🆕｜sd3

[dull horizon]

Super neat!!

[sage burrow]

How much ram? I have 16gb
Though it's very likey this is what is slowing me down!

[silver sluice]

results were inconclusive, take those measurements with a grain of salt FYI

[torn wharf]

gonna be hard to upscale very high with fp8

[silver sluice]

32gb of ram but when i had 16gb of ram i was still averaging 30-40 seconds per render. i9 13th gen CPU if that helps you compare

[coral sable]

it's not working just like you said earlier, but we still had to try

[sage burrow]

i7 here

[silver sluice]

yeah there's a lot of factors, im going to mess with the diffuser nodes for comfy and see if i can get sd3-flash to work for me

[sage burrow]

Please tag me when you post results 🙂

I'm really curious what the min requirements are for that, I didn't see it on the download page

[torn wharf]

the biggest factor is vram. if you overflow it, the system has to buffer with system memory and that'll be a magnitude slower

[raven fern]

im gonna start my own manga :3

[mortal kite]

This thing does not have as great of prompt comprehension as people think

I'll say "object <X> is on the left" and it still puts it on the right quite often. And it always makes spelling errors or misses a word when putting in text

[sage burrow]

I tried, using realistic photo style art. Was not popular, aside from my close friends lolol

[sage burrow]

You mean flash right?

[raven fern]

i was actually kidding, but yea, prob wont be popular, cause it's not just about images anyway, you have to then make something that makes sense with story, character development, setting, etc, a lot of work

[raven fern]

smol boi :3

[mortal kite]

These body standards are getting too high

[sage burrow]

via online flash, exact same prompt, just different seeds

[sage burrow]

How?! Too bad about the bg image lol

[mortal kite]

That was on purpose LOL

[silver adder]

sd3 2b

[sacred jewel]

Thanks>>>

[naive jewel]

This uses 'dpm_adaptive', this means it ignores the set number of steps and adaptively determines its own. Check the console window it is probably really using >30 steps not 4.

[sage burrow]

You're welcome 🙂

[sage burrow]

Oh!! I'll have to do a test where I put it at 30 (my usual), then try it at 4, and see which is faster. Somehow though putting it at 4, reduced the time the generations took by 75% about

[limpid drum]

I think I am doing OK with the LLM aiding the prompts...

[naive jewel]

The speed of dpm_adaptive is usually more dependent on the cfg.

[raven fern]

missed the L

[sacred jewel]

CFG 1

VERY FAST == 3secs

[sacred jewel]

CFG 4, same steps of 4

9 secs

[sacred jewel]

Same settings, CFG changed to 8,

15 secs

[sage burrow]

I think he's cheating, but it'll do

[raven fern]

finally some hard muscles Becky 🙂

[raven fern]

i have no idea about this game

[mortal kite]

This is the best I could get so far...

[sacred jewel]

Base Model

[sage burrow]

Took WAY too much effort

It's a horror story, that's all 😉

[sacred jewel]

And with the extra fingers, she can smash you even worse than a regular bodybuilder

[mortal kite]

Cheat day

[torn wharf]

one strong and one skinny leg. you know what that means.

[sage burrow]

Zombies, by Bosh
Bosch probaby isn't the best artist to name until the model is more perfected 😉

[silver adder]

🙂

[sage burrow]

So what are the content rules of this particular section again?

[mortal kite]

Zombays

[sage burrow]

I'll just say that SD3 is really good at horror.

[mortal kite]

He's behind me, isn't he

[mortal kite]

very nice

[sacred jewel]

Is that the famous Calgon?

[lusty wren]

Is there a good guide anywhere on how you're supposed to prompt differently with T5 for stable diffusion specifically? I know SD3 is better, but I'd like to leverage the new quality as much as possible.

[mortal kite]

well, I tried

[sacred jewel]

Trying to get the hands right... same seed... prompt tweaking

[upper snow]

well, there'd have to be an Unstable Diffusion 1 first, but anyone who's been here since the start knows how that turned out lol

[mortal kite]

This is closer to what I was trying for...

"Mooooom I'm reeeaddyyy"

I think they want you to spoiler any image that is bad gore

[sage burrow]

LOL that is also my question, but decided to save everyone lol. I was thinking of posting them on my alt DA account and the link here lol

[sacred jewel]

So, interesting experiment.

using the singleCLIPTextEncoder with SD3,

This Prompt:

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, 1woman \nbored, annoyed, unimpressed,\ns, nipshiny_skin,\n\nsolo, ear_piercing,\n\nby rakisha, by totesfleisch8, by meesh, extremely detailed background, BREAK, ↑ trending on (artstation:1.2) ★★★★☆ ✦✦✦✦✦, painted concept style

and this modified prompt removing most of the "stuff"

generates literally identical results...

1woman bored, annoyed, unimpressed, nipshiny_skin, solo, ear_piercing, by rakisha, by totesfleisch8, by meesh, extremely detailed background, trending on artstation,

[mortal kite]

oh man, this would have been perfect except for the AI messing up the window

Hey that's pretty good

[sacred jewel]

The one thing that DOES MAKE A DIFFERENCE?

Ready for this?

"trending on artstation"

LOLOLOL

amazing... didn;t expect that

[sage burrow]

all the same prompt, but different seeds

[sage burrow]

That is quite the difference!

Excessive... ah well lol

I think they forgot some bits with their censorship 😉
Or they crawled my MJ account, one of the two ROFL

[mortal kite]

omg this actually freaks me out

[bitter hearth]

skurry skurry

[kindred mica]

Not safe for sleep

[sage burrow]

Prob want to spoiler that one, don't want SD to get into trouble with discord

Thank you 🙂

[bitter hearth]

how do you spoiler something

[sage burrow]

phone or computer?
Start to post the image, but before you hit send, there's this little eye on it that you can click. Hover over it to see it

[bitter hearth]

oh! good to know!

[mortal kite]

When you look at the closet door at night

[sage burrow]

Discord (overall, not just here) has some strict rules about public being able to see certain things. So in very public places, spoilers are good :>

[bitter hearth]

Nice information all, thank you

[mortal kite]

I think I see that, I walk the other direction

[bitter hearth]

He is just a humble demon trying his best!

[sage burrow]

But he has cookies! 😄

[mortal kite]

I'm trying to swap the girl and the zombie but the AI ain't having it. So the zombie is in bed and the girl is looking in . I guess this sort of is that.

[mortal kite]

Well, I mean, pastries. Ok I'll listen

[bitter hearth]

nom nom nom

[mortal kite]

Guy on the right looks like he could actually fly

cheat on his taxes?

[mortal kite]

SD3 is really really good at pixel art, it seems to listen better with that style even

[mortal kite]

Best SD3 I've seen in a while

[sage burrow]

Completely innocense... 😉

[sage burrow]

So checkout my twitter account 😉 @mortal kite

[mortal kite]

I like the glowy alien lol

[sacred jewel]

ONLY difference is "trending on artstation"

[bitter hearth]

Gimme prompt

[upper snow]

uhh, are you sure you aren't just getting an unconditional image? the only things in that prompt that would actually be likely to do anything would be "woman", which the model is very likely to generate anyways with no prompt.

[sage burrow]

FOr ultra the SD discord Artisan section (you do have to get a subscription though). Or Glif for nearly ultra

[mortal kite]

of vhs video still of a teen girl in a dark hallway at night. The little girl is running toward the camera screaming. A hideous tentacle monster zombie is behind her chasing her. Film grain. Vhs tape.

