🆕｜sd3

[noble coyote]

Super Flux is really sharp - pin tack sharp! And a very simple workflow

[winged seal]

I am a little worried about the viability of training SD3.5 vs flux, mainly because I think Flux is so effortless to train because it knows so much so well. Sure, Flux has some lackluster default aesthetics, which I would say SD3.5 has better, but SD3.5 also looks like a hot mess in some areas

Training aesthetics on a smart model is way easier than training concepts/quality into an aesthetic model

time will tell, honestly

[turbid grotto]

lol sd3.5 is better than flux at nudity

so it is not censored

it just lacks dpo, as Lykon said

[winged seal]

@noble coyoteI have plans to train my own Flux model for general photographic realism. Professional, amateur, just generally realistic looking images. Something a lot of people in the AI community mess up 😅

[winged seal]

Oh lord, I forgot about him

Mine time away from the SD server has healed me lmao

This server is a lot more clean and organized than I remember. Did they trim it down and streamline it?

[turbid grotto]

I have been checking for sd news for a while and didn't notice anything, only flux)

[old pike]

how can i add img2img inside comfy?

[old pike]

how can i add img2img inside comfy?

[winged seal]

those are cool images in concept, but they just look so... Ugly and un-refined to me. I hope thats something that training can fix

very noisy

are they just not fully refined?

[noble coyote]

Straight out of SD3.5

[craggy crest]

the way to find out for sure is ... go train it and see

[winged seal]

I have plans to when I know of training tools with good and proper support. I have the datasets to do it no problem

[turbid grotto]

yea, it is unrefined and not directed to certain style, that is the point of base, for us to finetune as we want

[winged seal]

if it trains even half as effortlessly as Flux, we are in for a treat

[craggy crest]

Prompt: hills of glass by artist "Kentaro Miura"

[noble coyote]

I'm going to try Super SD3.5L and see if I can get a sharper look.

[craggy crest]

Prompt: flowers galore; Rob Gonsalves; hyper-realistic,hyper-detailed, fantasy, cosmic art; elegant, intricate, detailed, extremely textured, colossal, monstrous.

[winged seal]

I would argue a base should have good concepts/details/understandings of things. Aesthetics are effortless to change, but having to re-teach a whole 8B param model what people look like, and fixing huge deformations, thats more of a foundational thing

Changing a coat of paint is a lot easier than rebuilding the whole car

[craggy crest]

Prompt: an ornate box, open lid, a spider crawling out of it; Rob Gonsalves; hyper-realistic,hyper-detailed, fantasy, cosmic art; elegant, intricate, detailed, extremely textured, colossal, monstrous.

[winged seal]

Oh nice, an improvement in the typical SD digital art blur. I saw that improve with cascade as well

[mortal mesa]

people will take the shittier base car an work on it when the license suites them better

[craggy crest]

Three bubbles with smoke in different colors, by artist "Yoshimasa studio lighting"; ray tracing global illumination, octane rendering, 8k resolution, Unreal Engine 5, hyperrealistic photography, complex details, minimalistic, back lighting

[winged seal]

the prompt adherence leaves a lot to be desired, but the aesthetic is consistently "SD"

[craggy crest]

these are straight out of 3.5. just pure base model. nothing else. 40 steps

[craggy crest]

give me a prompt and i'll run it

[winged seal]

Lets try it against my validation prompt for my flux realism tune, just to see

A cinematic wildlife photograph of a black leopard perched up in a tree in the jungle. Foliage, plants, vines, tree bark, detailed, photorealistic, photographic style

I wonder how long until we have usable training tools

SD3.0 was distinctly EXCEPTIONAL at photographic realism, so I wonder about 3.5

[winged seal]

I assume its aesthetic neutrality will make it worse

ah yeah, thats a huge reduction over SD3 for sure. Much better than base flux for photographic realism tho

[craggy crest]

got another prompt?

[winged seal]

Landscape photograph of a broad and dense cobblestone village in Europe, detailed, dense houses, fountain to the right side, carniferous trees

[winged seal]

the more I look at this, the more concerning issues I see. I SERIOUSLY hope it will be easy to fix the deformations and incoherence

[winged seal]

the improper look of the black leopard also makes me worry about subject overfitting

[winged seal]

oh, this one is not too bad actually. Not too far off of flux at all for this one

[craggy crest]

it is fluid and very trainable. this is base model, no tweaking.

[pseudo owl]

Yeah it’s definitely not fair comparing a bigger finetuned model to a slightly smaller base model.

Only fair comparison is base model and another base model really.

[turbid grotto]

okay, that sounds fair but I still think it is okay to fix, especialy when model is so unrestrictive

[winged seal]

Just because a base model doesn't mean its gonna be easy to train out huge issues. Trust me, I trained over a thousand LoRA's for SDXL, and that model was a hot mess to fix certain issues in

[noble coyote]

An excellent catch-all prompt = Stock Photo or Surreal Stock Photo

[winged seal]

The unrestrictive thing is a bittt of a moot point now, but I do like to see that they are being less closed with this model. They kinda didn't have a choice haha

[winged seal]

this image shows the most promise so far, I'd say

[winged seal]

that one is less promising haha

good variance tho

one thing I DON'T see a chance of improving is SD3's limited/low res generation cap

but, hopefully upscale will be as good with SD3.5 as it was for SD3 medium

[winged seal]

that might help bridge that gap with flux

[split bramble]

Love seeing all the 3.5 examples, keep em coming. 🙂

[craggy crest]

here - https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main - go get the code and look through it

[winged seal]

very high res generations have been something I have had a lot of fun with using Flux

[craggy crest]

got a prompt? i'll run it if you like

[winged seal]

I already had a look through most of this. Why the send?

[split bramble]

Not really, just a busy and haven't had time to try it, so this is giving me a good feel for it's capabilities

[winged seal]

I'm not sure what the code has anything to do with a physical limitation of an architecture/model

[split bramble]

I'm always interested in it's recognition of characters... Sounds like it's been heavily obfuscated

[craggy crest]

in case all you'd seen was the repo comfy posted this morning

[winged seal]

ah ok, cool, thanks

[craggy crest]

it has not. give me a prompt

[winged seal]

I thought it was in response to the low res worry

[split bramble]

Well I generate a lot of Tifa Lockhart and Aerith Gainsborough 😝

[winged seal]

Flux is absolutely in a league of its own when it comes to high res generations, however, I imagine that upscale workflows will get SD3.5 there as well

tho, running upscale workflows are a lot less seamless than just running one native high res gen, so there are tradeoffs

[craggy crest]

who? i dunno if it knows who they are, but let's try

[teal fossil]

Honestly I didn't even think about SD at all anymore... I'm very happy to see you pulled through and relased 3.5. We'll see if large and medium can actually keep what the comparison graph promises.

[winged seal]

I haven't been able to have any fun just yet upscaling with flux, since the ultimate upscale node in comfy is suddenly and inexplicably broken in comfy now

[split bramble]

LOL probably not

It probably knows batman 😉

[craggy crest]

on the left is elon musk, on the right is mark zuckerberg,

Prompt: Tifa Lockhart and Aerith Gainsborough shaking hands

[winged seal]

I agree as well. I do not need SD3.5, as Flux is already phenomenal, but I would love to see more competition and be proven wrong. Its liberating to not rely on one company for good results

[noble coyote]

Rene Magritte and Lowell Herrero are a great artistic combo in a prompt

[split bramble]

I like sketchy, ink sketch style stuff.

