🆕｜sd3

[turbid grotto]

backed in vae

[wispy epoch]

8 steps

[craggy crest]

he's got a VAE loading

[turbid grotto]

how is it comapred to just dev + 8 step lora?

[turbid grotto]

I also tried decoupled one from civit and worked too

[muted dove]

Not tried that, due to using fp16 models and trying to reduce lora usage

[turbid grotto]

but I had different workflow, very simple one with single positive and negative and basic clip L instead of that carzy name

[craggy crest]

here. click on this image, then click the words open in browser, then right click on the large image and save as. then drag and drop it into comfy and queue prompt. tell me if you have issues

dont' change anything

[sullen moss]

I think they just have a massive dataset with adult content, which the open-source community is afraid to use.

[craggy crest]

seriously? you're talking about openAI remember

[bitter hearth]

Dev with hyper is closer to Dev than Schnell

you can do Dev with Schnell lora at partial strength to get a kind of hybrid

[craggy crest]

SD3.5

[bitter hearth]

there are two kinds of Schnell lora though, ones that leave guidance blocks as Dev and ones that turn the guidance blocks to Schnell

[wispy epoch]

saved as webp, and nothing happens in comfy when i drop image into

[turbid grotto]

yeaa yeaa, but it also got that specific style somehow, i thought it could be from cgi

[icy drift]

Do you have it running locally?

[craggy crest]

you didn't follow my steps. first you have to click on the image, because it's Not the original image. that'll open it in a viewer. THEN under the image you'll see the words open in browser. click those words. that'll open it full size and the original image THAT image is what you right lick and save as. THAT image has the workflow in it

[craggy crest]

no

[noble coyote]

Yes - use the w/f buried in the PNG

[turbid grotto]

what is schnell lora? I wanted to try 4step dev but found nothing

[bitter hearth]

there are some on civit and hugging

[sullen moss]

Has anyone tried upscaling or inpainting yet with 3.5 L?

[bitter hearth]

they subtracted schnell from dev to make a lora

[turbid grotto]

so, just huge dataset and tons of resources?

[craggy crest]

all of flux IS basically a lora. they froze it with dpo

[bitter hearth]

and if you adjust the strength of schnell lora, you can make the dev checkpoint more or less schnell-like

[turbid grotto]

thanks, I will look into it

[bitter hearth]

possibly, its hard to know
I was hoping for the paper by now

[icy drift]

Wait, to confirm, is this SD3.5, or OmniGen? (I know it's my OmniGen prompt, just making sure before I cancel my install.)

[craggy crest]

i know for a fact that's what was done. it's also why it's not trainable. it is frozen

[wispy epoch]

now it works, i miss one step, but now i see default workflow with your prompt, and defaul checkpoint loader

[craggy crest]

and you're never going to get a paper - robin's not about to write down what they did

[noble coyote]

3.5L

[icy drift]

LOL almost fell for it. 😛

[turbid grotto]

omnigen seems to be slow, but it's features are very interesting, would like to try locally

[craggy crest]

so are you getting good images now?

[noble coyote]

I thought OmniGen was someone's name 😄

[icy drift]

It's an LLM trained to generate images.
https://github.com/VectorSpaceLab/OmniGen?tab=readme-ov-file
You can give it images as part of the prompt and tell it to do stuff with those images.

[noble coyote]

Try Searge Flux LLM

[craggy crest]

https://www.meta.ai/

[bitter hearth]

Omnigen works very differently yeah

[wispy epoch]

i need to change checkpoint loader cuz i use gguf models

[craggy crest]

go talk to meta - it generates images, it just doesn't think it can

[craggy crest]

how about don't use gguf, okay? cazuse 3.5 doesnt' use unet

you'll need to wait till gguf for 3.5 is released with the right archtechture

[wispy epoch]

12gb video and 32 ram idk

[turbid grotto]

but it worked for me

[craggy crest]

it'll run, it'll just be slow

[wispy epoch]

funny

lets download full model

[turbid grotto]

I have same setup and can run both, sd35l and t5 at fp16

[muted dove]

Almost got some good text from it 😄

[bitter hearth]

the Omnigen paper cited Meta Chameleon as inspiration

[craggy crest]

you might try Comfy's Fp8 model https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/tree/main

[turbid grotto]

what a crazy workflow

[noble coyote]

Not!

[turbid grotto]

he should be fine with fp16 as I am

[bitter hearth]

LOL it turns out Omnigen is literally using the SDXL VAE, that's cool

[craggy crest]

he should be, but he doesn't seem to want to run the workflow i gave him, with the models it requries, and not change things

[bitter hearth]

yeah you only need FP8 if you are trying to get a very specific speedboost
in Comfy its called -fast

[icy drift]

It's been stuck at 0/50 steps for ages. Don't think it's gonna work. 😕 (Local OmniGen.)

[wispy epoch]

downloading default 3.5turbo to test it on my setup and your workflow

[noble coyote]

Suggests too small VRAM

[noble coyote]

Oh!

[icy drift]

It's not even trying. No idea why. Maybe once they release the Diffusers version. It's a 3.8B parameter model. Even at full f32 precision, it should use around 16GB VRAM, with plenty left for the latent and context.

[turbid grotto]

I wanted to try it with 12gb 😦

so, I shouldn't try

[icy drift]

I'd wait for the diffusers or for someone to make a Comfy node. (Not sure how hard that would be.)

[turbid grotto]

sad that often cool things being left with no support

[craggy crest]

my workflow is designe for 3.5 large, not turbo, but it might work with it

[icy drift]

It's been out for 5 hours. Code has been out for 2 hours.

[craggy crest]

people are asleep. patience

[bitter hearth]

wait for Diffusers if you want support is what I would say

oh actually its in Diffusers already lol

[icy drift]

😮‍💨

[turbid grotto]

oh I it's been a day already

[craggy crest]

good lord! clear up space

pack stuff into .zip files if you have to

[noble coyote]

Just installing a new C:\ drive - 4Tb SSD - up from 2Tb, which is overfull!

🥳

[muted dove]

Anyone had any luck with Latent upscaling SD3.5 images? Mine end up like this, with x1.5 upscale

[bitter hearth]

seen a few people get that effect

looks like we have to tile

[muted dove]

It seems to hate latent noise injection too...

Ahhh! Just realised I had the wrong CFG set on 2nd stage. This is with noise injection 🙂

[bitter hearth]

ah that's great I got scared lol

it worked with DPM++ 2S so I knew it could take at least a little bit of extra noise

[icy drift]

My SSD M.2 ports are underneath my graphics card, which means I have to disconnect it if I want to add anything, and it has that crazy power cable that can only be connected/disconnected like 12 times before it fails.

[bitter hearth]

my CPU cooler means I can't change ram lool

[muted dove]

I think this is a common design fault. You need to change it for water cooled (small head) to free up the other slots. I have the same problem so had to discard the old RAM to upgrade.

[bitter hearth]

yeah I think small head water cooled is nice

[craggy crest]

time to buy a data center and a few racks

[icy drift]

"But if it aint broke don't fix it" so I'm very hesitant to mess with my PC. (Even though I've built dozens / worked in IT.)

[noble coyote]

I have this vast PC case - it seems to be three or four times larger than the motherboard!!! SSD access is luckily very straightforward

[icy drift]

I have an external SSD with nothing on it at the moment, and I could get more. It's not even really that slow if I want to load models from it. But I just cleared out HF's cache and deleted some old LLMs. I'm really just hoarding old stuff I know I'll probably never use again.

[noble coyote]

All my checkpoints and LoRAs are on a 1Tb external SSD as well

[icy drift]

Sometimes I wonder if these models will be outlawed in the future, and sites like HF and Civitai shut down, which makes me a little hesitant to delete them, even though I know that's almost definitely never going to happen.

[craggy crest]

nope

won't happen

[woven sapphire]

Is img2img possible with SD3?
If yes, how do I do it?

🫠

[noble coyote]

Input image/change widget to input ... feed to KSampler

[muted dove]

It always gets the same letter wrong! 😦

[woven sapphire]

Well, still nice to have your own copy

[sullen moss]

SD 3.5 with FLUX pro 1.1

[woven sapphire]

I actually prefer the first..