For some reason it still made a little girl, not a teen

Negative: "ugly, disfigured, malformed, low quality, watermark, poorly drawn, bad proportions, out of focus, bad anatomy, disproportional, amputee, missing feet, missing foot, missing fingers, missing arm, watermark, signature"

[upper snow]

pony tag strings definitely will not do anything to SD3 and it absolutely isn't going to know furry artists (until my team fixes that problem of course)

[sacred jewel]

How the heck would I know? I am just a normal regular NON model expert, dude 😛

[bitter hearth]

You wrote little girl..

[mortal kite]

Stephanie has her first visit to the I.T. department

[sage burrow]

The only reason those are still stuck to so many prompts is that for experimentation between SDXL and SD3, I like to keep the EXACT same prompt, even the irrelevant bits.... (so that was my fault)

[bitter hearth]

i cutefied it

[upper snow]

i never started using pony because i already have enough trouble with clip context windows without a 42 token string at the start of my prompt

[mortal kite]

oh you are right I still had "little girl" in another part of that prompt

[upper snow]

baking quality alignment into the model from the start is a mistake imo. it's so easy to do after the fact, and you can do it in a much less intrusive way. i managed to get decent quality alignment with a single negative embedding that's 6 tokens.

[faint breach]

what even are the point of style up tags?

[mortal kite]

its decent. Still gets confused if you have multiple subjects in the scene

[sage burrow]

Pony took me FOREVER to figure out after I was used to 1.5! It's so completely different, or something. I mean even with the scores

[bitter hearth]

I don't care about anatomy or artists names in it, so it's basically the best model for me atm lmao

[sage burrow]

People say it was a training error or something. But also if you use some, but not all, you can get different results

And the ultra fine detail!! ❤️

[upper snow]

quality alignment. if you train a model on a lot of booru images... well, they almost all contain useful information, but most of them look like shit, so the mean of the distribution is going to look like shit too. so you want something to bias it towards images that look good.

the need for them is probably a lot greater on pony because the uncond model is completely dead

[sage burrow]

Dalle 3 is exceptionally good if you want to make images for loras 😉

[faint breach]

but with dpo, if you're prompting all the low quality stuff to be ignored, it's just not useful then. it might as well just not be in the set then right?

if you want the model to learn low quality why not just use low quality

[upper snow]

no, the images still have useful information. even better -- if it's in the model, you can target it with a TI embedding, then put that in the negative. then, that will push you away from the low quality data and towards everything else

[faint breach]

yeah maybe i might make a pony embedding instead of using a default prompt

i experiment with the pony lineage some times but its annoying when you're trying ot avoid nsfw

[sage burrow]

I recommend either Autismmix or one of DucHaiten's pony models, straight up pony is too difficult to get good results

[bitter hearth]

At least on the api I get a very high variation of colors, if i want stuff super contrasty i can put "shot from iphone" or something else if I don't

[mortal kite]

Her first nightmare

[faint breach]

Thanks for the insights

[sage burrow]

I prompt the clothing in those cases, works 80% of the time. I have an entire gallery of shorts failures lololol

[mortal kite]

doesn't Pony have "option_safe" or such, it won't create lewd then

[upper snow]

another thing that caption dropout helps with... if most of the distribution is NSFW and that's represented in the uncond model, you'll be pushed away from NSFW pretty hard unless you ask for it

[bitter hearth]

"rating safe" is a tag that avoids nsfw on it, never really had a problem with that tag in place, just have in mind what your prompt can do, if you're trying things that could closely relate to nsfw you can pump the weight a bit or do some "rating explicit/questionable" in negatives

[faint breach]

It works better on base pony than it does on the realism variants i've tried. i think it has an underscore too

[sage burrow]

I don't know if it matters but "rating_safe" or rating... nm lol

[mortal kite]

rating_safe yes

[bitter hearth]

Well, those merges usually do that ye, break the original model

[sacred jewel]

Those all look conventional not fission... 🤭

[mortal kite]

sorry I had to

[sage burrow]

this lack of refereance image stuff is making me sad 😦

[faint breach]

lol i'm not going to keep complaining about pony tags this is sd3

[mortal kite]

I dont care about politics but it knows how to draw them so well 😄

[faint breach]

i thoguht image 2 image workflows worked fine. do you mean contorlnet and ipadapter? yeah only diffusers file format that isn't compatible with everything else is available

[sage burrow]

I remember on the competitor's forums, that the staff were complaining that by far, the most prompted thing/person was Trump ROFL

[sacred jewel]

Anyone have a SD3 with ControlNet available to share? ANd where the heck are the controlnets anyway? LOL

[sage burrow]

they were on huggingface a day or 2 ago!!! (404 now)

[faint breach]

i generally tune him out. i've had a lot of practice living on his doorstep. He was funnier as an entertainer clown man on The Apprentice. Innocent like ozzy crying for Sharon is.

Everything changed when the birther nation attacked

[sacred jewel]

I wonder if Hiroshima and Nagasaki do anything ... hmmm brb

[sage burrow]

furry fail! 😄

[faint breach]

were images of either made popular ever? i know the aftermath had images published. They sure didn't set up cameras to take pics of the event

[sage burrow]

challenge accepted!

[faint breach]

I'm pretty sure both cities are rebuilt today too

[faint breach]

there was only the bomber that had the range to get there. they didn't exactly have a naval fleet with air craft carriers

[sacred jewel]

Enola Gay?

[faint breach]

i guess these are the only pics of the event

[faint breach]

yeh the super fortress mofo

[mortal kite]

Ooops damn portal opened again

[faint breach]

looks very similar to the image i linked. the two japanese detonation photos clipped together

[sacred jewel]

Bockscar?

[mortal kite]

ah good 'ol demoncrats amirite

juuuust kidding around 😄

[bitter hearth]

Someone tell me what the AI wants to code

[sage burrow]

I'm more than happy to make fun of all political parites 🙂

[mortal kite]

made this the other day lol

[raven fern]

time for snu snu?

[mortal kite]

nah that's enuf of snu snu for a while ain't it?

[sage burrow]

is this what your mystery code made?

[mortal kite]

Looks like me preparing for work

[bitter hearth]

my code was a fail then cause this gen didn't work how I wanted

[bitter hearth]

The trump lovers are in chat

[raven fern]

i personally dont love any political person :3

[sage burrow]

It's extremely annoying to me, that so many folks I talk with blame the more nsfw censorship, on the political party that they personally don't vote for. I mean the people who feel that censorship is bad.

[mortal kite]

nah here we go

[dull horizon]

Is pretty unifying honestly on people being afraid of AI and regulating it. It's pretty non-partisan -- around the entire world, honestly

[bitter hearth]

Me on a good day be like

||Copium||

[sage burrow]

Just before SD3 came out is when I kept seeing/hearing that

[mortal kite]

wtf

[livid stirrup]

i blame taylor swift

[placid swallow]

Taylor is know for hating images of her laying on grass.... 🤨

[sage burrow]

Gemini is my fave help me figure out stuff bot 🙂

You get a 6 months free trial if you have google drive premium or whatever it's called

Some people say Claude is better though

[bitter hearth]

Literally me
|| after few years of self care and gym maybe||

[raven fern]

is that a chad?

[bitter hearth]

You guys should watch the Nvidia presentation that happened this year

[sage burrow]

It is! Recently, any online news about ai, more than a week old is irrelevant quite often!!!

