🆕｜sd3

[faint breach]

but where the kitt scanner?

[craggy crest]

prompt it in i guess. i was just trying to get the overall look

[torn wharf]

lol you got a hot looking pontiac for sure. i love that it does another one so perfect on yours. and mine is quintessentially the 1982 model

but i also can't get the kitt to show up

[craggy crest]

the scanner is a set of horizontal flashing lights. you might have to describe it like that

[faint breach]

I'm running a bunch of gens and it seems like its 3-4 seconds faster on average

[sage burrow]

Should I dig out bluestacks? :>
What can a macro help with that a randomizer prompt and multi queue don't take care of?

[simple thistle]

so it's look like it's faster on the 4000 series but maybe not on the 3000 series and earlier

[craggy crest]

??

[simple thistle]

--force-channels-last with SD3

[craggy crest]

i'm gonna have to set up a new tab on my spreadsheet and go through all the modfiers again

[faint breach]

So they're making a new Knight Rider series or movie i heard. With a new KITT.

Val Kilmer was one of the only good parts of that last Knight Rider reboot. They should use the AI version of his voice for the new KITT

[sage burrow]

😄

[craggy crest]

where's the frog?

strangely whimsical surreal winter by photographer "ansel adams"

[faint breach]

awww sick. i called it a red chaser light bar than queued a dozen. one came out looking so good

[craggy crest]

oooo go run that through lumalabs!

[torn wharf]

it does the various years so well!

[silent ginkgo]

SD3 medium, sgm_uniform, euler, 30 steps, 1024x1024, batch_size 4 repeated 4 times - RTX 4080 without:

Prompt executed in 29.50 seconds
Prompt executed in 29.33 seconds
Prompt executed in 30.09 seconds
Prompt executed in 29.95 seconds

and with --force-channels-last:

Prompt executed in 28.39 seconds
Prompt executed in 28.22 seconds
Prompt executed in 28.33 seconds
Prompt executed in 28.56 seconds

[craggy crest]

strangely whimsical surreal winter moose by photographer "ansel adams"

[simple thistle]

so a small but noticeable difference

[upper snow]

comfy, where's the easiest place to inject a profiler that would cover the main inference loop?

[simple thistle]

do you want to only cover the diffusion model?

[upper snow]

yeah, pretty much

[simple thistle]

or the whole sampling loop?

[torn wharf]

i'm using the ode sampler too. it has really erratic sampling times between images

[simple thistle]

https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/model_base.py#L106

diffusion model is called here

[upper snow]

the forward pass and everything that happens between forward passes

[simple thistle]

that call I linked is the same for all models supported by the base

[simple tundra]

I wonder has anyone successfully trained SD3 with dreambooth lora? I have tried the example in diffusers but the results is not satisfactory

[torn wharf]

theres a bunch on civit

[simple thistle]

I have seen a few loras

most are not very good but people are probably going to figure out how to get good results soon

[simple tundra]

how many images should I prepare to train a lora?

[faint breach]

for a while with sdxl, a guy was posting an empty lora over and over, but ever example image gallery was a different artist that sdxl knew. so every gallery looked like a vastly different style. but the file was nadda. he was one of civit's most published users

it was like 100s if i remember it right. i'm just thinking, same thing gonna happen to sd3 scene sooner than later lol.

[weary snow]

What setting ?

[simple tundra]

and which training script is often used to train a lora? is this a good one https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora_sd3.py

[digital thorn]

2nd pass with sdxl and back to SD3 again

[faint breach]

simpletuner people using too. it uses diffusers. i'm not sure what it brings to the table on top of just using the diffusers script

[simple tundra]

I see, thanks a lot

[viral plaza]

retroactive damage control. that wasn't even posted publicly, it was in a private group with community members but they wanted to share and i okay'd them to. The removal of the "beta" label was just a mistake we made in hindsight. All the theorizing is wild, people needa chill.

But maybe put some pressure in the decision makers because… lol trust me dude everyone at stability who interacts with the community is going back to internal chats to push the public opinion into the heads of the decision makers as best we can.

have someone make an official announcement with a realistic roadmap man i wish lol

[faint breach]

its okay lol. the videos are often not what i'd expect but they're juicy vids

[dusky thistle]

so i've been training some epicly bad loras - i really need to just mod that script to allow me more than One PRompt to Caption Them All... but yeah, it's looking like it responds pretty damn well to training so far

[faint breach]

love this glitch

[upper snow]

@simple thistle yeah, you got a few blocking calls lol

[craggy crest]

did you prompt it to have the car drive off? you can use a text prompt and an image at the same time with luma

[simple thistle]

if you disable the latent preview where are the other ones?

[craggy crest]

Sir Broccoli

[upper snow]

there's two somewhere outside of the forward pass, I also see a bunch of them in the get_area_and_mult method in samplers.py line 12. there's also one that I know is part of k-diffusion, if you patch the to_d function to have division by a scalar and we assume that you'll never have two batch items sampling on a different sigma value then you can avoid that one by placing sigmas on CPU fron then on.

[craggy crest]

@viral plaza congrats on fixing the fitting issue with Afremov!

[patent acorn]

fitting issue?

[craggy crest]

kahlo is still a serious issue, however

[simple thistle]

yeah it doesn't actually make sense for the sigmas to be a tensor

[craggy crest]

yeah. use Afremov in a prompt anywhere close to the front and all you get is his painting. that's not the case with SD 3, thankfully. i can get his influance but not his painting now. Kahlo, though, is way overfit

[upper snow]

keeping them as a tensor as fine, as long as they're a 0d tensor and ideally on cpu (which is itself a cursed concept that only zuckerberg's finest could come up with)

[patent acorn]

oh so thats an artist or

[craggy crest]

https://afremov.com/ yes

[patent acorn]

no wonder lying on grass are more consistent now

[upper snow]

the issue is actually not with to_d itself btw, it's all the control flow operations in the sampler loops in k diffusion. actually I know for a fact that this is what those item calls are

[craggy crest]

lots of data on him and - it's all the exact same thing

[hallow lion]

sdxl to sd3 seems nice

[upper snow]

the only things sigmas are used for in k diffusion are scalar pointwise ops, except for the to_d method, which expands the tensor for no fucking reason

[hallow lion]

sd3 is good at giving it a gloss. sd3 right now needs sdxl as a starter lol

[simple thistle]

let me try removing something

[hallow lion]

so to preserve to promt adherence thing we can use pixart sigm with sdxl and the give it some vibes with sd3

[craggy crest]

you need to start playing around with these settings in comfy - the ModelSamplingSD3 node value. the cfg value in the ksampler. the sampler and schedulers in the ksampler

[hallow lion]

uni pc sgm_uniform and a higher CFG than sdxl seems good

sgm_uniform is the new karras guys

[craggy crest]

try using the simple scheduler, and also the ddim_uniform scheduler. and stick with the denoisers for the samplers

but set your cfg to 2 not 4.5 - and set ModelSamplingSD3 to .3 and see what happens

[digital thorn]

Feels like it has been since Turbo was released.

[upper snow]

this is also a blocking call, it would be more appropriate to pin the cond tensor and change that to .to(device, non_blocking=True) (which should be completely safe from CPU to GPU). also you should consider using torch.Tensor.expand() since that creates a view instead of X copies of the tensor. more memory efficient that way

There's nothing nearly as bad as was in A1111 here though. A1111 had gradient checkpointing running this whole time since SD1.5 came out lol

[upper gust]

tried some seamless texture

[simple thistle]

https://github.com/comfyanonymous/ComfyUI/commit/f2e844e0542ae98b6bfcd438fbc8d22e66f178c9
not sure if that improved anything but you should have one less blocking copy

[late walrus]

agail

[upper snow]

yeah to really benefit from it you have to get rid of all of the blocking calls. which is absolutely possible, because I did it to A1111's codebase. the problem is right here, at the left side of GPU 0 SM efficiency:

gpu is basically idle at the start of the step, because it's waiting for the last step to complete before it can ask what the value of sigma is in a sampler for control flow purposes, then you're making blocking transfers to the GPU, and only once all that is done can it start working on the next forward pass.

If you remove all of the blocking calls, it'll queue up a bunch of steps ahead of time and the GPU will never be idle until it's done.

oh, almost missed that one. same issue was in the LDM code base. you need to create that tensor on device.

[rain palm]

why does it have to be like this?!

https://youtu.be/sx1G9IvTa0s?si=TbpjdAnUgw08mxs6&t=1205

[rain palm]

people are comparing SD3 with ||SD2|| which means SD3 will be just as dead.

why are they doing this? literal self-termination...

[mortal mica]

i really like sd3. ii hope people can figure out how to finetune it and itll be really really good

[bitter hearth]

people are stupid
AND want attention

[viral plaza]

Anyone in this channel that's just scrolling past the technerd talk, the magic graphs drhead are posting are work to optimize SD3 so it runs faster for everyone. Give this dude massive shoutouts he's doing awesome work

[patent acorn]

everything has to be white woman booba thumbnail

[simple thistle]

https://github.com/comfyanonymous/ComfyUI/commit/1281f933c1c38ac0491ff2f86cbcd2ec90743ce3

here that should fix that one

[viral plaza]

and helping optimize both auto1111 and comfy with no bias or nonsense, just better for everyone, it's awesome

[mortal mica]

colors are vibrant, it can go full white full black. prompting is fun and gets all the details. it gives me exactly what i want. just wish the bodys and stuff wernt deformed . if that gets fixed then im super happy

[rain palm]

triggered?

[upper snow]

but i submitted a pr for that one 😦

[simple thistle]

I should have copied that from the ldm implementation

[graceful osprey]

[simple thistle]

oups

[rose hawk]

what's "DonutsDelivery" ?

