🆕｜sd3

[sterile pendant]

in your case, that just means that SD3 udnerstands the prompt directions very well

which is actually a good sign

[bitter hearth]

My prompts are usually very vague, I do like "style, character, place, doing this"

[placid swallow]

same thing with streets and alley ways if you don't specify geographic location you tend to the the same background

[bitter hearth]

A single piece of information for everything

[sterile pendant]

well the models cant read your mind, so they go with the law of averages

[bitter hearth]

Plug a neuralink and run SD4

[placid swallow]

ya that's what i'm saying it feels like the average from the set is very specific and not varied as much as the other models but that's probably what they keep pointing to with 2b training

[sterile pendant]

reminds me of this:
https://www.reddit.com/media?url=https%3A%2F%2Fi.redd.it%2F0joyb4oqx2j21.jpg

from way back showing composite average faces by ethnicities

[compact niche]

hi

[placid swallow]

that's why it feels like you have to tell it every nuance of the image

[sterile pendant]

that's the whole point, it CAN mostly handle that

[compact niche]

lol

[sterile pendant]

otherwise, it will go to a de-facto law of averages version

[placid swallow]

I guess i find it's average boring

[sterile pendant]

well blame the internet and the dataset then lol

[bitter hearth]

Kitty bear

And some Aztec words

[compact niche]

lol

[placid swallow]

aztec for rawr

[compact niche]

saved

thx

[sterile pendant]

maybe try adding in random tags like (facing forward | turned to the side). i can't remember the exact format for comfyui, but it will randomly pick between the options. it might be that they need to be [a | b], maybe it's curly {}, dunno, but you'll know when it works. just test a prompt like "a (red|blue) ball" but try the different () {} []

[placid swallow]

i do

2b just feels more stuck into certain things than previous models for me

[sterile pendant]

maybe, but again, you have to not be lazy and need to specify things more. which is a good thing. in previous models, they didn't comprehend as well, so you got a lot more image variance between seeds. i know people meme about it hard, but it really is a git gud issue as one of the devs said

think of what a full body portrait photo means inside your visual thoughts

99% of people will think the same thing: facing the camera for a professional portrait photo

[placid swallow]

I'm not saying it's wrong I'm saying it doesn't like to vary unless told which is different and good, just cumbersome when you have to tell it every single thing to change

[sterile pendant]

blow the prompt up with an LLM then. i've seen good results with even prompts as large as 500 words (in theory, you can go all the way up to 10k words, but that would just drown everything out)

[placid swallow]

seems like it's averages come to a point rather than a bell curve

[sterile pendant]

for a simple prompt direction like that, that's exactly what you want. it's on you to then add in the details and specifics to flatten the curve out more, which will give you far more variance

jfc, i can't believe i just called visual thoughts "head movies"

[bitter hearth]

brain scenes

[hallow lion]

My Datsun!

[hallow lion]

Gaga

[sullen moss]

I'm currently testing DALL-E with more lenient filters and realize that the gap is very large between SD3 2b and the current DALL-E. This is very disheartening 😔

[terse belfry]

Dall-E is more artistic. And it's very good with anime and Artists' styles

SD3 is more realistic

[hallow lion]

SD will be better

[terse belfry]

SDXL is good for ART and everything. And with some finetuned models it's miles ahead of SD3

[bitter hearth]

DALL-E with more lenient filters?

how did you get that

[sullen moss]

DM me

[sullen moss]

It's not about realism, but about understanding and constructing complex concepts, especially when it comes to the anatomy and expressions of characters, as well as their interactions

[bitter hearth]

one undertained model vs trillion dollar model

I find 8b api way wayy more fun to use than anything I'd get from dalle, 1.5 or sdxl

the newest jugger is good

juggernaut XI preview
I made a bunch of sci fi cities with it

[desert garnet]

just gotta wait for juggernaut SD3

[bitter hearth]

yeah wow

[desert garnet]

[terse belfry]

Dall-E using GPT for that. It can be achieved with SD3 as well.