[bitter hearth]

that is one pissed off raccoon

[silver adder]

whatever this bugatti is doing under the sea

[bitter hearth]

Just search for jensen huang keynote computex 2024

[placid swallow]

just two trash pandas doing trash panda things

[sacred jewel]

MAGATs cult is alive and well 🤭

[bitter hearth]

Can you guys try to get animals wearing masks? Had a hard time on that with api

[placid swallow]

like a dog wearing a cat mask?

[raven fern]

@kindred mica were you demoted? i dont see the pink :3

[placid swallow]

best i can do is a red panda watching miley cyrus while eating his salad

[bitter hearth]

Some pretty amazing AI stuff in there, it definitely shows how much more insane things AI is doing in the real world, our little image gen fun is small compared x3

[bitter hearth]

Yeah anything, even tried a generic "black mask" it wasn't working well

[placid swallow]

hmm i got it to do a astronaut chef's helmet

[bitter hearth]

I see a lot of this in AI, models placing characters face outside the window cx

[mortal kite]

ya..ugh

[placid swallow]

ya I dunno I put red panda wearing a ski mask to commit crimes and it just makes him happy in all the photos no mask

[sage burrow]

and my failed gen (not an animal)

[bitter hearth]

Not an animal, it's nightmare fuel

[placid swallow]

well he's got gloves now

[bitter hearth]

happy to commit crimes, doesn't need a mask we all know red pandas are horrifying creatures now

[placid swallow]

lol wearing a paper clown mask just put him infront of a circus tent

[placid swallow]

typical sd3

bingo skunk wearing a gas mask while robbing a bank

[placid swallow]

looks like a rainbow 6 character

[sage burrow]

does sd3 have inpainting yet? Use that for the mask?

[raven fern]

sd3 barely has a license, let alone inpainting :3

[low stone]

scary snake guy

[low stone]

even scary bee creatures need to go shopping.

[placid swallow]

omg it's creepy bee guy can we take a picture with you

[upper gust]

hammerhead shark

[placid swallow]

missing a few chromosomes in that hammer there

[sage burrow]

🤣

[placid swallow]

I think my sd3 has gotten lazy and started using photoshop to make my prompts

[sterile pendant]

Differential diffusion works just fine with SD3. Just drop in the node between the model and ksampler and make sure to blur the mask

[sage wharf]

Create A luminous Russian girlfriend, her skin shimmering with delicate opalescent hues, stands amidst a misty Rome in a mesmerizing oil painting. The soft, dappled light filters through the ancient buildings, illuminating her flowing silver hair and intricate, iridescent gown. The artist's meticulous brushstrokes capture every intricate detail. This breathtaking masterpiece transports the viewer into a mystical realm of ethereal beauty and wonder.

[raven fern]

ah yea i haven't tried that yet with sd3, will try tomorrow

very sexy

for sure

[tough oriole]

I heard someone made a good lora with SD3. Is training going well?

[raven fern]

i like the sketch lora

[sage wharf]

full body of robert pattinson walking through the fire and flames, handsome, athletic, cowboy hat, desperado style black suit, using a yamato katana, the lonely shepherd, kill bill origins, natural look, eastman kodak 35mm vision camera negative film stock, Vision 500T 5279, anamorphic panavision lenses, twilight breaking dawn, red dead redemption aesthetics, sergio leone spaghetti western style, tenet time entropy inversion cinematography by christopher nolan, directed by quentin tarantino

[raven fern]

i remember that movie

[placid swallow]

i like how it has the guy in the back ground staring like "wtf" is that guy wearing makeup

[raven fern]

"a clown on the street? what the hell man..."

[sage wharf]

Please teach how to generate this images in this server by mobile phone

[placid swallow]

https://discord.com/channels/1002292111942635562/1237461679286128730

no click on the artisian-faq and read it to find out how to use the server to generate images

[fallow torrent]

SD3 vs SDXL

[red nexus]

nm found it thanks!

[placid swallow]

you're welcome

[rain current]

https://huggingface.co/spaces/jasperai/flash-sd3

[patent magnet]

Flash SD3:

[upper snow]

single concept lora paper authors on life support, sks now has meaningful information encoded on it. although that is DEFINITELY not an SKS.

[noble coyote]

I'm using Ollama with IF_AI_Nodes - they seem setup to do 3 paragraphs ... it wasn't down to me - I mean, I didn't request 3 paragraphs! Hope this helps?! Havaniceday 🙂

[sacred jewel]

Anyone can tell me how to run Flash SD3 on ComfyUI locally?

I tried it as a LoRA and it does not seem to be working 😦

[upper snow]

i guess that VLM captions make it much better at dealing with complete nonsense. I just put in the WHAT IN FUCK IS DONE TO THIS POOR RIFLE copypasta and got this. these are overall very coherent guns, and they also reflect the prompt I suppose

[faint breach]

i've never used any of these "unique" tokens like sks and my loras have always worked and train easy. i've often found that youtubers relay a lot of information confidentally, that's just not how it is in practice.

Unique in quotes since so many loras use them

[upper snow]

i actually did an analysis on tokens to find the one that had the absolute least impact on the direction of the resulting clip embedding. i found that the gender neutral couple emoji 💑 had the least impact.

[faint breach]

emojis as tokens that sounds fun

[upper snow]

t5 doesn't support tokenizing emojis though sadly

[faint breach]

i tried prompting with emojis in t5. it wasn't as good as i thought it'd be. i thoguht it would

[faint breach]

can't we just train loras with only the clip layers and not t5?

[upper snow]

don't train text encoders with sd3, i wouldnt even do it for loras

[faint breach]

oh yeah because transformers or something

[upper snow]

because MMDiT specifically. it is unique to this model having its own space for conditioning

[rich venture]

Prompt?

[rich venture]

Looks great

[faint breach]

i was just thinking. A bunch of researchers that left stability built the mmdit architecture? Where did those guys go?

I hope their next project doesn't go like RWML did , and it's closed forever

[upper gust]

an photo of Female Knight standing tall, clad in gleaming gilded armor adorned with an abundance of ornaments that shimmer and shine. Her armor is a masterpiece of craftsmanship, boasting intricate details and engravings that seem to dance across her imposing physique. In one hand, she wields a massive sword, its blade etched with runes that pulse with energy. The sword appears to be an extension of her unyielding spirit, forged from countless battles and victories. She wears an towering shield in the other hand Castle background

[sterile pendant]

Depends on if you are you trying to train in new data/concepts that don't exist in the base model or not. If not, then you actually only want to train the tencs to get better at mapping to the existing data/concepts in the model.

[upper snow]

HAHAHAHAHAHA. guess what prompt got me this

[placid swallow]

gorilla war fighter

[upper snow]

it's the fucking navy seal copypasta. gorilla warfare.

[faint breach]

i haven't explored copy pastas too much. i been doing poetry and song lyrics as prompts haha

[placid swallow]

Batman: EU

[faint breach]

tracksuit makes him russian batman.

[upper snow]

Own a musket for home defense, since that's what the founding fathers intended.

[faint breach]

when the musket first came out, that mustve been like dalle 1 for them. just so remarkable

we're all making laws for ai today that are going to be grossly misinterpreted in 200 years

[teal fossil]

This one is excellent. Which sampler are you using? Is it upscaled? If so - how?

[vernal vine]

Background: A dark, foggy night scene.Main Image: A ghostly silhouette centrally positioned.Text: "HauntedMinutes" in a creepy font, placed at the bottom of the image.Colors: Use dark greys, blacks, and white or light blue accents for ghostly elements.