[simple thistle]

wasn't checking my emails have too much stuff open

[graceful osprey]

BTW - embeddings for sdxl probably works with sd3

[viral plaza]

not probably, they do work!

[craggy crest]

youd have to ask that question of the person that posted that video

[steel helm]

i thought that's you?

[graceful osprey]

🤿

[rose hawk]

I'm davcha on civitai

https://civitai.com/user/davcha <--- me

[simple thistle]

loras for SDXL should also work on the SD3 CLIP

[rain palm]

the person that made the video just summarized benchmarks and saw the similarity to what happened to ||SD2|| , which isn't that far off.
i'm asking the StabilityAI team why they have to hurt themselves like this. pulling another ||SD2|| is going to make ppl lose interest in their models.

[steel helm]

lol i saw the i'm thyri thing and thought it was you clarifying, good catch

[craggy crest]

it might be good if you scrolled way WAY back through this discord and look at all the images that have been posted

[dusky thistle]

just trying to learn what i can from this

[rain palm]

why do i have to scroll way way back?

[rose hawk]

yes that was me. It doesn't appear like that for you?

[craggy crest]

cause there are that many screen pages of really good images to look at - which sort of proves that the nonsense being spread around about Sd 3 is, well, nonsense

[steel helm]

it does, i didn't look, give me a sec to finish fixing it

[upper snow]

the main lesson is that profiling is fucking awesome.

[rain palm]

why don't you generate a person with the same prompt on base-sd3 and base-xl and compare them?

[shut sable]

instead of guessing why gpu slow, see why gpu slow

[craggy crest]

i have.

[rain palm]

let see 'em, with the metadata

[steel helm]

ok updated, i just took the reference files out of the space

[craggy crest]

SD 3 is miles better than SDXL - IF - you know what you are doing - but IF you just try to drop in a prompt and don't know how to adjust various values, no telling what you'll get

[craggy crest]

naw, we've had enough trolls in here demanding this same thing the last 48 hours. not interested.

[dusky thistle]

my coding background is from some time ago, and was more from the binaries/C/asm/security/tcp/ip fuckery side of things... much newer to this angle

[rain palm]

so what, i have to use braincells to use SD3?

[shut sable]

yes, thats The Point

[craggy crest]

yeah, you might have to do that.

[rose hawk]

you can write links like this link in markdown. That looks a bit better btw. but that doesn't really matter

[craggy crest]

if you want a point and click, i hear that Runway does an okay job

[shut sable]

dont expect that just throwing in "big bobba anime uuoohh" will work perfectly, i've said this before; most of you need to learn to prompt properly (for sd3)

[rain palm]

i don't think the Stability team develops an image model specifically designed to have ppl perform braincell gymnastics.

[steel helm]

imma just move it to another server. it's fine. thanks for the input

[craggy crest]

and SD 3 also requires a little tiny bit of common sense in other places

[shut sable]

"braincell gymnastics" my pal you just have to explain what you want, its not that deep

[craggy crest]

you're right, i guess. you don't think.

Stable Diffusion is probably not for you. Bing image creator is probably a better choice.

[rain palm]

you don't get the point right?
make a simple prompt for a human in SDXL and use the same prompt in SD3. the output will be more accurate in XL than SD3 if you use the same prompt with just the base model.... it's called a benchmark.

[craggy crest]

(getting the idea yet that we are sick and tired of people coming in here to complain about something they have no experience with?)

[shut sable]

news just in: model that has 3 text encoders needs more text to be better, more news at 7
also, as i've said before, you. need. to. learn. to. prompt. correctly.

[craggy crest]

news flash - SD 3 uses a totally different neural network, and a totally different architecture. you get to learn it from the ground up - it doesn't work like other versions of stable diffusion.

it doesn't even speak the same language

[rain palm]

how difficult could it be to prompt/draw a horse with 4 legs, or a person with 2 arms?

[craggy crest]

extremely difficult if you are an alien, you don't speak human languages, and the person talking to you has no idea how to communicate in a way you understand

[shut sable]

very easy

[astral grail]

were you involved with the big performance improvements that were merged to A1111 a few weeks ago?

[upper snow]

I did some of them, yeah

[craggy crest]

my name is SD 3 - and i am not human

[mossy star]

all hail robit

[dry wave]

but prompting behaviour can be trained. If sd3 needs good prompts to be good, then we might be able to finetune it such that it works with simple prompts, too.

[rain palm]

full screen? i don't see the loaded model.

[astral grail]

but you're not the guy called huchenei or so on github who seems to have led that investigation into A1111 performance improvements?

[craggy crest]

it actually works much better on simple, concise prompts, than long winded rambling prmopts

[mossy star]

just dont confuse the model and you're golden

[shut sable]

?

[craggy crest]

prompt: a horse

[upper snow]

He did most of them, I found a few other things. He was mainly focusing on the SDXL side and a lot of what I caught was on the LDM side covering SD1.5 mainly, and also the one in k-diffusion

[dry wave]

yeah, but there is something strange. Like certain prompts just doesn't work and I'm very optimistic this can be fixed with training

[graceful osprey]

let's say that - it doesn't work well with tag style prompt.
especially medium

try talk with it

[rain palm]

yeah, now try that with the model they released on civit.

[craggy crest]

it doesn't speak english. you have to learn how it thinks, which is different from every other model of stable diffusion, and use the terms it understands

[shut sable]

that... is the model.....

[mossy star]

i use a combo of tags and phrases tbh

[upper snow]

Also put live preview on a cudastream which works great as long as you don't mind an occasional preview that is just noise. So that's not merged yet. But I'm very proud of it.

[graceful osprey]

and... there is a reason for that.
probably can be fixed with further finetune, yes.

[craggy crest]

who is this "they" that released it? not stability, i don't think

[rain palm]

no, the API version is different, the released model makes worse images.

[astral grail]

nice, it surprised me why suddenly people started to profile A1111 and managed to significantly optimize it... I mean, why so long after the A1111 release, why didn't this kind of profiling not already happen a year ago but now suddenly multiple people work on it at once?

[upper snow]

few problems can't be fixed with large volumes of furry artwork

[craggy crest]

i'm using the api on a different discord. it makes fantastic images. i think someone's scamming you

[shut sable]

ah yes the api version

[dusky thistle]

i'm finding it to be very responsive to training in my limited tests

[patent acorn]

i do not like boob bait on oyutube thumbnails sorry ❌

[rain palm]

yes, the API version is better, download the community-released SD3 and you'll see.

[shut sable]

that. is one way to put it, i prefer weebshit though

[craggy crest]

SD 3 via API - prompt: a galloping horse, mane flowing in the wind

[graceful osprey]

yep. VERY responsive so train it with care 😉
It should be easy for any kind of alignment task - but will be harder to blend in totally new concepts

[rain palm]

dudes, download the community model from CivitAI and make the same prompts
that horse will have either 3 or 15 legs.

[craggy crest]

i only see 4 legs on that horse

[mossy star]

@dusky thistle do you have a paper i can read to help me understand scheduler equations or maybe a simple example of one i can put in your node? please and thank you

[craggy crest]

no thanks. that's NOT SD 3, i have no idea what it is. go get the actual check points from the stability.AI huggingface page

[rain palm]

https://civitai.com/models/497255/stable-diffusion-3-sd3?modelVersionId=552771

[hallow lion]

any loras out yet?

[shut sable]

i dunno man, this looks like 4 (also this isnt API if you read the code lol)

[craggy crest]

nope - here you go https://huggingface.co/stabilityai/stable-diffusion-3-medium download your checkpoints from THERE

[dusky thistle]

i made a really shitty one

[rain palm]

are you dense? i said, the API version and their released community model are different.
get the community released SD3

[muted dove]

I have been reading through it, in awe. It's great to see the masters at work! 😊

[craggy crest]

are you dense? the huggingface page are the open sourced weights. i have no idea what's even in something suspiciously posted to civitAI.

[shut sable]

are you dense? i literally posted proof that this is the sd3 we all have, dont believe me? check the space.

[rain palm]

yes but yall are generating through the API
O_O

[dusky thistle]

i don't really... i know what i know just from screwing aronud. it's pretty intuitive honestly, make sure you watch the previews as they progress, pay special attention to things like a mutation disappearing or getting stuck in place, vs how large the sigma was at that point or how rapidly it was cahnging

[shut sable]

if this is "the API version" to you then theres no fixing this case of brain damage

[craggy crest]

fact is, if it's NOT on the official stabilty.AI huggingface page, i don't want to run it on my machine. there might not be malicious code in it but that isn't where they released it.

[rain palm]

lol. that was fixed with safetensors.

[craggy crest]

but hey, someone that's addicted to CivitAI and it's super sketchy stuff isn't going to listen to reason

[craggy crest]

you hope. - why are you so resistant to getting the real checkpoints from huggingface? is it because yo uknow they would actually work, and then you'd be embarassed?

[rain palm]

why's it just so difficult to try the same model but without the StabilityAI API? get the model from whatever source you want, huggingface or whatever, i don't care just download the community model and run it in whatever WebUI of coise and tell me you get the same results.

[craggy crest]

talk yourself blue in the face, @rain palm we've been listening to this nonsense for24 hours now. no one cares any more

[signal shuttle]

Nice burning car

[rain palm]

ok so if you've heard it for 24 hour, perhaps ppl are having a legit point?

[shut sable]

why's it so hard to understand that i am using that model

[rose hawk]

btw:

GatedRepoError                            Traceback (most recent call last)
Cell In[1], line 4
      1 import torch
      2 from diffusers import StableDiffusion3Pipeline
----> 4 pipe = StableDiffusion3Pipeline.from_single_file("g:/ai/models/stable-diffusion/sd3/2b_1024/sd3_medium_incl_clips_t5xxlfp8.safetensors")
      5 # pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16)
      6 pipe = pipe.to("cuda")

GatedRepoError: 401 Client Error. (Request ID: Root=<guid>)

Cannot access gated repo for url https://huggingface.co/api/models/stabilityai/stable-diffusion-3-medium-diffusers/revision/main.
Access to model stabilityai/stable-diffusion-3-medium-diffusers is restricted. You must be authenticated to access it.

