🆕｜sd3

[noble coyote]

My PNGs contain the w/f

[short thicket]

Sweet! Thanks!

Some of these loras have really funky effects on the models lol

Notice the white line on the right edge of the image. I've seen this with a few of them. Totally ruins the models. I dunno what they used to train on.

[short thicket]

I think I identified the lora. You can see the effect in some of the example images towards the bottom.

https://civitai.com/models/661699/style-of-h-r-giger-flux-295

[noble coyote]

Super Flux

[sacred jewel]

What idea is that?

[noble coyote]

https://www.youtube.com/watch?v=3WViQ2pL6ks

[sacred jewel]

Thank you!!

[fathom turtle]

An Asian female star,with artistic texture,a beam of sunlight slanted on the body, --ar 9:16

[hallow lion]

https://www.youtube.com/watch?v=9jXdA2MPm60

[noble coyote]

SANA has grown out of PiXart Sigma 🙂

[split bramble]

What is that? Is it even released?

[bitter hearth]

just another t2i model, i doubt it's not a flux killer though but a fun research model to tinker with. Will probably come out within 2 weeks

[split bramble]

ah

[split bramble]

gotcha, thanks

[short thicket]

1.6b and .6b, definitely not flux killers but I dig that people have options for lower vram.

[sacred geode]

how much longer until SD3.5? Been waiting for it as Flux gives me the runs.

[pseudo owl]

Yeah it does not beat flux dev but it’s incredibly small and fast while being really good for its size. There are 2, 0.6b and 1.6b params.

Both models have great prompt following, great text rendering, and decent image quality according to benchmarks in the paper. Humans might have issues tho.

[craggy crest]

as long as it takes. you could just use SDXL for now, it's good and i don't think it'll give you the runs

[sacred jewel]

using the fp16 GGUF model

[hallow lion]

Here we go again..

Give them at least 3 weeks this time to sort thigns out

[split bramble]

what's she holding?

[rapid pivot]

Holding the looking at the camera prompt

Or away from

Cx

[sacred jewel]

Spy Camera

[split bramble]

ah i see

that's flux?

or that sana thing?

[sacred geode]

I train and no one downloads XL trainings. ratio is 30 to 1 flux to XL. I don't need waifus.

[sacred geode]

Yes, here we go again as they are late so something is holding them up since they announced it, or did they announce it knowing they are months away?

[craggy crest]

sure they do, they're still downloading 1.5 trainings. is the problem that you want something for your use? or to sell to a fickle customer base?

[pseudo owl]

Sana did not come out yet, will come out later.

[split bramble]

ah

[sacred geode]

sell? LOL, I wish. I am not here to praise, nor bury SAI only be glad someone out here still wants to see a working model from them.

my hope is they don't screw it up again as they did with SD3 AND that it is not like Flux. Make it more like SDXL quality, ease of training styles in, and speed with the prompt adherence level of t5. If they can't do those there is always flux.

[craggy crest]

my hope is that whatever is released, we don't have another raid on this discord by elements that are here to just stir up trouble and see if they can kill the company. we'll see

[sacred geode]

That happened? Probably did knowing the base I have had to deal with. I left discords over it as so many became beguiled with Flux to the level of madness. The AI community is pretty sick and I started to see that with Pony.

I like to use Flux, or Hunyuan, to SDXL so xl can clean it up but they get the adherence.

[craggy crest]

that happened

[craggy crest]

try using sd3-2b-medium as a refiner for sdxl

[sacred geode]

I haven't touched sd3 since it first came out and the quality issues in the gens as well as the deformed people I deleted it and went right back to XL. Now, unhappily, on Flux. Like flux for text and conforming adherence but XL has the better quality.

plus the speed is horrid in Flux

[craggy crest]

here's a little known fact - because SD3-2b-medium and flux use the same neural net, and the same code - flux has the same issues as sd3-2b-medium. now ask yourself just what did black forest do to mask those and why is flux so inflexible?

[sacred geode]

yes, I want to know the first and I think the last one is due to dev being 85% of pro and distillation.

Schnell is even worse at 70% I think it was

[craggy crest]

i put the time in to learn both models - SD3-2b-medium is a fantasic model, but you do have to learn it from the ground up

[sacred geode]

I don't need to do time as a good model just works. Out of the box (never tried to train it as back then nothing yet could) it was bad when XL was next level in quality over it. Adherence SD3 was better but flux, or hunyuan, is much much better than SD3 for that. LyCoris said it is due to the t5 used for sd3 and hunyuan but I have not spoken with him about flux. Being flux is 12B t5 it would be why. btw, that t5 is also a curse as the more it knows the more inflixible it seems to be.

[craggy crest]

you do need to learn how the AI thinks. remember - both flux and SD3-2b-medium do NOT use unet, and thus everything you learned for previous models isn't going to work the same way

[sacred geode]

oh, and flux has major issues (out of the box even) with concept bleeding

[craggy crest]

oh flux has far more issues than that

[sacred geode]

a s tonne, I know

btw, BFL < SAI. All the bad devs I cursed at SAI seem to be BFL now. just an fyi

[craggy crest]

bfl are the original eluther researchers. nothing wrong with any of them

[sacred geode]

if you say so, but I have my opinion of them. A trainer dev asked them about how flux is just not trainable as it should be and they said they have no want to make it more trainable. Pro is why.

[craggy crest]

it isn't trainable and they can't make it trainable. it's frozen

and pro is not why

[sacred geode]

well, they haven't even released the shift value used while SAI was open about it

[craggy crest]

they can't.

[sacred geode]

yet SAI could? why not?

[craggy crest]

that's something i can't talk about. but they can't release that, or much of anything else. be happy flux makes you pretty pictures if you stay in its narrow range

[sacred geode]

As I said I prefer XL over Flux for pretty pictures but the adherence goes to Flux hands down.

[craggy crest]

i walked through flux's latent space, one token at a time, after it released. for a full month, 12 to 16 hours a day. it has an incredibly heavy bias toward: women, dogs, fantasy, and anime cat girls. if you give it a term that has an unclear subject, instead of getting random subjects that depict the term, you'll get one of those four things, extremely well drawn, and frequently a mixture of them. and even in a lot of cases where the subject is clear, out of a run of 4, three will be women or dogs

so of course it does pretty pictures

it has no choice

[sacred geode]

Flux is heavily laden with big eye Anime.

[craggy crest]

anime cat girls

[sacred geode]

probably, but I never prompted for that

even in training it everything wants to turn into anime and I do realism

[craggy crest]

you don't have to prompt for it, it's so heavily biased toward it, you're gonna get it anyway

[steel beacon]

Okay all!!! This is fantastic. It's a new FLUX LoRA that makes it Turbo. 8 steps for a fantastic image. It's Alpha, but totally worth testing. https://huggingface.co/alimama-creative/FLUX.1-Turbo-Alpha/commit/b2db8dcbd15fb095cffd8ab530499e47883466e7

[sacred geode]

pass

[craggy crest]

here's something for you to try. pull up flux dev, put in just the word pail for the prompt, no negative prompt, generate 10 times, see what you get

[sacred geode]

I do cfg 1 so can't negative prompt it

[craggy crest]

okay good.

[steel beacon]

It's really an amazing LoRA. I just found this earlier but I put it through the paces for a few hours.

[craggy crest]

read this channel's name. sd3. go post that in general chat

[sacred geode]

pail you say?