[placid swallow]

no upscale https://up.reselltek.com/u/xSudOV.png

[noble coyote]

LLAVA and/or ZEPHYR. I am using IF_AI_Nodes and the two input methods - i2i or t2i. Each method returns a prompt which is then fed into Clip ... Olllama examines the input image and makes a prompt; for t2i Ihave used this phrase "You are a creative director and I am going to give you a project you need to work on. I want you to come up with generative ai prompts in grammatical sentences - and in a natural-language format related to this project. Develop entirely random prompts on entirely random subjects. Be as anarchic and subverting as possible with combinations of subject and theme. Start with "PROMPT = " This will be used as a generative ai prompt. Keep its length to 77 tokens."

[teal fossil]

Then your prompt was great and the noise-Gods were on your side.

[wraith star]

may I share some SD3 models here？

[placid swallow]

i will allow it

[teal fossil]

Go ahead! Also share how they were trained, please. 🙂

[placid swallow]

I'd vote for that

[teal fossil]

I still can't comprehend why politicians (nr.1 aim for Fakenews and political deepfakes) were not purged, but everyone else was.

[patent magnet]

I've heard he is a very funny guy

[upper snow]

i think most celebs are mostly in through clip base knowledge and politicians have so many pictures of them that it is impossible for the model to not learn their likenesses

[patent magnet]

Yeah, tragic and sad... 😕

[upper snow]

surreal. bee movie script

[placid swallow]

not sure it knows who Harley Quinn is either

[upper gust]

harley quinn at home

[placid swallow]

even got the Temu Harley Quinn

[ashen pilot]

#🆕｜sd3 On a gloomy day, a little dog with a wine glass on its head comes to find a little girl. The wine glass is filled with milk, realistic style.

[wraith star]

https://www.shakker.ai/modelinfo/090553f4ee5843fca21949fae75d7b55
you can have a try

Version Description:
Base film: The 10G version includes clip and T5 version SD3;
Fusion: Fantastical fusion, SDXL Realistic Scene v3, paintinglORA (fun version, generated models? indeed generated. Is it effective? A fantastic answer!)
Workflow: Comyfui
ps: As long as people are not painted, there will be no terrible body parts!

you need to download to comfyui

[harsh rock]

If we shifted from women laying on the grass to cats, I have lots more to say:

Discord wobbles in a weird way when I try to upload a picture. What's that??

[patent magnet]

It's because of the timer

[harsh rock]

The one above is the upscaled version with SUPIR. This is the original generated by SD3:

[patent magnet]

nice one

[patent magnet]

Supir and others cause it to loose some detail because of the SD3 VAE, but I prefer that way for several use cases. Even if the SDXL based upscales lowers the tones and details .

[harsh rock]

Discord compressed my SUPIR version. It's more detailed than that.

[patent magnet]

And there is that. Yeah.

[harsh rock]

What's the difference here? Which is which?

[patent magnet]

I mean, the discord compression makes things worse. But I've opened your 2 images and the Supir one looses quality in details and colors. And I still prefer the lower quality (upscaled) one depending on the situation.

[cosmic crown]

discord does not compress images, you have to click "open in browser" to view the original

[patent magnet]

I don't know anymore, there was a time it did compress and then it didn't. I don't care. I did open his images in the "Open in browser" and compared.

[placid swallow]

adobe super resolution on the right

[somber jackal]

A girl, walking on the grass, full body, smiling

[sonic matrix]

A girl, walking on the grass, full body, smiling

[floral apex]

A girl, walking on the grass, full body, smiling

[harsh rock]

You are right. Open in browser shows the 4K image as it is

[noble coyote]

"B.L.T." 😄

[zenith spear]

hello

[viscid pivot]

Oof Monkey also resigned from stability

[bitter hearth]

Got a link or something where did you see it

[viscid pivot]

Yes I'm trying to copy the link from my phone from Reddit, it's a post by comfy and monkey is commenting

[bitter hearth]

I'll try and check on Reddit, is it the same one with that comfy post?

[viscid pivot]

No it's a new one called "next step for comfyui"

Man I hate that fkin app, every other one easily let's you generate a link

[wet sparrow]

step 1. hire Comfy guy
step 2. Comfy guy leaving saying releasing the best model is not SAI management priority
step 3. part of SAI team leaving to Comfy

kek, love this guy, hope comfy will prevail

[viscid pivot]

https://www.reddit.com/r/StableDiffusion/s/AVRHh55dtX

[viscid pivot]

Yea SAI really killed themselves. It's like management tried to destroy SAI for some reason. If they lose all their Devs, how are they gonna hire new ones especially with their bad rep lol

[bitter hearth]

[viscid pivot]

Just waiting for the bankruptcy news now. I'm no expert in this field, but I'm expecting nothing else at this point 😂

Hopefully then someone will come out and really say what the hell went wrong

[wet sparrow]

so many strange / bad decisions in a row yes
I would assume that it's all about trying to find some funds for SAI, just extremely poor execution

but yeah, we don't have too much information on that

[viscid pivot]

Poor is an understatement of titanic proportions. It was criminally poor execution, sabotage one might say. I am not sure some average clown like me could have made worse decisions lol

[viscid pivot]

They also hire Devs who fully believe in open source, promise them the world and then force them into corporate shenanigans. What could've gone wrong?

[fervent stirrup]

a pig

[frigid mango]

BUT! They seem to be very confident about comfy and the future which gives me 2 ideas: New Base models from other companies are going to be more integrated AND they know how the SD3 model works and may have an idea how to fix it

[storm saffron]

The problem with anything to do with SD3 is that license that's still not been clarified

[bitter hearth]

lmao shrek lost some weight xD

[craggy ridge]

He's been working yes 😁

[viscid pivot]

I'm not sure they are interested in fixing it, it's not their problem anymore. I think they will just add basic support for it and be done with it. There are way more promising alternatives out there that should be focused on

[frigid mango]

But they work on it and I really think, there is something coming. Nevertheless: There are plenty of alternatives out there. And they are getting better. Just imagine them having community support and they will be as good as SD in a miute

[frigid mango]

Exactly what he also mentions on Reddit - sounds great IMO

[strong bramble]

#🏞｜general-with-images
Image Prompt:

Create an image of a modern, sleek office environment where a User Interface/User Experience (UI/UX) engineer is deeply immersed in work. The scene should include a high-resolution computer screen displaying a sophisticated design software interface, with multiple design elements and a color palette visible. The UI/UX engineer, dressed in casual professional attire, is shown intently focused, with sketch pads and wireframes spread across the desk. The background includes minimalistic decor and a whiteboard filled with flowcharts and user journey maps. The lighting is bright and natural, suggesting a creative and productive atmosphere. The overall theme should convey innovation, creativity, and attention to detail.

[noble coyote]

Comfy has formed a not-for-profit named Comfy.org. He and his team remain dedicated to making ComfyUI "the best open-source free software available anywhere!"

[noble coyote]

The team includes Dr.ltd.Data, and pythongosssss

[pastel latch]

i smell a gpt prompt

[noble coyote]

Ollama's version of P R O M P T Create an image of a modern, sleek office environment where a User Interface/User Experience (UI/UX) engineer is deeply immersed in work. The scene should include a high-resolution computer screen displaying a sophisticated design software interface, with multiple design elements and a color palette visible. The UI/UX engineer, dressed in casual professional attire, is shown intently focused, with sketch pads and wireframes spread across the desk. The background includes minimalistic decor and a whiteboard filled with flowcharts and user journey maps. The lighting is bright and natural, suggesting a creative and productive atmosphere. The overall theme should convey innovation, creativity, and attention to detail.

[noble coyote]

Ollama

[random dawn]

dall.e 3 ?