[bitter hearth]

its kinda obvious flux has that extra oily shine

[noble coyote]

OmniGen workflow anybody, svp? 😄

[bitter hearth]

some call it too plasticy

[woven sapphire]

Ok...is this for stable artisan?
I haven't really done something like this so...🥲

[muted dove]

I've never noticed it quite so bad as that comparison though

[bitter hearth]

sd 3.5 is pretty damn good

[noble coyote]

ComfyUI

[woven sapphire]

Oh thanks 👍🏼

[bitter hearth]

i like how sd 3.5 doesnt over do realism

[noble coyote]

SD3.5 is a tad gritty and grainy for photorealism; but it is majorly good at art imho

[bitter hearth]

yeah i would honestly prefer sd3.5 to be slightly more fine tuned for realism but not as much as flux

just to make the image a bit crispier

[noble coyote]

Use my Super SD3.5L workflow for the crispier image

[bitter hearth]

sure, link?

[noble coyote]

#🆕｜sd3 message

The w/f is in the PNG

[bitter hearth]

loaded it, what's up with the note at the bottom tho

[noble coyote]

You can add a 4th KSampler (Advanced) and set to 40/30 - 40 for an even sharper image

It stores my prompts - u can safely delete

[bitter hearth]

rendering one image now

[noble coyote]

😄

[bitter hearth]

slight issue tho

im using large turbo

[noble coyote]

Oh - probably not compatible

This is definitely a non-Turbo w/f

[bitter hearth]

yeah its not

[noble coyote]

Nice image anyway

[bitter hearth]

but just to give you an idea.. what i've been getting with turbo....

[noble coyote]

Cool

[bitter hearth]

those are sd 3.5 turbo

these are flux schnell

was using gguf q4 model for flux

this too is flux schnell but you can see how strong the tones are

comparing it with sd3.5

if sd3.5 can be fine tuned a bit to make the images bit more vibrant thatd be perfect

and most of the flux images are half body, cause it struggles with full body, which btw is much easier with sd 3.5

[lavish sparrow]

mild headache?

[bitter hearth]

your multi sampler pass to make images crispier is interesting, im going to download the large model and test it out

[noble coyote]

Olivio Sarkas did similar wit Flux - his idea originally - kudos!

[bitter hearth]

oh. then i can try the flux right away

[swift sundial]

how long does it take to generate one image using sd3 or sd3.5?

[fossil pagoda]

I found this but not sure if its a good working one https://civitai.com/models/879251/stable-diffusion-35-large-gguf?modelVersionId=985076

[noble coyote]

I'll post a w/f

[lavish sparrow]

it is time for war!

(i'm against pineapple on pizza tho)

[noble coyote]

Super Flux

[bitter hearth]

yeah.. i'll grab q8... ty

[noble coyote]

Yummy!

[bitter hearth]

i was editing your previous wf

[lavish sparrow]

what happens if you don't put pineapple on it, to quench the evil inside

[noble coyote]

No problem - create, develop, extend - into the known Universe!!!

(I'm sounding like Toy Story!)

[bitter hearth]

thanks a lot for the flux wf ... made it much easier 🙂

[noble coyote]

SD3.5L_Super

[noble coyote]

You're welcome

[lilac latch]

Horrible anatomy.

[noble coyote]

These folks show the typical anatomies of the peoples of Indonesia

[lilac latch]

Nah, so far, so good.

[sullen moss]

Wait for SD 3.5 XL . lol

[noble coyote]

A more Western look SD3.5L_Super

[lilac latch]

But it's something at least.

[lilac latch]

Let's wait for community finetunes.

[sage burrow]

Has there been any online 3.5 lora or checkpoint creators released yet? 😄

[icy drift]

If anybody wants to finetune OmniGen, it is MIT LICENCE which is straight up one of the most "you own this" licenses in existence:

[noble coyote]

Do you have an OmniGen workflow at all? 😄

[icy drift]

I don't have it in Comfy, and I couldn't get the local code working. Only used in HF spaces demo, and I was surprised by the quality for the "woman lying in the grass" prompt.

Also, OmniGen has controlnet and IPadapter built in as soon as it's in comfy.

[noble coyote]

SD3.5L_Super - quite good on the whole - 3 x KSampler

[fossil pagoda]

🐸

[sullen moss]

Hmm...

[muted dove]

SD3.5 with 2 x Ultimate upscale on main character

[rancid sundial]

SD3.5 generate a image if world war 2

[muted dove]

#artisan-faq

[noble coyote]

I've missed your excellent art since I abandoned PiXart-Sigma 😄

[lavish sparrow]

yeah, it was good while it lasted, but got superseded by flux and now sd3.5

[noble coyote]

PiXart-Sigma has morphed into Sana?

https://8876bd28ee2da4b909.gradio.live/

[lavish sparrow]

yeah, i thought so? sana is the pixart-team right?

[bitter hearth]

the sd3.5 large model is going to be great for fine tuning, but not as good as turbo distilled

[noble coyote]

nVidia - so prolly the PiXart guys

[bitter hearth]

yeah

[bitter hearth]

Q8 gguf sd3.5 turbo with default workflow on comfyui ...

didnt use any condition zero out or any of that sd3.5 stuff

and that q8 is fp16

[fossil pagoda]

I wish it would do better skin texture or is it the prompt

[bitter hearth]

wrote a basic prompt

no filtering a gorgeous Viking warrior princess.

i could add some quality prompt to make the image look more realistic

and i was actually testing out the gguf model

[fossil pagoda]

[rapid pivot]

Can I eat it?

[fossil pagoda]

Running the prompt on my setup

Different noise setting

[bitter hearth]

same prompt?

what res are you using ?

[lavish sparrow]

and now some generic fantasy art!

[noble coyote]

Sana Demo

[lavish sparrow]

A hauntingly beautiful portrait of a cracked porcelain doll with an otherworldly allure. The subject's kintsugi-like skin is adorned with intricate cracks and fissures, yet its sensual gaze still exudes an unsettling charm. Specular lighting accentuates the reflective surface of the porcelain, creating a hyperrealistic effect that is both captivating and unnerving. The anime-inspired proportions and high-resolution 8k clarity render every detail, from the cracked porcelain to the hooded features, with unnerving precision. The overall atmosphere is one of eerie disturbance, evoking feelings of foreboding and malevolence - as if the doll might come to life at any moment, its cursed existence lingering like an unseen presence. the actual prompt

[bitter hearth]

i already stole it

[wispy epoch]

and default checkpoint loader?

[bitter hearth]

through chatgpt 🙂

[bitter hearth]

gguf diffusion loader

[lavish sparrow]

but not quite...

[bitter hearth]

i shared the workflow #🆕｜sd3 message

[wispy epoch]

thanks

[bitter hearth]

mine is close to your original

[lavish sparrow]

yeah, but using vision models usually doesn't get you the prompt that actually made it...

[bitter hearth]

i know but chatgpt is pretty close with image description

[lavish sparrow]

can't deny that tho

[rapid pivot]

where glass rose

[lavish sparrow]

a photorealistic painting,I'll leave you choking on every word you left unspoken Rebuild all that you've broken, and now you know I'll leave you choking on every word you left unspoken Rebuild all that you've broken, and now you know Every wound will shape me, every scar will build my throne,I didn't say a word I looked back into my eyes
this is why i like song lyrics tho -> you can't chatgpt those. these are the same prompt, just different seed

[bitter hearth]

so sd3.5 turbo model is originally 16gb but this gguf is 8gb.... isnt that nice

[lavish sparrow]

"this isn't the kind of reaping i signed up for"

[muted dove]

The prompt from @lavish sparrow

[noble coyote]

Now I recall - the man who uses pop lyrics for prompts!!! LOL

[old pike]

🙂

[lavish sparrow]

i suppose there's ... a few of us, but i'm one of them, yeah 😄

[old pike]

who was in charge with the artist studies back then? would be interesting to see for 3.5

[noble coyote]

Eface's prompt in Sana Demo

10 toes on 1 foot!!!

[noble coyote]

Lyrics prompt in Super_SD3.5L

[lavish sparrow]

wow, this one so nice

[noble coyote]

Lyrics prompt in Super_SD3.5L

Acuity!!! Walkers!!!

[lavish sparrow]

"WTF DID I JUST READ"

[lavish sparrow]

@rapid pivot now with extra flame too!

this one sparks joy ^^

[zenith hemlock]

L Stability AI. Flux.Dev is still way better

[lavish sparrow]

care to elaborate, or is this just something you think/feel?