[steel beacon]

I asked CS1o where I should post FLUX info and was told that sd3 was sort of combined with FLUX and to post here. But alright.

[craggy crest]

just the word, pail

[sacred geode]

just one word?

k

[craggy crest]

you're gonna get a lot more people seeing the link in general than in here.

[sacred geode]

oh, that ugly 16x16 grid?

[craggy crest]

?

[sacred geode]

the 1024sq gave me a blond woman

[sacred jewel]

Flux.1 Dev

[craggy crest]

that's actually a cool texture effect for certain situations

[sacred geode]

yes

well, pail gave me all kinds of images

that is just two of them

[craggy crest]

did you use the same word for both encoders?

[sacred geode]

oh, I only use t5 the L I leave blank

[craggy crest]

run it again and this time put pail in both of them

[sacred geode]

k

LOL, I got a pail

[craggy crest]

run it 10 times, and see if you get anything that's got french fries

[sacred geode]

with my loha

so weird

not yet, but a lot of nice pails

none yet

mostly just empty pails

[craggy crest]

but you are getting pails. it's a concrete, solid, subject. but what if you give it a term that has no clear subject? switch teh word pail for both encoders with Hand-drawn and see what you get - as it's an unclear subject you should get a large range of images from landscapes to who knows what - run it 10 times too

[hallow lion]

The pail master

[sacred jewel]

First seed

[sacred geode]

I told it something for a prompt to test my training and it was not what I expected. I have noticed flux falls into holes.

[craggy crest]

go try the next test. use Hand-drawn for both encoders. just that term. 10 times

[sacred geode]

if you expect it to think outside its box it will do random weirdness to off thewall stuff

[sacred jewel]

second seed

[sacred geode]

Hand-drawn so it is asking hand-drawn what. fill in the blank. It pops a fuse

[craggy crest]

maybe. run it 10 times, post the images

[sacred geode]

hand drawn first seed

there us 5

*is

[craggy crest]

so you got - dogs - 1 that's a dog and 2 that are dog-ish/fantasy creatures - and one anime cat girl

[sacred geode]

I can't post 6

nsfw

[craggy crest]

and that gives you a glimpse of what was done to mask the issues flux has and why it's not trainable

[sacred geode]

wdym?

[craggy crest]

and why BFL can't release training code

[sacred geode]

just a shift value as we are currently using SD3's of 3.1518 I think it is

[craggy crest]

you should have gotten quite a range of subjects, all in a hand-drawn style. you got dogs, women, fantasy, anime cat girls. with one dog being just a dog, and the others a mixture of dog+fantasy creature

[craggy crest]

remember that flux and SD3-2b-medium ARE THE SAME THING

except after robin left SAI, he did things to flux, before releasing it

[sacred geode]

10 as 6 I could not show

[craggy crest]

now try using hand-drawn as the first term of a longer prompt

[sacred geode]

for l and t5 again?

[craggy crest]

for testing stuff, yes, give all the encoders either the exacxt same thing, or prompts crafted for their strengths. don't leave one blank unless you want to fluff the result with random dice rolls

[sacred geode]

this one is cool

[craggy crest]

now try it with your lora

[sacred geode]

WOW

[craggy crest]

interesting lora you made

[sacred geode]

a different training

[craggy crest]

try just the term Javelin with both encoders

[sacred geode]

here is a different training I was working with and this one is really nice

[sacred jewel]

What was I to expect?

[craggy crest]

when testing something, you shouldn't expect anything. those came out nice, except it looks like you got a dead carcass in that top one

[sacred geode]

javelin

[craggy crest]

is that with or without your lora?

[sacred geode]

without

[craggy crest]

what do you get, with?

[sacred geode]

let me try

with my lora

[sacred jewel]

I meant, what were you trying to prove with your test... I am not sure I understand what my results mean... as far as your point of "flux is x, y and z, etc."

[craggy crest]

what did you train that lora on, anyway?

[sacred geode]

death, destruction, corpses

[craggy crest]

oh, take a look at the content. is it all over the place? or is the range of random subjects fairly narrow

[craggy crest]

seems to have given you that - don't see any javelins though

[sacred geode]

probably on a dead knight somewhere, lol.

[sacred jewel]

hmmm... I guess it is related? I am not that intuitive 😛

With my LoRA on, no trigger

[craggy crest]

see what you get with your lora and just the term Harmonic

[sacred jewel]

Same prompt (hand-drawn) with my LoRA, new seed

[sacred geode]

okay, I was right these loras are overtrained but I was testing the trainer. here is 0.5 weight (the above is 0.8 where I normally run them)

[craggy crest]

now that's really nice

[craggy crest]

that's a very strange gun he has

[sacred geode]

I did not train on guns

Not sure where it came from

[craggy crest]

neither did he, apparently. he's not likely to survive shooting that

[sacred jewel]

My LoRA, and "harmonic"

[sacred geode]

LOL

[craggy crest]

sort of a drone-ish thing

[craggy crest]

try the short phrase Harmonic scale

[sacred jewel]

I guess... not sure how to interpret. My training was heavily specific to a topic of spy cameras and spy tech and a 1950s theme with a muted color palette...

[craggy crest]

you got wires at least

[sacred geode]

here is with my lora at 0.8 and I did not use the activation word.

[craggy crest]

if you don't use the activation word, it's not supposed to use the data in the lora

[sacred jewel]

With and without the LoRA enabled (prompt: harmonic scale)

[craggy crest]

the second one looks like soundwaves - not sure what the first is an attempt to draw

here's a longer prompt with a very unclear subject: delicate cosmic quilling, points on a curve, waves, ripples, crescent wings, epic maximalistic nebula

[sacred geode]

it does, even on XL, but at a massively reduce rate. calling it is akin to 10x weight, or higher just depends. I have one lora I released that did have an activation word but I never released the word because it made the effect too heavy even at 0.1 (for XL).

[sacred jewel]

Trigger word added

[sacred geode]

why did sd3 make Cronenbergs while flux normally doesn't?

[craggy crest]

not sure what you're asking

[sacred geode]

Rick and Morty reference

[craggy crest]

black forest did things to flux after they left SAI - part of which is the so-called distilling they did

[sacred jewel]

With and without LoRA, same seed

[sacred geode]

which is why I am having all kinds of issues training it for styles

[craggy crest]

because flux is not trainable.

[craggy crest]

which one do you like better?

[sacred geode]

disagree, as people are training it for likenesses with ease. concepts bleed too much and styles are hit or miss

[craggy crest]

they are creating small models that do provide some sort of image generation direction. but that isn't the same thing as actually being able to train a lora to update the weights of the model

[sacred geode]

how are they getting the various ages of Cheech and Chong to perfection?

[craggy crest]

They are creating an object that is used during the generation process with information for that specific thing. it's not updating flux's weights however, i'ts just providing data for that specific thing

YOU are trying to train a style, YOU need to actually update weights

and YOU are hitting walls

[sacred geode]

big time. though it did help to unfreeze the t5 it lead to massive bleeding

especially background stuff

[craggy crest]

you are hitting what are essentially titanium walls. flux is not trainable. it's frozen. and most of the loras out there don't do anything that a well crafted prompt for flux wouldn't do

try this prompt: 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

[sacred geode]

why did they lock it down like that so we can't really train it? Now I know you are right because I was seeing that what I trained an LLM could create with a prompt. A lot of the people I mentioned earlier who are in a spell over Flux would fight with you over what you said, but I know otherwise.