[bitter hearth]

an image like that ? from dalle 3 ? no definitely, thats at least dalle 4

[random dawn]

Peak AI !

[cursive nymph]

nah thats Dall F at this point

[sage burrow]

Has anyone else been able to get Ollama to work for SD3 and comfy?

[low stone]

Chibi baby Siberian cat, posing with dual lightsabers, background with a chaotic urban street, chibi cat's face, baby superhero pose, full body shot, full-length view, wide angle, perfect composition, hdr, soft light, epic

[opal bay]

so hands are fine....as long as they are cat hands.

[bitter hearth]

do NOT mess with kitty

I have been wondering...
Did team who made 2B (and also is researchers who already left SAI)
Instead of only doing an abliteration (negative reward if prompt produced is NSFW, training without image), they instead did part of training on a bunch of nsfw, but with the loss function kind of inversed, like, instead of noise reconstruction, trying to make as much noise as possible on nsfw prompts....

And so we get those mangled bodies, bc noise, structural level

[sage burrow]

The images don't turn out bad if you prompt N SFW, they turn out without any uhm, specific bits 😉 🍆

They just "forgot" to train a few things I think 🙂

[bitter hearth]

"army of cats invading the streets of new york" x3

[bitter hearth]

So many kitties! 🩷

[sage burrow]

Along that theme, I guess he's from the Bronx lololol

[sage burrow]

Well that escalated quickly! 😉

[noble coyote]

I have yes!

Interesting (free) extension for ComfyUI-Photoshop https://exchange.adobe.com/apps/cc/3e6d64e0?pluginId=3e6d64e0&workflow=share

[sage burrow]

I mean other people, I'm just convinced you have coding superpowers LOL

[sage burrow]

There's a GIMP one as well fortunately, for doing inpainting!!! 🙂

[noble coyote]

My first (SDXL) image via ComfyUI-Photoshop - the first one is the base image; and the second is the ComfyUI image masked over the original. As you can see, it "doesn't do text!!!"

[bitter hearth]

lmao

[bitter hearth]

evil kitty

[noble coyote]

Sucks its own blood!!! 🥳

[patent magnet]

actually it's an anime bat 😆

[patent magnet]

Here's a few more

[low stone]

Nah, this is sd3 with cosxl-sdxl refinement stage. As long as the denoise on the sdxl stage is high enough with the res_momentumized sampler, it'll fix the hands etc.

[patent magnet]

One of the most difficult prompt for any model and of all time, the infamous scorpion prompt. SD3 almost gets it:

[bitter hearth]

cat paws

[opal bay]

clever!

[sage burrow]

Me and my army are taking over this place, hand over ALL your catnip and cardboard boxes!

[sage burrow]

what's the prompt?

[patent magnet]

Try putting the word "Scorpion" and see if the model gets it right. It never does. Most models get a lobster type thing, nothing like a scorpion. Every SD model fails this prompt. Even MJ and Dalle did at one point (I haven't tried in the recent versions)

[bitter hearth]

This is what it would be like if I were the one doing these

[sage burrow]

Anthro scorpions are also difficult!
Will try a scorpion in SD3 now 🙂

I'm not sure what it is precisely, but I'd stay indoors that day 😉

[bitter hearth]

it just need a friend

but that aint me!

[sage burrow]

Here's a couple 😄

[solemn raven]

Dev list getting shorter and shorter. How long until SAI collapses under it's own incompetence?

[patent magnet]

It always reminds me of a dancing lobster

[solemn raven]

Will be funny when flowwolf and crystalwizard are the last two people in here insisting SD3 was perfect to an empty room.

[sage burrow]

He's the star of a scifi B movie I think

[bitter hearth]

oh goodness

[desert garnet]

they will be the last 2 ppl on the building washin the used empty toilets 🤣

[frigid mango]

Lizard warrior of the realm

[sage burrow]

Couldn't resist, sorry

[small sequoia]

@magic turtle CONFESS U GOT MUTATED PEOPLE TO PHOTOSHOOT ON GRASS ON PURPOSE FOR THE DATASET

[pastel latch]

https://tenor.com/view/hoblos-uhhh-uhhh-brother-brother-uhh-gif-1627688090339014668

[patent magnet]

with some hair ^^

[small sequoia]

Somebody’s been watching fallout

[bitter hearth]

I never finished fallout

[frigid mango]

coming for help

[small sequoia]

Finish it or it will finish u

[desert garnet]

Zendaya without makeup ( shes reptilian) 🦎

[small sequoia]

I’m not joking though

It will finish u, the marketing team will come to get u if u don’t @bitter hearth

[frigid mango]

forgot muh shield muh lady

[sage burrow]

I had to look her up

[bitter hearth]

Do I need to delete my statement 🤨

[small sequoia]

Now ur admitting to hiding

[bitter hearth]

no!!! I am a good Amazon Citizen

[small sequoia]

Just like the vault people… ur done for rip

[desert garnet]

https://tenor.com/view/huh-pepe-gif-8145044083223085060

[small sequoia]

WHO DO U THINK THERE MARKETING TEAM IS?

[storm saffron]

ChatGPT

[small sequoia]

We can’t even generate examples of this moment… cuz it’s been censored from the model

[small sequoia]

plus or free?

[bitter hearth]

oh plus now, for sure

[magic turtle]

Some people dig it. Don’t judge man.

[desert garnet]

oh dang its emad 😮

[magic turtle]

I lost my 2fa again so got a new account

[sage burrow]

What has been censored from the model? (mean besides the obvious anatomy)

[desert garnet]

https://tenor.com/view/real-and-true-cat-walking-gray-cat-gif-14865986020845776272

[magic turtle]

nice to meet you sd3 ceo

[small sequoia]

I’m talking about characters from shows and stuff

[desert garnet]

🫡

[past thicket]

Even with all the chaos of the last week, I'm still having fun with SD3 medium!

[magic turtle]

its a solid model. just prompt it in a different language or something

[past thicket]

I wouldn't say solid, but there's some gold to be found.

[ember ravine]

Was it trained on multilingual stuff?

[desert garnet]

Thank you emad very cool 👍

[sage burrow]

this model sucks, there's not enough anthro ladybugs in the dataset!!!
<sarcasm>

[magic turtle]

mebbe

[dull star]

oh guys look its the very real emad

[magic turtle]

updated my screen name

[desert garnet]

https://tenor.com/view/osaka-azumanga-daioh-azumanga-meme-azumanga-daioh-meme-gif-27109112

[sage burrow]

Oh! I"ve only tried presents so far... I don't watch TV or movies, so this might be a stretch, but I'll try 😄

[solemn raven]

He means another language like morse code or braille

[magic turtle]

french will do

[solemn raven]

skill issue if you aren't prompting in braille

[sage burrow]

Johnny Depp?

in a 70's disco bodysuit ROFL

[frigid mango]

damnit

[sage burrow]

I see no lack of celebrity data in this model 😉

[sage burrow]

Oh wait, you might mean nsfw stuff of celebreties? That's illegal these days, so just put some shorts and a top on them 🙂

[craggy ridge]

Well at least he didn't took an arrow to the knee

[desert garnet]

thats johny deeper

[frigid mango]

Actually only working with this, so prompt understanding is.... meh.

[bitter hearth]

Now it makes sense, the wall of dots in negative is actually a very smart braille negative

[solemn raven]

exactly

[frigid mango]

https://huggingface.co/spaces/madebyollin/sd3-with-taesd3-previews

[dull star]

oh wow already

I suppose the point of this is to have a realtime VAE like the regular TAESD, but with 16 channels and stuff

[bitter hearth]

I tried it

[bitter hearth]

What the fish doing

[frigid mango]

it needs some time to get it doing okay stuff, but okay is the upper end. Mac User, waiting for Draaw Things...