[zenith hemlock]

I will share it when I do more detailed tests, this is my current opinion by looking at outputs

[lavish sparrow]

this is somethign they have stated themselves too... you know? You're going to be to more specific than that.

but it doesn't fail at shattering faces, making the world a jigsawpuzzle that gets sucked into a void, or other weird prompts i made -> something flux did fail at

so just being apretty image isn't all there is. prompt comprehension is a thing too. and that's where it's better than flux

if your prompt difficulty is : "person standing" flux will win, pretty much 100% of the time

[zenith hemlock]

Oh, I just realized you have a Stable Society role, so I can't expect you to be objective about SAI 😄

[bitter hearth]

everyone knows Flux looks better\

the idea is that SD3 is not distilled

which will make it much better than Flux, given time

[lavish sparrow]

"a black heart"

[lavish sparrow]

i got that for winning an art contest.

i've also tested kwaikolors, hunyuan, pixart alfa and sigma

[zenith hemlock]

Flux dev already has a de-distilled version, and although it is still experimental, it works extremely well

[gusty trail]

What?

[lavish sparrow]

and now we have another model that seemingly works pretty well, but is just like pixart or kwai -> a base model

so, as far as basemodels go, i like sd3.5

[bitter hearth]

at the moment I don't think its aesthetics are very good
I've been following the various de-distill attempts quite closely

[zenith hemlock]

this one looks cool ngl

[pseudo owl]

It is not fully de distilled at all, it still has many biases of flux. Only thing is it supports negative prompts and cfg.

[lavish sparrow]

i really like how well sd3.5 responds to negative prompting.

[zenith hemlock]

It doesn't even work properly on the negative prompt, youre right

[bitter hearth]

if the de-distill works then we can switch back over to flux :shrug:

[zenith hemlock]

I don't know, maybe I developed a dislike for StabilityAI after being disappointed while eagerly waiting for SD3 at the time 🤷

[bitter hearth]

well I guess it wasn't possible to be a BFL fan before they were founded

[zenith hemlock]

yeah LOL

[bitter hearth]

I thought pixart model was coming but then the pixart team made Sana

which is different

[lavish sparrow]

well, pixart got hired/taken over by nvidia

i suppose that's their new model

[bitter hearth]

yeah I think we are getting just Sana from them

[bitter hearth]

this is what Sana looks like by the way
its not a great image quality but its super fast

[muted dove]

But if it's not good then you have to start refining it, which defeats the purpose of it being fast and you may as well have used Flux or SD3.5 to start with 🤷🏻‍♂️

[bitter hearth]

yeah I am having trouble thinking of how Sana would be useful

[bitter hearth]

I think there are some scientific models that use diffusion now that it might help

weather maps apparently

[lavish sparrow]

bio-engineering.

source: company i work at

[bitter hearth]

wow nice

[gusty trail]

I would give it a try. The idea of sana is amazing.

[bitter hearth]

yeah I will definitely try it

[lavish sparrow]

sniper-wolf? XD (the cuter edition)

[bitter hearth]

the VAE itself will be useful separately anyway

[lavish sparrow]

I'm feeling bubbly today

[gusty trail]

I always want a MoE pipeline to generate images and Sana seems good at this.

In size and speed

[bitter hearth]

useful for "edge" devices like phones as well

[lavish sparrow]

ww2 fairy

[muted dove]

Looking for teeth?

[lavish sparrow]

shattered frame of mind

careful around killer-bunnies

[muted dove]

https://tenor.com/view/diamond-hands-gif-24678515

[lavish sparrow]

waifu design schematics

[muted dove]

Some concerning hands there 🤔

...and here!

I think I'm going to have to try and add a Flux detailer 😦

[lavish sparrow]

fuck who broke the universe jar

[muted dove]

https://tenor.com/view/cleanup-isle-three-simpsons-apu-gif-21988221

[fossil pagoda]

My current ComfyUI workflow for anyone who needs it. It has GGUF and Lora support and an optional upscaler

[lavish sparrow]

HO. HO. HO.

[fossil pagoda]

Isnt that the base workflow

[noble coyote]

Oops! My bad

[real abyss]

please help a noob get started with stable diffusion😆
how do i start generating images here?

[lavish sparrow]

a lot of people here generate on their own computer, but you need a powerful gpu to do that

[real abyss]

i have a powerful gpu

[noble coyote]

Super_SD3.5L w/f

[real abyss]

i mean is there a channel where i generate it ,like in midjourney?

[random isle]

Anyone get SD3.5 working with 8gb vram?

[lavish sparrow]

there used to be bots, but uhh... i haven't been here for a while, so i dunno if there's still generation in discord

[gusty trail]

artisan?

Not sure still work or not

[noble coyote]

Yes and here's my w/f

[fossil pagoda]

You can try the GGUF versions found here https://civitai.com/models/879251?modelVersionId=985027
Those should reduce VRAM cost

[random isle]

@noble coyote thanks I'll test it out. Do I need any launch parameters?

[noble coyote]

It should all be taken care of ...

🥳

[lavish sparrow]

sexy zombie taco. i dnno why. but now it is a thing.

alien understanding of coffee

"u broke my heart"

[weak trellis]

Does anyone know... if clip g l and T are the same for sd3 and 3.5. I havent downloaded any new clip and just used what was there from when sd3 released

[lavish sparrow]

suspicious spy-bird

[lavish sparrow]

yeah, you can use the same

[muted dove]

SD3.5 on left and Refined with my own Flux merge on right

[weak trellis]

https://tenor.com/view/theyre-the-same-picture-the-office-pam-the-office-us-gif-20757621

[lavish sparrow]

too bad it failed on the headdress

and it lost some of the painterly quality the og had

[muted dove]

Yes, but it's first attempt. I'll play around with the levels and see what happens.

[lavish sparrow]

this is a death-mauler 9000, do not be fooled by it's cute looks

[noble coyote]

Super_SD3.5L

[fossil pagoda]

Christmas?
Easter is coming

[lavish sparrow]

"don't worry mom, it's just a phase"

[muted dove]

Better?

An interesting "fail" with denoise too low

[rapid moon]

do stabilityai/stable-diffusion-3.5-large fit in 4090 memory?

[muted dove]

Yes

[lavish sparrow]

weird gundam.

[muted dove]

Missing a hand, but... 🤷🏻‍♂️

[muted dove]

Don't play with fire kids!

[vapid radish]

SD3 Large struggles with these prompt adherence challenges quite a bit:
But it did sort of get it in the end after about 10 gens (but missing the liquid).

[lavish sparrow]

first time i got my burgers so... disgusting 🤢

[lavish sparrow]

how many steps?

[winged seal]

Hopefully AIT should have proper and full SD3.5 training support soon

It definitely needs a LOT of fixing, but I am down to give it a shot

[vapid radish]

20 Steps, but I doubt using more will have any effect, but I might give it a go.

[lavish sparrow]

it'll give it more time to get things right... 20 is on the low side

[fossil pagoda]

I run 30

[lavish sparrow]

50. because uh. i'm addicted to getting a lot things right.

[dusky thistle]

Yeah 20 isn't enough

30 should be your floor for most things even ODE

If you're doing SDE with this then it benefits even more to add some extra

[lavish sparrow]

SDE doesn't work with sd3.5 i think? it just burns the image

[lavish sparrow]

anime, girl, progression chart left to right, human to robot, cute to horror - first model that got this right, pure prompt.
well done sd3.5 team ❤️

[lavish sparrow]

is the AI saying art is obsolete from now on?

[muted dove]

AI says nothing 🤐

[lavish sparrow]

mess with the gecko, u get the stabbo

[errant dust]

It is proving to the complainers that it did not scrape anything, it learned straight from the source

[bitter hearth]

I've seen such weird images from AI that definitely never would have existed in the world before 😄

[errant dust]

and in such fascinating and fun volume

[vapid radish]

I partly figured out what the issue might have been I think.
I was using a workflow with the Dual Clip loader, so the T5 Text encoder wasn't in play, but it is still not all that prompt adherant now I am using all 3 text encoders.

[lavish sparrow]

yeah, t5 is kinda big...

[errant dust]

I used to have such fun feeding Midjourney v3 the vaguest prompts and it would come up with mindboggling and compelling images. I'd share and wow friends and family alike with some

[cunning lintel]

First old 8b api, second each time sd3.5. Seems 3.5 untrained 3.0 api or something 😬

photo of an alien holding a barbie sitting in his ufo, hard surface, retro futuristic, gothic, art by h r gigger, android jones,james jean, alyssa monks, john martin, highly detailed, fine detail, intruiging, #absurdres , nikon d850 '#photography, sharp focus, dof cinestill

In sd3.5 it's not sitting inside the ufo

A cadaverous boy with sunken, black-rimmed eyes, alabaster skin, and a rictus grin, adorned with a top hat and a tangle of vines, in the eerie, whimsical style of Ray Caesar and Tim Burton, with intricate, high-contrast details.