[desert garnet]

tbf sd3 better and bigger than dalle,its just that ppl suffer from skill issues

[bitter hearth]

not sure that sd3 is bigger

[sullen moss]

No, because SD removes all vulgar content from its datasets, while DALL-E used petabytes of pornographic content for training

[terse belfry]

SDXL vs SD3 (Right)

[bitter hearth]

dalle could be something crazy like 40b

[terse belfry]

If u need rly good, non-shiny realistic traditional looking ART, SDXL and Dalle-3 is still the best models for that.

[bitter hearth]

sd3 makes better steven seagals

[desert garnet]

they banned Seagal from dalle 😔

[bitter hearth]

[desert garnet]

https://tenor.com/view/seagall-closeup-meanlook-gif-20975873

[hallow lion]

Where are the 4GB of Vram? 😦

[bitter hearth]

in the camera

lmao

[desert garnet]

macho Seagal

[bitter hearth]

steven seagal is gone

[sacred jewel]

Why do people think that DALL-E is trains the same way as SDXL? DALL-E 3 is like SDXL's best and favorite refined version with every favorite LoRA baked in. It's a look.

Try getting a shitty image from DALL-E. You can't. It's always artistic and beautiful. It caters to a look. I am not sure that's very flexible.

Meanwhile, SD models are ALWAYS a great base to start from ...

If you like DALL-E , stop using SD3 immediately. SD3 2B might be lacking in some areas but it's still far more "normal" and flexible than other models. In [my] opinion for [my tastes].

[terse belfry]

There are things where SD3 is better than Dall-E3. So I'll keep using both for different needs.

[desert garnet]

haha you funny guys

[terse belfry]

SD3 is betterthan dalle at Watercolor, SD3produces more realistic watercolor paintings+ Great at photorealism and Digital art style

Dall-E 3 vs SD3 (Right).
SD3 is more superior when it comes to realistic watercolor look.

Prompt: Real watercolor painting of a blueberry cake slice with a mint leaf on top

[desert garnet]

try with this prompt : Real watercolor painting of a girl sleeping on the floor next to a giant cake slice

[bitter hearth]

[terse belfry]

Well.... Of course dalle will do better here. I was just comparing the watercolor looks.

Dalle-3 x SD3 (right)

Real watercolor painting of a girl sleeping on the floor next to a giant cake slice

[desert garnet]

the cake is a lie 😔

dalle doesnt look watercolor

[bitter hearth]

they are turning humans into cheesecakes

[bitter hearth]

You too don't

[terse belfry]

Real watercolor painting of a girl sleeping on the floor next to a blueberry cake slice on a plate with mint leaf on top of it

Dalle-3 x SD3 (right)

[desert garnet]

american flag cake

[sterile pendant]

Have you ever actually done watercolor painting? Most people just think it ends up looking like simple shapes with little blending and few colors used, but it can be extremely detailed if you want it to be

But I'm not about to preach art literacy at 9am lol

[desert garnet]

yes depends on how wet the tip is

[mortal kite]

hehe that's what she said

[sterile pendant]

It depends on a bunch of factors more than just that, but at least you know understand the tip of the iceberg

[desert garnet]

you can also use a big wet brush if u wanna make it look messy

[sick cedar]

Jugganaut has moved to Pixart i think.

[sacred jewel]

When SD3 hits, it hits well...

[sacred jewel]

SD3 really really likes 5 legged horses... oof..