[rain palm]

get ComfyUI or SD.Next and try again.

[craggy crest]

i'm gonna guess that zingo has no idea how to adjust settings in comfy. just drop it in, drop in a prompt, and cry

[shut sable]

ye, you need a hf token and to have sd3 access to use that / use the HF space

[mossy star]

omg i think i realized - sigmas are comma separated lists of values, right? i was thinking a literal math equation

[rose hawk]

I need to be authenticated to load a local file? Is this UbiSoft ?

[shut sable]

its downloading the diffusers version, try just using the space https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium

[mossy star]

if thats right then i am a huge idiot lmfao

[rain palm]

you know the fundamentals of a benchmark? it's supposed to be an even comparison without changes to settings.

[rose hawk]

I want to do it programatically, so I can mess up the model quickly

[signal shuttle]

The model on hugging face is gated, apply for access or use this https://huggingface.co/adamo1139/stable-diffusion-3-medium-ungated

[graceful osprey]

well... hmmm... it's calculated

[dusky thistle]

yup, we often use an equation (usually do) to generate them, but that's convenient, not necessarily because that's the best way to do it (align your steps is a great example)

[craggy crest]

i'm a programmer. are you? i'm done with you. you don't know what you're doing, you won't admit it, you won't listen to anyone. you just want to sit here and whine.

[rain palm]

hahah

[mossy star]

AAAAAAAA lmao we are so back

[shut sable]

news flash: model that expects long text suffers with short text

[untold valley]

Has SD3.1 released yet with wamen laying on grass?

[dusky thistle]

everyone shut up and admire this beast

[craggy crest]

the programmers do have to eat and sleep

[mossy star]

thank you for the help again, i will use this newfound knowledge and return at some point with more images

[rain palm]

again, get the community released model, do the same prompt on StabilityAI and SD.Next.
that's all.

[craggy crest]

are you really that dense? how many times does he ahve to tell you that's the model he is using

[shut sable]

that is what i am doing. i have stated this 3 times already. i can bring a horse to water but cant make it drink.

[dusky thistle]

i embed all my workflows, i'm usually screwin garound with something sigma related in them

[untold valley]

It shouldn’t be to hard… same thing just more/better training data. They probably even already have it.

[craggy crest]

since they're not sure why it's an issue yet, it's slightly harder than you realize to figure out what to do to fix it

[shut sable]

anyways i hope you have fun with dall3 or whatever normie model you use

[rain palm]

he's using it on StabilityAI
use it ALSO in SD.Next

[shut sable]

https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium

[craggy crest]

@rain palm just go play with your models and leave us alone.

[untold valley]

They don’t know what’s wrong? Really

[craggy crest]

yeah, they don't know what is wrong. yet. really.

[shut sable]

using it on the hf space is the same as running locally, you'd get that if you had any sense of competency

[craggy crest]

you're trying to explain 1+1=2 to someone that doesn't know what 1 even is.

[shut sable]

"you can bring a horse to water but cant make it drink"

[craggy crest]

😉 but you can prompt it trying

[untold valley]

Ok y’all can be a little more helpful and ppl would be more receptive not everyone is technically minded. @rain palm the demo basically downloads the huggingface SD3 model and runs it in the cloud. It’s the same as if you download to computer and used auto1111 or comfyui. However I don’t know and don’t care to check what settings it is using. There may be some sauce. Regardless the model had it’s legs cut off. For unknown reasons

[shut sable]

amazing.

[faint breach]

i keep coming back to the kitt prompt. 1/30ish has a proper light bar randomly. all of them are so good tho

[shut sable]

the model was lobotomized due to "safety concerns"

[untold valley]

@rain palm you can check here. Mind you this is the diffusers version but should be the same.

[untold valley]

Why man? It boggles my head. This is 2.x all over again, SDXL was good right off the bat. A tiny bit censored but it was by all intents and purposes a successful model.

[shut sable]

lawsuits basically

[rain palm]

that's what i said, then i got the fire for saying it

[craggy crest]

except it isn't. people need to stop saying that. that isn't true

[gusty trail]

Misinformation a lot

[rose hawk]

Exception: data did not match any variant of untagged enum PyPreTokenizerTypeWrapper at line 960 column 3

[untold valley]

Do you know if SDXL got lawsuits cuz that dataset was okay. Not extreme. Workable.

[upper snow]

chilling effects of potential regulations, to be more precise

[craggy crest]

reddit assumptions and wild conjectures being repeated by everyone that hasn't gotten a bunch of naked women they didn't want

[untold valley]

Then you need to get on a horse and pull that carriage of information and explain it without looking like a psyops.

[upper snow]

unfortunately there are a lot of people who use these models for bad things and having the model unable to do them natively helps keep that attention off of the makers of the model

[shut sable]

i dont feel that theres another explainable reason, well other than this was supposed to be the beta model

[rain palm]

ofc, it's better to force them out on the streets than keeping them at home generating AI images.

[untold valley]

IMO cat is out the bag since 1.5 SDXL basically exploded it. Why be shy w SD3?

[faint breach]

no light scanner tho. what a borked unsaveable model

[craggy crest]

explainable reason for why i have to battle with it to put clothes on females on some settings, or what?

[dusk silo]

https://i.imgur.com/PlM9E4g.jpeg

[mortal mica]

could almost fool me thinking this was a real image

[craggy crest]

Prompt: splashes of essence

[shut sable]

i havent gotten that issue yet mind sending me the prompts

[rain palm]

where's the driver?
the kidneys are messed up, lines on the road doesn't align, reg plate is a mess, car is dry even tho it's raining.

[craggy crest]

i'm going through a massive prompt spreadsheet. next time one gives me fits, i'll DM it to you

[shut sable]

dont actually dm me it please dont

[mortal mica]

why i said almost 😂

[craggy crest]

will you PLEASE go complain somewhere else? NOTHING is going to be okay with you. Go gripe on twitter or something

[untold valley]

I want to believe this. But initially didn’t they say this was done and moving on to 8B? So the assumption would be this was final model. And if they followed SDXL.09 released as beta and within a month 1 was out. Stands to reason if this is really a beta why don’t SAI say new model comes out within a day unless model really is borked and they have 0 way to fix.

[rain palm]

i just gave him feedback... O_O

[craggy crest]

how about giving some positive feedback in that case?

[craggy crest]

no, they did NOT say it was done. in fact, it's been said several times it is NOT done and was rushed out because the community is impaitent

[upper snow]

probably one of the main things they're concerned about is people generating nudes of real nonconsenting people and then sharing them either because they lack the sense not to or because they actively intend to harass that person, which, well, if they were keeping that contained to their home it would not be nearly as much of a problem.

[faint breach]

just perturb the model its ez

[shut sable]

heres proof that this was supposed to be sd3 med beta

[rain palm]

why? everyone else is positive, someone has to be on the other side of the argument otherwise everyones idea of what's good becomes cemented and non-changing.

[craggy crest]

@viral plaza talked about that earlier and explained exactly what happened

[faint breach]

wrong spot for it. borrrked

[shut sable]

i doubt theres reason to ping him again (also hi again :3)

[craggy crest]

and since he was involved in that decision, he happens to know what he's talking about

[shut sable]

he advocated for this being called the "beta" model fwiw

[dusky thistle]

cut the employees some slack... issues usually come from the top

[rain palm]

true

[faint breach]

such a broken model. i'm really sad we have this now

[craggy crest]

i know. the community wanted to remove the lable. i can guess why, too - with all the screaming and hollaring about why wasn't somethign released, i can just imaging how ballistic the crowd would have gone if they were told "here, here are the weights but this is still beta"

[untold valley]

This is not really proof. It’s a message after the fact dated day after . This would be a PR move id like to see one of them talking about it being a beta before it was release.

[shut sable]

why would this be a PR move if it was shared in an internal chat

[craggy crest]

yo uwant more proof than the words of the person that approved things?

here - GO READ THIS POST #🆕｜sd3 message

[faint breach]

how fat is it gonna make the light scanner the silly broken model. theres no fixing this.

[mortal mica]

hehe

[craggy crest]

try "off-road lightbar"

[shut sable]

anyways its getting late and the tensors are getting tired, good night

[rose hawk]

just ran the example on the diffusers page and got this: "smiling cartoon dog sits at a table, coffee mug on hand, as a room goes up in flames. “This is fine,” the dog assures himself."

[craggy crest]

that tokenized well

[mortal mica]

first try

[rain palm]

worlds most advanced generative AI btw...

[faint breach]

3 dozenth try

[rain palm]

didn't you say byebye like 2 minutes ago?

[shut sable]

Yes. Let me ruin my sleep schedule. [if you cant tell, they're baiting]

[craggy crest]

go. away.

[rain palm]

shush, get Zingo out of here!

[craggy crest]

trolling hours are over. go play with bing or something

[untold valley]

Didn’t think this mess could be worse. So there’s actually internal drama too and decisions makers not really listening. Guess SAI is in some type of internal conflict. Guh. Feel sorry for the ppl in SAI who work and try their best but are also handicapped by the actual decision makers.

[craggy crest]

like most companies

could be worse. could be openAI

[untold valley]

Guh. Such a mess this was.

[faint breach]

bing is pre good at kitt . gonna have to give kitt poitns to bing here

[craggy crest]

yeah, those are really nice

[rain palm]

ClosedAI

[craggy crest]

i hear they are hiring, they might like to have you join them. you've got the right mindset

[rain palm]

they wouldn't have me, i'm too critical.