In sd3.5 it's a generic skeleton face :/

dark gritty pinhole photograph, dusk luminous faerie realm landscape, crouching figure in tattered faded clothes on the side of a broken overgrown path, crumbled ruin archway, ivy and wildflowers, warm color grading, ancient stone structures, moody aesthetic, trending in magical fantasy

In sd3.5 it's overly glossy generic style

But still trying, these are old fun prompts that worked in for me in 3.0 api, seems i need another way to get this out of 3.5

[lavish sparrow]

multiclassing - cuz wizards run out of spells

[errant dust]

anyhow, SD3 is being praised wildly out in the tech world. TMG claimed it was better than Flux 1.1. Seemed a tad overdone. I agree it does trumpt it on a few obvious fronts though

[lavish sparrow]

tmg?

[sage burrow]

And the awesome thing is, all our kids mj images and prompts are still sitting there a couple years later 😄

[bitter hearth]

some of these are kinda subjective

[lavish sparrow]

subjective?

[bitter hearth]

meant to reply to Aliquip

[errant dust]

https://www.tomsguide.com/ai/stabilityai-releases-stable-diffusion-3-5-a-step-up-in-realism

[sacred jewel]

I am not ashamed of stealing good prompts 🤭

[errant dust]

I have no idea what that even means. 'stealing a prompt'

[cunning lintel]

kinda, but the alien is not in the ufo, the style in 2 is not like the artists mentioned (it defaults to skeleton face), and in 3 the glossy path is not what i'd expect from (pinhole) photography

[errant dust]

I do so regularly to improve ideas and build a visual (in my brain) library

I always tweak and modify them though

[cunning lintel]

i steal everything i can, and try to embed prompts in all images i post (wasn't in these, the 3.0 ones were old, the 3.5 not set up right), that's the fun part 🙂

[bitter hearth]

the alien sort of is in the ufo, its just that the UFO is more like a sci fi chair

[errant dust]

Yes, every image I post here has the prompt and whatnot embedded

deliberately

[lavish sparrow]

https://tenor.com/view/meme-our-now-gif-10391955707070564011

[bitter hearth]

also I agree it did not get the pinhole camera right
artists should probably be a lora thing if they are named

[lavish sparrow]

flux detailer?

[errant dust]

SD3.5 is great at comics and cartoons. Flux is quite bad at it out of the box.

and SD3.5 inherently understands art styles. Flux was a colossal idiot for those

[lavish sparrow]

progression charts are good ^^

[errant dust]

impressionist oil painting for Flux means photorealism

[lavish sparrow]

agreed. one of my few gripes with it

[sacred jewel]

https://tenor.com/view/office-dwight-halvsies-half-split-gif-25611347

[errant dust]

This is not to bash Flux as bad. It is simply acknowledging and working with its strengths and weaknesses

[lavish sparrow]

yup. but not being able to do other styles other than the baked in style really well... kinda sucked

[lavish sparrow]

for prompting idiots like me

[fossil pagoda]

[muted dove]

Yes

[errant dust]

For example, SD3.5 has failed miserably (might be a prompt issue) with my classic cappuccino with the words Coffee is my World written in the milk foam with coffee swirls. SD3.5's output had the words, but they just looked like printed letters

[fossil pagoda]

Whats the prompt, i wanna try it

[fossil pagoda]

I mean this looks like it works?

[errant dust]

The words are supposed to look like they were made from the coffee swirls, and not like printed words on a milk foam background

[hallow lion]

It's a Rich to Rild story.

[hallow lion]

The age one is good at least, poor guy finally dropped that bagpack later in the life.

[errant dust]

TO be fair, Flux is better but not perfect either. The best by far was Dall-E 3. Ideogram and MJ were a solid second

Flux output

[errant dust]

One thing that SD3.5 has been consistently terrible at, and it is both specific and kind of weird, is a hand holding a pen or pencil

[hallow lion]

yes but sd35 has the potential to become better than flux.

if the people take to it and refine it

[errant dust]

I am not condemning any of them

[hallow lion]

which they shoudl this time

😄

[errant dust]

I test and learn what they do well or not

[lavish sparrow]

young to old, bling to homeless. damn

[errant dust]

Yes, text has been a bit of a sore point for 3.5. But it might be that I am using fp8 and not the good version

Here is a good result by Sd 3.5 on a prompt that has proven quite problematic for the AIs

You can see the hand holding pen issue here too though

[hallow lion]

Toung, Loonelleg, IbIIgnt, Leght to thungots... those are the stage of your life.

[errant dust]

Like I said, just a matter of knowing the tools and what they do best

[fossil pagoda]

Switched the prompt slightly, prompt issue for sure on your end

[muted dove]

It's numbering and colouring these correctly for me every time. It is missing the liquid, so I've rewritten the prompt and trying again...

[errant dust]

Pretty but the idea is a cppuccino, meaning milk foam on top. 🙂

[fossil pagoda]

My point stands

[hallow lion]

right now the best thing is omnigen tho. tho its not out yet

[errant dust]

well, it would stand if the output were what I expect

if you cannot get a correct result either....

[hallow lion]

gens look decent and if it can do all thos ethigns they say thats amazing

that means way less noodles

XD

[noble coyote]

I once did an image of a cappucino with almond syrup - but no matter how I tried - it was always hazelnut syrup!!!

[errant dust]

that is awfully specific. lol

[noble coyote]

O, this isn't Starbucks then?! 🥳

[errant dust]

This is Ideogram 2 just now

Exact same prompt

[lavish sparrow]

semi-fail

[errant dust]

and the letters really look like they are made in the foam and not printed on it

[noble coyote]

You specifically asked for foamy letters?

[errant dust]

Here is the prompt: "On a clean white background there is a cup of cappuccino, and in the foam the coffee swirls spell the words, "Coffee is my world". Around the cup we see an elegant coffee spoon with an ornate handle as well as some grains of coffee."

[noble coyote]

Ah so it intuitively 'foamised' the wording?

[errant dust]

yes

it is the challenge

[noble coyote]

Good one

[errant dust]

Another interesting variant is to do this with a pizza and have it produce an image from the toppings. A world map, or a short phrase like this

[noble coyote]

.. afk ...

[errant dust]

will it be just colorful words on a pizza? Or will they look like they were formed from pizza toppings?

[errant dust]

I just noticed the chart says that SD3.5 Medium will be 2.5B parameters and not 2

[dusky thistle]

since bottle prompt adherence is being discussed:

four bottles lined up on a table. from left to right, they are numbered "4" then "3" then "1" then "2". from left to right, they are red, blue, green, and orange. the setting is a magical forest from a tolkien novel. in the sky is an alien UFO abducting a cow with a ray of light.

first seed.

[errant dust]

on a sidenote, I see artifacting in images of 1280 x 1280, suggesting that larger sizes are not supported

[noble coyote]

...

[noble coyote]

Astonishing!

[dusky thistle]

this is what i got from the api with that prompt on may 4th

[noble coyote]

Yes, the cow-grabbing spaceship is actually believable!!!

🥳

[devout schooner]

One thing I noticed on the announcement, it seems to directly imply that the upcoming 3.5 Medium is somehow trained differently than the other versions, and also supports higher resolutions? Not sure if that's what was intended though:

Stable Diffusion 3.5 Large: At 8 billion parameters, with superior quality and prompt adherence, this base model is the most powerful in the Stable Diffusion family. This model is ideal for professional use cases at 1 megapixel resolution.

Stable Diffusion 3.5 Large Turbo: A distilled version of Stable Diffusion 3.5 Large generates high-quality images with exceptional prompt adherence in just 4 steps, making it considerably faster than Stable Diffusion 3.5 Large.

Stable Diffusion 3.5 Medium (to be released on October 29th): At 2.5 billion parameters, with improved MMDiT-X architecture and training methods, this model is designed to run “out of the box” on consumer hardware, striking a balance between quality and ease of customization. It is capable of generating images ranging between 0.25 and 2 megapixel resolution

That would be great if it's actually the case

[errant dust]

Intriguing also is that their self-made chart suggests it is toe-to-toe with the Large Turbo in quality and adherence

[lavish sparrow]

@noble coyote art by Alan Kenny

i see u posting stuff like this at times, thought you'd want to know

[devout schooner]

Large Turbo has some straight-up advantages over Large I noticed, even, it can actually do traditional hi-res-fix denoise passes without creating weird jagged artifacting around the edges of the image, while 3.5 Large still can't really

I wonder what it is about SD3 in general that causes that even, it's never really been clear

3.0 Medium had the same problem

Perturbed Attention Guidance seems to ĥelp a lot with it but not completely fix it

[noble coyote]

Streamline Moderne stylee

Thanks .. Googling Alan Kenny images ...