Flux spell never worked with me, and I lost a lot of people I knew over this as they refused to accept reality.

the inferior quality they were willing to accept was the final nail

one trainer I work with remain in my camp but they have been fighting everyone and lost more people due to this Flux miasma

[craggy crest]

that's something you'll have to ask black forest

[sacred geode]

I have my suspicion called potential lawsuits. I was told by an insider about these devs at BFL as they were the ones complaining about the money flow, etc... SAI. This is why I suspect pro for cash and to prevent potential lawsuits while the tech flounders in general.

[craggy crest]

the code is open source. lawsuits are not an issue

it has everything to do with the core issues of the neural network they decided to use

[sacred geode]

well, even SAI was being sued and even if you win the resources to fight sucks.

you win but still lost. Courts

[craggy crest]

the black forest team developed the code. they'd have to sue themselves

it's open source. everyone has the right to use it

[sacred geode]

No, I mean artists, etc... suing as they did to SAI, or were about to and why emad rushed 2.0 with less artists in it.

and liknesses. the connections could be trained back in though

[craggy crest]

sure, and the artists lost because they're entire lawsuit hung on somethign that wsan't true. however, the only people that could sue black forest over flux would be... black forest

and that would be silly

[sacred geode]

this is all unsettled territory and until settled China wins as they don't give one pee pee about IP, or liknesses, or much of anything.

[craggy crest]

please understand something - the BFL researchers created the code. it's their design, their work, and if anyone has a right to it, it's them.

but it's open source so no one can take anyone to court for using or modifying it

open source is not uncharted territory

[sacred geode]

you are not grasping what it is I am saying

[craggy crest]

i am. lawsuits are not possible. only a stupid lawyer would take the case and a judge would toss it out within the first 5 minutes

[sacred geode]

SD is open source as well but SAI can get sued for scraping, or having actor likenesses in it, etc...

[craggy crest]

you can sue anyone, for anything - but it's expensive so most people dont' do that unless they're pretty sure they can win

[sacred geode]

this is why they made that opt out deal

[craggy crest]

it's not

that's spawning's work

[sacred jewel]

With...

[sacred geode]

well, if I had elon musk money, or Google, or Sony, and went after BFL guess who wins? draw it out for an enternity which is why I hate the legal system as the one with the most money 99.97% of the time wins cause you can't afford to outlast them. Out of court settlements etc...

[craggy crest]

black forest wins. They. OWN. THE. Code.

they created it. it's their work. they own it. however it's open source so you can't go to court against anyone that uses it

[sacred jewel]

Without LoRA and With my LoRA

[craggy crest]

they're both not bad - but the smoke's supposed to be different colors

[sacred jewel]

Blue Future LoRA

Dark side of the Future LoRA... same prompt

Dystopia LoRA

[craggy crest]

try this 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.

[sacred jewel]

Big impact LoRA

[craggy crest]

now that one worked right. the smoke is three different colors and it looks good

[sacred jewel]

Without / With my LoRA

[craggy crest]

i think i like the larger spider better

[sacred jewel]

Yeah you're right..

[craggy crest]

these are sd3-2b-medium

comfy workflow is in them if you want it

[noble coyote]

Flux + LoRAs

[narrow epoch]

Blue Future LoRA

[proven pecan]

Looking at the artisan channels and still nont impressed. Really disappointing.

[pseudo owl]

Sana demo: https://ea13ab4f5bd9c74f93.gradio.live/
Sana github: https://github.com/NVlabs/Sana (no code yet, vae out, model out too?)

It's actually pretty nice, not flux level but great for the size. It has great prompt following, surprisingly decent humans(not flux level, around sdxl I guess?), and nice text capability. Its undistilled and supports negative prompts.

left is Sana(looks like photoshop lol), right is flux. Sana got everything right, 4 cow pictures, ufo out, cat on couch. Flux got cat and ufo I guess but not space nor 4 cow pictures
prompt: A highly detailed photo of a white cat sitting on a brown couch in a living room. Behind them, is a window, and 4 cow pictures, one in each corner of the window. Outside the window is mysterious space and a ufo.

Can do woman in grass too

[pseudo owl]

You'll see that Sana isn't actually too great with text, can only do small phrases(like the above). Apart from that, its actually pretty nice.

[turbid grotto]

WHAT

[dull star]

if this becomes apache 2 this might replace 1.5

its so small and doesn't use a large encoder like T5-XXL

and low step count needed

[sage burrow]

Uhm? This is what my extremely detailed description of 2 characters produced 😦

[turbid grotto]

demo overloaded

[hallow lion]

is it out? What? comfui ready?

[turbid grotto]

no code yet

[hallow lion]

is it good, whats happening

oh 😦

[turbid grotto]

but something is happening

[hallow lion]

yeah i get that feelin

[turbid grotto]

It is 1.6b and 0.6b, so it is more like competition to sdxl, sd3m and sd1.5

[hallow lion]

probably not a "flux killer" like the sensationalist clickbait-y youtube crowd depicts it but I'm excited. come on youtubers chill out

every day a "sora killer" or some other killer... come on now

[turbid grotto]

yea, they make false expectation about everything

[cunning lintel]

well, all it did for me is not taak any youtube crap seriously, don't think i've looked at non instructional vids (it's nice if you need to do some weird diy thing you never done before) there or years

[turbid grotto]

same, but a lot of non technical people have no choice

[hallow lion]

Remember when we were excited for 512/512 text to image XD

fk that sht! We need 4k now, and not with upscaling either!

[cunning lintel]

the very first midjourney was magic!

and those images look sooooo bad now, but for the longest times i was happy MJ could do 'a scary sunflower with a face"

[turbid grotto]

for me it was sd1.4, and I was shocked

[cunning lintel]

i think it was at the same time, i had that sunflower in mj, and sd couldn't do it 😢

[turbid grotto]

Sana:

seems like it has bigger color range than sdxl

[cunning lintel]

Style: A photograph with a red Cinestill-inspired aesthetic, shot on a Nikon D850 with a shallow depth of field, emphasizing the subject's haunting, android-like features. Highly detailed, with fine textures and intricate details, evoking a sense of unease and intrigue. Scene: A small, battle-scarred mech, reminiscent of H.R. Giger's biomechanical art, with glowing, crimson eyes, holding a vintage Barbie doll in its long, slender fingers. The mech's skin is a deep, iridescent blue, with intricate, cybernetic implants. The Barbie doll, dressed in a flowing, black, Victorian-era-inspired gown, seems eerily out of place amidst the futuristic surroundings. The mech stands in a vast, desolate, post-apocalyptic landscape, with crumbling structures and overgrown vegetation in the background. The sky is a bruised purple, with wisps of smoke and dust swirling in the air.

pixart

[turbid grotto]

hm pixart is better there

btw demo doesn't work anymore

[dull star]

yeah sadly

it's just overly aesthetic

there's no image without bokeh unless its like a landscape shot or 2D/painting

[turbid grotto]

I also got just heart but for explicit prompt, maybe they gave specific instruction to text encoder for this demo

[sacred jewel]

Anyone know the comfyui extension that lets you delete a group and all the nodes in it in one shot?