[craggy crest]

you'd fit on their board

[rain palm]

either i'd be lonely as a whale or i'd get fired before i even parked my car.

[untold valley]

OpenAi got a chain on their necks tho. Microsoft hold the whip.

[rain palm]

there's two sides of that story and idk which one to believe.

[craggy crest]

they asked for that chain, it's their own fault

[untold valley]

And they will get Gov 100% involved. CIA, NSA, FBI, Homeland etc all of it will be up OpenAi soon if it’s not already there

[signal shuttle]

People really need to use ClipTextEncodeSD3 node in their workflows in my experience it improves prompt following and coherency a lot

[craggy crest]

all you ahve to do is read their contract with microsoft.

[rain palm]

i thought contracts were confidential?

[craggy crest]

you see the news that a military general has just joined openAI's board?

[craggy crest]

you need to read twitter more often. especially when Altman decides to post the contract with microsoft

[unkempt tusk]

Is there any plan to make some com around sd3 ? Back in my time damage control was faster 🫡

[rain palm]

🤷‍♂️ i don't have twitter

[untold valley]

I can’t imagine how many secrets it already owns anything typed in their website gets added to their database.

[craggy crest]

can't help you there. microsoft owns all of OpenAIs tech until openAI achieves AGI. IF they achieve it.

[craggy crest]

everything that you post online goes through one of two military gateways - one on the east coast of the USA, the other on teh west. EVERY single thing you put out on to the internet is already on multiple computers on on Gov. computers.

[scenic shadow]

Not like AGI is actually a term that has any REAL meaning. It’s just gonna be whatever they decide it’s gonna be

[craggy crest]

it has meaning. ignore the news media. if it didn't have specific meaning, it wouldn't be in the microsoft contract with openAI and it is

[unkempt tusk]

Only if you live in the US

[craggy crest]

nope. if you post anything to the internet, it routes all around the world and it goes through those gateways.

[rain palm]

i heard one story where Sam got fired for keeping secrets, then got re-instated and the board-members changed to downvote the existence of AGI so Microsoft could keep their GPT4-based Bing model as their deal with OpenAI would end at the AGI achievement.

i've also heard the other story about the contract you mentioned.

[unkempt tusk]

That's not how the internet work

[mortal mica]

donut

[craggy crest]

you should'nt believe everything you hear. and i'm not going into all the details of that mess. i lived through that and almsot didn't sleep all weekend while reporting on what was going on.

[viral plaza]

yeah that was part of the reasoning for why it was removed
ftr at the time i didn't really have strong feelings one way or the other on Beta or not, but in hindsight ... yeah removing it was the wrong call

[craggy crest]

sorry, charlie, but that is exactly how the internet works. and it's designed that way on purpose. to keep communications from being disrupted

[dusky thistle]

you're really stuck in quite a quandary half the time on this stuff

[rain palm]

that's why i said idk which one to believe. i keep neutral and don't care about the drama, i care about what i've got and what i've got is meeh.
it's getting better but still not quiet there yet. and the fact that it's closed source makes it potentially even more dangerous than if they kept to their slogan "Open"AI.

[craggy crest]

don't kick yourself, it's a no win situation. no matter what you'd decided, the community would be screaming in rage anyway.

[scenic shadow]

What is the exact meaning? That it can learn unsupervised? Thats the only thing I could find officially.

[noble coyote]

What settings in SD3-Medium will give the best text? 🙂

[untold valley]

SDXL had worked with the 0.9-> 1.0 transition. But I believe above all else it helps tremendously to have community involvement with SAI staff like yourself. Thank you for some of the clarification you have presented.

[craggy crest]

that it's able to do things it wasn't specifically trained to do - at the moment all AIs are specifically trained for their functions. AGI is Artificial General Intellegence. not Artificial specialized intellegence

[unkempt tusk]

Too tired to argue I'll just say that I disagree what you say isn't accurate but I think I understand what you mean

[ruby marlin]

I don't really understand how it could still be in beta, I remember a while back they were saying it was just about finished and they just had to check if the safety features were compromising the model quality

[untold valley]

Well they did say this was done until 8B was done too so 🤷🏻‍♂️

[rain palm]

so the small model comes last? 😢

[viral plaza]

if you care for dates so much... well i'd probably get in trouble if i posted internal chats but hopefully nobody yells at me for a screenshot of my own post. This the day that decision was made

[untold valley]

It came first.

[rain palm]

huh? we got the medium model, where's the small one?

[craggy crest]

i wanna be back on the beta test team 😦

[untold valley]

I humbly apologize and thank you.

[untold valley]

I was wrong and that would indeed have been the best route

[ruby marlin]

Anyway the issues with the model are so specific that it's hard to believe it's from undertraining, it is actually a great model outside of a couple things

[rain palm]

dUo, you said the small model came first.. where is it? we just got the medium model like 3 days ago, and the large model coming...

[craggy crest]

medium is small

[rain palm]

the matrix just had a glitch... medium is small?

[hallow lion]

i always tought medium shld be 4B not 2B

[bitter hearth]

400$ on the best restaurant

[rain palm]

isn't there going to be an 800M model?

[viral plaza]

do note takyon is somebody i'd trust to be pretty knowledgable on that topic (he's an infosec professional). I think the relevant key here is the modern internet has end2end encryption (HTTPS) where theoretically it's impossible to decrypt the traffic in the middle unless they've snuck in a backdoor. (Which... whether or not they have done that is an open question well beyond the scope of a random discussion in the sd3 channel lol)

[scenic shadow]

Couple running ideas is that it got anti trained on some concepts for safety and that that part was too heavy or there are just very specific gaps in the training data.
Speculation ofc

[rain palm]

i've always heard SD3 will have 3 models.

[craggy crest]

medium, large, and ultra

[bitter hearth]

Small, medium, possibly a medium large and large
0.8, 2, 4 and 8

[ruby marlin]

I agree but feels more like a deliberate tradeoff than due to something being released while still in beta

[rain palm]

oh man X_X my RTX3060 is crying....

[ruby marlin]

Maybe I'm reading too much into the semantics lol, I guess there is an art to the safety features that takes time to refine as well

[viral plaza]

small is 800M (equivalent to SDv1), medium is 2B (close to SDXL), then there's 4B and 8B. imo they should be "large" and "huge" but I think we're going with 8B is "large" now and 4b is just deprio for now

[craggy crest]

prompt: transmission fluid,oil,triadic alcohol ink marbling on encaustic;intricate hyperdetailed background by artist "inga moore ingrid lamberts"

[untold valley]

How to get hired as SAI spokesperson. My qualifications: can use chatgpt and can bs on discord.

[viral plaza]

there's also cloneofsimo is training a 6B supposedly so idek what name that'd get

[rain palm]

so will there be a smaller model than the Medium one?

[viral plaza]

"SD3 Ultra" is a workflow based on 8B

[foggy cloak]

How is this proof? This was after the model was released was it not?

[viral plaza]

there is one yes

[craggy crest]

i was just giving her three words to use

[hallow lion]

We got the "Tiny" model 😄

[viral plaza]

whether it gets finalized/released idk, but it exists

[rain palm]

oooooh ok

[untold valley]

@foggy cloak Look here Alex gracefully posted proof before date

[rain palm]

well if it exists what's the "finalizing" part about? can't they just throw it out there?

[untold valley]

The truth shall set you free
-FBI

[rain palm]

a 1.5 sized model with 16 channel VAE is what i want... 2GB 🙂

[scenic shadow]

And have the same fiasco happen that the 2b model had?

Something tells me they are gonna be a bit more specific on the training before it’s released.

[rain palm]

at least if it's a "small" model it has an excuse.

[craggy crest]

https://x.com/fofrAI/status/1801884682427965655

fofr (@fofrAI) on X

Princess Peach in Mario attire with SD3

a render of a short princess peach, she is standing in blue denim dungarees, with a pink long-sleeved top and white gloves, she is wearing a pink cap with the letter "P" on it in a white circle, on a simple backdrop

Twitter · Jun 15, 2024, 7:49 AM

[foggy cloak]

👍 good to know. Still gotta wonder why they took the "beta" label off when it was this bad.

[viral plaza]

technically Medium is 2GiB if you save it in fp8 lol

but yeah you'd want the 800M (Small) model for SDv1 equivalent param count

that one is definitely better than SDv1 last I looked, but not great on any other comparison

[untold valley]

Human error.

[rain palm]

and 16 channel VAE drools

[noble coyote]

Get it free @Glif; or $10/month@ClipDrop

[craggy crest]

search this discord for from:mcmonkey and read what he said about that

[viral plaza]

(before anyone @'s me with some nonsense I mean better than SDv1.5 base model, not your favorite v1 based finetune)

[rain palm]

ofc, not comparing finetunes to basemodels

[winged seal]

Honestly, you can do some pretty incredible things with SD3 already if you know what you are doing

[mortal mica]

I agree sd3 is pretty amazing in a lot of ways.

[foggy cloak]

I mean the model is great in many ways but the downsidess are also massive

[winged seal]

I absolutely hated it when it first came out, mainly because it has issues that it never should have had, but I've used it enough and had enough amazing results from it to where I can honestly say that I think it is an okay model.

Whether it's a good model remains to be seen based off of if it's critical flaws can be trained out or not

[rain palm]

the SD3 basemodel is 4GB large, and someone on Civit put up a 15GB model of it, how does that work?

[noble coyote]

PiXart-Sigma and SD3 = PiXart-Omega at all?

[steel helm]

it had all the large clips included in fp16

[rain palm]

X_X

[craggy crest]

as is the case with all tools

[noble coyote]

SD3 is a wild colt that'll need wrangling to the ground! But once it's "broken-in" - we will all have become Prompt Wranglers!!! 😄

[foggy cloak]

Brother, other older tools from the same company didn't some of the glaring downsides of SD3

[craggy crest]

you running it in comfy or some other way?