[dull star]

is there a site or thread on artists in SD3.5L yet (Yes it just came out yesterday I know)

[foggy heron]

cup

[sacred jewel]

Horizon Zero Flux LoRA

[silver sluice]

do you know a list for SD3.0?

[dull star]

no

there might be one for medium but large I have no idea

[cunning lintel]

they were all unlearned from api :p

[dull star]

thanks

[cunning lintel]

No i dunno, but all my styles that worked in api broke

[dull star]

okay so I'm not alone

I won't claim that I did extensive research but Zdzisław Beksiński used to work well

or at least make something I really loved

but now paintings (other than maybe watercolour) look too bright and photo-like

like.... flux...

[cunning lintel]

i posted #🆕｜sd3 message there about it

[dull star]

welp, thank you

So it was ruined

I literally wanted the large model BECAUSE of the painting capabilities

[cunning lintel]

but generally whatever i try, it's gone. it does do styles kinda sorta still though, but not very specific anymore

[dull star]

and now they took that away

[cunning lintel]

same for me, not so much painting but just different styles in an image

[dull star]

this was in the api

I love how this looks

it was a Zdzisław Beksiński prompt

he was like the new Greg Rutkowski for me

[cunning lintel]

i wonder how much 3.0 was lykon's dreamshaper styled finetune and he was kept away from 3.5 or something, but even then artists seem gone too

[dull star]

but now it just makes super bright and contrasty paintings

and idk why they needed a newer version to this when the VAE/speckled look isn't even fixed from the api

(yes to remove artists too that was another reason)

but seriously

I wonder if medium is just gonna be another flop

[cunning lintel]

A grotesque, hunchbacked woman with a bulbous, crimson nose, skin like cracked porcelain, and a toothy, sinister smile, sporting a tattered, lace veil and a bouquet of dead flowers, in the haunting, dreamlike style of Ray Caesar and Tim Burton, with rich, velvety textures. 

went from 1st to 2nd, soo generic, no fine texture or anything

[dull star]

second is SD3.5?

[cunning lintel]

yup 1st 3.0 api

it's all totally different

[dull star]

:(

[cunning lintel]

they aren't lying it's 3.5 and not 3.0 🤣

[dull star]

yeahhh

[noble coyote]

Super_SD3.5L Workflow

[craggy crest]

yeah, you read the annoucement correct

[dull star]

these are all api images

look at the TEXTURE

its screaming at me

[cunning lintel]

Claustrophobic close-up showing that In the depths of a cave system, a wraith stands guard, the wraith has a gaping hole in its chest with an eldritch creature inside, tentacles crawl out. illuminated by the faint light of crystals. The scene is styled in the manner of Virgil Finlay, paying homage to the influential Weird Tales artists August Derleth and Frank R. Paul 

3.0 2x then 3.5 2x, i have no words :/

i kinda hope i'm using 3.5 all wrong 😢

[dull star]

it just turned into "digital illustration of"

same...

[noble coyote]

Are you using ModelSampling SD3?

[dull star]

I definitely am

and still got oversaturated/contrasty messes

[cunning lintel]

the standard diffusers workflow

[dull star]

this is a painting, trust me!!!

also I get oversaturated messes at above cfg 3

and below I sometimes get weird anatomy

[noble coyote]

SD3.5L does do a woman's "duck-quack lips!!!" 😄

[dull star]

welp

no cleft chin, but we do get those lips

there's no such thing as a free lunch

[noble coyote]

Flux does a mighty clefty chin chin

[dull star]

it turns everybody into Henry Cavill

[noble coyote]

Quak quak

Nice bob cut

[dull star]

yeah the paintings are super nerfed

[devout schooner]

If it supports twice the res as Large as the announcement says I don't see how it could be
Might even be viewed as better by people even if that's not the intention of SAI lol

[dull star]

sure

especially since its gonna be relatively fast

@cunning lintel

this is SD2.0

[cunning lintel]

uf, for none generic style it's all backwards

[dull star]

its a dumb ass model

[devout schooner]

I hope they intend to like address why it's like that if it is lol
People will be confused for sure

[cunning lintel]

MJ seems the only imagen caring about style 😢

[craggy crest]

with any new model, you need to start at square one and explore it. learn how it thinks. put in a single prompt, lock the seed, try all the different sampler and scheduler combos. try different values for shift and cfg

[dull star]

at least its a damn painting

not a photo with a paint filter

[craggy crest]

you can't assume even a new iteration of a model will work the same as the previous iteration, much less and entirely new version

[craggy crest]

try adding 'brush strokes' or 'pallet knife' to the prompt

[dull star]

this isn't SD1.5 -> SD2 where we have a completely different CLIP model (Clip-L -> Clip-H)

But you're right, we should experiment

just in case there might be something good left in this model

[craggy crest]

no, this is sd3.5 which is a completely different neural network, different data set, 3 encoders instead of 2 - you need to learn how it works

[craggy crest]

give me your prompt and let me see what i get out of it

[dull star]

No like SD3 API vs SD3.5

the api only used 2 encoders?

I just put SD2 as an example

[craggy crest]

the API is not even a consideration here.

[dull star]

is this a watercolour

[craggy crest]

Prompt: soft watercolor painting of a soldier standing at the edge of a battlefield. heavy canvas texture, ragged edges

[dull star]

cause I heard that works well

[old mango]

forget about sd3.5, try https://blendic.ai

[dull star]

I want paintings

[craggy crest]

give me your prompt and let me play with it then

[dull star]

rough impressionist painting of a ww1 soldier with a helmet, looking onto the sunset in the Grand Canyon. The primary colours are orange, red and blue. The colours of the painting are faint and the brush strokes are thick and detailed.

this is from the API

I would look up my previous prompts but Glif DELETED ALL OF MY RUNS

💀

rough impressionist painting of, A man in a forest, sitting on mud, which around a pond. It takes place at night with stars in the sky. There are candles around the man and he is looking up at the sky. the painting has large strokes and has high contrast between the colors.
(from the api)

[cunning lintel]

😢 i checked mine too, uuuffff... but that's on me not copying it sooner

[dull star]

luckily I still have some left over from my post

[devout schooner]

Large Turbo gens women who look a LOT like Flux Lady lol
presumably just as a result of also being distilled

[noble coyote]

Not got sweaty/oily Flux skin

[craggy crest]

Acrylic Paint, heavy impasto brush strokes, rough linen canvas texture: faded, expressive impressionist painting of a WW1 soldier with a worn, M1917 helmet, gazing wistfully onto the muted, warm sunset in the Grand Canyon. Employ chiaroscuro, sfumato, and drybrushing techniques to capture the rugged, weathered atmosphere

[dull star]

did you use unipc? it has weird artifacting

also this does look a bit closer

but still looks too much like a photo

and jfc so many keywords just to get somewhat close-ish

[craggy crest]

no. but i can go try it with that sampler if you like

[dull star]

thank you for your effort though

[dull star]

nah I remember unipc artifacting like this actually

[craggy crest]

not nearly as many as you might ahve to tell a human artist to get what you want

using specific painting terms carries very specific data - and directs the AI to what you want rather than it having to guess

[dull star]

its just that SD3 API did these by default and not with a super contrasty look to everything

I'll try this prompt and mess around with the CFG

[craggy crest]

the API isn't just the model. it's the model and a whole bunch of settings. you're going to have to play around with everything

[craggy crest]

and shift - i tend to keep shift at 2 most of the time. cfg at 4 and steps at 32 to 40

also try giving each encoder a different prompt, crafted to that encoder's strengths

[dull star]

https://tenor.com/view/edp-i-mean-its-all-right-gif-26702511

yeah I give up on this one

[craggy crest]

i think the main problem is that the AI has just seen too many photos of faces - the rest of the image comes out looking like a painting. but that's something i'd deal with in photoshop in post production. and remember, this is BASE model, so it's also something you could train a lora for

[sage burrow]

Tails are still difficult for ai 😦

[errant dust]

for some really bizarre reason, SD3.5 is convinced pens and pencils write on both ends

This is not an outlier

[noble coyote]

"The other end is for when you want to repeat yourself!"

[silver sluice]

do you think SAI purposefully trained it on a bunch of double-ended pencils?