[split bramble]

Yeah that's rough 😕

[craggy crest]

Prompt: sketchy, ink sketch style Tifa Lockhart

[winged seal]

Lets try iconic IP. Can we try "Toothless from How To Train Your Dragon on a rock to the right side of hiccup at night time, dark, underexposed, dreamworks style"?@craggy crest

[spark quail]

thing with that prompt is that tifa is a character in a video game and the model is probably confusing it with real life* toy figurines. might have to specify art or game or cartoon, mabey anime

[teal fossil]

Well... if 3.5 is finally actually capable of good finetuning... Flux suffers a lot because of the bleeding issues.

[winged seal]

oh yeah, thats valid

[split bramble]

It has a vague understanding but definitely obfuscated

[vital moth]

you can't use stable diffusion webui for sd3 and sd3.5 as generator?

[craggy crest]

that was one of the main focus points of this release - make sure it is trainable

[winged seal]

I have been training flux as of late and have found it exceptionally easy to get good results with, without bleeding. I am not sure where people see issues with that

I will say, I was using Kohya for a few weeks before AI Toolkit, as it had issues on windows. Kohya sucks MAJORLY for training flux, and will not give anywhere near as good results with the same settings

[split bramble]

Just comfy right now I think. And probably swarm.

[craggy crest]

they're going to have to have a chance to code in support for it

[vital moth]

not bad

[winged seal]

not too bad actually

[craggy crest]

swarm runs comfy, and comfy released support for it this morning

[vital moth]

ohhh...

[vital moth]

big fan of httyd lol

[split bramble]

Honestly I'm exhausted with maintaining all these front ends and models. I'm at a point where I just want to pay for a service online that has better hardware than me anyways. 😝

[winged seal]

I wonder how long until we have viable training options for SD3.5. I will try it when its stable

[craggy crest]

fun prompt

[craggy crest]

now

it is stable

[spark quail]

interesting that it seems to still struggle with background characters like hiccup here, i noticed that with other image gen models as a whole too. mabey this is more precisely describing what the devs meant when they created the refiner for sdxl

[winged seal]

Its not that bad at toothless, very nice

[winged seal]

in what, specifically?

[split bramble]

How about Dumbo?

[teal fossil]

I still didn't try AI toolkit yet. I definitely have to try that one next.

[marsh violet]

3.5 medium won't be much better than 3 medium

[noble coyote]

Down Under Manhattan Bridge Overpass

[marsh violet]

[craggy crest]

it's a base model - so a lot of stuff like that is probably something you want to train a lora specifically for

[winged seal]

I have trained over 1000 LoRA's for SDXL and 1.5 in Kohya, and let me say, the results I get with AI Toolkit are ASTRONOMICALLY better

[split bramble]

How's the 3.5 speed compared to Flux Dev?

[winged seal]

like, here

[noble coyote]

Similar

[craggy crest]

wrong

[winged seal]

base flux vs a training I did with 1K images for 12k steps in Kohya

[noble coyote]

My 64Gb RAM and 8Gb VRAM manage a 1024x1024 in under two minutes both Flux and SD3.5

[winged seal]

vs a training I did with 30 images in AI Toolkit at 1000 steps

[winged seal]

(sampler was changed)

[noble coyote]

Not quite! 🙂

[winged seal]

I am using massively less data in a fraction of the time of Kohya, and getting astronomically better results

[noble coyote]

That is ON Manhattan Bridge Overpass LOL

[winged seal]

I will not be using Kohya for Flux again

[split bramble]

Y'all running locally or online?

[mild grove]

https://jaex.getsharex.com/2024/10/yHcCSyo7KS.png

[winged seal]

flux base vs my result after 3k steps on the same 30 images that fixed that dude above

Flux learns like a BEAST

[split bramble]

That's a good one

[winged seal]

no images of big cats in the dataset, either

[marsh violet]

It won't be,
3 med is 2B
3.5 med is 2.5B
Flux is 12B
I don't believe there will be much of an improvement between 3 med and 3.5 med

[winged seal]

I am inclined to believe the same, but moderatly hopeful to be wrong

[craggy crest]

you are wrong.

[winged seal]

SD3.5B is most certainly less aestheti-slop than SD3 Large they had in the Artisan's a while back, thats for sure

but what kinda worries me is that SD3.5 doesn't have very great aesthetics OR general information. Flux made up for its bad aesthetics by having huge amounts of information deeper inside

[spark quail]

@craggy crest havent downloaded the model yet, could you try out tdg8uu's old dragon prompt? he always made some crazy good ones.

realistic, Ice dragon, desolate, intricately detailed, artistic lightning, particles, beautiful, amazing, highly detailed, digital art, sharp focus, trending on art station

[winged seal]

screw it, I can try it on my realism flux finetune for shits and giggles lol

[spark quail]

yummy

[craggy crest]

drop the trending on art station from your prompts. that's never been anything but a dice roll for random data. give me a minute and i'll run this

[spark quail]

hahah i never used it myself, looks like tdg did though and at least the ones that he shared always looked great

[winged seal]

Thats pretty cool actually

[mortal mesa]

quite square

[winged seal]

I think thats just the... res? lmao

[spark quail]

hmm looks like its again meshing a drawn artstyle with more of a 3d artstyle work

[craggy crest]

sure, cause i didn't bother to change the AR from what i was using

[winged seal]

yeah, mine is gonna be 1024x1024

tho MAN flux does higher resolutions exceptionally well

Thats the biggest thing I am confident no amount of training SD3.5 is gonna allow it to do

[craggy crest]

your prompt isn't very specific for style though

[spark quail]

it does dragons pretty well though, i remember sdxl was giving me tons of 4 eyed, 4 horned dragons

agreed, im surprised tdg made that pic above with that prompt lmao

[mortal mesa]

listen, we are collectively trying the new thing, not fanboying

[spark quail]

i usually use longer prompts

[craggy crest]

as opposed to one eyed, one horned flying purple people eaters?

[winged seal]

lmao

sorry my generation is taking so long lmfao

[winged seal]

@craggy crestWanna try something for me real quick?

[craggy crest]

sure

[spark quail]

now im interested to see a one eyed one horn one

[winged seal]

Can you try a 2048x2048x gen? I wanna see if its usable at all or if it just dies

[winged seal]

Jesus those ground textures irk me so much lmao

huh, interesting... Result from my Flux Realism training lol

[winged seal]

to be fair, my realism training shouldn't really damage other styles since I know how to train

[spark quail]

definitely looks like a deviant art pic lol

now that is spooky

[craggy crest]

i can't, sorry. i don't have a massive machine

[spark quail]

very nice details on that eye though, crazy good actually

[craggy crest]

i'd just use magific to upscale anyway

[winged seal]

ah, no worries. Maybe I will try it some later time

the viability of high res with SD3.5 is intriguing to me. I doubt it can make it past 1536x with how unstable it is at 1024x

[craggy crest]

i usualy don't do more than 512x512 - and then upscale with magnific or topas the few i want larger

[noble coyote]

There's a lot of MB3D Fractals about this SD3.5 look!!!

[winged seal]

I have been consistently blown away by how effortlessly coherent flux is past 2048x

[mild grove]

https://jaex.getsharex.com/2024/10/hKnfycqdFK.png

[spark quail]

oh that reminds me, is 3.5 trained on larger images? i recall a more recent model of SAI's could do 512x512 images but they weren't great

i figure theyre trained on the same images from 3.0 but its been so long since that release for me lmao

[winged seal]

I also wonder if SD3.5 will accept images of all resolutions with no problems like flux. Its insane how diverse of an image res range flux can train on and generate

from 256x to nearly 4k all in one model

[split bramble]

I'm ready for a model that can reliability do two or more known characters in the same image from one prompt.