[foggy cloak]

Local install, running in comfy

[craggy crest]

what settings have you played around with?

prompt: hairy eyeball macro photo

[vapid radish]

Those were different times though, the AI image community was a lot smaller with way less community developers and money behind it

[trim arrow]

How big is an sd3 model?

[steel helm]

because that's obviously what people generate all the time right lol

[rain palm]

approx 4-5 GB, thus far

[trim arrow]

That ain't bad

[noble coyote]

36-24-38

[steel helm]

only if she's 5'3"

[trim arrow]

I don't get the joke

[foggy cloak]

bigger community, more devs and money = worse or sidegrade?

[craggy crest]

model - human female that shows off the new fashions.

[rain palm]

McMonkey, can you look away for a minute?

[trim arrow]

...oh. God I'm dumb

[steel helm]

lmao

[craggy crest]

🙂

[vapid radish]

I am using it there, that's what gets me excited, but it is not the same as running locally, too much censorship (I tried to make a picture of a wet beaver and it blocked my prompts!) . If I wanted to use a free online service I would use Bing Image creator as it is better.

[rain palm]

here's some sfw hentai for yall...

[trim arrow]

I wanna.mess around with sd3 but I suck at prompting lmao

[steel helm]

even the chickens have long necks in base lol

[rain palm]

that was techincally not an SD3 generation tho

[trim arrow]

I'm still prompting like 1.5 because of ponyXL so it's gonna be difficult to try and translate that to sd3

[craggy crest]

just use short, concise prompts with SD 3. it works best

[steel helm]

i don't think it even does that res that clear haha

[steel helm]

i am loving it technically, we'll work out the rest, it's just some bit flipping in the weights i think

[trim arrow]

How long would it take to generate an image for sd3 medium? About the same as sdxl or a little faster?

[rain palm]

in my custom 1.5 model i rarely ever go beyond 20 tokens, and i get crystal clear images and have solved bad eyes and almost bad fingers too.
it gets the nails wrong but counts to 5 fingers per hand pretty much 9/10 times.

[steel helm]

yea bout sdxl, maybe a little faster

[craggy crest]

on my machine it's about the same speed

[trim arrow]

I should...mention. I have a 3060 gpu

[steel helm]

i ran it on a 3060 12, it was about 5 or so seconds faster stock without toys

[trim arrow]

God this cooldown is a menace. I understand why but damn 😭

[trim arrow]

I got a 3060 12 as well. I just wasn't sure

[craggy crest]

it should still be about the same as SDXL is on the machine you're running stuff on

[steel helm]

all the clip stuff offloads to ram, it uses like 5gb without the vae decode. it's tiny lol

[trim arrow]

I have ... 48? Gb of ram. I forget. It's a funny number.

[steel helm]

prolly faster if you change that, but why lol

[craggy crest]

prompt: claymation mountain scene with ducks

[trim arrow]

Two 16 and two 8. It's my s/o's build

[idle parrot]

duck lying in grass

[rain palm]

ops, thought you said 4GB lol

[trim arrow]

If I had 4gb I wasn't even gonna TRY.

[rain palm]

you can have 4 and run SDXL tho

[idle parrot]

SD15 works on 512 MB vram 😄

[trim arrow]

I have 8 gb in my laptop and sdxl loading in models crashes out. I doubt that.

[rain palm]

Forge UI

[steel helm]

yea you can run it with no vram. it's just trading speed each time you offload a chunk or something

[trim arrow]

Already tried forge and it crashes out.

The laptop has a 1660ti if that's anything.

[craggy crest]

prompt: duck lying in grass

[rain palm]

then something else is wrong
"Forge has very low VRAM requirements in comparison to Automatic 1111, and other interfaces, but you'll still need a minimum of 4GB of VRAM for SDXL image generation, and 2GB of VRAM for SD 1.5 image generation!"

[hallow lion]

so its really nice finishing up the detail but i def wasnt thinking sd3 will be used as a refiner model when i heard about it XD

[steel helm]

is what it is until we fix it I guess lol

[trim arrow]

Yeah idk man, I have a 1660ti which may be the issue. But I absolutely can't use sdxl at all.

[idle parrot]

the anatomy...

[craggy crest]

it's a duck.

[rain palm]

try an LCM XL then

[idle parrot]

censorship

[trim arrow]

I'll eventually try it.

[craggy crest]

that's what ducks look like when they're sitting on grass. ducks do not, as a rule, lay down

[patent acorn]

https://cdn.discordapp.com/attachments/1230206273451069540/1250615133508472903/view.png?ex=666b959b&is=666a441b&hm=97f9bd3017e7bfca8086ca063d8f5bf038bcfeae7e116008c9237bc08e1b8044&

[rain palm]

wait, i lost a braincell ...
i was thinking about pruned models, the models come in various sizes, perhaps you've been getting the larger models and just hit the vram limit.

[rain palm]

legendary anatomy

[rain palm]

here's something that'll trigger some of you 😛
generated with SD1.5...

[craggy crest]

prompt: Prismacolour Oil pastel Bokeh Glowing Cel shading Ghibli Cute pastels, bokeh digital painting fluffy clouds sparkles, beautiful

[rain current]

https://civitai.com/models/511463/afros-sd3-experimental-sdxl-merge?modelVersionId=572928
Is it SDXL or SD3? Because I am testing the model with clip t5xxl_fp8, and it is giving me good results... I don't quite understand how it works... Woman Laying in Grass

[rain palm]

it says SD3 in the description, so it's using SD3 as the base, and XL merged into it.

[craggy crest]

it's a mixture of both models.

[noble coyote]

"Copies selling fast - get yours soon!!!"

🥳

[craggy crest]

Turtles seem to be somewhat of an issue

[trim arrow]

for sd3, what do i put for oranization or affiliation if I'm just a solo person? I need to put one to access the repository on huggingface... do I just put N/A?

[craggy crest]

just put NA

prompt: Food coloring ink,wet splashes, drips, drizzles, sparks

[trim arrow]

do i get the incl_clips? I don't know what that stuff means...all i know is that clip=prompt or something

[craggy crest]

hang on a second - what are you going to be running SD3 with?

[trim arrow]

comfyui mostly. on the 3060 12

[craggy crest]

https://youtu.be/9zfF7Jt-JnU?si=BvaFcisZiVPDw72t watch olivo's tutorial on installing SD 3

[trim arrow]

Ahhh, ty.

[craggy crest]

and this one's good as well https://youtu.be/Di1KqPXxx2Y?si=S2O9fXSnSETKVKD6

[trim arrow]

my first go at sd3 with a default comfy workflow! Very pretty.

Hah, i suck at this.

[noble coyote]

Nice! You're a natch!

[trim arrow]

Trying to make what I have been making has produced... undesireable results though.

Just flatout looks bad compared to what I've been making with pony xl.

this is what i've been making with ponyxl

[severe forge]

how to create a picture?

[noble coyote]

The thing with SD3 is that it has a greater nuance of natural light - flat surfaces such as cartoons and anime - they do not fully benefit from that aspect of SD3!

[mystic vortex]

why i don't see this channel in list on the left panel?

[viral plaza]

right click "Brows Channels" at the top left

[noble coyote]

Click on STABLE MODELS again

[viral plaza]

oh or that too, category

[patent acorn]

https://www.reddit.com/r/StableDiffusion/s/9FEmAkXtOt

you cant generate capybara.. ur serious??

[lean mortar]

I bet you sd3 can't do this

[mystic vortex]

thanks guys 🙂

[patent acorn]

someone needs to try generate a capybara rn

[mystic vortex]

1.5 vs SD3 :

[noble coyote]

Cap E Barra

[robust kelp]

Wow. You guys have really outdone yourselves with this one. The dog/human hybrid is nuts.

[patent acorn]

try a sleeping capybara

[bitter wadi]

Wen SD3 ?

[dull star]

Biya Spigex
⭐⭐⭐

[mystic vortex]

Hitler works better than Eva Green...

[patent acorn]

didnt know it could do hitler

[bitter wadi]

C'mon buddy. Stop this

Not sure what you're trying to prove. This is dumb.

[noble coyote]

This is waaay cool!

[bitter wadi]

Colors are so good, but the model is so broken. Sad

[compact forge]

Lovely image

[shell plaza]

It s good to make characters honestly

[bitter wadi]

I want SD3 in Omost ! Should be awesome

[shell plaza]

Can't wait to make some lora on my 4090

[bitter wadi]

There was a fix for the broken gradient checkpointing on diffusers. LoRa training still not okay for a 4090 ?

[shell plaza]

I'm waiting for onetrainer update, I use that for my sdxl and pony Loras

[dull star]

🤨

[storm saffron]

I'm training a lora on a 4060ti 16Gb right now (with a single instance prompt, granted), but it's possible, I'm working on a way to cache the T5 and VAE embeddings first so it uses even less memory.

I'm still trying to figure out optimal settings for LoRA training.

[dull star]

I hope onetrainer just takes care of it all without us having to mess about with optimizing manually

it will probably speed up the rate at how many loras are being made for SD3 rn

[gusty trail]

It is a good take to learn how the trainer work

[storm saffron]

I don't think I'm smart enough to learn how it all works... I just bash together bits of code and hope for th best 😄

[muted cargo]

6.5gb is really not that much for a "non optimized" training session then.

[gusty trail]

If you are interested, here is my implementation of SD3 lora training script. https://github.com/lrzjason/T2ITrainer I also reference many others, like kohya, freon, etc. The script is rough but it works.