[noble coyote]

Its the Ninja Star Pencil which worries me!!! 😄

[craggy crest]

tails are also inconvient when you wear pants

[craggy crest]

they do if you sharpen them on both ends ;)

[errant dust]

Seems obvious

[silver sluice]

they overtrained it on double-ended pencils as a little easter egg for the community

[errant dust]

and pens

[sage burrow]

It's good at character turnaround sheets 😄

[silver sluice]

in all seriousness tho for a pencil I'd adjust my prompt to talk about the red rubber eraser on the other end of the pencil, or the clicky thing for pens, just to avoid that weird edge case

hey @sage burrow long time no talk, sort of dissapeared after Flux came out

[sage burrow]

It was hiking season 😄

[errant dust]

Didn't expect you to take "Go take a hike!" literally

[silver sluice]

@sage burrow do you have a quick review on the level of explicitness that can be generated with SD3?

[noble coyote]

Don Quixote

[silver sluice]

i know it's able to do nipples now but can it do more? i'd test it myself but i rebooted my computer and now ComfyUI is giving me a weird error "tokenize not found" when I try to use the simple workflow, which was working and tested prior to the restart

[noble coyote]

Make a large nipple with a human attached

SD3.5 - question for you: Why do men have nipples?!

[sage burrow]

I haven't trained any loras yet for 3.5. Without loras, for men, you can get some suggestive stuff but that's about it. For ladies, you can get a bit more but it looks odd. I censored the badly done nips

[craggy crest]

that is the worlds worst censor job

i find it interesting that if someone wants to know if the AI can do ... that ..., the person they ask is... @sage burrow

[devout schooner]

Turbo seems better at Anime than non Turbo for some reason, not sure if you tried it
The image is a lot more resolved and less painterly I find
Probably just a byproduct of distillation

[sage burrow]

I haven't tried turbo yet, after other turbo models not being as goid as the base, I just skip turbo now

[sage burrow]

Probably because I know which prompts most likely coerce a model to do nsfw lol

[devout schooner]

It's not as good at super complex prompts but there's no resolution support difference like there was before
And it has a really distinctly different overall style than normal Large

[silver sluice]

okay so bottom linne this is the takeaway for me
SD3.5
nipples - ✅
genitals - ❌
lewd scenes - ❌
is that accurate?

[sage burrow]

Horror character turnaround sheet 😄

[craggy crest]

other than just telling it 'bark like a dog' or something?

[sage burrow]

Different styles are fun, I'll have tp try it 😄

[craggy crest]

@dull star (20 minutes later as i forgot a had a batch of 10 set on comfy) - click on this image to open in the viewer. then click on the words Open in Browser to open the original image. then right click, and save as, then load the image into comfyUI to get the workflow. if you look at the original image, it's a painting. the workflow has my settings

[dull star]

thank you

[sage burrow]

More like "bark like a dog, make a vocal sound which dogs are known for, bark bark, dog dog, yelp loudly"

[silver sluice]

I'm thiinking this file: https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/sd3.5_large.safetensors doesn't have the CLIP model built in so that's why iit's not working with the simple example. do you guys know if there's a variant that has CLIP built in?

[craggy crest]

welcome. just make sure you change batch to 1 - cause 10 takes a while

[craggy crest]

prompt: a small kitten staring at the camera, roaring like a lion

[craggy crest]

??? what do you mean, built in?

[sage burrow]

Medium is able to work on consumer computers... I wonder what most consumers own?

[craggy crest]

so does large

[craggy crest]

try downloading it from here https://civitai.com/models/878387/stable-diffusion-35-large

[sage burrow]

10 minutes per image on mine 😦 <same as flux>

[silver sluice]

sd3.5_simple_example.png doesn't work with that model but it works with the advanced example where you manually load your own clip models, and it's having an error during CLIP processing so i'm thinking maybe iti's not embeeded inito that file

[craggy crest]

i'm still not sure what you actually mean. the example workflow loads all 3 encoders in their own node. if that's not what you're talking about, please explain

[silver sluice]

theres 2 example workflows, the simple one which just uses the Load Checkpoint node and the advanced one which shows off various ways of loading your own CLIP models, simply put the simple one doesnt work with the file given, can you confirm thats the case for you too?

[frail shoal]

Did anybody test to use sd3.5 with T5 and one of the clips only? Is it that important to have the 2 clips ? Considering that flux uses just one ?

[craggy crest]

my workflow is in this, download it and look at how i have it set up. 1. click on the image to open it in the viewer. 2. click the words Open in Browser to open the original uploaded image 3. right click on the image and save as, then drag and drop it into comfy

[craggy crest]

yeah. me. and it doesn't work well - it'll just use random data if you don't give the encoders actual prompts

and flux uses 2

[silver sluice]

@craggy crest your workflow doesn't help or answer my question

• i posted the image for the simple workflow which demonstrates I know how to use and load images as workflows
• I specifically mentioned multiple times that when manually loading CLIP models with their own node it's not a problem
• I'm simply asking if you've managed to get the simple workflow to work that JUST uses Load Checkpoint which internally attempts loading the CLIP model from that given safetensors file. it's a yes or a no, yes I used it and i got it to work, no i can't get the smple workflow to work either

[craggy crest]

and i've said over and over i have no idea what you mean by clip models

[silver sluice]

you know the little yellow input/output dots on ComfyUI denoted as CLIP?

regardless of whether you understand or not what i mean by clip models can you just confirm if you can get the simple workflow to work? in other words can you get it to work using just Load Checkpoint

[craggy crest]

so your safetensor's file has clip. yoru encoders all use clip. are you talking about not being able to connect the clip connectors in your workflorw?

i have had no problem getting the workflows to run

[silver sluice]

so can you try loading this image/workflow and confiirm it's not a problem ?

I'm talking about not being able to use the Load Checkpoint node

[craggy crest]

your negative and positive prompts need to connect to the conditioning node

[craggy crest]

that's your model.

[silver sluice]

noticie how Load Checkpoint has a yellow CLIP circle on it?

[craggy crest]

yeah. that's your model .

[silver sluice]

my concern is that the siimple workflow isn't working because the file doesn't output the CLIP from there

[mortal mesa]

so you need comfys fp8 scaled checkpoint that includes the clips, there confusion averted

[silver sluice]

and that's why iit's not working for me

[craggy crest]

yeah, it has clip in it

[silver sluice]

EXACTLY

can you pls link me to that

[silver sluice]

that contradicts what Kagi is saying

[craggy crest]

i know. that's not where you connect clip form in that workflow

[silver sluice]

there's an alternatiive sd3.5 checkpoint that includes the clips in it right?

[craggy crest]

you connect it from the encoders

[silver sluice]

wrong, that's how you connect clip from in the SIMPLE EXAMPLE

[mortal mesa]

https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/tree/main

[craggy crest]

comfy does have a discord - he posted the information in his annoucements yesterday

[silver sluice]

ill try that out thanks again would've never found it on my own

https://comfyanonymous.github.io/ComfyUI_examples/sd3/

The sd3.5_large.safetensors file (put it in your ComfyUI/models/checkpoints/ directory) does not contain text encoder/CLIP weights so you must load them separately to use that file just like in the following example:

Oh I should've read further down into the article, that's a my bad

For convenience there is an easy to use all in one checkpoint file sd3.5_large_fp8_scaled.safetensors (put it in your ComfyUI/models/checkpoints/ directory) that can be used in the default workflow like any other checkpoint files.

@craggy crest maybe you didn't know that which is why you didn't answer with the right link?

[craggy crest]

this part right here "encoder/CLIP weights " is what i was asking you if you were talking about or not. it's the encoders you are wanting embedded in the model, just an FYI - which isn't the best idea as you then can't use your own encoders if you want to

[sage burrow]

Do you have any idea what this may be about?

[mortal mesa]

you can use your own encoders, just load them in a node and dont use clip from model

i use comfys fp8 but load encoders seperate to be able to use fp16 t5, cuz that checkpoint has fp8 in it

[craggy crest]

did you forget and feed the gremlins after midnight?