[winged seal]

ok, this is fun haha

[winged seal]

I like that a lot

I have done a lot of upscale work with images in the past, and I have to say that native high res gens do have a differen't chjarm to them, cause they will try overall more dense features and textures

like a bush at 256x pixels and one at 2048x pixels will have drastically different leaf densities

[craggy crest]

Prompt: on the left is a green cat with a red ball of yarn. on the right is a blue dog with a green bone. in the middle is a yellow pyramid

[split bramble]

I mean of known characters without getting their attributes confused

[winged seal]

256/512/768/1024 in flux. You can see how the feature size and natural detail/informtational density changes. Upscaling that 256 image to 2048 will make it sharp, sure, but the detail/feature density will be off scale

[spark quail]

not specifically high res but this is my experience with 1.5 too, which is what i mainly use. 512x640 images turn out much better than simply 512x512, which we then scale up

[winged seal]

like look at the leaf to the size of the leopard difference

and the size of the bokeh

there's certain things that scaling up lower res images to higher res won't be able to cary over

@split brambletried to gen Vegeta next to Goku, but I spelled Vegeta wrong lmao

[winged seal]

omng, littlest petshop lmao

[craggy crest]

:)

[winged seal]

I used to have a friend when I was growing up who LOVED those lmao

I wonder how Sd3.5 does with fine semantics, like plant species and such

[noble coyote]

SD3.5L artistic look VERY GOOD!

[winged seal]

Flux does an EXTREMELY good job with different tree types

[winged seal]

ohhhhh, that one looks very very good

thats probably the best looking of the ones I have seen from you

[craggy crest]

any specific trees?

[noble coyote]

Great colours

[winged seal]

specifically evergreen, carniferous, sprawling, weeping, stuff like that

[noble coyote]

The realistic photos I did have a strange granular graininess

[winged seal]

it does mixed carniferous trees VERY good

[noble coyote]

Art outpuit seems more solidified

[winged seal]

yeah, thats how SD3.0 was as well. Upscaling fixed it at least

Flux has that too sometimes, but it was easily trained out

@split bramblelmfao

cursed

[winged seal]

oh god, that is...

I thought we left that early ugly AI aesthetic back in 2022

[winged seal]

wowww

[winged seal]

yeah, those look insanely good

[winged seal]

oh lord

ok, so, extensive training needed I see lmao

[winged seal]

exact prompt please?

[craggy crest]

depends on what you want, really

Prompt: pine forest, deep shadows, sparkling stream - around us we can see lush ferns and small wildflowers. sunstreaks, early morning

[winged seal]

like, realistic images/photographs

[craggy crest]

a lot of that is probably nothing more than tweaking the sampler or scheduler till you get your prefered look

[winged seal]

oh, its 100% not that, its just missing data/improper aesthetic tuning

it happens, all models have it

[craggy crest]

okay. i like the look of this one, personally. but 3.5 is trainable. and it's fluid. it's very easy to work with

[winged seal]

I don't think any of that has been confirmed, like... It JUST released

nobody has trained it lmao

[craggy crest]

i just spent a month and a half beta testing this. it is confirmed

[winged seal]

same prompt in flux with my training. I forgot to increase the steps

[winged seal]

oh, well in that case, I am hopeful is. Lord knows SD3.5 needs to be trainable. Its a hot mess mixed bag still haha

[craggy crest]

it is. we made certain of that

[winged seal]

looking forward to it then. It has a lonnggggg ways to go to be competitive with flux for training, so I hope SAI put in some actual work this time

[noble coyote]

This is six-finger-slicing!!! 😄

[craggy crest]

the better to slice you with, my dear

[sacred jewel]

Really? OMG I didn;t get the memo...

[noble coyote]

Fee! Fi! Foh! Fum!

[winged seal]

I see some small glimpses of hope here, but man, still a lottt of issues. Its gonna need a ton more work than flux, but its gonna be a little more accessible when it has more training optimizations

@craggy crestalso real quick, you didn't say where I can train SD3.5 just yet. I might try it today, even

[gusty trail]

simpletuner or diffusers

[winged seal]

I will say, the shapes and forms here are very nice and good looking. The fundementals are here for this image, even if the aesthetix/textures/details are fucked

but, those are easier to fix than core issues

[winged seal]

oh lord. Looks like I am not training it then haha

[craggy crest]

i think you're going to have to wait a few days for the guys that create trainers to put out public ones for 3.5

[winged seal]

I will wait for more accessible trainers, then I'll give it a proper try

[gusty trail]

Not much different compared to sd3 M

[craggy crest]

yea, if you want to go the diffusers route, all the code and stuff you need are on the SAI hugging face page i posted the link to earlier

[winged seal]

I am WAY too inexperienced with diffusers, and I really do not want my inexperience to lead to bad results that give me a general bad taste

[winged seal]

I'll likely wait for AIToolkit, its my trainer of choice as of now. Or One Trainer

[winged seal]

Probably not Kohya, definitely not simple tuner, and likely not diffusers

[winged seal]

I would rather write off all of SD3.5 than have to use simpletuner. Until then, I will keep making datasets for flux and save them for SD3.5 when its training is available

[winged seal]

@craggy crestoh real quick, since you seem to have insight into D3.5

[winged seal]

how damaging is multi res training for it? Can it accept multi res inputs, or does it explode like SDXL?

[craggy crest]

no idea. haven't tried

[winged seal]

fair enough. I am gonna assume its not good with multi res given how unstable it is. Most of my datasets should be just fine for that anyways

[gusty trail]

Boot Up my sd 3.5 training script.

[winged seal]

Honestly, I might use SD3.5 to generate some image bases, use Flux do un-fuck them, and then aesthetic tune Flux using it so we have cool aesthetics and a model that can do good details AND extreme resolutions with fantastic prompt following

[winged seal]

is this training script available to public, by chance?

[gusty trail]

https://github.com/lrzjason/T2ITrainer Here is the repo but sd3.5 is not public yet. I need to test it before update the repo.

[winged seal]

ah, alright, thank you! Looking forward

always love alternatives to trainers

[gusty trail]

Basically it is just diffusers script with cache embedding and cache latent

[craggy crest]

someone's gonna get to learn diffusers ;)

[winged seal]

how hard is it to implement model quantization when training?

[craggy crest]

first find out if you need to do that

[winged seal]

I will not be doing that. Definitively not

[winged seal]

you do, there is no chance you don't

[craggy crest]

don't take something apart before you learn how it works

[gusty trail]

Very small dataset with fp16 weight

[winged seal]

16GB Unet with a 10GB TE, and a .5GB VAE

[craggy crest]

you don't know that, yet. you haven't worked with it yet

[winged seal]

cached latents/TE I presume?

[gusty trail]

yes

[winged seal]

then yeah, quantization is for sure needed for more

tho, that is not bad actually for double cache

[craggy crest]

you cant' know that until you work with what you're starting with

[winged seal]

dude, thats not how it works lmao

you don't get to train 16GB in less than 16GB unless you either quantize it, offload it, or do block swapping

[craggy crest]

yes it is. in every single area of life. learn it first, then take it apart and modify it

[winged seal]

Let me just load 16GB into 10GB. I am sure it will work

[craggy crest]

comfy put out an fp8 version this morning if you're all that worried about size

[winged seal]

dope, a quantized version lmao

thats all I was asking about

[craggy crest]

grab what he created then and play around with it first

https://blog.comfy.org/sd3-5-comfyui/

[winged seal]

Oh, I have exacly 0 interest in generating with the base. I am just curious if the base is worth training over flux. I'll be waiting until I can test that for myself

Its gonna have to offer specific benefits over Flux for me to switch over, but I am more than happy to test said benefits

[split bramble]

@winged seal , get me Cloud, Aerith, and Tifa all in one image and you're golden 😎

[winged seal]

Thats not really stuff I am for or do, personally

I don't really care for that in image gen models cause its easy to do with other means

[sage burrow]

I just looked, flux didn't really seem to go much further with community loras past a few days after its release. (Compared with earlier models i mean). I'm not sure why, but sd3 may have more potential.