I started a fresh install and I can't find that custom node package 😭😭😭

Thanks 🙏

[lucid swift]

this model might be good for the comunity because its so small. people will be bale to train it very easaly locally

[lucid swift]

sdxl is 6.6 b paramtere with the text encoder. but only 2.3 without it

1.5 is 890 million parameters but with clip its 1.47b parameters

[turbid grotto]

6.6b is combined with refiner too

[lucid swift]

yes its not a flux killer. here is a ideogram vs flux vs Sana prompt "A photo of a fursuit of a fox. The fox is wearing a blue and orange striped shirt and a pair of glasses. The fox is standing on a wooden platform. The background is a forest."

[lucid swift]

oh

sorry then i was a bit wrong because nobody uses that xD

i think ideogram is still often better then flux

[turbid grotto]

wow, ideogram looks super real

[hallow lion]

the first one is phenomenal minus the questionable hand situation

[lucid swift]

yes ideogram is good

[hallow lion]

The Sana is just like aggravated...

😠

[turbid grotto]

ran same prompt with sana and got more adherent results first try, however it is not fursuit

[lucid swift]

but i think like ideogram is the only model that has good furry stuff in there training data. so the other models just have a dataset problem

[lucid swift]

maby my prompt was bad idk

[hallow lion]

Well we can always try to make a furry lora for flux

[lucid swift]

training flux is very hard. but a lora is realistic

the cool thing for sada is that training will be easy

we will be able to make full finetunes locally

[hallow lion]

flux gym is faily easy

i made some celebrity loras on it

they looked good

[lucid swift]

with what gpu?

[hallow lion]

12 vram 32 ram

[lucid swift]

daim thats good

[hallow lion]

yeah

took about 6 hours or so

[lucid swift]

daim

[hallow lion]

ye lol

its an overnight deal

[lucid swift]

i think sada is bigger then 1.5 but only slightly

why are ur messages disapering

[turbid grotto]

yea, and it is smarter than sd1.5 \ sdxl, however flux noticably smarter

curious how 0.6b going to look

[lucid swift]

yes but flux is like 12 times as big

i thikn the 1.6 is the only realistic one to use

[hallow lion]

Nvidia knows their stuff so it must be decent

[lucid swift]

the 0.6 is just to small

its from the pixart alfa team they got higherd by nvidia

[turbid grotto]

fair, hope community won't ignore this model

[hallow lion]

u never know 😮

some models are embraced some are ignored

[turbid grotto]

but this one has a chance, as it is fast and light, also encoder is lighter than T5

[hallow lion]

Rooting for it

[turbid grotto]

I want to try sana with tensorrt and 8step hyper lora, it might be 3x speedup

[lucid swift]

yes and its good that its not useing t5. t5 is old. models shuld use modern llms

[lucid swift]

this model is alredy very fast. i think the demo is just slow because a ton of people are using it

[turbid grotto]

this demo just is unbelievably slow, can't Nvidia allocate some more gpus for this super fast model?

[lucid swift]

but i think the really next big thing is this one. but it might be to hard to finetune. https://arxiv.org/pdf/2409.11340 because its 3.8b parameters. its based on the phi mini llm. its a gpt style image model. it can do some very cool stuff

[lucid swift]

they can but are not lol

[lucid swift]

it can do stuff like this

[turbid grotto]

Oh yeaa, I am very exited about this

[lucid swift]

btw i found a demo of it on hugging face

[turbid grotto]

they should release model this month, according to their comment

[turbid grotto]

gimme it

[lucid swift]

but its qute slow. be carfull i think you can only runn it like once or twice https://huggingface.co/spaces/Shitao/OmniGen

and put the right image name toknes into the promt like deskibed in the space

[kind acorn]

画一张吸血鬼

[lucid swift]

zecihene einen vampier

[lucid swift]

if you look at the botom of the image you see that most of the time waiting is just beeing in a queue

[turbid grotto]

for me it is just "processing" and quite long

[turbid grotto]

Thanks btw!

I ran out of quota, did not even test with references 💀

but that demo is great sign, seems like model is at final stage

[lucid swift]

i said that you can only test it only onece or twice 💀

[lucid swift]

yes!

imageine if they make it work iwth lama cpp

that would be huge

[turbid grotto]

I missed it

[turbid grotto]

I am not into llm much, what would it bring?

[lucid swift]

llama cpp is a very optomised way of running llms. it can run on the cpu but its still very fast

so it would work like on every device even if it doe snot have a nvidia gpu

[lucid swift]

wow sama canfollow very long i asked chat gpt to deskibe the furry picture from erlyer and used that as a prompt " The image shows a person dressed in a detailed fox fursuit. The fursona has bright orange fur on the head, arms, and body, with white accents on the snout, chest, and tips of the ears and paws. The character has black fur on the edges of the ears, as well as large, exaggerated paw pads in black on the palms. The fursuit features large, expressive eyes covered by a pair of blue, semi-transparent glasses, giving the fox a playful and intellectual appearance. The character is wearing a casual blue-and-orange striped long-sleeve shirt, matching the fur's orange color. The fox has a cheerful, welcoming expression, and its paws are raised toward the camera, as if greeting the viewer with an enthusiastic wave. The background appears to be an outdoor forested area, with many tall trees visible, though they are slightly blurred. The image is set on what looks like a wooden bridge or platform, and the overall mood is lighthearted and friendly.pts but i odnt like the style "

[turbid grotto]

this model in particular would be great on cpu, especially with low step lora

[turbid grotto]

I wonder why most models has so little knowledge about fur suit?
Maybe majority of images are amateur photos from various contests or meeting, so they are being filtered out due to not meeting aesthetic score?
Just assumption...

[turbid grotto]

I finally made Meissonic 1b work but it is not much faster than SDXL. 13-15s for 1024px with 20steps, however it needs more steps for good quality

[muted dove]

Pirate influencer

[wicked roost]

画一张吸血鬼

[arctic birch]

create a dog with a knife

#🌠｜show-and-tell create a dog with a knife

[noble coyote]

Can I get SANA for ComfyUI yet?

SD3.5 anybody? https://huggingface.co/stabilityai/stable-diffusion-3.5-large

[bitter hearth]

wow awesome

[frosty berry]

Awesome

[steel helm]

eh prompt adherence def isn't sota. gotta go work. i'll play later

[bitter hearth]

can use as a second pass over auraflow

[steel helm]

nothing to see here is my first impression, i'll give it a chance later

[compact forge]

GANG GANG

[dusky thistle]

wooo!

[gilded silo]

the taste of vindication is so sweet

"all good things come to those who wait"
what did we wait for then

[coral sable]

omg hype <333

so excited

[past thicket]

Not blown away yet...need to play with it a bit more

[eternal ermine]

What are min requirements to run the new 3.5?

[cobalt moon]

Look like 12GB is a safe bet

[eternal ermine]

Ok, tyvm 🙂 Hopefully i will get that one day xD

[rotund stump]

I love the tongue in cheek of choosing girl laying in grass in the announcement post

[eternal ermine]

so 12GB is for that 3.5 medium model?

[errant dust]

I didn't catch that. Well spotted!