[terse belfry]

SDXL x SD3 x Dall-E 3

Prompt: Real watercolor painting of a cat sleeping on the floor next to a blueberry cake slice on a plate with mint leaf on top of it

[mortal kite]

ugh. Hands

[bitter hearth]

6

[mortal kite]

yes it always makes five fingered hands too, it's obessed with 5 of things

[mortal kite]

wut

[bitter hearth]

I tried for a snake with many heads didn't go well

It must of been you, cuz sd3 don't make no mistakes yo

[bitter hearth]

For sure, I didn't use 500 words from chat gpt ultra extreme llama 90b to describe the image that's why it's skill issue on my part

[sacred jewel]

Yeah this one is even worse.. .typically 5 tho... 4 is rare 😦

[sacred jewel]

You need to up your skills

[bitter hearth]

git gud!

[mortal kite]

I think he's a bit short tho

[runic tusk]

Either that horse is really short or he's kneeling.

[bitter hearth]

https://tenor.com/view/horse-horses-owned-gif-21828610

[runic tusk]

Look at all that salt.

[runic tusk]

FOO RRD, my favorite truck brand.

[bitter hearth]

Do elephants on the back

[mortal kite]

now the dogs are too tall

[bitter hearth]

They are on top of stuff!

[mortal kite]

it wants to give the elephant 6 legs....

[mortal kite]

I don't know what's going on the left side

[terse belfry]

Says someone who is using comfyui.... I'm not using it unfortunately, have to use limited HF spaces for it. There's not much options to make ur generations super good.

[mortal kite]

just because it's complex doesn't mean its "super good"...

[runic tusk]

Batman's ashy...needs to lotion up.

[bitter hearth]

It's all salt cx

[mortal kite]

looks like snow

[terse belfry]

The generated pic with comfyUi will obviously be better. So...

[sacred jewel]

Minor detail...

[mortal kite]

nah.

weightless

[bitter hearth]

Salt out of this world

[bitter hearth]

Lmao

[mortal kite]

hey look its the Twitter belt

[runic tusk]

This woman lying on grass picture was too sexy not to set spoiler tagging on.

[sacred jewel]

A potato will outrun A1111 and ComfyUI... I am sure of it... Potato3.0

[mortal kite]

I get the feeling ComfyUI is mainly used by people that think they know what they are doing 😄 It really feels like the trend right now - there are math guys that really know what's going on, and 80% of the AI community thinks they know what's going on "oh add this to your prompt" "oh use this node in comfy" when it's just all random BS LOL

[sacred jewel]

Oh COME ON!!!!

[mortal kite]

constant conjecture and hearsay

[sacred jewel]

https://tenor.com/view/damn-damn-damn-damn-damn-right-dame-time-gif-18029769

[runic tusk]

Anyone can use ComfyUI; it's really not that difficult.

[bitter hearth]

It's like a cactus

[mortal kite]

he got one of those fungal infections

[sacred jewel]

I am using one of them newfangled LLM things...

[terse belfry]

ComfyUi vs HF space SDXL

[mortal kite]

You can see how much better actor Chris Tucker is

[bitter hearth]

Wish version

[terse belfry]

My hardware is limited, So i have to use spaces and sites for generations.

[mortal kite]

Stop being Poor 😄

[runic tusk]

Skill issue.

[sacred jewel]

This is a JUDGEMENT FREE zone brother... (or sister)

[mortal kite]

license plate

[bitter hearth]

It's okay to be poor, we rich are the yin, you're the yang, we need you!

[bitter hearth]

Elephant consumed

[bitter hearth]

Bloo

[bitter hearth]

Your thing really loves to double those trunks

[mortal kite]

maybe put elephant trunk in negative

[mortal kite]

or turn up cfg

[terse belfry]

Neyitiri

[sacred jewel]

and legs...

[bitter hearth]

Not needed to say but elephant cat didn't go well

[terse belfry]

Leech Cat

[sacred jewel]

4.5 cfg did not improve things 😦

[bitter hearth]

What is your negative

[desert garnet]

luma+sd3

[bitter hearth]

luma is pretty cool

@desert garnet have you used luma on Steven Seagal gens yet

[desert garnet]

im out of free gens 😔

[hallow lion]

Holly molly... we're almost there...