[craggy crest]

you can't know what the model is capaable of, and thus know what you actually need to train, till you use the base for a while

[sage burrow]

Has anyone run sd3.5 large on a 4060 yet? Did it take less time than flux dev?

[winged seal]

I think a lot of people just don't know how to train it/don't have the resources to. Its really quite easy to, and the quality is much higher than SD3.5 (for now)

Sd3.5 will likely be popular for its accessibility, but I do thin Flux will continue to be higher quality

[winged seal]

it will likely take longer for now, until it gets proper quantization

[craggy crest]

comfy put out an fp8 version this morning. if you can't get the one that we released to run, use his

[craggy crest]

did you try comfy's yet?

[winged seal]

I have seen enough to know it needs a hell of a lot of fixing. Not much else is needed

[craggy crest]

looking forward to your loras. you might play around with the base model while waiting for a trainer

[shell plaza]

sd3.5

can do mermaids

[winged seal]

FP8 is bigger than 8GB, so its gonna have to offload in some way. Having a smaller quant that doesn't need to do that will be a lot faster

FP8 will be good for people with more VRAM tho

[craggy crest]

how tiny a machine do you have?

[winged seal]

I'm not talking about me, just generally

[craggy crest]

thought you were talking about you. not that many people any more with teeny machines - most have been upgrading their hardware the last couple years

[winged seal]

that is insanely out of touch. A monumental majority of people are still on normal consumer GPU's. I know, I work with dozens, if not hundreds of people who are actual normal people

Focusing on it being accessible for other people is about the least somebody can do in this community. Its why I refused to support flux until it was easily accessible for most people

[craggy crest]

you must not be reading the same flood of 'hey, i want to upgrade, what machine should i get' posts that i have been.

[noble coyote]

Is SD3.5 open to being "GGUF'd"?

[craggy crest]

should be

[winged seal]

The dozens of posts about that don't account for the MILLIONS of people using these image gen models dude. Again, astronomically out of touch

[craggy crest]

stop tossing out insults

[winged seal]

Its not an insult, its an observation. Most people are on 8GB or less systems. Its like 100% factual

[craggy crest]

most people trying to generate images that are on those systems are also using online options...

[winged seal]

I already made the distinction by saying people with smaller GPU's looking for local image gen, where they can actually control outputs. The entire community is lucky if more than maybe 5% of users have 24GB GPU's

The cheapest one you can get is a used 3090 for like $600 if you are lucky. Most people are on 8GB or less, with some on 10/12/16

[sage burrow]

I hope mage adds sd3.5 large soon 🙂

[craggy crest]

i'm sure they will

[pseudo owl]

Sd3.5 large can fit on 8gb vram with nf4, quanto4bit, or gguf 4bit. It should be similar quality to fp8 as well.

[sage burrow]

Fortunately these days the size of one's own system is nearly irrelevant, since the online resources are faster, and coming down in price. Fal for example

[noble coyote]

I'm on 8Gb VRAM already with full 3.5L

[pseudo owl]

Lower 4bit quant will be a lot faster bc full 3.5l will use shared ram hence very very slow images. But full 3.5l will be very slightly better quality.

[winged seal]

is that really the dystopian future people want? I know I and many others who sure as hell don't want that

[gusty trail]

I don't understand why people using sd3.5 with 240 sec per gen on local. They could be faster using onlinse service.

[errant dust]

anyone know why I am getting this error? RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument scale_a in method wrapper_CUDA___scaled_mm)

[winged seal]

yeah, this. I think 6bit GGUF will be best bet for SD3.5 on 8GB, with low VRAM mode

Maybe 5 bit for some extra headspace

[wispy epoch]

nice nails but hand xdd

[noble coyote]

Just refresh and start again - usually clears it

[raven fern]

so is it possible to train lora already? 😮

[glad yew]

How big is the diversity of SD 3.5 styles and multi-character interactions? Did they remove the names of all or most classical and modern artists?

[raven fern]

also cant wait for controlnets

[pseudo owl]

Its better in flux in styles and artists I believe but flux might be slightly better in multi-character.

[winged seal]

kinda I guess? None of it is properly validated just yet

[sullen moss]

bad

[winged seal]

ohhh, this is an annoying one. Are you using ultimate upscale in your workflow by chance? That is a very big tigger for that error

[raven fern]

according to their technical paper summary or whatever, it seems to have better score for prompt adherence than flux, but a bit lower score for the aesthetic

which i guess makes sense since flux is 12B

[errant dust]

No, using the default workflow in Comfy

[hallow lion]

I have a good feeling about this...

[raven fern]

thats a cool pic not gonna lie

[errant dust]

That is the workflow I am using

[raven fern]

it's funny how they introduced the model with a woman lying on grass LUL

[hallow lion]

Might as well rename the channel SD35 XD

[noble coyote]

This is the picture at Comfy.Blog

[errant dust]

Exactly, and is the workflow I am using that is yielding the error

[glad yew]

Lykon showed that its anime-style quality was going to be so good that it could create images that look like screenshots.
Can SD 3.5 do that? Dall-E can.

[raven fern]

some dude on huggingface says he has problems with the vae? gives black images, is that just him or is it fine? im gonna try it later, busy now

[noble coyote]

VAE baked in

[errant dust]

I'd be happy to get a black image, since that is a black image more than my nothing at all

[noble coyote]

Show me your workflow 🙂

[raven fern]

so are you guys using the baked in all in one version?

[errant dust]

That pic is my workflow

[noble coyote]

Let me see if my PC can do it too?

[pseudo owl]

What do you mean? Even sdxl can create great anime images, sd3.5 can do it better then base sdxl tho.

[raven fern]

im curious about 3.5 medium also, they mention it can go to 2 megapixels, and is slightly different than the large model? idk

[raven fern]

oliver queen :3

[winged seal]

hmmm, I am not sure then

[raven fern]

haha nice

[noble coyote]

You've got no Model Sampling - set it at 3

[raven fern]

does it do anime well?

[lavish sparrow]

i remember super mario rpg differently xD

[raven fern]

@simple thistle is probably generating tons of fennec anime girls as we speak :3

i also hope they release svd2 one day

[errant dust]

Where do I fit it?

[raven fern]

@dusky thistle did you try some clown sharks with 3.5? :3

[winged seal]

Ok, general takeaway. SD3.5 is far from flux, but its kinda fine if its something else. It remains to be seen if it can train anywhere near as good as flux, but maybe it can

Its got a more artistic base, but its also way less coherent and consistent, so time will tell

I look forward to training it, even if my expectations aren't that high

[glad yew]

A StabilityAI employee (Lykon or other) showed that SD 3 Large would be able to create anime with this quality that looks like a screenshot (this image is not generated)

[raven fern]

as long as it can train some nice loras, il be happy

[errant dust]

@noble coyote With this I still get the same error

Can you share your WF?

[raven fern]

what error are you getting?