[craggy crest]

if you're going to do that, there's no point is using a model that has them embedded

[mortal mesa]

cool but thats whats avaliable

[silver sluice]

right, i appreciate you trying to help and trying to figure it out with me, I know it's not the best idea but Load Checkpoint is how I have my workflow setup to work on my website so I dont want to do some crazy rewriting of the code to have it be compatible with the new clip loader nodes

[craggy crest]

should be a number of new versions, quantaized ect, by the end of this week - the speed at which this community is moving

[mortal mesa]

yup

[craggy crest]

you have to load a checkpoint for everything. that's your model. you're not generating without it. but it's a lot better to use the encoders alone, not embedded in it

[silver sluice]

like i said this isn't me arguing about what's better or not, this is just me responding by saying it's in reference to integrating to existing code

I have code that takes a workflow and adjusts it for an image generation so the workflow must use load checkpoint

[craggy crest]

i'm not arguing either. just trying to point out that you might want to rethink what you're doing and try something different, it might work better

[craggy crest]

okay, could you at least change that to "load checkpoint node'

[silver sluice]

just trying to make sd3.5 work with my system at the most basic level and then slowly mprove support

[bitter hearth]

im already using gguf quantized version of the Large turbo which is half the size of original turbo model

[silver sluice]

like my system supports the t5xxl and the clip_l prompt encoder nodes so I have the LLM generate 2 different pieces of texts and injects it into each fiield

[silver sluice]

that's a good idea actually I should Q_8 the large model for even more performance

[craggy crest]

meanwhile, if you want to use SD3.5 - clip_G is the encoder that does all the work.

[silver sluice]

i thought it was clipg/t5 that did the majority of the work together

[craggy crest]

if you don't use it, you have removed the core

[bitter hearth]

they already have the q8 version of the large too, reason im not using the large model cause it's not distilled as the turbo

[craggy crest]

no, clip_g does the work. clip_l does the artsy, ambient, fine details and text. and t5xxl has your prompt comprehension. you need all three

[mortal mesa]

there is a merge also

[bitter hearth]

Large https://civitai.com/models/879251?modelVersionId=985027
Turbo https://civitai.com/models/880060?modelVersionId=985219

[silver sluice]

oh dude you're the best thanks

[bitter hearth]

because turbo is distilled its got finer output, which im enjoying atm

[silver sluice]

is your large based off the one that doesnt include the clip?

like is it large + t5 fp8 or large without clip?

[bitter hearth]

you will need all the clips, yes

[silver sluice]

i'll pass on that then

ill just Q_8 the version that has the t5 fp8 built in

if you wanna post those too i can wait for your copies

i prefer to have one on civit so i can link my models to a version online

[mortal mesa]

so you can merge your own file to what you want also, not particularly hard

[silver sluice]

i know how to quantsize a file but not how to merge it so i would just take this one: https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/sd3.5_large_fp8_scaled.safetensors and run it through the q_8 converter

[bitter hearth]

so you want to embedd the clips as a single checkpoint?

[silver sluice]

yes pls

you'll probably see me post a bunch of images under that model's page too

if you see one guy posting a bunch of images with that model it'll probably be me lol

[bitter hearth]

but you realize, by doing that you wont have the flexibility of using improved clips

[silver sluice]

i'm very well aware

like i said im using a fixed workflow that demands Load Checkpoint so there's no room for changing that at the moment

[bitter hearth]

for example look at what im using

[silver sluice]

yeah I'm using that exact same seutp for flux (minus clip g)

[bitter hearth]

bottom two were released later

[silver sluice]

i have the VITL14 text detail and I also have the SMOOTH one

[bitter hearth]

thats the thing, they are improving the clijps too

and its moving fast

[silver sluice]

i also have longclip which is even better than the VITL TEXT DETAIL one

[bitter hearth]

there is no particular advantage in embedding the clips into a single checkpoint

[silver sluice]

im all for subbing out clip_L for better models by all means it's just my sdxl workflow isn't setup for that at all

[mortal mesa]

its this easy

[silver sluice]

the major reason I need to have CLIP come from Load Checkpoint is because the workflow supports sdxl/sd3/pony base models so I can't sub out the CLIP models for pony specifically

none of the fne tuned CLIP Ls work with Pony

[bitter hearth]

i see

[silver sluice]

so it's a major rewrite to detect the base model and adjust the workflow dynamically based on all that

[bitter hearth]

it's fair i guess... no harm

but everytime they change the clip and not the model you'd have to update the whole checkpoint

[silver sluice]

yeah im willing to do that

[bitter hearth]

and people will have to redownload the whole thing

[silver sluice]

so am i to expect a sd3.5 large q_8 + t5fp8 to be posted on your page soon?

[bitter hearth]

i keep my wf organized for easy click load/access

[silver sluice]

i keep my wf organized as code as part of the system

// Import workflows
const { output: vaePlusLoraOutput, workflow: vaePlusLoraWorkflow } = require("../../comfyGalleryAppFrontend/src/components/img2img.json");
const { output: fluxOutput, workflow: fluxWorkflow } = require("../../comfyGalleryAppFrontend/src/components/fluxWorkflowKsampler.json");

const workflows = {
    sdxl: { output: vaePlusLoraOutput, workflow: vaePlusLoraWorkflow },
    flux: { output: fluxOutput, workflow: fluxWorkflow },
};

[bitter hearth]

which UI are you using?

do you people notice something about sd3.5? usually with these models, including flux when you add a lot of detail into different objects in an image it can somewhat reduce the quality of the character's face, which is not so harsh with sd3.5

[errant dust]

@silver sluice Well, if you were able to quantize the fp16 build at Q_8 that would be exciting, but fp8 already comes fast and so lossy, I am not sure of the benefit

[errant dust]

ok nvm

[bitter hearth]

looks like screenshot from a game

[errant dust]

I saw someone already GGUFed the whole lot already

dang

https://civitai.com/models/879251?modelVersionId=985027

[compact forge]

wow

[bitter hearth]

he wants to quantize the model into a single checkpoint unlike gguf diffusion model where you have to add clips node, i dunno if that's possible with quantized gguf

[errant dust]

Yeah, no clue. why is it such a big deal though?

[bitter hearth]

no idea

but he mentions he wants to make it convenient to load up checkpoints between sdxl and stuff

[errant dust]

I mean, having GGUF is a big deal for performance purposes, but being an all in one or piecemeal makes no diff

ah

I just keep a pic handy for each workflow I want to use

[bitter hearth]

yeah i tried to tell him, when they update the clips, as they do, he will have to recompile the whole model and ppl will have to redownload the whole checkpoints just for a clip

[errant dust]

and a quick drag and drop

so I have an SD3 pic (with WF), and a FLUX pic (with WF), etc

[bitter hearth]

how do you show images within comfyui for the workflows?

i usually name my files and load them from workflow tab

[errant dust]

Well, suppose I wanted a FLUX wf and it is not already loaded. I have an image labeled and just drag and drop it into Comfy

[bitter hearth]

oh that

i thought you had it set up for images to show up within comfyui like a1111

[errant dust]

I think the last time I used A1111 was with SD 1.4

I have no idea where it is at now

[bitter hearth]

dont see anyone using a1111 for months now

people still use forge but i prefer comfyui

[errant dust]

I have heard about Forge, but not tried it or had any reason to

[bitter hearth]

its similar to a1111 but personally i like comfyui too much to change ui

[errant dust]

The sheer modularity and flexibility of Comfy just trumps its flaws

[bitter hearth]

performance, convenience, efficiency, all those are plus points for me with comfyui 🙂

[mortal mesa]

very active development

[bitter hearth]

yeah

[errant dust]

sure. The downside is that godawful local menu setup to add nodes.

[bitter hearth]

once i got the hang of noodling around with comfyui im hooked to it

[errant dust]

where you can find a checkpoint loader of 10 different kinds found all over the place

[bitter hearth]

sd3.5 medium gguf https://huggingface.co/city96/stable-diffusion-3-medium-gguf/tree/main

just about 2gb

[dull star]

I tried the 3.5-Large gguf Q8

its wonderful

say... compression....

@cunning lintel are you using fp16 or fp8 T5?

[bitter hearth]

turbo isn't bad either, for speed and quality both

[dull star]

T5 isn't usually the thing giving the style as opposed to clip-l and maybe clip-g

[bitter hearth]

downloading medium now, server seems jammed

[dull star]

but just in case it might contribute to the inferior painting capabilities in our tests

[tardy crag]

Any fps test on a 4090 on the sd3.5 and sd3.5 turbo?😯😯

[errant dust]

3.5 Medium is not out yet, but he did do the 3.5 Large also

[bitter hearth]

he has all 3 quantized

[errant dust]

SD3.0 Medium

[bitter hearth]

but not sure how he got hold of medium before 29th

[errant dust]

He didn't

[pseudo owl]

What do you mean by fps? If you mean frames per second, Its much slower then sdxl and even sdxl cant generate 1 fps on a 4090.

[bitter hearth]

his page link shows it

and its exactly 2.29gb

refering to q8 medium

[errant dust]

His page shows the link to SD3 medium, not 3.5

[bitter hearth]

ohh yeeh.. overlooked the version

[pseudo owl]

Mochi-1 is crazy amazing video gen(this is text to video) for open source, testing mochi on fal.