[desert garnet]

another one

[terse belfry]

Yeah, but it takes eternity to generate 1 vid. Waiting for gen-3 now, will be out in few days.

[desert garnet]

it was faster on the first day,i guess they nerfed it

[bitter hearth]

How long do they take

[desert garnet]

i have one thats been stuck for 1hr

[bitter hearth]

Actual process: 1 minute, but you can buy this premium service!

[terse belfry]

The demand is very hight atm, it took me like 5hrs to generate 1 vid.

[bitter hearth]

That's pretty looong

[terse belfry]

4 hrs in queue already...

[bitter hearth]

Gaga!

[sacred jewel]

NVIDIA GB200s are ExpenSiVo 😛

[remote holly]

Yes is the draconian queue

[pseudo owl]

hmm all 3 text encoders kinda seem to ruin the image, is t5 alone the best?

[mortal kite]

awesome frame, was that in the prompt?

[past flame]

I wonder if the AI would understand Vader and/or Anakin stating his opinion about sand

And know the reference

Maybe that one model that uses a full 2B llm for text encoding knows?

[mortal kite]

inpainting isn't quite a powerful as I remember with SDXL. with SDXL it would always make the new inpainted object fit perfectly into the scene. It seems a lot more finicky with SD3 to "hide the mask" and it doesn't always just fit it in

[past flame]

It wasn't perfect with SDXL either

The best state of inpainting I think was SD1.5 with inpaint controlnet, you could turn any finetune into a great inpaint-capable model without changing anything about the model's capabilities

[mortal kite]

I usually had really good success with it. Anyway, at least I figured out how to do it in Stable Swarm UI, watched a 90 minute tutorial vid last night on how to use Stable Swarm

[past flame]

Personally I prefer just ComfyUI

[mortal kite]

it's probably more because the terms constantly change between UI's. Some UI's call it "add object" some call it "unsampler prompt" or something.

[past flame]

StableSwarm just adds questionably convinient UI layer on top of Comfy

[sacred jewel]

Nope... (but then again, I did not check the LLM result... let me check )

[past flame]

Yeah, one of the reasons why I use Comfy, it uses technical terms for everything so there is no confusion once you learn the basics

[mortal kite]

I just don't know COmfyUI that well, there are a lot of nodes, plus in Stable Swarm, if I do go to the Comfy tab, it doens't seem to always equal what Stable Swarm is doing (even if i say "import from generate tab) and pressing the 'Queue image' button in the comfy tab doesn't seem to do anything

[past flame]

Probably because of the way StableSwarm is written

[sacred jewel]

Here is the prompt generated by the LLM

Best quality, masterpiece, Epic, cover-art, : The image is a captivating scene set within the enchanting realm of Renaissance, where a beautiful girl dressed in exquisite garments, adorned with delicate embroidery, gazes towards an ancient castle perched atop distant hills. Alongside her are two playful kittens and a loyal companion, a friendly dog by her side. The sun sets upon the horizon as soft clouds gather above, casting a mystical glow upon the landscape below. Intricate brass elements accentuate the setting, while fine details of nature merge seamlessly with human creations. Every stroke is executed in the mesmerizing style reminiscent of J.C. Leyendecker, his signature artistry captured through vivid colors and intricate patterns. The composition exudes a sense of wonderment, inviting viewers to immerse themselves into a world where fantasy intertwines with reality.

[past flame]

Anyways ComfyUI seems harder than it actually is, this is all the parts of a basic workflow:
1) Loading AI model into program 2) Positive prompt and negative prompt 3) Empty latent image (a format that lets AI understand an image) for AI to generate into 4) The node that runs the AI, it's same as normal a1111 settings 5) Turning latent into a normal picture

[mortal kite]

how do you do inpainting in Comfy? Like how do you draw the mask out

[bitter hearth]

you don't have to draw masks you can use object recognition models or face recognition models

[compact niche]

hey where u got this model? i tried to use pony too but this turned out ugly each fricking output

[noble coyote]

Latent/Checkpoint/+Clip and -Clip/VAE in/KSampler/VAE out/Save

[past flame]

You replace that input "empty latent image" with this, loading an image with a hole in it and turning it into latents

[runic tusk]

Visit the examples page:
https://github.com/comfyanonymous/ComfyUI_examples

[past flame]

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

[runic tusk]

All kinds of examples to draw from. 🙂

[mortal kite]

ok..but does that mean I need to draw the mask with an external program? Kind of a hassle..