[terse belfry]

The girl was meant to be underwater ...but i like how it came out🤣

[small sequoia]

No that wasn’t what I was suggesting

[patent magnet]

Reminds me of Gyo by Junji Ito, great story

[compact forge]

😂 https://www.youtube.com/watch?v=yDlHykxYJGA

one of us

[terse belfry]

Oh, thx for remind me of junji ito! Will try to generate something in his style. Love Tomie

[mortal kite]

be sure to share here gotta see that

[cinder junco]

@upper snow With your custom node for ODE samplers, do you have any guidance on which scheduler is best? I tried a number last night and they all seemed to give identical output except for one (I think it was “simple”). I just saw this repo and the corresponding arxiv paper for GITS (third link in the repo readme), which appears to be a new scheduler optimized for ODE diffusion models. https://github.com/zju-pi/diff-sampler

[frigid mango]

any other open apis without registration out there?

[viscid pivot]

damn that guy followed all of the discord drama hahaha

[mortal kite]

ok I can't be doing Junji Ito. It's too much nightmare fuel

[patent magnet]

So SD3 does know something about his style

[sage burrow]

I have done so many in his style before, but not with SD, what an awesome idea 🙂 🙂 🙂

[terse belfry]

I'm still trying.

[mortal kite]

um. Hm. Maybe don't do it

[terse belfry]

Yeah, junji ito style is 👍

[sage burrow]

I find it easier sometimes if you say their name, then also the name of one of the characters they play

[terse belfry]

This one's pretty good

[mortal kite]

hahah. Ok...I think I need to stop

[terse belfry]

Pls don't 🤤Love this!

[mortal kite]

gonna need to purge my hard drive after this morning 😄

[sage burrow]

How is that different from the Flash version on there? (in layman's terms please :D)

Done with Taesd via huggingface - SD3

[frigid mango]

you got a link to the flash version?

[storm saffron]

That's a big gerbil

[sage burrow]

This one is awesome! OK I need to get out my old Junji Ito prompts 😉

[mortal kite]

I have to get my Bible and bless my PC now

[frigid mango]

As I understand - being anything but a pro, expert or even informed - its a very small encoder therefore showing the genration quikcly but a lot of tokens harm him by not taking care of them

[sage burrow]

https://huggingface.co/spaces/jasperai/flash-sd3

[sage burrow]

But, but reference images for later 😉

[mortal kite]

eyebleach

[storm saffron]

Wow flash SD3 makes SD3 even WORSE

A herp doing a derp

[frigid mango]

I would say theres a tat more control by CFG and step choice and therefore better results especially for longer prompts (only very quick testing)

[sage burrow]

if you click the advanced settings button, you can change the seed, and try all the various seedds with your prompt, to find the best one. Each is VERY different.

[patent magnet]

I Wonder... what if... Junji Ito + girl laying on the grass?

[sage burrow]

Thank you 🙂 Fortunately now I can do side by side comparisons, with both sites.

[sage burrow]

a DVD screengrab, of a horror movie, of a lady laying on the grass

JUst looks like a lday that got into a bar fight is all (all limbs are fine)

[frigid mango]

https://huggingface.co/spaces/philipp-zettl/stable-diffusion-3-medium wanna try a third one? I wonder if there is a differencs to the Taesd3 version

[frigid mango]

okay, those are the same

[terse belfry]

Getting there slowly

[sage burrow]

Taesd3 web, Jinji Ito 🙂

[frigid mango]

Just looked into the SDXL channel after a while. This here looks a lot like kindergarden 😄

[sage burrow]

I like this one the least of the 3. I thin I like the Taesd3 the best 🙂

[frigid mango]

taesd is fun because you see how the image generates which gives some nice insights 😄

[sage burrow]

The flash one has the instantaneous change the seed and get a new image, thing though!

[frigid mango]

for non-human i got the feeling less cfg around 3 is good and for humans you can go up to 10

[patent magnet]

🤣

[bitter hearth]

[terse belfry]

Perfect

[bitter hearth]

Lmao

[sage burrow]

Junji Ito and Banksy

[desert garnet]

https://tenor.com/view/skeleton-boomerang-skeleboom-skeleton-boomerang-gif-26191631

[bitter hearth]

Inflatable toy of... Steven Seagal

[desert garnet]

https://tenor.com/view/rio-ferdinand-rubs-hands-rubbing-hands-excited-anticipating-gif-15922156

[bitter hearth]

@desert garnet

[desert garnet]

Sensei 🙇‍♂️ thats why hes the goat 🐐

[signal shuttle]

Have you guys seen the video on SD3 by latent vision?, its pretty interesting

[gentle wolf]

Lol I'm watching it right now. "Let's confuse the little bast*rd." Boom it worked.

[signal shuttle]

there is so much good information in that video, really explains SD3 well without bashing it too much

i hope the video can restore people's hope in SD3 2B

[gentle wolf]

No. He shows hacks to make SD3 work. It's still too much work.

[gentle wolf]

If we can automate the scrambling, then we have a solution

[wild coral]

The architecture for SD3 is so good, especially it's VAE. I just wish we had better weights for it, and a more clear license. :/

It really had the chance to be the last image model we'd ever need.

[silver adder]

(Octane render, Unreal Engine, by Weta Digital, Cinematic, Photography, DSLR, Nikon D750, Short Exposure, F/2.8, Cool Color Palette, Kodachrome, 32k, Ultra-HD, Super-Resolution, HDR, DCI-P3, Natural Lighting, Flare, Cinematic Lighting, Contre-Jour, Beautiful Lighting, Ray Tracing Global Illumination, Ray Tracing Reflections, Ray Tracing Ambient Occlusion, RTX, VFX insanely detailed and intricate, hypermaximalist, elegant, ornate, hyper realistic, super detailed) A whimsical robot finds an unexpected garden in the heart of the city, with the robot being curious on exploration and its interaction with the vibrant flora and fauna. The scene is complete with vivid colors and intricate details.

use this alternative detail prompt if it gets truncated: (hyperdetailed, masterpiece, 108mp, hyperintricate subtle details, cinematic volumetric lighting, imax, extreme contrast)

[frigid mango]

This is nice!!

[silver adder]

thanks!

[frigid mango]

would use it as wallpaper

[gentle wolf]

True. In 8k would be nice

[silver adder]

i used 8b for that, 2b gives somewhat ok of a result

[signal shuttle]

looks nice, did you use a upscaler?

[silver adder]

yeah. i used a glif app that has one equipped with it

[fleet meteor]

So beautiful image 🥺 🤖 🤗

[finite osprey]

trying different stuff with oil painting

[silver adder]

some 2b images using the same prompt

[terse belfry]

My hope for SD3 restoring already. Kinda like the results im getting. The only problem now is the anatomy so far.

[solemn raven]

#🆕｜sd3 message

[signal shuttle]

If the anatomy was the only problem in SD3 then it can be ironed out using a finetune, but there is so many weird issues and quirks with SD3 that it might take a long time to fix them

[finite osprey]

2B

[signal shuttle]

Comfyui or SD.next?

[finite osprey]

comfy

upscaled

[solemn raven]

missing a wheel but looks nice

[finite osprey]

got destroyed during the war

heres another one

[frigid mango]

Still, this would not be my wallpaper 😄

[finite osprey]

yeah well it's less "artsy"

[terse belfry]

Hate the License as well

[silver adder]

that image looks crisppy

[bitter hearth]

thats super nice

[finite osprey]

why do I get the feeling this is a woman laying in grass? 😄

[terse belfry]

That was the attempt lol. Wanted to make it in style of junji ito...