[bitter wadi]

Hô, indeed 6.5gb sounds already very good !
Nice

[cedar gale]

SD XL --> SD3 img2img, seen better but we lost the lose limbs. 😛

[patent acorn]

do hitler in public execution

[vapid radish]

There is actually a ComfyUI workflow to do that: https://civitai.com/models/511228/sd3-boobs-unchained?modelVersionId=568187

[teal fossil]

So loading the Clip of SDXL LoRA's actually works in Comfy (just don't load it onto the Model) and it definitely has a strong effect... but not quite sure how much it actually depicts the training it had.

[patent acorn]

SD3 cannot do capybara. -10/10

[mystic vortex]

does anyone have a workflow SDXL+LoRA>SD3>img?

[steel beacon]

Can one get SD3 to work on automatic1111?

[noble coyote]

Thankyou @shell plaza for the prompt idea - SD3 composite - text taken from several different creations! "Apres Moi, La Deluge!"

[gusty trail]

The icons are neat

[storm saffron]

I think that's the trouble I'm having, I'm trying a brand new concept, and it doesn't want to do it. 😐

[runic seal]

you can use SDnext until auto11111 is updated

[vapid radish]

Not until Vishnu updates the code, could be a few weeks but hopefully sooner.

[steel beacon]

Thanks for the replies, Kyra and @vapid radish ! I'm not familiar with SDnext, I still have an AMD and it's so frustrating I went on hiatus until the process gets improved or I get Nvidia.

[patent acorn]

wait what? it did twitter and facebook icons perfectly?

[faint shard]

I guess..

[patent acorn]

instabook

LOL he got timeout

[formal bramble]

What does this image tell us? 😅

[shell plaza]

Yeah it's quite good

[signal shuttle]

the CliptextencodeSD3 node is powerful when you know how to use it

[compact forge]

no coconut doggy 😦

[storm saffron]

How do you use it?

Also what order do I need to load things in the TripleClip? Does that make a difference?

[compact forge]

good quality photo of a dog riding an elefant 👏

[upper gust]

dragonfly vs ant

[signal shuttle]

I recommend putting all quality tags ontop while a mix of quality and your prompt and the last just your prompt without any quality tags

[storm saffron]

My lora is getting close now, I can smell it.

[bitter wadi]

lol, this is ridiculous. Are you still with these magic tokens ?
Give up buddy, this is going nowhere

[signal shuttle]

here is the result of that

[signal shuttle]

these "Magic tokens" have improved image quality by a long shot in my experience

[bitter wadi]

nb: i don't care as you don't get people using it too.

[compact forge]

clip g, clip l, t5

[bitter wadi]

Are you aware about what Macaronic Prompting and similar methods are ?
I think you misunderstand the concept of tokens in this model.

[quartz mulch]

That should be in 'Join the dark side of the force' posters.

[ruby marlin]

This is the type of prompt adherence we've been dreaming about for the past 2 years

[bitter hearth]

[cinder lodge]

Oh no I get it man. It’s an unfortunate situation that you were put in.

The wider public only has pieces of the information though which is what lead to the “damage control” perception without an actual announcement.

Like you said though it was only in a private group, there wasn’t any “hey guys this is the situation” from SAI so people are making their own assumptions.

[alpine axle]

did some initial controlnet tests with this https://huggingface.co/InstantX/SD3-Controlnet-Canny_alpha_512

[bitter wadi]

Looks good 👍

[alpine axle]

it's only trained at 512, but kinda promising

[latent fiber]

tiger|sd3

[twilit cobalt]

the "trick" for correct anatomy

[pastel latch]

Really doesnt understand whales too well

[bitter hearth]

[calm zinc]

That wasn’t our employee, it wasn’t KandooAi. That was an imposter I reported to the stable diffusion mods. That’s probably why all their posts were deleted.
Make sure to verify people pretending to represent other people. Especially in the world of Ai.

[signal shuttle]

what prompt did you use?

[hallow lion]

kinda cool being there when a new tehcnology starts off but i really wish we were at the point when any image you made was just normal and usable. this feels like early nineties internet, like tis this cool new thing and aesthetic and kinda clunky and uh can i download the internet or what is that anyway... -

[calm zinc]

None of this is true.

[pastel latch]

eta on cracking the ft?

[compact forge]

great hands, great leader

[bitter hearth]

Girl darth Vader as a famous model walking down the streets of a very nice beach, photo in 4k, professional shot

2/10 brain power used don't judge

[lilac bridge]

Testing SD3 for making items for dnd 😄
So far not bad. Gotta keep experimenting

[gentle wolf]

can you add a famouse childrens cartoon bear to the prompt?

[hallow lion]

He looks so stable.

[calm zinc]

We’re fine. In talks with SAI. We’ll train SD3 if it’s trainable. Nothing has changed. Just waiting for tooling to get to a good place. Meanwhile be prepared for a limited release for people to test Juggernaut XI. 😙

[pastel latch]

how to be a part of that?

[hallow lion]

well dats good newz

[compact forge]

he eats US nuclear missiles for lunch

[gentle wolf]

I'm not sure of the effort. If there's a 3.1 because 3 is beta. We should probably wait before wasting GPU times, right?

[bitter hearth]

Mcmonkey said if there's an update stuff from our training now would be usable

[gentle wolf]

huh, interessting. how?

[calm zinc]

Not true. That person was an imposter and we reported them.

[severe phoenix]

rundi coming in clutch 🙌

[signal shuttle]

really

[hallow lion]

seriously this was a beta only? They never mentioned that ever... Theres gonna be a 3.1? like with sdxl thjeres the 0.9 and then the 10?

[gentle wolf]

apparently

[bitter hearth]

I'm smooth brain I don't think he went into details about it

[calm zinc]

This is not true. That person was not KandooAi. It was an imposter. Simple as that.

[bitter hearth]

Update is a possibility

[acoustic kite]

i just want the ability to use poses without getting exploding humans

[desert garnet]

it was the early access bro,we getting the official release + battlepass in next update

[wind talon]

yeah they only mentioned this after the release, very scummy

[bitter hearth]

This @gentle wolf

[pastel latch]

think the wrong intended bear

[bitter hearth]

Battle pass ?! I'm getting it

[gentle wolf]

sounds good. but not checkpoints trained I guess.

[desert garnet]

free SD3 lootboxes 4 everyone 👍

[calm zinc]

@prime perch @craggy crest @torn wharf that “employee” was an imposter pretending to be @cosmic mantle for some unknown reason. We’re fine. Chilling on the sideline waiting for tooling. We’re in talks with SAI. We’ll play by their rules. We are excited to train. All that stuff going on is a big nothing burger. 🍔

If you ever need anything verified simply @me or @cosmic mantle

[pastel latch]

uh....

[signal shuttle]

well goodluck to you guys

[severe phoenix]

wait does this mean anyone who wants to train sd3 requires specal tooling provided by sai to train sd3???

[wind talon]

that's nice to hear, are you guys considering looking into alternative models like hunyuan, pixart and lumina?

[calm zinc]

Follow our socials/discord for updates.

[hallow lion]

He is fatter in each image. Turbodong is that you?

[gentle wolf]

once again it shows how amazing discord is, to make things worse and spread fake news and have information lost in the wall of trolls

[signal shuttle]

No, its that people haven't made tools to train SD3, and rundiffusion is waiting for someone to make tooling

[cosmic mantle]

Main Focus is SD3 (Stable Diffusion feels like home for me 🙂 ), at least thats my thoughts...but yeah we are also looking into other alternatives 🙂

[signal shuttle]

I am in love with CliptextencodeSD3 node, all gens have been nothing but amazing

[bitter hearth]

God himself in chat

[calm zinc]

Yep we’re just chilling. Still working on SDXL and will be for a few months I presume.
Jugg XI is done
Jugg XII is in development
Both SDXL models
Still trying to find partners and help with development moving forward. The future is unclear but we’ll be fine.

[cobalt moon]

oh hi creator of Juggernaut

[bitter hearth]

Do you guys like cheese cakes

[cosmic mantle]

https://tenor.com/view/kto-lbow-hi-hello-hi-there-gif-25347432

[compact forge]

you're doing different prompt for each clip?

[faint shard]

Mvp models

[wind talon]

good to hear! i think sd3 has shown that having all your eggs in one basket is a bad idea, best of luck to you all

[lilac bridge]

Oooooh SD3 is quite good at img2img

[gentle wolf]

SD3 is good in a lot of things

[cobalt moon]

just not laying...

[lilac bridge]

Don't say that or else people will post people laying in grass xD

[calm zinc]

But a little secret if you want to try Jugg XI right away. It’s available on our platform inside Fooocus right now….
DM me and I’ll get you an hour or two of server time.

[gentle wolf]

xD

[pastel latch]

might be able to help with that

[cobalt moon]

I mean it have flaws, of course. After all these complaints and oversights

[calm zinc]

@cosmic mantle is his official account

[signal shuttle]

I am putting all quality tags ontop while a mix of quality and my prompt and the last just my prompt without any quality tags

[pastel latch]

done

[lilac bridge]

Eyyy you're doing something similar to me :DD buuddyyyy

[signal shuttle]

I recommend changing empty padding from none to empty prompt, it made thing better in my experience

[calm zinc]

Oh yeah?

[pastel latch]

dm

[lilac bridge]

oooh will try, thanks!

[cobalt moon]

@calm zinc I might sound a little bit like putting oil on a fire ( you have the right to reject answering this though )

Do you have opinions on this Lykon vs AstraliteHeart rap battle?

[signal shuttle]

also try doing this to your negative prompt

[hallow lion]

everyone looks confused and disporportionate

[lilac bridge]

Could you please copy pasta that for me? Some of the letters are hard to read

[signal shuttle]

negative string 1. "deformed and malformed with worst qualities. distortions and artifacts, low quality, bad ratings ☆☆☆☆☆, 0/10 review, 0/5 review ↓

[lilac bridge]

Since you're using the same prompt in every field, the multiprompt node isn't necessary right?