[runic tusk]

No

[past flame]

This one, personally I use DPO variant because it's a very slightly better version when it comes to understanding prompts: https://civitai.com/models/288584

[bitter hearth]

comfy can draw masks for you
you just need to give it recognition model

[mortal kite]

Ok, I'd have to look up some tutorials on the tube on Comfy. Still, seems overly complicated to me for the things I ever want to do

[runic tusk]

It's really not. It's just different.

[past flame]

Also remember you have to add the whole following text to the positive prompt with Pony models (it's because of a training error sadly, next version just score_9 will be enough)

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up

[bitter hearth]

if you inpaint a lot its way easier to not have to draw mask
comfy is actually easier in a few ways like that

[mortal kite]

I've used node based tools before like in Blender...and most music applications are similar.

[runic tusk]

You grab a workflow example, drop it right into Comfy to load it, then go from there.

[compact niche]

thx

[sacred jewel]

In the image load node, you draw your mask and feed that into an Inpaint VAE encode node.

[sacred jewel]

no...

[mortal kite]

apparently it's a topless beach

[past flame]

This extension lets you right click on a Load Image node to draw mask or basic colors on it: https://github.com/ManglerFTW/ComfyI2I

If you want to play around with other extensions, ComfyUI can easily run nodes for stuff like object detection

And you can automate inpainting different things on your own

[runic tusk]

Which is essentially exactly how the face and hand detailer nodes work.

The do a little detection magic, draw a box for inpainting, and then go to town.

[mortal kite]

hahah pretty cool I was able to just swap the text with inpainting

[past flame]

Anyways I once made this schizo inpaint pipeline, first the nodes on top left can auto downscale an image if it's too big because I felt like doing this manually using math nodes instead of getting an extra extension for it later it turned out I already have it and I used lineart controlnet so that the model is more-less aware of the image composition while generating even at the very start of the generation which helps avoid the issue where the inpainted area doesn't fit, and reference controlnet so that it automatically copies the arstyle of the original image perfectly

The results are very nice, as messy as it looks here

[compact niche]

whats better? at the end or start

[past flame]

I believe it was trained with score tags at the start

[compact niche]

ok!

i still get pretty bad quality

[past flame]

I recommend joining Pony Diffusion Discord server, or at least reading the original Pony readme, they have some extra instructions and such

What are your results? Also maybe let's move to DMs to stop spamming here, and to not have slow mode

[compact niche]

xd ok

[mortal kite]

make sure use correct sampler Pony is very picky about it

[sacred jewel]

added "framed" to prompt

[bitter hearth]

we could very easily see whats wrong if you give an image with png info

you have to get the prompting exactly right for that model

pony is a weird combination where
the fine tune training was done very wrong in many ways
but because the dataset is so massive, with decent tagging
it still came out as a very useful model in some ways

[mortal kite]

Still. Foooocus has the best inpainting tools I've ever used. Even Stable Swarm I can't really get a "add this object and blend it with what's already there" working quite yet. Maybe I'll figure it out. It usually wants to replace the masked area entirely with new stuff

I am thinking perhaps the "Unsampler Prompt" does this but have to test it out

in Fooocus you could choose to add new item and it would layer it over the top of stuff, etc. Seemed way easier to deal with

it looks like the "trick" to having it layer more "appropriately" is if you use the same seed, at least the underlying image stays the same inside the mask...Well anyway, experimenting

[remote holly]

Sd3 have inpainting ?