[errant dust]

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument scale_a in method wrapper_CUDA___scaled_mm)

[raven fern]

isnt there a comfy example that should work out of the box from the comfy sd3 examples page

maybe try that

did you update your comfy ui?

[errant dust]

This is the default copmfy on the announcement page. Now with model sampler added

same error

yes

[dusky thistle]

yeah, the SDE samplers are working with it

[raven fern]

nice

[craggy crest]

comfy released an exmaple workflow on his blog page. but SAI also released one on the hugging face page, and they both work out of the box

[dusky thistle]

i'm trying to clean this repo up a bit cuz it's gotten to be bloated with experiments over the last few weeks with nailing down the math for SDE sampling with flux/2b/wetc

[raven fern]

yea idk why that dude has problems then...

[errant dust]

and giving me the error described

[craggy crest]

i missed the error. what is it?

[errant dust]

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument scale_a in method wrapper_CUDA___scaled_mm)

[raven fern]

what is your system specs?

[craggy crest]

oh. welllllll that's something you'll find other people reporting in the past in the #🤝｜tech-support channel

[errant dust]

Dude, I run SD3 Medium and Flux with no issues

[craggy crest]

you using AMD?

[errant dust]

RTX 4060

[raven fern]

are you using the scaled versions?

[errant dust]

Yes

[raven fern]

can you try without scaled

[errant dust]

with 8GB VRAM, it has crashed before even trying, But sure

[raven fern]

8gb vram... man these days that is not great my dude

[craggy crest]

i would suggest asking @naive sparrow cause it sounds like your CPU is trying to take over

[errant dust]

It runs Flux without issues

[raven fern]

cause comfy is probably like offloading to cpu and thats why you have that error

with the scaled versions

[craggy crest]

https://stackoverflow.com/questions/66091226/runtimeerror-expected-all-tensors-to-be-on-the-same-device-but-found-at-least it's a python error - that's a post on stack overflow that might by helpful

[raven fern]

@errant dust can you run the flux scaled version?

[craggy crest]

it's not the model. it's a python issue

[raven fern]

yea but im just curious

[craggy crest]

he said he could

[raven fern]

well he said flux, not sure if its the scaled version tho

[errant dust]

I have no idea what a Flux Scaled version is. I run the OOTB models, super slow, or the GGUF builds

[noble coyote]

This w/f OK with me ...

[raven fern]

can you run this version of flux? https://huggingface.co/comfyanonymous/flux_dev_scaled_fp8_test

if you cant, then your hardware just cant run the scaled versions at all

[winged seal]

Flux can be used on like a 4GB GPU now with how well optimized it has been

my 8GB 3060ti runs it just fine when I am training on my 3090

[raven fern]

you didnt upgrade to the 40xx cards? :3

[signal shuttle]

It could run, just very VERY slow

[wispy epoch]

possible to run on 4070 super 12gb? someone tested ?

[winged seal]

yes, it is

[winged seal]

its not too bad for the quality you get

[raven fern]

@bitter hearth can you run 3.5 on 4gb vram? :3

[signal shuttle]

It could be sped up using tensorrt but i don't think any 4GB card supports tensorrt

[pseudo owl]

Since sd3.5 large is smaller, it can run run on even less vram then flux lol. It's also faster. 6gb is needed for ok speed tho.

[winged seal]

that would take like an hour to compile, no joke. On GPU's with lower VRAM it takes longer and longer

[winged seal]

yeah, it is smaller, so it will be faster, but the quality difference is much bigger than the size difference IMO. I would way rather wait for Flux than just go with SD3.5

[raven fern]

so wait, 3.5 uses clip g as well? man thats a lot of text encoders

[winged seal]

yup, for like no reason lmao

just like SDXL, you can totally not use g, and it works just fine

[raven fern]

yea

[pseudo owl]

yeah it should be the same with sd3.5 large too, it was the same with sd3 medium.

[raven fern]

the first thing im gonna generate is some CCTV footage :3

those are always fun

[winged seal]

I am just not sure if they undid their lobotomization of T5XXL

it seems like they at least attempted to in their write up

[raven fern]

not a huge fan of t5xxl personally, this is why im curious how Sana will do it

[signal shuttle]

Didn't google make T5XXL?

[simple thistle]

update and it might be fixed

[raven fern]

comfy 😮

[winged seal]

I believe so, but I meant the way they trained normal 3 Medium lead to T5XXL being damn near useless. I just hope they undid that

[wispy epoch]

sad

[raven fern]

im not an expert on licenses, does sd3.5 have a "good" license now?

[winged seal]

yes

its license is quite good

[raven fern]

nice

[winged seal]

tho, the license isn't much of an issue for Flux now either

[raven fern]

imagine flux releases their next model like tomorrow LUL

[errant dust]

It is updated. I did so multiple times. Will do it again, so sure

[winged seal]

Flux is monumentally ahead of SAI, so I mean... I wouldn't be surprised

they also have funding out the ass compared to SAI

[raven fern]

are you updating just comfy or the dependencies as well? cause maybe you need a better torch version

[winged seal]

I have business partners that work with BFL

[errant dust]

Just the COmfy. I don't use the embedded

[raven fern]

yes but i heard latest torch versions are better for some stuff, you might have to update that tho, like at least version 2.4 i think

[errant dust]

I have it already

[raven fern]

kk

maybe sd3.5 medium will work for you :3

[errant dust]

The newest update is rendering! (something for the moment)

[raven fern]

😮

[winged seal]

Medium will likely be a way better option sthan large. Large isn't really all that good for its size

[winged seal]

:D

[errant dust]

and it worked! Woohoo!

[raven fern]

nice

[raven fern]

now go and generate some "stuff", you know what stuff im talking about cough

[errant dust]

That is crazy fast though

[signal shuttle]

I bet 10 dollars that when SD 3.5 medium releases its going to get trashed by reddit because it isn't outperforming Flux dev

[raven fern]

the classic bottle example

[errant dust]

I am a mega square. Meaning I have zero interest in boobies on japanese cartoon goils. 🙂 But yes, now for some fun and experimentation.

[raven fern]

sure

[winged seal]

I mean, large isn't either, but Medium will at least be viable for most people

[merry zodiac]

But where sd 3.5 small

[raven fern]

smol 😦

[winged seal]

likely not gonna happen

[signal shuttle]

I think anything below 2B is just too small to be good

[raven fern]

wait i forgot, how many params was sd 1.5?

[winged seal]

0.85B or 850M

[turbid grotto]

turbo runs in 12 seconds on rtx3060 😮

[merry zodiac]

1B for all, 850M unet

[raven fern]

ah yea

[winged seal]

thats so not true lmao

there are 1.5 models that can still do certain things better than SDXL and hell, even flux

[raven fern]

yea i still use some sd1.5 finetunes

[errant dust]

So quick question: how many params is the Large here? 8B?

[pseudo owl]

They will not be good for general, just specific areas tho. But yes, a finetuned specific model can beat models way way larger. Same with llms where even 3b finetuned models can beat gpt4o in certain domains.

[pseudo owl]

Yeah, the mmdit only tho. It has t5xxl encoder which is like 3-4b. Everything combined should be 13b. Flux is 16b everything combined

[raven fern]

what about the quality of the pics?