[pseudo owl]

kinda cool honestly lol

[bitter hearth]

lmao

[errant dust]

It needs to be pointed out that this is very old-school prompting. It will still be ideal for the SD XL type models, but that sort of prompting relied on just throwing in a ton of key words and hope the parser made something nice from it.

[finite osprey]

Yeah i modified it a lot

the end result was more like an oil painting, so the unreal engine prompt didn't make any sense

[silver adder]

gotcha. i was just experimenting with those details that i used with the older models such as 1.5 and 2.1 back in the day to get the best result of out images that i generated there.

an app i'm using to generate sd3 images has an llm in between so it redetails some of the stuff in a more human tone and truncates it without getting rid of the most important things of the prompt (mostly)

[errant dust]

It takes some real readjusting, so it was not meant as a bash, just a heads up. When Dall-E 3 came out, the first to really emphasize natural language prompting, it took quite some time to 'relearn' prompting. Nowadays with MJ6, Ideogram, and SD3 (8B and 2B), natural language prompting is the norm.

To be fair, before it made no sense to bog it down with natural language as the parser would almost randomly choose which words to latch onto and which not. So avoiding anything 'distracting' was the underlying key to prompting.

[sacred jewel]

Qwen2 prompt

[signal shuttle]

Latent Vision has interesting tips for SD3 prompting, saying that we should start first with the most difficult elements to generate and then the easier elements and then background and lastly everything else, in this order

[errant dust]

This is not actually new advice, and applied to old-school prompting as well. It was just how you wrote it that was different.

Before, and still applicable today, you started with the main topic, then threw in some key info and details, then proceeded to stylistic desires, and finally into aspects of quality, detail, etc.

[digital thorn]

Who you gonna call?

[signal shuttle]

why is he giving a scuffed mutated middle finger lol

[digital thorn]

That's what i wanna know. But call center is busy

[errant dust]

Today's parsers are far more advanced, so if you don't respect this order, they can usually do it for you, and are thus far more forgiving. This means they respect your prompts far more as a rule, and are far more user-friendly by nature.

[floral apex]

his arms are fused too

[errant dust]

It's like the mutant creature from Altered States making a phone ad.

A great movie from 1980 that fused Huxley's The Doors of Perception with sci-fi extrapolation/speculation

[finite osprey]

zipper man in a dark forest

[digital thorn]

darth zipper

[finite osprey]

robots all have the same shape, it's hard to get another type

[finite osprey]

the upper body is almost the same

[zealous karma]

where is the best place to browse \ discover local image gen models?

[remote holly]

👀

[solemn raven]

huggingface or github

[finite osprey]

death metal

[silver adder]

buzz (sd3 8b)

[finite osprey]

this is like too sharp 😄

[zenith hemlock]

SD1.5 finetuned

[zenith hemlock]

@ancient cape and SDXL Epicrealism Hades

[silver adder]

seems like the upscaler did an extraordinary job with that one. makes me think if it's really meant for with close-ups

[hallow lion]

Touching grass tutorial.

[finite osprey]

try crocodile

[finite osprey]

am I the only one who enjoyed SC ?

[zenith hemlock]

Just because it wasn't cared about too much

[hallow lion]

I did too.

[noble coyote]

Having fun with this ComfyUI to Photoshop Plugin

[zenith hemlock]

I really liked Stable Cascade too, but it felt more like a transition to SD3

[frigid mango]

Stable cascade was just too slow on my mac

[finite osprey]

i was able to make 4k * 4k images without upscale on 8gb ram

ill try with SD3 when I have time

did anyone try double upscale with sd3?

[finite osprey]

lying in lava

[finite osprey]

that dude's expression

[craggy ridge]

"They pay us to do what?"

[finite osprey]

so, ahem guys, today's assignment is

[hot dawn]

The same Felicity Jones prompt after 4 epochs of training the wider dataset, way further along than when training in isolation. It shows that a lot of the learning was probably the prompt format and high quality photo tag (which is trained in a lot of other images), whereas the actual subject is quite easy to learn. Previously it took 16+ epochs to begin showing any likeness at all.

[bitter hearth]

pikachu after the LOBOTOMY and CENSORING of SD3 (meme)

[mystic vortex]

Pika on fentanyl

[bitter hearth]

how long did that take to train

[hot dawn]

the dataset was ~8k images and each bath was the same image repeated twice at a high and low timestep. About 4 hours per epoch, training only the transformer with bf16, stochastic rounding and fused backpass. Encoded prompts were pre-cached, though the images were encoded through the VAE again each time with flipping.

[hot dawn]

trying one of my own characters with more training data, there's no apparent likeness aside from maybe the outfit colour. So it seems it was probably riding on the original person being known to the model somewhat

[remote holly]

What's the prompt ?

[bitter hearth]

cctv footage of an army of cats inside a bedroom at night, nightvision view

[remote holly]

Thanks

[icy drift]

Any response from SAI to Civitai yet? News? Updates?

[bitter hearth]

No

[placid swallow]

last we heard they said nothing

[terse belfry]

All my attempts on ' a photogrph of a Female eating a banana' was a FAILURE...

[terse belfry]

This one is SDXL 🤣

[placid swallow]

i mean that's one way to eat a 🍌

[signal shuttle]

People eating food isn't one of SD strong suits

[terse belfry]

Lmao

[eternal drift]

Its my favorite of the models thus far.

[placid swallow]

people doing anything isn't a strong point

[terse belfry]

🤣🤣🤣🤣

[placid swallow]

using LLM output for the prompts definitely makes it feel useable, I like letting it run all night and seeing what it came up with in the morning

[icy drift]

I guess just ask it for deformed horrors. Makes sense. Strong suit.

[placid swallow]

it's like it has built in monster loras

[sage burrow]

Just done use humans 😉 ROFL ROFL ROFL

[sacred jewel]

Qwen2 prompting for the three encoders

[placid swallow]

I'm just doing phi3 medium with the triple clip

[brittle nexus]

Has anyone tried using natural language in the negative prompt? It's hilarious. `The image is a close-up, cropped mess. It's hard to tell what's going on because of the poor quality and weird colors. It appears to be a drawing, possibly anime-inspired, but it's poorly done.

The main focus seems to be a creature with two faces. Both faces have ugly eyes and deformed features. The hands are especially disturbing, with extra fingers, mutated proportions, and interlocked digits. There might even be extra limbs poking out of the frame.

The overall impression is one of pain and disfigurement. The artist seems to be trying to depict some kind of mutation, but the execution is clumsy and unsettling.`

[noble coyote]

Latent Vision has done an excellent video on SD3 - a neat way to use the negative prompt (which SD3 doesn't seem to like at all!); and a way to regularly generate at 1728x1024 resolution (low CFG apparently!) https://www.youtube.com/watch?v=OrST6Nq1NUg

Also from "Latent Vision" - a 'hack' to stop deformed bodies is to step away from the "dimensions must be divisable by 64". He suggests 1044x1044 - try it and see?!

[bitter hearth]

That one is more natural than the one i asked ai to make Cx
"The deformed creature, with its mangled limbs, disfigured face, and body, was twisted and distorted, its abnormal proportions making it look unnatural and aberrant, monstrous and grotesque, misshapen and unusual, anomalous and deformed, crippled and ghastly, its very existence a grotesque anomaly in the natural world, demons, horns, horror"

Grotesque anomaly in the natural world is so funny to me lmao
It works surprisingly well, sometimes it stops me from making ugly images on purpose

[brittle nexus]

It makes you want to put it in the positive

[hallow lion]

SD3 Civitai unban when?