[faint breach]

people have ported llly's code to comfy UI nodes. you could jigger that up to work inside stable swarm

[noble coyote]

Good-ish 🙂

[bitter hearth]

that black dressed girl is having a hard time on that tree

[naive jewel]

I'm experimenting with a 2 sampler workflow
generating a base composition first, with high shift and few steps
then dial in details with img2img using a different seed

[bitter hearth]

I'm addicted to pure black background help

[faint breach]

the seed doesn't really matter towards the quality or success of that process. it just provides initial noise. Your second img2img pass will start with different pixels so it effectively has different initial noise already.

[upper snow]

definitely not bigger than DALL-E most likely, DALL-E 1 is a 12B parameter model and openAI doesn't know how to make reasonably sized models

[naive jewel]

I swear I've seen the same seed resulting in a burned image sooner. .. I just tested this with the spacecat and of course it does not happen when I want to demonstrate it...😕

[faint breach]

heh. I know the feels. the seed really only determines the arrangement of noisy pixels at the start of generation though. It's easy to get stuck into confirmation traps when there's so many settings that affect everything.

[naive jewel]

It did not happen because I added an upscale to 1048x1048. This is refinig with the same seed and not upscaling:

[faint breach]

Hmm, so you still don't believe me that the seed only affects the initial noise positions. Alright. Not my problem to have.

good luck with your testing

[naive jewel]

I did not... guys you all need to chill... look at the first picture and compare it to the second one... you see a difference? The difference is the seed and the upscale. so if I change the seed or upscale I don´t get that burned effect. That is all I am saying.

[faint breach]

oh ok. good luck then. sounds like you're figuring it all out fine.

[spark quail]

more people need to play minecraft to understand what seed really is lol
the number itself isnt a setting but moreso that specific seed with those specific parameters

[runic tusk]

Aw shit, here we go again.

[terse belfry]

This looks nuts

[odd basalt]

Ppl should stop complaining and do more art… they are becoming to sound like a broken record… never seen so much of it in only the SD community… it’s crazy and above that SD told everyone there’s was not gonna be NSFW yet they act surprised about it… crazy… I agree

[terse belfry]

People who make money on NSFW stuff are definetely affected by this. Lol. I mean, if u rly wanna keep doing nudes, there are tons of SDXL finetunes for that, they produce rly HQ realisistic photos. My only cry for SD3 is not understanding most of the artists' styles and real 80s Anime. The rest is all good. The model is more than usable.

Ah, and the anatomy problems as well...

[pseudo stone]

LMAOOOOO

[terse belfry]

I prefer the custom one

Too much contrast for my liking. Lol

[bitter hearth]

what happened with the custom one it looks 240p

or is that not sd3

ooh

I'm dumb runs away

[frail shoal]

pixart sigma finetune on midjourney images. Results got better

A dark ghostly ice man, reminiscent of a White Walker, stands on a snowy mountain top in the midst of an ice and snow storm. He has long white hair, glowing eyes, and is swinging a long white glowing ice sword, attacking the camera. He also wields a glowing spear. His light armor is made of roots and leaves. The scene is set in darkness, illuminated by moonlight with cinematic and volumetric lighting, but with less backlighting than before. The composition is epic and photorealistic, with a bokeh blur effect. The atmosphere is mystical and menacing, capturing the essence of Nature's mystical beings and the devil spawn (masterpiece, 4k, octane render, volumetric lighting, perfect lighting, perfect picture, best quality)

[bitter hearth]

🤨

ok frenchy

[frail shoal]

?

[bitter hearth]

I'm giving you a hard time because you use a dall-e prompt and a character from game of thrones

That's like certain death in some circles

[frail shoal]

i just used the prompt someone else posted in this channel with sd3. I wanted compare the results with sigma. Last time i used another finetune. Now there is a new one finetuned on midjourney images. So the why of the post

[faint breach]

in the sand castle kicking club maybe?