[cunning lintel]

fp16, i used diffusers default config/download

[craggy crest]

it's clip_l that deals with the fine details, artsy, style, background, ambient

[silver sluice]

Pony = Main CLIP Switch set to 1
SDXL = Main CLIP Switch set to 2, SDXL CLIP Sswitch set to 2
SD3 = Main CLIP set to 2, SDXL CLIP set to 1

is there an advantage in quality using the triple clip loaoder over the dual clip loader?

[bitter hearth]

i see you are trying to incorporate different workflows into one

[silver sluice]

indeed

im finding the quality is better with triple clip loader vs double clip and the performance is about the same, im not seeing any longer times, also I just tried the longclip model in the triple clip loader and I'm pretty sure it worked great, running it again to double check it actually worked

just to correct my earlier screenshot here's my 'final' version. to recap the LongClip model does work with SD3.5 via the triple clip loader

[turbid grotto]

btw, 35l can be used on very low ram\vram systems becasue you can drop t5 and use clips only with q5 transformers which is about sdxl size or smaller

[craggy crest]

peple have already got gguf's out

FoFr's updated his repo https://github.com/fofr/cog-comfyui/blob/main/CHANGELOG.md "I've updated any-comfyui-workflow to have access to SD3.5L regular, fp8 and turbo weights."

[rapid pivot]

What is gguf

[craggy crest]

https://medium.com/@phillipgimmi/what-is-gguf-and-ggml-e364834d241c

[errant dust]

The TL;DR explanation is that it is a form of compressed format with minimal loss that is fed in blocks so even an 8GB GPU can handle a beast like Flux Dev which nominally needs 24GB to run

[errant dust]

There is a chart that shows the levels of loss, but suffice it to say, it is far better than fp8 and Q8 is less than 0.1 worse than the pure original

[silver sluice]

yeah i use gguf q8 for my flux setup and it works great, really speedy, high quality output

[craggy crest]

this bugs me

[errant dust]

It is the poster for a new Pixar movie: "A Bug's Life: Knights of the Round Table"

[sacred jewel]

https://tenor.com/view/point-hehe-haha-smiling-i-see-what-you-did-there-gif-21567119

[pseudo owl]

Nice gen from Mochi-1, I'm still amazed for open source, from Banadaco discord
prompt: You're sitting in the driver's seat of your car, driving along a winding mountain road. The scenery is breathtaking as you pass by lush green forests and sparkling waterfalls. Suddenly, a brilliant blue screen appears on the windshield, seemingly out of nowhere. The error screen slowly pulses

[tardy crag]

Oh ok... sdxl definitely can generate more than 1fps on a 4090

[pseudo owl]

I doubt with 25+ steps, unless you are talking about sdxl lightning, lcm, turbo models.

[rapid pivot]

https://tenor.com/view/pingu-the-thing-fire-burning-flamethrower-gif-21416314

[tardy crag]

yes, i was thinking on lcm, turbo,... thats why i was curious about the sd3.5 turbo model

[dull star]

you know an inpainting model would actually be really fricking good for SD3.5

[real terrace]

SD3.5 turbo

SD3.5 large

[sacred jewel]

very nice

[cunning lintel]

painterly things

full body, style of Simón Prades, Ismail Inceoglu, by Ashley Wood, obscure, loose and expressive, broad brushstrokes, grays, whites, subtle skin tones, moody atmosphere, otherworldly demonic woman, bioluminescent fist, Dark shadowy heels, golden eyes, dark torn scarf, golden sun tattoo, dark curled horns, body composed of swirling energy adorned with intricate patterns resembling swirling smoke and sparkles

I tried a few flux prompts in 3.5, they do better, as there i already had an llm rewrite the artists to explicit styllistic stuff

[craggy crest]

it's coming i'm sure

[cunning lintel]

Style: High-contrast, black-and-white style, blending photography and illustration, with subtle, painterly textures, reminiscent of a delicate, dreamlike collaboration between the moody, high-contrast techniques of atmospheric, mystical and symbolic illustration. Scene: A creature with the body of a craftsman and the head of a gentle, skilled artisan, stands amidst a workshop filled with strange, half-finished projects, its eyes shining with a warm, creative light, as if it is about to glue together the disparate pieces of a new, wondrous creation, and offer it to the viewer.

Funny how 3.5 creates 2 different styles, i like it 😉

[bitter hearth]

what are the key difference between q4 and q8 other than size?

im guessing quality . but then why make q4 when you can have q8

or is that q4 specifically intended for low vram

and if so, how bad is the trade off for using q4

[pseudo owl]

Yeah you’re correct, q4 is going to be less accurate while taking much less vram.

The trade off isn’t too huge, you might lose some tiny detail and maybe very little prompt following compared to fp16. For most images, it’s going to be pretty similar.

[bitter hearth]

ok, and im also guessing this model file size impact the ram rather than vram?

[pseudo owl]

Ram doesn’t matter much, you only need enough to load the model file and that’s it. Q4 will use less ram then Q8.
Q4 will also use a less vram too.

Very easy way to measure vram usage is find how big the model file is(Q4_1 is 5.27) and add 1 or 2. So the model file should take around 7gb vram.

[bitter hearth]

ok neat, ty, i think im gonna run some extensive tests with q4 , already got some good results with fp8 and q8

[pseudo owl]

There’s also the text encoders, vae, they are not included in the model file and will take a bit of vram too.

[errant dust]

Essentially Q8 is within 0.1 error rate from the vanilla build. Q4 is a big hit and Q5.1 is roughly 0.3

Q4 is likely about fp8 level

[bitter hearth]

doing some renders with q4 .. quality isn't too bad

[errant dust]

it is not blind, it is simply less than the full pure model

a bit like the table SAI posted showing Turbo to be X Elo worse

[bitter hearth]

yeah and i dont think i would call such texture bad at all ..

i had confusion that q4 would have blurry and poor quality output, which is not .....

[turbid grotto]

seems like people have good experience with training sd35l

can't wait 12gb optimization

[silver sluice]

Is there an analogy to that between t5 fp16 and 8? Like would it be fair to say it has an error rate of half compared to fp16?

[silver sluice]

Nice response but doesn’t quantify the error rate just said slightly

Thanks for the effort tho appreciated

[bitter hearth]

yeah well.. it says ...
The error impact isn’t linear—so it isn't accurate to say FP8 has "half the error" compared to FP16. The impact depends on the specific workload and model sensitivity.

[silver sluice]

I just like how the other dude gave out easy numbers, i understand irs not as easy to quantify. Just a ballpark would be nice tho

[bitter hearth]

if the system can take the workload fp16 is definitely a better choice but in likely scenario of consumer hardware fp8 fine

[silver sluice]

The “slightly” and “it depends” isn’t really a good answer but kudos for trying it def provides some good insight

[real terrace]

so you load these models instead of the large turbo and that's it, it uses the same clip models and text model?

[bitter hearth]

technically yes but thre is a trade off in terms of quality which may not seem like a big deal in most cases but would matter for profesional production

[bitter hearth]

also just to be precise in answering what you asked, the gguf are quantized / optimized for better memeory managements so yes these quantized models replace the original weights

[errant dust]

That is correct. fp8 in t5 is also worse than fp16, but there is also a GGUF build of T5

https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main

[bitter hearth]

how do you compare that encoder with Scaled version https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main ?

[errant dust]

fp8 is lossy, and the fp16 is the same, though the GGUF should load and be processed faster

[silver sluice]

@bitter hearth can you ask that LLM that was primed for an answer to try to approximate a ballpark answer as to the error rate amount

[hallow lion]

local sora soon... XD

[bitter hearth]

interesting... so by what you are saying gguf version is a better and faster choice

[silver sluice]

Like if original to q8 is 0.1 then is fp16 to fp8 like 0.1 or is it closer to 0.5? 1.5? Just curious

[errant dust]

all things being equal, yes

[bitter hearth]

there are side factors with model and workload

[errant dust]

Q8 is about 0.1 away from pure fp16. fp8 is around 0.5

[silver sluice]

Yeah just ask the LLM To factor that in and give us a ballpark see what it says

[bitter hearth]

it cant assess it like that

it depends on the model sensitivity and workload

[silver sluice]

Oh I see so in this case sd3 large starts at fp16, whereas other models can exist as fp32 like certain sdxl models so the 0.1 is the difference between fp16 and q8 got it

[silver sluice]

Is that what it said? Like even if hallucinates a number I’d like to see the logic for what factor it comes up with, like the LLM knows everything so the dependencies can be calculated internally from what it knows