[signal shuttle]

But generally models like SDXL and Flux and 3.5 outperform 1.5, 1.5 wins in very very specific use cases

[errant dust]

I am going to guess we will soon be seeing some nice GGUF models

[winged seal]

yes, for the unet

[raven fern]

nf4 too i guess

[errant dust]

GGUF was by far the higher quality of the two, and efficient. Q8_0

[pseudo owl]

nf4 takes way less vram then q8, while being faster then q4. but yes q8 is slightly better quality as its 2x bigger

[raven fern]

this shows 8.1B? maybe typo or they rounded down to 8 lol https://images.squarespace-cdn.com/content/v1/6213c340453c3f502425776e/3c173f1c-00d8-47b2-9d7b-f321a452935f/chart+(1).jpg?format=2500w

that elo score tho... just like my chess elo :3

[turbid grotto]

not bad for simple stuff but smooth textures

[raven fern]

yea

[turbid grotto]

needs finetunning

[pseudo isle]

sd 3.5 shit

[raven fern]

gonna be awesome to see controlnets, loras, finetunes and so on 😮

[winged seal]

they always work like that for transformers

its always a soft rounding

[raven fern]

ah

[pseudo owl]

Already some loras from shakkar lol. They uploaded like 4 hrs ago lol

[winged seal]

31.5B becomes 32B, 9.3B becomes 9B

[raven fern]

what kind of loras? 😮

[winged seal]

not very good ones I would assume 😅

[raven fern]

wait isnt shakkar the same group that did the flux stuff or am i mixing the names

[winged seal]

I will train SD3.5 when good training options are released

[raven fern]

nice

i wonder who will release training stuff first, prob koyhaa

[winged seal]

I am not too hopeful, but I am also not gonna fully write it off

[pseudo owl]

Yeah they made a flux controlnet union too, and lots of loras for it. They already trained a few loras for sd3.5
https://huggingface.co/Shakker-Labs

[turbid grotto]

🙏 😳 ❤️

[signal shuttle]

I think shakker had early access to SD 3.5 large since they released the loras a hour after the announcment of SD 3.5 large

[raven fern]

yea they prob worked together

[winged seal]

likely, yeah

[errant dust]

There is a big diff on my system though: with NF4 I was limited to 1024 x 1024 since anything more and it would be slow as a snail. Also GGUF is loaded in blocks and not all at once

[signal shuttle]

Like how latent vision had early access to the 3.5 large checkpoint

[sage burrow]

can you just select update comfy or update all to run 3.5? I haven't used comfy in a month or 2

[winged seal]

When I was in the SAI beta's, we weren't really able to train them, but with how shit SD3 was, so I presume they wanted to have something to go along with it so it doesn't flop as hard

[raven fern]

Becky? are you generating some sexy men? 😮

haha

[sage burrow]

I

'm at the trying to get the thing to run stage lol

[winged seal]

I'm quite the sexy man generator myself

[raven fern]

oh shit

waddup

[winged seal]

oop

[winged seal]

:3

[sage burrow]

grumble

[raven fern]

pngs are usually bad, i like json, it loads without failing

[winged seal]

rare 1% of AI users who give any shits about men

[errant dust]

Becky, use this

[sage burrow]

well also, I probably have to wait for some lora creation places to popup to create my custom loras 😄

[raven fern]

yea i cant wait to train some loras

[errant dust]

and be sure to use the fp8 model by Comfy, since the plain model won't load with less than 12GB Vram

[raven fern]

hopefully my boy Kijai releases something for sd3.5 how he did with flux

[errant dust]

https://comfyanonymous.github.io/ComfyUI_examples/sd3/?ref=blog.comfy.org

[winged seal]

Man, remember when image gen models were made for normal consumer hardware to run? Back when local AI was properly accessible?

[errant dust]

It says SD3 but is now updated to 3.5

[raven fern]

i will sail with her

[winged seal]

ah, as synthetic/plastic as ever

[raven fern]

but yea the plastic hurts tho

[cunning lintel]

nvidia has you covered 🤡 https://nvlabs.github.io/Sana/

[winged seal]

SD3.5 didn't fix that, thats for sure lmao

[vapid radish]

The skin or the boobs? 🙂

[raven fern]

LOL

[winged seal]

everything

[sage burrow]

I guess I hadn't updated my comfy properly before, or I had downloaeded the wrong thing called sample image

[winged seal]

lol. I don't have a limited system. I just meant for other people

[raven fern]

Sana gonna be interesting, hopefully "soon" means couple days max... 😦

[signal shuttle]

Nivida cooked with their LLM nemotron 70B model, yet with Sana nivida managed burned the kitchen down so hard they started a wildfire

[winged seal]

NVIDIA is actually starting to get pretty decent at AI models. Took them a while

[raven fern]

yea Nvidia doing some cool stuff

[errant dust]

NO, I had updated, mutipl.e times. Seems a late update after this solved my direct woes

[cunning lintel]

i really hope some mystical magical small model gets released that's good (so there will be usuable video without 4h100 😬 ) but for now bigger seems better :/

[winged seal]

Yeah, Nemotron is actually surprisingly good. Their other nemo models were hot garbage

[errant dust]

Now awaiting the GGUF builds. 🙂

[winged seal]

so glad gguf is available now

[raven fern]

what happened to the gguf master the bloke? he died? 😦

[winged seal]

would be even better if we get faster acceleration structures

[errant dust]

Anyone can build a GGUF model. I just lack the system resources. There is a page with the tools to do it

[winged seal]

exl2 would be crazy haha

[raven fern]

of course, but i always remember thebloke having all the latest models lol

[winged seal]

can you link please? I need to convert some of my Flux trains to gguf at some point

[raven fern]

aww

[raven fern]

im curious how many anime characters it will know

elden ring dlc 😮

[errant dust]

Here you go: https://github.com/ggerganov/llama.cpp

[winged seal]

If SD3.5 trains even 1/10th as good as flux, not too bad

[raven fern]

when is llama.cpp gonna support visual models... sigh 😦

[winged seal]

does llama.cpp just have direct support for ggufing flux now? Last I checked, you needed a special tool

[vapid radish]

were you prompting for a "Flux Chin"? or is that just Flux?

[winged seal]

unfortunately not soon enough

been waiting for the same

[signal shuttle]

When llama 3.1 405B fits on mobile phones, thats when Llama.cpp starts to support vision models

[raven fern]

molmo is very good at uncensored stuff

[raven fern]

405B that is WILD to me

[sage burrow]

grumble

[raven fern]

becky did you update?

[winged seal]

Wait till you hear about the bigger models lol

[raven fern]

haha

[winged seal]

Fatlama would put us all in a coma

[signal shuttle]

Isn't gpt-4 1T parameters or something?

[errant dust]

Then reach out to the author: https://github.com/city96

[sage burrow]

I did

[winged seal]

1.75T

Fat lama is also 1.7T IIRC, but hes making much bigger ones

[errant dust]

Use the workflow I mentioned and the all-in-one checkpoint

[winged seal]

all of the real progress is being made in accessible consumer scale LLM's tho, its pretty amazing

[pseudo owl]

Old gpt4 was 1t but the new better gpt4o is unkown. It's probably much smaller as its way faster.

[raven fern]

1bit 405B :3

[winged seal]

still over 100GB

[raven fern]

haha

gonna need RTX 9090

[sage burrow]

I'm using yur workflow

[winged seal]

I believe the new Obese llama is planned to be 3.8T

[pseudo owl]

405b is actually considerably cheaper then gpt4o, claude, gemini while being similar performance to gpt4o.

[signal shuttle]

Isn't this bloating paramater numbers at this point?, this gives no real preformence boosts does it?

[pseudo owl]

Yeah no performance boost(degrades quality), but why not add a few trillion parameters for fun?

[raven fern]

if it can generate the whole source code for GTA 6, its a good LLM :3

[signal shuttle]

I wonder how a MMDiT model would look like with over 100b?