[signal shuttle]

idk but it works better in my experience then the normal negative node

[sacred jewel]

It makes a difference. Each field goes to a different encoder

[lilac bridge]

right, but the example for the negative weights used the same prompt for each field anyways

[gentle wolf]

haha

[sacred jewel]

It's like saying the same thing to three different people. Each will have a different answer. It makes a resulting token difference

[cobalt moon]

classical SD political humour

[sacred jewel]

Test it yourself and see. Easy to prove

[lilac bridge]

Wasn't doubting. Was learning. Thanks :]

[calm zinc]

None. Nothing but respect for @lavish osprey
I don’t envy him at all. He’s got a hard gig and he’s doing a great job. I’ve chatted with the Pony creator and he’s said nothing but good things about this whole thing too. Professional and tactful. People don’t realize that there’s literally nothing “there”. It’s all drama made up by the community.

Everyone is literally doing the best they can. We all need to give a little grace to each other. I think in the end everything will work itself out.
How do I know this? Because we simply want pretty pictures, and we’ll figure out a way to get that.

[sacred jewel]

You think I know anything? 🤪. I am the dumbest person in the room. Always learning too 👊

[pearl condor]

https://huggingface.co/InstantX/SD3-Controlnet-Canny

[compact forge]

but booba bro?

[cobalt moon]

I truly agreed with you. I dont think they actually intended real harm to each other. Just that they have different sightings on how to train a finetunes.

[compact forge]

its always the booba that makes men fight

[twilit cobalt]

Something is actually happening in range of 60-80 steps with DPM++ 2M (SGM Uniform) sampler, but after 80 steps convergence is just jiggling a bit

[calm zinc]

I officially and on the record said, “Pretty pictures”.

[sacred jewel]

There's a boob right there

[pastel latch]

beauty standards are getting out of hand these days

[cosmic mantle]

Fun is allowed right ? 😄

[patent acorn]

Spy is boob

[cobalt moon]

boob is the real King of Civitai

[tranquil oracle]

How did Stability release such a pile of junk?

[desert garnet]

she needs shrimp jesus

[calm zinc]

I’ve had yelling matches with business partners over direction and decisions in numerous places. Then afterwards we all head to lunch.

Heated discussions are totally normal. It’s all business.

[cobalt moon]

yooooo it is not a junk, do more test carefully with different styles and different objects.

[pastel latch]

not kosher

[tranquil oracle]

SD3 is not something that, under closer inspection, reveals flaws. It's complete junk.

[sacred jewel]

https://tenor.com/view/laughing-too-funny-lmfao-haha-so-funny-jj-jameson-gif-19803792

[uncut pine]

When for automatic?

[desert garnet]

forgive for haram 😔 🙏

[signal shuttle]

I love cliptextencodesd3 node so much, single handedly changed my view of sd3

[tranquil oracle]

I tested with literally hundreds of styles, and it's super obviously junk.

[sacred jewel]

They uploaded it to Huggingface I think that's how they did it

[cobalt moon]

remember, like Pony, artist name dont worked on SD3 as they removed that out.

[tranquil oracle]

Huge mistake 😉

[pastel latch]

send workflow

[tranquil oracle]

Why? Are Stability the Taliban who remove everything that looks like art to them?
Total barbarians.

[signal shuttle]

here its super messy tho (Workflow is in the image), also it uses sd3_medium_incl_clips_t5xxlfp8.safetensors

[cobalt moon]

More like the dataset of the art are still there, just with different way of tagging it. like Picasso drawing to surrealism drawing with cubism shape

[patent acorn]

it cannot do knuckles sadly i want to make a meme 😭

[tranquil oracle]

It's fair use to train on copyrighted art. But they've even removed medieval stuff. It's really the new Taliban of Gen AI.

[tranquil oracle]

Picasso and "surrealism with cubism" is clearly not the same thing. Look at the results. SD3 is a sad joke.

[signal shuttle]

no they didn't

[sacred jewel]

Ah Adam and Steve again

[cobalt moon]

Obviously not, but it came close.

[desert garnet]

needs more shrimp jesus

[tranquil oracle]

Yes they did. This is just generic junk. You could prompt SDXL for specific paintings. SD3? All gone.

[calm zinc]

Literally right below this comment is me saying this account is an imposter. 🤦🏼‍♂️
Just go down 200px on that screen. Haha

[signal shuttle]

The purpose of a diffusion model is to make new images, not give you existing ones

[sacred jewel]

Shhh. That's a Secret

[tranquil oracle]

Netherlandish Proverbs. SDXL on the left, SD3 on the right. All completely erased and forgotten. (And it's not just this one painting, it's pretty much all of it.)

[cobalt moon]

( one could say that SD3 is not a fundamentally diffusion model lol, it is a DiT model, but yeah it is supposed to train that image to make new image )

[eager wraith]

cannot find wally

[tranquil oracle]

What? Why? If my prompt says "New York", I want New York. No difference with Pieter Bruegel. It's just insanity.

[past thicket]

I'm liking the weirdness so far

[patent acorn]

Mona lisa in the artstyle of picasso drawing
is this even picasso??

[spark quail]

youre overestimating what the models are in that case

[cobalt moon]

what's your real issue with the image actually. Netherland proverb prompt is pretty tricky for TEs

[tranquil oracle]

No, I'm happy with SDXL, it works very well.

[desert garnet]

it is,but it was when he switched to oil painting

[gentle wolf]

monalisa is probably way more often in the dataset than individual picasso

[patent acorn]

i see

[signal shuttle]

now you are changing arguments, what i am saying is its gives you real cities, but not existing paintings, that defeats the whole purpose of a AI image generator

[tranquil oracle]

It's not tricky. It's 100% solved. In SDXL. SD3 is just sabotaging it.

[cobalt moon]

You want a storybook artstyle rather than Religious Art one?

[tranquil oracle]

I'm not changing arguments. I'm pointing out blatant censorship, a Taliban-style art purge that apparently knows no limits, since it goes way beyond "copyright".

[cunning lintel]

This again, are people so thickheaded they think using styles/artists as reference is regurgitating art? It's specifying style, nothing more, nothing less. If i want a portrait commisioned in the style of picasso i tell the artist i want an portrait in the style of picosso, not a portrait that is surrealism with cubism 😵‍💫

[cobalt moon]

SD3 seemed to tried to get the shape of the object as close as possible without thinking of Picasso artstyle.

[gentle wolf]

the side effect of purging information from a model, is the destruction of similar ones. It even says so on the github page.

[spark quail]

what is there to point out, the research paper specifically stated that it essentially had a filter with its pretrained images

[hallow lion]

its not total junk and you can see the potential, we all kne wits undertrained... if its easy to train then skys the limit, its a good foundation

[twilit cobalt]

I get it. Censoring of "laying on the grass images" was due to amont of people advising to touch grass online

[drowsy phoenix]

try in style of cubism

[gentle wolf]

For example: remove all sex positions > bad person lying on grass SFW results

[tranquil oracle]

Yes, the best word for a Picasso painting is "Picasso", not a long description.

[cobalt moon]

SD1.5 is the great example, really

[desert garnet]

try with Pocassi

[patent acorn]

ill do it, here it ignores the photrealistic prompt

[signal shuttle]

You aren't understanding my argument, i am not talking about making ai images based on real artists styles, i am talking about the model giving a real existing image as output

[patent acorn]

Cubism

[tranquil oracle]

Why not? It's a painting made in 1559. It's part of art history and visual culture. SDXL can quote it. SD3 fails.

[gentle wolf]

removing existing artstyles is just making it harder to replicate, it's still possible. So why make it harder for everyone, just keep them in.

[tranquil oracle]

Of course these are fictional quotes, but they're 100% true to the spirit of the original. SD3 output is just generic medieval junk.

[cobalt moon]

Legal issues

[hallow lion]

SD3 YOGA-LORA when?

[tranquil oracle]

Why remove them in the first place? It's without culture, totally barbarian.

[cobalt moon]

You have to be realistic here that some artist can get very furious when their art is getting trained as a dataset.

If they rich enough, they could give SAI some traction by filing a lawyer letter to them

[signal shuttle]

but that would introduces a lot of legal issues

[cunning lintel]

that's not what anyone wants...

[drowsy phoenix]

yeah, i think this model is a finetune of its own styles

[tranquil oracle]

What? Legal issues? Quoting a 1559 painting? I beg your pardon?

[gentle wolf]

you can still copy style with photoshop. There's no change here. If you commercially release copyrighted stuff, you're still as liable as before AI

[patent acorn]

i figured out it can do lightnign mcqueen lol

[spark quail]

so many people misunderstand image models lol. youre not going to prompt an image that the model was trained on. its going to create something new -- thats what they do.

[signal shuttle]

Look at the message send by RLX, then you would understand what i am saying

[tranquil oracle]

Look at the image. It creatively "quotes" the original

[cobalt moon]

from what I understand the context here, is @tranquil oracle tried to replicate the old existing medieval image as close as possible or tried to replicate the artstyle for the new one

[twilit cobalt]

and in order to do that it should learn all the backbone concepts very well

[gentle wolf]

hot take, all art is stealing from somewhere. 😉

[tranquil oracle]

It can discern between Bruegel paintings, prompt for "The Hunters in the Snow", you'll get that.

[sacred jewel]

My man Simon Stalenhag is gone
Steampunk is also hard to get

[spark quail]

right, and theres something different each time. thats the point, youre not going to get new York everytime. its going to be different, and even more wildly different if there are more city images trained

[tranquil oracle]

It's the same when I prompt for New York. I want something that really looks like New York, not something new.

[cobalt moon]

... skyscrapers New York or "medieval" New York? bro

[gentle wolf]

Sometimes you want XY in New York, then it should look like New York and not like Mars

[tranquil oracle]

It'll still have the Empire State Building, not some generic office tower like SD3 in the case of Bruegel.

