seal]

I have heard many people say that masks are very useful

compared to this?

[dusky thistle]

finally lol

[craggy crest]

yes.

[winged seal]

oh god

so you think the left looks more like a real pic than the right?

[craggy crest]

takea really good look at this, @winged seal - it's massively washed out, or over exposed if you like

[craggy crest]

i think the photographer that shot the cat on the right didn't white balance his camera

[dusky thistle]

mine looks more real than either of yours

[winged seal]

what does white balance have to do with anything?

[craggy crest]

it's missing the clown hat

[winged seal]

this is great lmao

[dusky thistle]

these damn shark cats are fn everywhere in my area

[craggy crest]

the man's a photographer and doesn't understand what i'm saying?

[winged seal]

the charks

[winged seal]

No, I know what you are saying, I just don't know what you mean. Do you mean a neutral white is the wrong white balance, or did you mean its overexposed?

[craggy crest]

just go fix it

[winged seal]

fix what? lmao

the white balance is 5k neutral

do you think the image is too warm, or too cool?

I do get if you think its too overexposed for night time tho, cause yeah, that is really bright for night lmao

[untold valley]

its that thing when light hits the snow and messes up colors and stuff

[craggy crest]

the details are lost, the cat blends into the background, the entire image has a number of issues

and moonlight doesn't have a warm cast to it

[winged seal]

Yeah, that much is fair, I will say

but yeah, I guess national geographic is bad at taking pictures

[craggy crest]

cat mask

[winged seal]

cat mask :3

looks like a gremlin haha

but yeah, that picture of that puma is a real photograph by national geographic. its not dark or cool cause its during the day 😅

I'm gonna mess with SD3.5 medium after all

Anybody have any recommended comfy workflows for 3.5?

I have seen tons floating around

[craggy crest]

i hope they didn't pay the photographer cause it's lousy

they USED to be good. if this is what they are publishing now, they've really gone down hill

[craggy crest]

not the workflows i just posted in the images i posted - or, maybe those

those have SLG though, and i'm not sure you want to play with that yet

[winged seal]

oh, I didn't realize you were posting with workflow

I heard that Medium does better with higher resolutions that large, is that true?

[craggy crest]

every image i post, that came out of comfy, has embedded workflow. just click to open, click open in browser, then right clidk and save as

[craggy crest]

ask @dusky thistle - i haven't done any work on that end at all.

[winged seal]

I know how to do it, I just didn't realize you were purposely sharing that. i didn't want to just take somebodys workflow without asking, thats rude lmao

[craggy crest]

anything i post out, i post for others to use

[winged seal]

fair enough, I'll remember that

[craggy crest]

in fact, if i put it online, i also put it into public domain right then. use it if you want to

[dusky thistle]

the real magic of most workflows is just learning how the nodes work

[winged seal]

yeah

[dusky thistle]

it's just appyling ppls code

[cunning mesa]

Solid.

[dusky thistle]

zero reason to be secretive or hoard them imo

i shalre anything and everything

[winged seal]

I made that really popular workflow for SDXL when it came out, and then I kinda stopped sharing my more advanced workflows cause man, it was too much to keep up with. I was not ready for all of those people asking me stuff

[craggy crest]

i do as well - if i can help someone grow, cool. if they take an image i did, and sell it - good, i helped make their day better.

i can make more, and maybe they will improve their own skills

or at least afford a cup of coffee

[cunning mesa]

Comfy workflows might just be completely broken in a week after publishing it anyway, especially if they rely on some more interesting nodes.

[winged seal]

holy shit medium is small lmfaooo

[craggy crest]

did you download it from the SAI page on huggingface?

[winged seal]

got too used to working with 24GB plus models

[winged seal]

I'm planning on using it with GGUF to save as much memory as possible, why, whats up?

[craggy crest]

that's where the example workflows are, too

[winged seal]

ok cool, thanks for the heads up

jeez, people on 4GB cards should be able to run medium just fine

[craggy crest]

i believe they have been

[winged seal]

Q5 T5XXL, Q4 medium, and it should be very close to full accuracy if its anything like flux/large

[craggy crest]

unlike flux - sd3.5 has qknorm

[winged seal]

that helps with color issues, right?

[craggy crest]

it helps with stability

[winged seal]

hmmm... fitting

lol

[craggy crest]

https://arxiv.org/abs/2010.04245

[winged seal]

well then maybe it can go even lower. Flux works down to Q3 with minimal issues as is, so if it can work even smaller, thats dope

oh righttt, network saturation, not image saturation. Thats why I remembered "color issues" lmao

[craggy crest]

take a minute and go read the paper

@dusky thistle scale 2 - rendered layer 19 only

and scale 1, layer 19 only

[winged seal]

Did a very very fast skim over the headlines and some of the charts. Looks like a sort of built in error correction which helps fix small misalignments that can compound over the entire forward pass of the network?

Sounds... Very useful, actually haha

if I got that completely wrong, my bad 😅

I am not in much of a reading mood at the moment

ohhh wait, I have my old very very good SD3.0 Medium workflow I used for a while that a friend gave me. I got incredible results with that. i should see if I can dust it off and get it working again

[winged seal]

you doing a sort of masked/segmentational refined inpainting?

oh right, SD3.5 needs more steps than I am used to. i need to remember that before I have issues with it lmao

[craggy crest]

i'm putting a spreadsheet together for how SLG works.