[winged seal]

oh yeah, its 100% that

thats the whole point lmao

he just wanted it to be bigger lol

fat llama 1.7T is over 1TB

[errant dust]

Ok, time to make some apples to apples comparisons

[winged seal]

https://huggingface.co/mradermacher/FATLLAMA-1.7T-Instruct-GGUF

oh wow, its actually over 3TB

[signal shuttle]

I remember years ago when having anything above 5B was considered huge

[raven fern]

part 1 of 19, part 1 is like 50GB 🤣

[winged seal]

consumer grade 20-30B is the SOTA rigth now and its incredible

Nemotron 70B is pretty great too

[raven fern]

skinnyLlama when

[winged seal]

thats just the new llama 3.2 1b

[raven fern]

yep

[signal shuttle]

Fatdiffusion when????

[raven fern]

Black Lotus 😮

[rancid dawn]

redditor

or discord mod

[raven fern]

LOL

[errant dust]

It is like that film Annihilation where people blend into plants, except here it is with burgers

[noble coyote]

Uses Mistral

[signal shuttle]

Woah 3.5 went really low quality

[rancid dawn]

how many step you guys use?

[winged seal]

I wish image gen models were making the progress text gen models have made

[rancid dawn]

im using the new sd3 so im wondering which one would be perfect?

[split bramble]

57 seconds to generate an image, comfy fp8 version of SD3.5, on an RTX 3060, 12GB VRAM

[rancid dawn]

i mena in term of step

the more step the more is slow

the less well ye

[split bramble]

20 steps

Feels about right.

[rancid dawn]

ye it good but it kinda slow to generate

is there one other option i should change?

[winged seal]

nope, thats just how models this size are

[rancid dawn]

oh wait 10 gave me this

[split bramble]

oh sorry I wasn't trying to respond, just sharing my experience 😄

[rancid dawn]

ohhhh

no worries

[split bramble]

similar to flux dev if I recall

[rancid dawn]

flux dev?

[split bramble]

yeah

Using the basic comfy example workflow.

[rancid dawn]

oh

do changing the cfg do anything or nah?

[split bramble]

Sorry Mihoru, I'm a little out of the loop on what you were asking for and feel like I am misleading you with my random chatter. 🙂

Let me reset...

[rancid dawn]

it alright

[split bramble]

What are you trying to do?

[vapid radish]

Flux / SD3.5 Large
"How do u' like them apples?"

[rancid dawn]

im trying to increase the generation speed

i did set the step to 20 and idk what else i could change to increase

im not really familiar with setting

all ik is that step are like faster if there less but bad graphic
instead of being slow and good graphic when higher

[split bramble]

What GPU do you have?

[rancid dawn]

hmm

lemme remember

also have this while i search

oh wait

nvm

[split bramble]

3060 but which version? 12GB?

[rancid dawn]

hmm

i think?

[split bramble]

Ok - what model and how long does it take for you?

[sacred jewel]

🤪

[errant dust]

Definitely one of my favorite films.

[rancid dawn]

the ai model?

[vapid radish]

Using in a big negative prompt seems to really mess up the image (especially the background), can anyone confirm that? I'm going to use it without negatives for now, shame as that is one advantage SD3.5 had over Flux Dev.

EDIT: aww maybe it wasn't the negative, backgrounds just often get messed up, for some reason.

[split bramble]

yup

[rancid dawn]

dat

and it take i do say 30-60 sec wiht 20 step

[split bramble]

yep that sounds right. that's how long it's gonna take.

[rancid dawn]

aw gotcha

[split bramble]

If you want faster, use SDXL models, or 1.5 models.

[rancid dawn]

was thinknig there was a way to tweak it

[lavish sparrow]

wow, this prompt actually did so well 😮

[spark quail]

sounds about right
im using it with barebones ksampler settings and im getting ~30s on my 3090 with 20 steps

[lavish sparrow]

i'm using dpmadaptive sampling -> it's slow, but real good quality

[rancid dawn]

"may i interessed you into bitcoin?"

if you dont like it dont

refuse or else

a bullet willl suddenly appear on your forehead

[weak trellis]

can anyone explaing in short what this does and how to use it?

[winged seal]

Large is for sure not as refined/stable as flux, but it is at least more accessible for people on mid sized GPU's. I think SD3.5 medium will be a lot better tho for quality density

[split bramble]

Is there a way to add live preview to this simple comfy workflow?

without needing a degree in comfyology

[winged seal]

yeah

I think its this in settings?

not too sure

oh wait, but it uses a different VAE.... hmm

[dusky thistle]

ODE vs SDE with SD3.5L

[rancid dawn]

Santa when seeing all those decoration in october for christmas

[spark quail]

6 legs but pretty nice detailing

first one was only 5 steps, was testing something, this one has 20 but i somehow like the 5 steps one better

prettty good one the finer detailing

noticing some casual disturbances in the snow in these, didnt even prompt for it. pretty neat action footage

[noble coyote]

Use the new GUI and select the top icon in the left toolbar - the one which looks like a Refresh icon

[rancid dawn]

Golum

oh god

[spark quail]

im used to prompting with 1.5 so its usually a ton of 1 or 2 words followed by a comma; its nice being able to just type out whatever flows through my mind in describing a thing with sd3's text encoder tbh

[slate portal]

lower the cfg to 2.8

helps a lot

the default suggested 4.5 cfg blows out hte images

[rancid dawn]

i wanted a mp5

[spark quail]

yeah def gonna play with the cfg, im still just playing with low steps because i want to test out different samplers and schedulers

[split bramble]

ooo.k

I just get a spinning circle.

[dusky thistle]

SDE sampling with SD3.5L

[weak trellis]

Use efficiency Ksampler

It has preview built in

[split bramble]

I have KSampler, and KSampler Advanced.

[sage burrow]

thanks for the workflow @lavish sparrow finally one that worked (mostly). I lowered it to 30 steps.

Took just as long as Flux Dev

SD3.5 does not do NSFW right out of the box (no surprise, none did/do). Looks really amazing though!

[lavish sparrow]

you're welcome, it was pretty easy adapting it ^^

[weak trellis]

Grab them from manager

[sage burrow]

For whatever reason, I just kept getting errors wtih all other workflows

[split bramble]

and now we see why comfy is a rabbit hole... there's always another step you have to do underneath the step you're trying to do. 😛

[lavish sparrow]

i'm mostly using default nodes, except for the ones that aren't obviously, but uh... i suppose most should work

[sage burrow]

has anyone managed amazing hands with 3.5 yet?

[errant dust]

Yes. But we may not mean the same thing when referring to 'amazing' hands

[lavish sparrow]

who cares about hands xD

hands?

[errant dust]

It is also hard for me to really comment on such since I am using an fp8 model which will be inherently worse than the vanilla. In fact in Flux the fp8 was about as bad as NF4

[weak trellis]

Ergo proxy flashback

[lavish sparrow]

ergo proxy? now that's a name i haven't heard in a long while

was a weird anime tho

[weak trellis]

It was, I enjoyed it I like theme song a lot

[icy coral]

Just wait for fine tunes. Why don't they start releasing an accompanying fine tune specifically trained on realism instead of letting the haters go crazy and amplify the defects of a clearly undeertrained base model is beyond me

[sage burrow]

4 fingers and a thumb on each lol

[lavish sparrow]

for some reason, that scene where the chick pushed the dude of the deck of the ship and scribbled the note "oh, he's left/right handed" (can't recall)

[proud haven]

Just to triple-confirm, no SD3 controlnet is currently compatible with SD3.5, correct? That is my assumption but wanted to check

[weak trellis]

nope I think

[errant dust]

Oh mine have 4 fingers and a thumb. They even have a spare extra finger for good measure.