[bitter hearth]

2 weeks

[noble coyote]

^..^<

[hallow lion]

Wouldn't have it way other way.

[upper snow]

👀

[covert galleon]

welp thats a tad terrifying

[upper snow]

I'm not entirely sure on its surface if this is a general ODE solver or if it's like DPM-Solver++ (DPM++ 2M) and is hyper specialized for diffusion models and not necessarily flow models, but we'll take a look at it!

[sage burrow]

When their earning started to subside? 😉
Kissing, maybe

[noble coyote]

Large Format generations using Latent Vision's idea on small CFG numbers/no negtive prompting to make clearer output at the higher resolution. 1728x1024

[cinder junco]

My comprehension of the paper was low, but even though they talked about it as a sampler, they seemed to deal more with the noise schedule than anything else.

[placid swallow]

ya there's some good info in his video i'm running with the changes now to see how it feels

[terse belfry]

These are perfect

[upper snow]

ah, the latest one is specifically for probability flow ODEs (which afaik does not include rectified flow models). It might still work but the general ODE solvers will probably work better

[noble coyote]

One of those fat gentlemen even got nillpes

[noble coyote]

^..^<

[terse belfry]

Are there any finetunes of SD3 available yet?

[noble coyote]

There are some SD3 Controlnets available for A1111

[cinder junco]

Haha, I don’t know what those two categories mean! I guess this is a sampler rather than a scheduler, then?

[upper snow]

Hopefully, as rectified flow models start to take over, we'll start seeing more papers looking into better sampling methods for them specifically. But for now we can surely settle for using the same ones people design airplanes with.

[lilac latch]

SAI is done. I see no future in SD3 architecture.

[noble coyote]

1728x1024

[placid swallow]

i'm trying out th shifted 6.0 with no negs for the details and it seems like it does change

[lilac latch]

SD3 8B won't be released at this point.

[upper snow]

It is a sampler/solver yeah. The main difference here is that diffusion models are an SDE, and SDEs are incomplete, and you have to model the incompleteness using random perturbations (like the extra noise injected by SDE/ancestral samplers). Rectified flow models are ODEs, they are complete problems, and you can run them forwards or backwards.

[noble coyote]

Yes, change on an image-by-image basis - I mean don't leave it set at 3

[hallow lion]

This model is like a puzzle... Like a riddle!

[terse belfry]

As long as you're not making any NSFW content SD3 is great, can't wait for finetunes

[cinder junco]

Do you have any recommendations for the default schedulers then? Or do the ODE samplers remove the scheduler effect?

[mortal mesa]

i like it too but its a bit deeper than just NSFW

[noble coyote]

I'm trying to dream-up a Model Merge using PiXart-Sigma and SD3-Medium ... A model-merge node rather than a checkpoint mash

[upper snow]

The adaptive ODE solvers literally only take a start point and an end point. So just use sgm_uniform because no other scheduler will make a difference (or if they do it will be a wrong one. The start should be 1, the end point should be 0.)

[placid swallow]

let's see if this 3060ti can handle a 1728x1024 with upscale

[hallow lion]

Overload the negative with random characters, make a weird resolution, perturb it, trick it in the promt and it works! 😄

Emad left us a puzzle! It was his way of saying bye.

[noble coyote]

Latent Visions' Negative Prompt (I kid you not!) aaaaaaaaaaaaaaaaaaaaaaaa aaaaaaaaa aa aaaaaaaa

[cinder junco]

Cool, thanks. Might be useful to explain in the GitHub readme for non-experts.

[terse belfry]

With LORA's it will be perfect

[placid swallow]

accidentally did 1728x1032 but it came out pretty good, even if it took 2mins

[terse belfry]

Squirkittens

[noble coyote]

afk

[placid swallow]

6912 x 4128 after upscale

that's crazy it's so hard to tell it's not a picture from a camera now

[hallow lion]

Sadly there is another eye at the end of the tail

[bitter hearth]

Where

[pseudo owl]

how can you even spot that

[placid swallow]

ya it's still not perfect but the pixel definition are 🤌

[craggy ridge]

kitty evolves

[brittle nexus]

That too

[placid swallow]

that's just the mouse it's saving for later

[gusty trail]

With a trained AI's eye

[bitter hearth]

Look what i just made (surely not ai)

[sage burrow]

You say that like it's a bad thing? 😉

[placid swallow]

you have gifted drawing skill you should try to sell your stuff

[craggy ridge]

wow you are goood

[bitter hearth]

Oh my god that's so small xD eagle eyes stuff

[mortal mesa]

let it fly the world is a weird place

[bitter hearth]

@craggy ridge Here is your drawing! Haha sadly the api is not very bad at woman lying in grass to do the meme consistently

[placid swallow]

edges are falling apart at 1024 wide i'm gonna try 1032 again

no better

[bitter hearth]

Such a weird effect

Can fine tunes fix that resolution issue?

[terse belfry]

It's really good with watercolor style

[placid swallow]

halved the initial latent and it still comes out pretty good after upsizing to 3456 x 2048

[bitter hearth]

Gimme prompt

[bitter hearth]

So much blur im thinking it's sdxl

[keen palm]

Discord doesn't let me, don't know what word is banned

Photo of a group of people riding a roller coaster on a clear, sunny day. The roller coaster, named "RollerCoaster Tycoon 84," has blue and yellow cars filled with excited passengers, many of whom have their hands raised in the air. The wooden tracks and supportive beams of the coaster are prominently visible, with the bright blue sky serving as a backdrop.

[kindred mica]

those supportive beams

[kindred mica]

i didn't have any problem with censorship?

[keen palm]

I only sent the G clip text.

T5 clip text: The image captures a thrilling moment on a roller coaster named "RollerCoaster Tycoon 84." The coaster is filled with an enthusiastic group of passengers, all of whom are experiencing the exhilaration of the ride. The front car prominently displays the name "RollerCoaster Tycoon 84" in bold, yellow letters against a deep blue background. The passengers, a diverse group of men and women of various ages, are clearly having a fantastic time, with many of them throwing their hands up in joy and excitement as the coaster plummets down a steep drop. The wooden structure of the roller coaster is sturdy and intricately designed, showcasing the engineering marvel of the ride. The beams and tracks are detailed and create a sense of speed and motion as they lead the eye through the twists and turns of the coaster. The backdrop of the image is a flawless, bright blue sky, enhancing the sense of an ideal, fun-filled day at the amusement park. The vivid colors, dynamic angles, and the expressions of the riders all come together to create a snapshot of pure, unadulterated joy and thrill, perfectly encapsulating the essence of a memorable roller coaster ride.

L clip text: Photo, vibrant colors, dynamic composition, high contrast, sharp details, action-packed, lively atmosphere, detailed textures, energetic mood, high resolution, playful scene, immersive feel, bold typography, vivid imagery, fun and excitement theme.

I could send them in one discord message each, but not allowed together...hmmm

[cosmic crown]

funny, you use L and G exactly the other way round i do.

[placid swallow]

Could I interest you in some fresh mouse crepes or a succulent house plant?

[keen palm]

L is better for object and G better for style?

[bitter hearth]

Lmao i love this

[cosmic crown]

its what i use, but is it better? i dont know. i just find it funny you use L for style and i use G for style 😄
maybe we should use T5 for style instead? who knows? 😄