[cobalt moon]

12GB is for Large

[past thicket]

SD3.5 Large

[eternal ermine]

oh, that is much better than anticipated

[errant dust]

Bah, we know that with Comfy it won't be stuck to that

How is it for text and human anatomy. I don't mean clothesless, just not deformed

Regardless, super pumped

Flux had become the defacto monopoly on open models by sheer size and quality. Now there is competition

[bitter hearth]

this model is perfect its everything I was waiting for personally

[errant dust]

Is it supported by comfy? Meaning do we need to wait for any special modifications or will it use the standard SD3 nodes?

[bitter hearth]

comfy discord announced day 1 support

you might have to do an update but it will work

[past thicket]

It's killing my 4090 right now, hehe

[compact forge]

its running on my 4070 12gb

[past thicket]

Wide angle shot of a beautiful women holding a sign that reads "SD3.5 Large"

[coral sable]

should use hugging face workflow or is there better

[bitter hearth]

diffusers is fine

comfy also

the turbo version is looking pretty strong also

[wide pagoda]

I like weird shit like this better than when it outputs "good" art

[coral sable]

https://www.reddit.com/r/StableDiffusion/comments/1g9itzj/sd_35_large_released/

[wispy epoch]

SD3.5 anime

[brittle nexus]

[SD3] Nice tires, dude

[compact forge]

sd3.5 confirmed good? 😮

[bitter hearth]

yeah look at that red car

photographic quality as good as the best flux image there

[brittle nexus]

It needs to do more testing but it has the potential to be better than FLUX as it already beat FLUX in some aspects even when broken. And it is more trainable.

[noble coyote]

For the smaller VRAM user, SD3.5 Medium drops 29 October

[past thicket]

I'm not getting anything close to what i get from flux

[compact forge]

i mean SD3 was already better at photorealism than flux, so it is just a matter of anatomy+consistency

[little smelt]

Stable Diffusion 3.5 Large Fine-tuning Tutorial from SAI :
https://stabilityai.notion.site/Stable-Diffusion-3-5-Large-Fine-tuning-Tutorial-11a61cdcd1968027a15bdbd7c40be8c6

[patent acorn]

sd3.5 failed the horse-fish hybrid test

[bitter hearth]

how do you pass the horse-fish hybrid test though

[old falcon]

How much vram is needed per tier

[patent acorn]

the image must have the horse head and the body of a fish

[lucid swift]

*enis

xD

[patent acorn]

sosig

[lucid swift]

wow what is the prompt

[compact forge]

selfie photo of 18yo woman at the club, red head, pale skin

[little smelt]

weird anatomy

[patent acorn]

someone try this:
"in the ocean, there is a fish creature whose is the head of a cat and the body of a fish"

[patent acorn]

wait i did it

[bitter hearth]

nice

[compact forge]

2 cfg

[patent acorn]

NOO i exceeded my wuota

[winged flower]

Just type catfish 😆

[whole ferry]

lmao came here to say that, that's great

[patent acorn]

thats a different thing..

[rotund stump]

Now make her laying in grass

[sullen moss]

((

[compact forge]

ban this guy

[lucid swift]

xD

[lucid swift]

😼

[deft niche]

Sorry, something went wrong while trying to process genre. Please try again later.
Reference id: PSB8GUSY

[rotund stump]

Ok, so far so good

[lucid swift]

sorry wht are the corect cfg /step values again?

[past thicket]

https://www.youtube.com/watch?v=en-GMBIa-N8

[deft niche]

The Last.fm user GinesRico has no listening history on their profile yet.

Just created your Last.fm account? Make sure you set it to track your music app.

Using Spotify? You can link that here. This can take a few minutes to start working.

Please note that .fmbot is not affiliated with Last.fm or Spotify.

[compact forge]

selfie in grass

[winged flower]

Try full body shot in grass, curious about her limbs

[compact forge]

thats what i did but i also put selfie in it so it kind of made it large FOV to get as much of the body into it as possible which is kinda cool

[wild coral]

Anyone know if we still need this node for 3.5?

[bitter hearth]

yeah

[winged flower]

Pretty good

[compact forge]

yes it is in the example workflow

[compact forge]

with more steps (im doing 20) it might fix that hand too xD

[winged flower]

Maybe "full body starfish pose in grasss" lol

[lucid swift]

no i dont see it in the example workflow

[whole ferry]

is this 3.5 or 3.0?

[wild coral]

I just tried gening 2 images, same seed, one with that node one without. No difference. It doesn't seem needed

[dull star]

https://tenor.com/view/streamer-wow-shocked-amazed-yes-gif-17334640

[compact forge]

then you dont have the right one

[dull star]

IT CAN DO PAINTINGS

[coral sable]

its so good

[wild coral]

I just tried gening 2 images, same seed, one with that node one without. No difference. It doesn't seem needed

[compact forge]

3.5

[patent acorn]

the grass look smooth af

[lucid swift]

or you have the wrong one

[coral sable]

rip flux

[lucid swift]

naw competion is good and i am happy that both exist

[dull star]

I wouldn't count the chickens yet

[patent acorn]

in terms of license yes but uhh the dataset uhhh

[compact forge]

even from side it aint too bad

[wild coral]

I still prefer flux for that (This is Flux not SD 3.5)

[little smelt]

16 GB VRAM for Fp16, 8 GB VRAM for fp8. Not bad, not bad.

[winged flower]

Haha so good

[compact forge]

lets not prompt for starfish xD

[coral sable]

works smooth on 4060ti 16gb, idk about settings tho

[winged flower]

No more starfish, that'll give me nightmares

[lucid swift]

these are mine

[compact forge]

ok let me try with some more elaborate prompts

[dull star]

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

[wispy epoch]

hahah the eyes

[coral sable]

cool

[dull star]

idk how they messed it up cause the large model could actually do laying down women apparently

[fair gate]

How's SD3.5 compare to midjourney?

[compact forge]

eh not as good 😛

[dull star]

I bet this is the only thing people will post on subreddit for a week straight

trust me

[compact forge]

BRO 3.5?

[lucid swift]

yes

[lavish osprey]

it's advised

[lavish sparrow]

sd3.5 not that bad, i suppose?

[dull star]

if it proves to be good at paintings its gonna be my secondary model that's for sure

[lavish osprey]

models usually have trouble with upside down and sideways faces, so they're overfitted on not doing them. SD3.5 will attempt to do them and mostly succeed

[fast tapir]

Incredible, but the bodies shown so far are a little worrisome

[dull star]

flux literally cannot make paintings without a lora or very low guidance scale

[lilac bridge]

Always going to try to make the good boi UniPC work (34seconds on 3060 12g vram)

[lavish sparrow]

never used sd3 for bodies either xD

[wild coral]

I think it'll come down to how easy it is to finetune, as it's main competition Flux has been notoriously difficult in that regard.

[dull star]

8B vs 12B

[wild coral]

yupe

[lucid swift]

these giant models are very hard to train

[dull star]

sana is smart and very small

its the first decent yet SMALL DiT

it might overshadow 1.5 if finetuning proves to be good for it

[dull star]

but by default its no better than pixart (Sana)

[lucid swift]

what ahhaha

[dull star]

Sana, have you not seen it?

[wispy epoch]

anime

[lucid swift]

please look at hte sana images again they look melted

[dull star]

even worse than pixart?

[lucid swift]

what prompt

[dull star]

I didn't look that close tbh

just a bunch of examples and moved on

[wispy epoch]

~~aesthetic~~ anime 1girl, detailed eyes, barefoot, feet focus, sitting, kitsune facepaint, pale skin, bikini, white long hair, red eyes, white fox tail, ninetailed fox

3.5 large model

[lucid swift]

sana is not good with details

[dull star]

it did look smooth/artifacted that's for sure

[lavish sparrow]

nevermind me, just limit testing sd3.5

[wild coral]

Ma'am, I believe your head is on backwards.