[noble coyote]

Go to Manager/PreviewMethod = Latent2RGB

[errant dust]

Very cool

[lavish sparrow]

where i'm going, i don't need realism!

[slate portal]

LMAOOOO i had to

https://x.com/DataPlusEngine/status/1848849793344283122

DataVoid (@DataPlusEngine) on X

SD3L

i had to lmaoooooooo

Twitter · Oct 22, 2024, 10:11 PM

[raven fern]

haha that pic tho, you are "getting there" Becky :3

[raven fern]

that actually looks cool

damn lot of spaces huh

[sage burrow]

Even my computer flushed with the prompt I used though; it left some stuff out haha

[raven fern]

haha

[raven fern]

can you make John Wick's car :3

[sage burrow]

3.5 via toyworld. Turned out better on my own system, but also took 10x as long on my own system

[raven fern]

big boi

[gilded silo]

ai has peaked, all downhill (in terms of quality) from here

[lavish sparrow]

"Honey, did you forget to feed the flower?" - "THE WHAT?!"

[raven fern]

haha

[raven fern]

that can easily be like an enemy in elden ring

[craggy crest]

@ostrisai already has preliminary support for SD3.5-large lora training
out https://github.com/ostris/ai-toolkit/commit/3400882a8099645ce4c797f57ac258f1e1424ffd

[gilded silo]

all these people
yet almost none answered the questionnaire

[craggy crest]

?

[raven fern]

nice, ye cant wait for more support

[craggy crest]

give it 24 more hours

[lavish sparrow]

yeah, i'm like 3.5 more and more, i feel it's more responsive to prompting. which i like ^^

[dusky thistle]

third order RES SDE with CFG++

[gilded silo]

the "future improvement" has arrived

[craggy crest]

it's fluid. it's almost effortless to work with it

[dusky thistle]

third order RES

[raven fern]

well according to their analysis, it did score a bit higher than flux for the prompt adherence

[lavish sparrow]

yeah, it does so, in my limited testing

[craggy crest]

it uses 3 encoders, flux uses two...

[dusky thistle]

regular CFG for comparison

[sacred jewel]

This is interesting

SD3.5 Prompt:

[raven fern]

yep

[sacred jewel]

Can someone try it? I can;t imagine it is the model...

[raven fern]

a lot of these pics can be like cover art for some cd 🙂

[sacred jewel]

Flux with same prompt as above

[sage burrow]

3.5 via https://huggingface.co/spaces/stabilityai/stable-diffusion-3.5-large

[craggy crest]

in my tests, you can take cfg through a wide range of numbers and take steps down to 8, and still not have too many issues

you also have shift - though that's reported to be missing in comfy's workflow. it is, however, in the SAI workflow

[raven fern]

you have failed me for the last time

:3

[lavish sparrow]

spooky

[craggy crest]

Prompt: an abstraction painting of a cat, minimalistic, line art, isometric bilateral differentiation Width: 1024, Height: 1024, Steps: 40, Cfg Scale: 4.0,

[lavish sparrow]

fibonaci cat. perfect proportions!

all things end in cat.

[raven fern]

golden ratio cat lol

[gilded silo]

all these dreams, someone had to shape them

[raven fern]

wild colors there

[noble coyote]

Fibonacci, rococopunk, voronoi, fauvism, zentangle cat

[lavish sparrow]

https://tenor.com/view/doja-cat-star-wars-gif-25078126

[sage burrow]

It won

t make warped clocks. Yet, I'll have to create a lora

[sacred jewel]

Workflow embedded... I broke SD3.5L 😛 😛

[noble coyote]

... just about to!!! 😄

[raven fern]

Capra Demon's cousin

[pseudo owl]

There’s also a new sota video gen model if I didn’t mention that. Just released today too. Apache license.
https://huggingface.co/genmo/mochi-1-preview

[rapid pivot]

Meow

[lavish sparrow]

did u see my crystal cat? 😮

[raven fern]

dude what is happening these days, technology is moving so quickly, barely have time to play with one toy

[rapid pivot]

Actually good local video model??

[rapid pivot]

I see everything

[rapid pivot]

👀

[pseudo owl]

It’s not really every time that it’s moving quickly but today has been a crazy day. 2 new video gen models, sd3.5 large, Omnigen

[raven fern]

nice

[craggy crest]

release tsunami!

[noble coyote]

Works quite well

[dusky thistle]

perfect hands!

[pseudo owl]

CogVideoX is great too but this seems even better. Only requires 4xh100 right now lol.

Kijai is experimenting with fp8 mode so it fits in 24gb vram. Not fully fleshed out yet tho.

[raven fern]

woman lying on grass where? :3

[rapid pivot]

In my backyard

[raven fern]

hehe

[lavish sparrow]

oh, i'm not changing seeds 😮 oh well

[raven fern]

i can fix her

[rapid pivot]

I can't even count to 4xh100

[lavish sparrow]

do not be afraid, now that i think about it, looks war-framey

[sage burrow]

Lora definitly meeded. Fortunately I still have all my images from the flux loras

[raven fern]

the best i have currently is just 12GB vram 😦 , im waiting for 5090 to drop, so i can upgrade my pc anyway

[noble coyote]

Fibonacci, rococopunk, voronoi, fauvism, zentangle cat

[lavish sparrow]

sometimes it's one of those days

[sacred jewel]

The longer the prompt token quantity, the worse it gets... Workflows embedded

[raven fern]

mirror's edge sequel

[weak trellis]

Angewomon

[pseudo owl]

More results from mochi video gen model

[raven fern]

wait i forgot, is angewomon the booba from digimon? lol

[raven fern]

sd3.5 pony?

[dusky thistle]

cfg++ 2.0

[noble coyote]

Streamline Moderne, retrofuturist, fibonacci, rococopunk, voronoi, fauvism, zentangle cat

[slate portal]

no just smart prompting and 2.8 cfg

[raven fern]

but im wondering if there will be a pony version eventually

[slate portal]

doubt we need it at this point

[raven fern]

now add some fractals

[brittle nexus]

Close enough 😄

[dull star]

lol

[raven fern]

haha

for the mochi video thingy "The model requires at least 4 H100 GPUs to run", aka, this is only for the 5 people out there

[dull star]

I hope the fact that paintings are mediocre is because of my sampling settings

anything above CFG 3 is a super saturated mess

and below is losing coherence

[raven fern]

i mean you know what they say man, one man's garbo is another's gold :3

[dull star]

you can rent gpus on runpod and stuff if you really mind privacy

but sadly since it cannot possibly be finetuned it will be forgotten

people REALLYYYY want corn

no matter how shit and unprofessionally bad it looks

[rapid pivot]

Astra is doing pony on another model

[dull star]

auraflow

[rapid pivot]

Auraflow???

[noble coyote]

Septo-prompt (7 x one prompt of - fibonacci, rococopunk, fauvism, voronoi, zentangle)

[rapid pivot]

Ye

I remembered right

[dull star]

auraflow is perfect for it cause its cartoony

[rapid pivot]

Actually smart for once

[dull star]

I won't care for it cause it probably wont do photoreal stuff

im just gonna wait for amateur photography lora for SD3.5

this model is SOOOO much faster than flux

[raven fern]

but flux is a beast :3

[dull star]

it sure is

[rapid pivot]

Beastly boring

[dull star]

its 4B parameters ahead

[rapid pivot]

[dull star]

but it's just so goddamn slow

I'm gonna experiment on where its superior

but for paintings its equally bad at the moment