[winged seal]

sounds good, I am curious. I have heard the name thrown around a lot, but not seen any real examples of what it does

[craggy crest]

you can go lower, but i stick around 32 to 40 steps, and cfg 3.5 to 4

[winged seal]

sounds good. Its not gonna be very slow to inference, luckily haha

[craggy crest]

slg - skip layer guidance. you skip some of the layers - from 1 to however many. which is what the images i've been posting are part of.

you do that to tweak the look or adjust things like hands that aren't quite right. it's experimental, but was include din 3.5 for people to work with if they wanted to

[winged seal]

very interesting. thats actually effectively how Flux Lite was made. They cut out a good chunk of the layers that were found to affect outputs minimally, which makes it full compatible with dev for training, but only 8B params instead of 12b, which makes it a lot faster

[craggy crest]

as an example. this is skipping layer 19. with scale at 2 and with scale at 3

[winged seal]

Oh, i like the way that improves the textures and dynamics of the image. very interesting

[craggy crest]

scale can take decimal points but i'm just doing three renders - scale 1, scale 2, scale 3

[craggy crest]

look at those two, and then look at this one. this is scale 1

that's all that's changed, jsut the value for scale

[winged seal]

very nice. Its starting to have actual detail in the background which is what I am usually after in my trainings

lite on the left vs my most recent training of it on the right. Background fidelity is always one of the first things I greatly improve/fix in models

Prompt: A wide photograph of a blue pug wearing a pair of sun glasses with its tongue out while laying down on a beach in Peurto Rico. Behind it are various colorful Mexican inspired houses in various shades and hues.

[craggy crest]

to repeat myself - you don't need to train anything, you just need to learn how to use 3.5 correctly.

[untold valley]

yeah but the dog is blues clues

[winged seal]

While your image does look better and closer to what I am after, its still nothing like what I want. Training will definitely still be something I am aiming for

[winged seal]

the prompt asks for that. This training I did greatly improved the prompt adherence of Flux lite

[winged seal]

A photograph of a carved pumpkin that is smiling with a purple and silver witch hat on and a broom to the right side. The Pumpkin has round eyes and a single tooth on the bottom right side. Behind the pumpkin is a forest of autumn trees and leaves at dusk, dark, cinematic, jack o lantern

you can see there, the training improved the background, lighting, prompt adherence, a whole bunch. it was a great little test

[winged seal]

background looks pretty solid, very nice

[muted dove]

Found a nice little perch here

[winged seal]

dog isn't blue, but eh, base flux doesn't get that either lmao

[craggy crest]

photoshoot tired the dog out

[winged seal]

yeah, hes melting haha

[craggy crest]

nothing a treat wouldn't fix

[winged seal]

that hat looks fantastic, totally missed the tooth tho, and the background is super artificial. But its stuff like that I wanna train to be better, so its no big deal. I am impressed with these results as they are

That hat especially looks really damn good

gives me hope for pulling out a more photographic alignment with some training

is that still medium?

cause if so, those results look wayyyyy better than what I saw of large

[craggy crest]

large

[winged seal]

oh, interesting

I guess I just saw bad ones then lmao

[craggy crest]

my pumpkin also looks like a real pumpkin - yours not so much

[winged seal]

the weakest thing in that image is for sure the super artificial background gaussian blur, but all base models have that it seems lmao

[craggy crest]

possibly. maybe somene had the wrong sampler/scheduler pair or something

[winged seal]

yeah, teh pumpkin looks solid too

[untold valley]

this is 3.5?

[winged seal]

uhhhh

thats not fun

guess I need to update

[craggy crest]

trying to use a lora that wasn't trained for the model you've got loaded?

[craggy crest]

tail's attached in a strange spot

[muted dove]

It is SD3.5 turbo, for the initial image, and then refined with Flux

Some very short broom handles in these images

[craggy crest]

it's got no hands, so that probably doesn't bother it

[untold valley]

Think it would hurt?

[muted dove]

If she bit you? Yes!

[untold valley]

[muted dove]

A few holiday snaps.

[noble coyote]

Werf!

[winged seal]

@craggy crest You here? Got a question

just curious what causes SD3.5 to look super messy/splotchy. Is that something I am messing up myself?

its super compressed and weird and splotchy whenever I try to generate pictures of people

it doesn't look as bad there

[bitter hearth]

is this M or L

[winged seal]

M

its just like really splotchy for some reason

[bitter hearth]

I don't think DiTs work well around 2B param
Hunyuan-DiT also has issues

[winged seal]

plenty of smaller ones worked fine and trained well

[bitter hearth]

could you give an example?

[winged seal]

Pixart Sigma was great for training. It wasn't SOTA, but it didn't have any issues like this

and that was 900M

ahhhh right, SD3.5 doesn't inference as fast as I would assume cause it has CFG, right

[bitter hearth]

hmm I don't agree about Pixart Sigma

[winged seal]

yeah, its getting worse, hmmm

[bitter hearth]

I don't think there is a 2B DiT that I feel achieved a high level of aesthetics
I am not sure it is possible

[winged seal]

I mean, the original 3.0 Medium had absolutely fantastic photographic capabilities without these issues. I might have something set up wrong

[bitter hearth]

the original 3.0M did have better textures than this yeah

[untold valley]

LoL are those stretch marks?

[winged seal]

no, its compression artifacts lmao

I changed the subject of the prompt and its looking better now

wow, look at that lip bite lmfao

[bitter hearth]

since you can fit Flux Dev or SD 3.5L on GPUs with 8GB VRAM, I think smaller DiTs are fairly niche models

they could have some use for mobile or edge applications

[winged seal]

I have high VRAm and I much prefer small models, cause they are way faster, more efficient, and more accessible to other people

[bitter hearth]

I agree with one but not the other
3.5M definitely more accessible, for people in the 1-4GB VRAM range
its not that much faster though, the 8B pruned version of Flux dev runs at 50% of the speed of 3.5M

not requiring a negative gives 2x speed up for Flux which closes the gap

and you can fit 8B flux on 6-8GB VRAM GPUs, which is most of the market