[dull star]

Oh it looks quite good. Is 3x faster than flux dev for me and it also seems to be capable of anatomy and some nsfw from the get go

someone wrote on the subreddit

[lucid swift]

wo this looks cool

[dull star]

oh I like that

ghosts of people

[lucid swift]

this was the prompt "A cursed, low-quality image inspired by disturbing found footage, featuring grainy textures and corrupted digital artifacts. The scene has a dark, eerie atmosphere, with unnatural, warped figures barely visible through static and glitchy interference. The clouds in the background are shadowy and fragmented, adding to the unsettling feeling. The text "Embrace the chaos inside your mind; it makes the clouds blush" is overlaid in a broken, distorted font, blending into the eerie and glitchy visuals. The overall effect is strange, off-putting, and deeply unsettling"

[wild coral]

normal 3.5 Large is about the same speed as Flux Dev for me, haven't tried the turbo yet though

[lilac bridge]

The seconds per iterations hurts so much xD

[lavish osprey]

wait until people start finetuning it. There are no limits. In one of my "tests" I tuned this for 120k steps at batch size 8 on 100k images and it didn't break

pretty high LR too

[past thicket]

I'm looking forward to some nice fine tunes for sure

[lucid swift]

cool

[compact forge]

the aesthetics of this image is insane from such a simple danbooru tagged prompt

[lavish osprey]

this is for you. Free and commercially available, easy to tune. No DPO (so hands might be wonky at times), but we might release a dpo tune in the future as a lora plugin, if the community doesn't make it first

DPO would have made this harder to tune

and also restrict the vector

limiting range and styles

[compact forge]

It is perfect lycon!

[bitter hearth]

I noticed that it said first party control nets are coming
that sounds like great news because control nets are so expensive to train

[lavish sparrow]

cool swords ❤️

[lavish osprey]

1 prompt 1 image 1 pass

[bitter hearth]

the grass is good its very sharp

[lucid swift]

its good at anime :§

[mortal mesa]

i forget is there preferred sampler/scheduler steps ect

[bitter hearth]

will do fine with DPM++ 2m, DEIS or UniPC, SGM uniform with shift

[raven fern]

she can kill me any day

[icy coral]

Very nice one, that's from a single prompt with no inpainting?

[lucid swift]

wow cool that both styles work

[raven fern]

new elden ring boss 😮

[coral sable]

some framing artifacts on 1344x768, but damn is so good

[lavish osprey]

yes, prompt by @fluid pecan

[raven fern]

im busy now, will try sd 3.5 later, looks good so far

[lucid swift]

i like this one

[raven fern]

i also hope they release Sana soon, looks interesting

[bitter hearth]

I'm excited about Sana yeah, it looked better than SD 1.5 base and is fast

[fluid pecan]

same prompt better image 🙂 was it done on turbo?

[lavish osprey]

no, regular

I just replaced Elsa with Rapunzel in your prompt

[fluid pecan]

dpmpp_2m?

[rancid dawn]

do you guys have a workflow i could use iwth comfyui for the new thing?

[icy coral]

A bit of style bleeding on the hair but at this point it's just cherry picking, I don't think I've seen flux get a similar concept either

[fluid pecan]

Look at this

[rancid dawn]

this is insane i love it

[fluid pecan]

perfect transition

[noble coyote]

Three-kneed beauty! "Mrs Jake-the-Pake!!!"

[bitter hearth]

that looks even better than latent couple / attention couple

[lavish osprey]

can I steal it?

[fluid pecan]

you must

[lavish sparrow]

sd3.5 is better at grasping concepts than flux is, it seems

it did the ethereal swords, and cracked face better than i got it ever running with flux

not to mention, the amethyst actually looks like amethyst now

[errant dust]

That's super cool. Can you share the prompt?

[icy coral]

At this point I gotta ask, what's the holdup with the hands compared to models like Flux? Is it the smaller size or is it the base model being undertrained?

[fluid pecan]

a scene that blends photo realistic and anime elements. A photo realistic man is taking a selfie with an anime woman. The scene is in a cafe, early in the warm spring morning. The photo is candid and dreamy. Everything is realistic except for the girl. The mix of reality and the anime girl makes the scene surreal but happy

[signal shuttle]

wow they really released sd 3.5 large, i thought that this day might never come

[coral sable]

was worth the wait

[noble coyote]

Its good - but I couldn't tell 3.5 or Flux?

[little smelt]

Yes, we've become completely spoiled. These days before, if you didn't get psychedelic stuff, you were already happy) but now - "the skin texture isn't worked out enough" 🙂

[lavish sparrow]

people would pay money for this!

[sacred jewel]

Flux with my LoRA

[bitter hearth]

ah yeah the hidden camera on the right LOL

[signal shuttle]

There are loras already for SD 3.5 large made by Shakker-Labs

[bitter hearth]

can't tell if Shakker got early access

or was just fast

[rancid dawn]

im sad was thinking the turbo would had been faster so imma use the large instead

[noble coyote]

Searge Flux LLM

[lavish sparrow]

classic 80's wizard multiclassing action!

[lilac bridge]

Is there a list of compatible samplers for 3.5?

[sage burrow]

Sd3.5 large 🙂

[noble coyote]

Wow!

[signal shuttle]

I think they got early access because they released the loras two hours ago, they couldn't have trained an entire lora in an hour

[fair finch]

golder sword

[bitter hearth]

acceleration loras like turbo are never really "faster" per step yeah they just need less steps

[turbid grotto]

SD3.5L has unique faces!!!

[rancid dawn]

i tried turbo my entire pc got frozen had to restart
i have like a 3060 idk why it not working

aw man

[cunning lintel]

🥳 New sd3 and the big one first too

[rancid dawn]

maybe i should wait for medium?

[tardy geode]

does sd3 use negative prompts?

[hallow lion]

oh shit it didnt even take 2 weeks since last time someone asked

XD

[turbid grotto]

I am running non turbo version with T5 encoder, everything at full precision on 3060 with no problems, it is slightly faster than flux, despite having slowdown from cfg

but 32gb ram is almost maxed

[hallow lion]

Alrighy then there goes another 16 gigs of hdd sapce

[rancid dawn]

hmm i see

what you using? the large one?

[pseudo owl]

Yes you can input negative prompts I believe.

[turbid grotto]

only the largest things

[rancid dawn]

gotcha

gotta try again

[lavish osprey]

Protip: Try init images

[pseudo owl]

It’s going to be very slow, full precision can’t fit on 3060, what’s happening is the model is leaking into ram. This makes it really really slow.

Quantization should fix this issue I believe. 8bit should be basically same quality while improving speed massively.

[turbid grotto]

It is very good at styles and diversity, like, a lot diverse than flux, but still has coherency problems, but I think it is fixable

[rancid dawn]

@turbid grotto on comfyui or stable diffusion?

[turbid grotto]

you mean comfy or a111\forge?
always comfy :3

[rancid dawn]

alright

im not used a lots of comfy

i either use inokeai or stablediffusion

cuz comfy seem hard to use

[turbid grotto]

same thing was for flux and I got no speed benefits from fitting model in to the vram

[turbid grotto]

you just have to use ready workflow

[rancid dawn]

ready workflow?