[spark quail]

if you wanted the same pictures of new york everytime, the tool you are looking for is google images, not an image model

[cunning lintel]

Ok, guess we agree that it's a defect the possibility to refer to styles/artists hardly works in 2b 😉

[tranquil oracle]

I want variations, just like the Netherlandish Proverbs variations I posted. If you asked someone: Which painting is that? they could easily guess it. Same with New York.

[spark quail]

now if you wanted to create a scene specifically of new york with something in it, like godzilla or something, youre probably going to need a lora for that

[twilit cobalt]

Image models are not just blending between "compressed images", it's actually concepts learned in the process of training. But to prove that to general audiece, models should exhibit this understanding of anatomy. The thing with art, that it always has been human-centric, because it's how people communicate visually, with the art

[tranquil oracle]

It's seems like a deliberate defect, caused by some absurd "policy"

[pastel latch]

interesting so addition of encodesd3

[sacred jewel]

SD 3 knows Man Baby is a thing 🤪

[patent acorn]

yeah. those misinfo are made up by anti AI

[cobalt moon]

is this what you tried to getting ( from Hugging Face )

[tranquil oracle]

SD3 is just total trash when it comes to style adherence, detail, creativity of composition, lighting, saturation, everything.

[patent acorn]

well its a base model so

[tranquil oracle]

(SDXL on the left, SD3 on the right)

[bitter hearth]

where did those negatives and stuff come from ?

[twilit cobalt]

The initial reaction is rooted in fear of being replaced by the technology. But essentially I still stand on the ground, that it's only tools for expressing creativity, not the replacement

[tranquil oracle]

... so why regress behing the previous base model, SDXL? Absoluteluy no reason for that.

[desert latch]

BIG THINK: Maybe some SAI stuff are from medieval and think that train artist ] 1515 can cause copyright infringement

[patent acorn]

im generatin for fun. do i generate slops? lol

[signal shuttle]

the cliptextencodesd3 node?

[bitter hearth]

no the stuff like 5/5 stars...

[twilit cobalt]

Every creative intent usually comes from desire to have fun

[cobalt moon]

bro forgot the 16 channel VAE and the way of your prompt, it is better if you can show us your prompt on this. Details maybe, Style Adherence, what are you even tried to create?

[signal shuttle]

this github https://github.com/DataCTE/sd3_prompting

[patent acorn]

ok then someone shouldnt let those safety team cook
wait does those have anything to do with removing the artstyle stuff

[cunning lintel]

Only SAI can tell, i can also imagine it's in the same bin as how words weight weird, the long prompts somehow caused certain words to override all else, and maybe demand specific incantations similar to whatever cogvlm produced; too strong a reliance on one exact sentence, too little on basic captions. The model is too much tuned to very specific phrasing of inputs. At the same time other words, like artists and specific styles, were hardly learned while training. Either way the real issue his: how did this release pass quality assurance?

[tranquil oracle]

"portrait of x by y and z". super simple, allows for endless mixing of styles, works perfectly well in SDXL. Totally broken in SD3. For no reason.

[shell plaza]

Anyone tried the fresh canny control net for sd3?

[alpine axle]

1024 version of the controlnet also just dropped

[sacred jewel]

Nice

[shell plaza]

Insane developing speed btw

[pastel latch]

not seeing much of a difference though

[sacred jewel]

How do you know this for sure?

[tranquil oracle]

As I said: the Gen AI Taliban.

[signal shuttle]

you are using the same prompt in all fields, try putting all quality tags ontop while a mix of quality tags and a prompt and the last just a prompt without any quality tags

[sacred jewel]

Gotcha 👀

[cobalt moon]

... I mean SD3 handling of prompts is dramatically different from SDXL. T5 and CLIP are presented you know ( assuming you use the individual TE + nonincluded SD3 )

T5 never like an one-shot sentence if I wont wrong. It is much better if you split up like. The portrait of X. The portrait was drawn in Y style. The portrait have simple background

[tranquil oracle]

I'd rather use Java bindings for tensorflow than this uselessly verbose type of prompting

[patent acorn]

the fact that its possible to generate dark humor content lol

[tranquil oracle]

Bad taste model. (Again, left SDXL, right SD3)

[cobalt moon]

like Lykon said, it is skill issue if you don't wanted to change your prompting style for SD3 just to see if it is better. And you here baffling your prompting style can't make a thing in the new model.

[tranquil oracle]

He said I should talk to SD3 like I'd talk to a lover 😅

[cobalt moon]

More like you should rather use LLM than your brain's language module for prompts

[signal shuttle]

true, we changed our prompting style a lot when switching from 1.5 to SDXL

[pastel latch]

ill play a bit more with it see if i can find the sweet spot for this workflow

[sacred jewel]

Do you know why there are 31+ flavors of ice cream?

[signal shuttle]

ok tell me if you manage to improve it

[cobalt moon]

Hm. I can't say things for sure in SD3, especially prompt adherence/coherency since everyone have their own standards on that.

[compact forge]

he murdered them

[mortal kite]

The Cropduster

[tranquil oracle]

Sorry for the rant. But SD3 is really infurating.

[signal shuttle]

Have a burning car

[cobalt moon]

Yeah nobody give a thing on you still using SDXL if you dont like changes.

[mortal kite]

do you want to know how damn long it took to get this image? It was ridiculous

[cobalt moon]

plus for real though SDXL is just sufficient for most works.

[desert garnet]

needs more santa

[cobalt moon]

truly a firepost

[sacred jewel]

https://tenor.com/view/more-cowbell-dance-dancing-funny-gif-19483045

[violet gorge]

what happened to stable cascade 😭

[desert latch]

BTW I remembered there were server bots on this service,

[cobalt moon]

researcher join SAI, research Cascade, release it, then off they go. It was never intended to hit the popularity anyway by SAI

[cunning lintel]

All I hoped for was better prompt understanding, SD3 doesn't have Ideogram level prompt understanding, but suddenly it is nothing but a minor inconvenience. Well played SAI 😉

[errant dust]

You know, I thought the whole theory of perturbed layers to undo some later safety training to be pandering to some absurd conspiracy theory, but after trying the perturbed model I believe the theory has teeth

[spark quail]

the research paper actually found that the devs found no improvement in "better" images when using t5 (altho idk if they thoughy about the average sd user's prompting skills) so it should be fine to just not use t5. it makes a huge difference if you wanted to prompt for text in the jmage tho

[pastel latch]

looks about right

[balmy robin]

what are cascade requirements? tried in colab and failed 😦 t4 15gb vram. maybe i should try forcing fp16?

[desert garnet]

when u dont get bananas on xmas

[mortal kite]

The prompt understanding isn't really that great IMO. It still doesn't really understand that well unless it's a basic simple thing

[hallow lion]

sd3 is like that rough cut george lucas showed his buddies before star wars was finished, eveyrone laughed only spielberg saw the potential

[violet gorge]

Does anyone know exactly what the difference is between Stable Cascade and Diffusion models?

[errant dust]

The discussion can be found here: https://new.reddit.com/r/StableDiffusion/comments/1dfuicw/perturbed_sd3_experiment/ But I will show a before and after all the same. Note that Perturbed does not universally fix the Elephant people, but it sometimes does to a degree

[mortal kite]

people still keep trying to cope

[cobalt moon]

although again there is possibility of SD3.1 but at this point they ( SAI ) will just rather fix the other 3 than another 2B

[spark quail]

the architecture of the models. read the papers if you're interested, theyre fun reads.

[hallow lion]

cascade has a sausage in it, hence it works better

[patent acorn]

4/5 ★★★★☆, photorealistic, (Monalisa:0.9) in real life
it seems to be impossible to make photorealistic mona lisa. RIP

[hollow shoal]

/generate Create a tiktok avatar in 3d cartoon of a young black boy

[errant dust]

This was an original image with the base SD3, and has a very long description since Lykron was telling everyone the issue was their competence. You'll note it does not even display the warrior the description starts with. The second image is the Perturbed version. Far from perfect, but it is unquestionably an improvement.

[patent acorn]

in new york. XD

[hollow shoal]

what are the prompts to generate images here pls?

[faint shard]

Lol

[hallow lion]

anything you like

[patent acorn]

its on local lmao
try hugginface space

[errant dust]

People are generating from their home PCs

[faint shard]

This isn't a generate channel. Like midjourney.

[patent acorn]

MIDjourney

[desert garnet]

sd3 mid

[bitter hearth]

steven low

[patent acorn]

MIDalle

[pastel latch]

getting better

[balmy robin]

the model is here and example workflows in the folder https://huggingface.co/stabilityai/stable-diffusion-3-medium

[desert garnet]

[bitter hearth]

@desert garnet

my random creature prompt only produces shit now

[desert garnet]

needs more muscles 😔

[desert garnet]

monalisa if she lived in San Francisco

[patent acorn]

fucked up monalisa lmaooo
what about venezuela

[desert garnet]

idk why but she now has a hairy hand

[fathom pelican]

when I try to use the perturbed model I get the error Error occurred when executing CLIPTextEncode:

'NoneType' object has no attribute 'tokenize'

any idea why?

[patent acorn]

ongggg

[bitter hearth]

@desert garnet ?!

[compact forge]

mona lisa as a slutty slutty _____

[patent acorn]

no difference

[compact forge]

SD3 is saying she was always slutty

[desert garnet]

dont peek

[patent acorn]

oh

[livid plover]

wow sd3 is so safe and good, it is hot like a dumpster fire!!! I love it xD

most advanced model out there))

[desert garnet]

exclusive footage of biden at congress

[bitter hearth]

[livid plover]

those pictures aren't safe for you guys, watch out!

[jolly abyss]

Solved it.

[wind basalt]

Is there any fine tune of sd3 yet that is actually good