[pseudo owl]

You need q8 with some cpu offloading so basically offloading the text encoder, vae when not in use. It will still be massively faster than the model leaking into ram.

[turbid grotto]

https://comfyanonymous.github.io/ComfyUI_examples/sd3/
just drag image into comfy

[rancid dawn]

alright

[mortal mesa]

mmm Comfy is beta testing an all in one installer/app should help with peoples comfy woes

[sullen moss]

Testing complex scenes

[bitter hearth]

this is more of a hardware isue

[rancid dawn]

how so?

[silver bridge]

So what’s peoples opinion on SD 3.5 rn?

[bitter hearth]

the models are just grids of numbers they can't cause your pc to freeze

[rancid dawn]

i see

[signal shuttle]

a 3060 is slower compared to a RTX 4060 Ti 16GB or even a regular 4060 8gb vram

[hallow lion]

love the announcement image lol

[silver bridge]

Same XD

[vital moth]

is it normal for 3070ti, or im cook

[signal shuttle]

Fp8?

[turbid grotto]

very diverse, unique, styles, not overfitted, but anatomy needs finetunning

[silver bridge]

Yh I’ve seen anatomy needs finetunnng but the styles looks good

[vital moth]

I don't know lol, I just run it, I'm not a professional

[pseudo owl]

But luckily the anatomy isn’t too bad or anything, not flux level but still decent.

[silver bridge]

I’m just waiting for the dev of the app I use for MacBook to run SD to support the 3.5 large

[turbid grotto]

every face is different

[signal shuttle]

There are versions of SD 3.5 large for both 24gb vram cards and 8gb vram cards. see which one you downloaded

[mortal mesa]

defiantly very poor performance something not right, i get like 10x better on a 2080 ti

[turbid grotto]

stability solved 1girl

[silver bridge]

1girl laying on the grass XD

[vital moth]

uhh uhh ueah

[turbid grotto]

you probably using shared memory, you somehow have to make it not being used, try adding --lowvram argument

[noble coyote]

My 1st SD3.5L output

[vital moth]

Ohhh, yes, I remember it

[signal shuttle]

Are you using Forge?, i don't think Forge supports SD 3.5 large yet

[silver bridge]

Since it’s non distilled I can’t wait till I see civit Ai loras and models etc

[vital moth]

nope, I don't think im using forge

let me change some setting fast

[signal shuttle]

Then auto-1111?

[vital moth]

yeah, I don't know what is this lol

[silver bridge]

Tho I will say it’s a cool effect

[mortal mesa]

unresolved noise, needs more sampling

[turbid grotto]

try 5 cfg

and sgm uniform

[vital moth]

I need to fix my performance first lol

[silver bridge]

So SGM Uniform is needed, I’ll have to use 2m Trailing on MacBook when the app is updated

[turbid grotto]

I am not sure, it is just default in homfy that works

[signal shuttle]

To anyone trying to run SD 3.5 large locally, just use comfyui its way easier to get it up and running

[silver bridge]

I’ll be using “Draw Things” as the application

Cuz I cba using the others

I just need to wait for the dev Liu Liu to update it

For 3.5

[mortal mesa]

ide knock down resolution a bit too, to a 1MP size like 1280 x 720 or something

[vital moth]

I don't think it help, but I'll give it a try later

[signal shuttle]

3.5 large

[silver bridge]

The classic spaceman riding a horse

[pseudo owl]

Crazy day today! 2 new video gen models as well as sd3.5 large.

[sullen moss]

closeup testing

[dull star]

can't wait to test

if its really faster then I'm excited

especially if we get amatuer photography or similar loras

[dull star]

probably more texture than flux

flux has smooth faces

and im not glazing right now this is just how I remember

[turbid grotto]

it is a bit faster than flux because it has cfg, with hyper lora and cfg 1 it would be a lot faster than flux with hyper

[sullen moss]

more

[dull star]

thank you

[pseudo owl]

There is also already a sd3.5 turbo model which seems to supports cfg as well.

[turbid grotto]

oh yea! but I prefer 8 steps lora for full model

[noble coyote]

Super SD3.5L Workflow (after an idea by Olivio Sarkas) - the w/f is in the PNG

[turbid grotto]

O my god, look at this, no bleeding bitwin opposite styles,
from LatentVision

[pseudo owl]

Yeah now that’s a lie lol, if flux can’t follow a bit of your prompt, either your prompt sucks or is way too complex.

[obtuse kernel]

What's latentvision

[signal shuttle]

A youtuber

[dull star]

yeah I'm sus of that comment

[turbid grotto]

but he is right, flux really overfitted and far behind sd3.5L at diversity and styles, however better at anatomy, probably due to overfitting

[noble coyote]

matt3o

[pseudo owl]

Yeah that’s true, but I was talking about the adherence to complex prompts part.

[noble coyote]

matt3o made the IPAdapters

[turbid grotto]

okay, fair

stability released training guide

[vital moth]

yes

but I think I mess up with command arg

[lavish osprey]

Film style, realistic, perfect structure,Depicting a high-resolution movie still of Gundam fighting in a combat setting. This photo shows a real, intricately detailed Gundam robot fighting on the battlefield, with dynamic effects and glow effects. The Gundam is mainly made of metal and plastic and has a complex mechanized design. Its body is mainly gray with shades of red and blue, and it has a prominent orange face mask on its head. The robot lifted its left arm and held a huge, damaged shield with obvious wear and tear, presenting a mixture of red, white, and black colors. The shield seems to have undergone intense combat, with cracks and shattered edges that make it look rugged and worn. The background is filled with dramatic, fiery explosions, dark gray smoke and sparks, creating a strong sense of movement and chaos. The explosion is vividly depicted as a mixture of orange, black, and gray, adding a sense of urgency and conflict. The entire scene is set in a realistic urban environment, with neon lights in the background indicating a modern high-tech environment.

[noble coyote]

Super SD3.5L Workflow

[lavish osprey]

they stole from me ||that stole from @fluid pecan ||

[dull star]

for some reason the comfy workflow doesn't have the sampling

maybe its automatic

also this is def way faster than flux

this is like as fast as medium if not slightly slower

[lavish osprey]

the one on our HF repo has it

[noble coyote]

Super 3.5L w/f - getting striations!

[bitter hearth]

@lavish osprey i've been checking your profile everyday for sd3 news, thank you for this

[old pike]

can you please tell me what im missing: im getting: 'NoneType' object has no attribute 'tokenize'.

im getting this error in comfy

[dull star]

I have no idea

[old pike]

can you help me lykon please

[lavish osprey]

been super silent until today to work on the release, I should have time to be a bit more active from now on, but I might also take a small break for a bit

[old pike]

ive put all three clip models into the clip folder

but getting this error: im getting: 'NoneType' object has no attribute 'tokenize'.

[sacred jewel]

Is this a crop or the actual generation? You have prompt that can share?

[noble coyote]

wrong checkpoint/clip combination

[old pike]

thanks alot! can you tell me which of those i should keep for the sd 3.5 large checkpoint? please

[turbid grotto]

great model, like it's aspects, but what do you think about anatomy and how flux achieved it? Is good anatomy only possible due to heavy overfitting\overtraining?

[turbid grotto]

all of them! but you can use only T5 or only clips (without T5) to save time on prompt processing and memory

[old pike]

what im doing wrong then?

[noble coyote]

@lavish osprey's prompt - very dark (too dark?) striations again!

[old pike]

this is how my ui looks like

[dull star]

not bad

and it runs faster than flux!

[bitter hearth]

have you seen the new mochi model? https://huggingface.co/genmo/mochi-1-preview

[lavish osprey]

specific DPO for anatomy helps it a lot, but will restrict your model to the vector range of a lora and make it impossible to finetune

[bitter hearth]

open weight t2v model but you need 4xh100 to run it

[old pike]

but why im getting the error? ive choosen sd 3.5 large and all clip models are in clip

[mortal mesa]

use triple clip loader and load the G/L/T5XXL, dont use the clip from model

[dull star]

no but now I have, it looks cool

[lavish osprey]

are you using the workflow on our HF?

[dull star]

and it's apache 2!

@bitter hearth and it seems now you have access to sd3!!!!

isn't it epic

[old pike]

how please. thanks alot for pointing it out

[noble coyote]

I am using one I made myself - it is probably causing the striations!

[dull star]

you weren't lying

[errant dust]

I'm waiting for a GGUF build to be made so i can run it on my 8GB card

[noble coyote]

OK, I'll try this - thank you

[turbid grotto]

sorry, this is for comfy, I don't know how things are in other interfaces

[old pike]

im inside comfy

[mortal mesa]

double click blank area and search for triple

[old pike]

thanks!!!

[turbid grotto]

thanks!

Also, do we have to use ModelSamplingSD3?

[old pike]

my final question: where to connect this triple loader to

[lavish osprey]

that being said, if we did dpo now (or someone from the community made it), the model would still be trainable since you got access to the original distribution

[lavish osprey]

yes, 3

and 2~4 for training

[mortal mesa]

the 2 text boxes, just like it is now

[turbid grotto]

it wasn't in default workflow 😳

[lavish osprey]

on HF? yes. You talking about the comfy examples?

[noble coyote]

Better - brighter - no striations!

[bitter hearth]

assumed so, it wasn't in comfyui example workflow tho. Do you need the timestep for the negative prompt? i remember needing it for sd3m

[turbid grotto]

yes, comfy workfow from https://comfyanonymous.github.io/ComfyUI_examples/sd3/

[lavish osprey]

thanks, I'll tell @simple thistle and @viral plaza

well I just did

[bitter hearth]

nvm just realized, the sd3.5 large files have the right workflow

the comfyui default one is different

[turbid grotto]

what it does btw

[lavish osprey]

thanks for reporting it. For now use the one on our HF repository

[dull star]

yeah they def nerfed the painting knowledge of this

[turbid grotto]

okay!

[dull star]

but it's on me for spreading info about Zdzisław Beksiński

[old pike]

needed 1 minute to see that sd 3.5 is an upgrade, for one reason; TEXTURE

[steel helm]

2 generation setback in anatomy and adherence (on par with SDXL). only thing this has going for it is copyrighted works and styles that it includes really. still have to fix all the stuff that we already moved past. not impressed. I'll prolly skip it

[viral plaza]

Shift defaults to 3 so not needed normally. Swarm tosses it in anyway by default

The negative prompt timestep thing isn't needed it's just a trick that can sometimes help

[icy coral]

SDXL is nowhere close to it. Cope

[vital moth]

is my setting right?
it just work very slow with 3070 gpu

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--opt-sdp-attention --medvram --opt-sdp-no-mem-attention --no-half-vae --opt-channelslast
git pull
call webui.bat

or what else should I try

[viral plaza]

comfy's example workflow should be all good ref #🆕｜sd3 message

[lavish osprey]

I wasn't aware Shift had a default. Thanks!

[noble coyote]

I added Model Sampling as it was not in the ComfyUI* w/f

[old pike]

im getting dark images, did i messe dup a node anywhere?

[noble coyote]

Thanks for the heads up

[cunning lintel]

Each their own, not everyone likes or needs the same thing, i'm delighted not to see smeared smoothed images (dunno how to describe flux's style really)

[turbid grotto]

yes, wrong clips

[noble coyote]

Your clip should be l and g

[turbid grotto]

and T5 optionally

[old pike]

thanks!

[dull star]

yeah I give up tbh

paintings are def nerfed

[old pike]

which sampler do you use?

[dull star]

tried many

euler, heun, unipc

and some schedulers too

(ones that don't break)

[hallow lion]

Not bad huh guys? And it will be trained too coz the license is good.

[vital moth]

how long does it take for each image?

[vital moth]

I'm on RTX 30 series

[hallow lion]

turbo 20 secs regular 3 mins lol on 3090 12 vram 32 ram

not sure the regular needs 40 steps tho...

[vital moth]

bc it took me 4min for 1024x576 20 step, I don't know is it normal or not :/

[turbid grotto]

no

[noble coyote]

Prompt = Surreal Stock Photo

[turbid grotto]

75s with 3060 for 20steps

[signal shuttle]

When did the 3090 have a 12gb vram variant?

[vital moth]

whaaaat

[winged seal]

Decided to join back after months of not being here, just cause I am mildly interesting in SD3.5

[vital moth]

do you use default setting?

[icy coral]

You really should release your base models together with a fine tune just so you can shut down some gratuitous and baseless criticisms

[turbid grotto]

yes

[noble coyote]

Just under two minutes 1152x896 64Gb RAM, 8Gb VRAM RTX2070

[winged seal]

Its definitely not as good out of box as flux, but if it can be trained properly this time, I can see it getting more popular in the community

[vital moth]

I just repull whole project...

[rotund stump]

Ok, so I guess 5 minutes isn't normal for 4 steps on a 4080

[noble coyote]

Missed you - some of my best work in your workflow

[winged seal]

Oh hey there haha

[hallow lion]

3060 i mean XD

[winged seal]

I just started getting into Flux training. Having some absolutely WILD realism results with extremely little data. Coming for the flux realism crown this time around

[vital moth]

Some settings are absolutely wrong

[vital moth]

but why??

[winged seal]

SD3.5 could be interesting, but its gonna be a while before its viable and supported, so I am gonna focus on flux instead. I held off of flux for a very long time cause I felt it was not aligned with the community, but after seeing how much work was made in making flux accessible, I feel better now

but yeah, I am interested in SD3.5. I had heard about it for a while, just glad its out now

[mortal mesa]

prepare to be surprised

[winged seal]

By?

[mortal mesa]

the speed of work behind this release

[turbid grotto]

it knows breasts

[winged seal]

@noble coyoteResult from today with my Flux training

[hallow lion]

Call the ponies.

[winged seal]

Flux is exceptionally easy and reliable to train, so I am curious how SD3.5 will be

[noble coyote]

Nice! Have you tried Super Flux? https://youtu.be/3WViQ2pL6ks

[winged seal]

I am not opposed to competition. I just know SAI is far from the most reliable company 😅

So I definitely don't want to get hopes up, especially when I see huge issues with SD3.5 still

[winged seal]

Nope, I only use flux dev base, none of the other things as of now. I migth start using Libre flux, just to see, as it was trained by a research partner of mine

[languid scaffold]

Same. I've tried countless times to train a sdxl lora, just won't train at all with the way i train "exotic" models for flux. As in unconventional "hairstyles" for instance