[bitter hearth]

I like Dall-e myself, not so sure about Midjourney. Seems like something you have to use on Discord is a major NO from me no matter how good it may be. I now want to try that prompt for myself however.

[frail shoal]

sd3 2b did not do a good job last time with this one, however 8b from the api nailed it

[frail shoal]

seems 2b might have some oversaturation issues

[bitter hearth]

Do Darth Vader eating a sandwich while contemplating the absurdity of his own existence.

[frail shoal]

sure

[bitter hearth]

best of 4

You really don't get that whole contemplation feeling however with 2b

At least you have a whiskey glass with water in it.

the poor man can't eat his BK Chicken with that mask on

You have the little tin cup of flowers in the 2nd image. Which maybe says " I am ok with my life right now"

the mask is messing up his speech as usual

[craggy crest]

what is it with sandwiches? yesterday, it was barbarians looking for sandwich thieves, today, darth is inviting you all to lunch

[frail shoal]

Darth Vader eating a sandwich while contemplating the absurdity of his own existence (masterpiece, 4k, octane render, volumetric lighting, perfect lighting, perfect picture, best quality)

could not manage to get the contemplating part

[faint breach]

2nd one has Warhol eating a burger vibes

[frail shoal]

first one is from midjourney finetune. I switched for the other gens because it gave me giant sandwitches every time lol

[bitter hearth]

He is thinking hard in these

[compact niche]

beautiful

[compact niche]

thats funny because he has mask and he cant eat with mask -_-

[craggy crest]

(darth vader using a lightsaber to spread mayo on a piece of bread)

[faint breach]

can't even do a fork laying down

[odd basalt]

I’m not into NSFW and ppl don’t really make money on it… unless they are pron bottling on pron hub and that’s even more disturbing because of the c- exploit complaints and the abuse of NSFW is what has ppl wanting to potention for it to stop… I’m almost positive it’s illegal too… glad SD has made it a safe model … which they actually said that it won’t be good at anatomy it’s more for photography landscaping model… the data sheet shows everything… it’s a great model once u understand it…

[craggy crest]

i did say that the issue affects everything, not just humans

[odd basalt]

Looks more like Leonardo AI generator

[faint breach]

i cant get it to keep the burger in the bag, but this is so much closer to the video setting. i love prompting sd3. it always surprises me.

doesn't seem to know the classic heinz bottles so well though and leans towards modern consumerism big bottles

[odd basalt]

Here’s what my custom bot said >>>> The metadata extracted from the image includes the following details:

•    sRGB: 0
•    Gamma: 0.45455
•    DPI: (95.9866, 95.9866)

This metadata provides some information about the color space (sRGB), gamma correction, and resolution (DPI), but it does not contain specific details about the tool or software used to create the image. Therefore, it does not indicate whether Leonardo AI was used to generate this image.

[faint breach]

holy you want some burger with that ketchup there anni?

[hallow lion]

Heinz Kernzz day fuct eviecckot burot!

[faint breach]

day fuct it up

[vapid radish]

That is pretty good. Dall.e3 with an upscale still does it the best I think.

[mortal kite]

You have to keep CFG to 5 or less or yes you will get high saturation. Unfortunately higher CFG also gives you better prompt following. So I say we should have post-image step to adjust saturation and correct it in the image later

[odd basalt]

I figured as much and chatgpt strips meta data as well I noticed or it should have at least grabbed the ss metadata

[terse belfry]

Are they really confirmed that improved version of SD3 is coming? Like SDXL Came after SDXL 0.9 ?

[dusky thistle]

no

[terse belfry]

Love the mood

[lilac bridge]

inteeeresting. Using multiple ksamplers is helping to get around the bad anatomy 😄

[frail shoal]

dalle3 is the best, but censored. Now tha you said it, i should test it now with that prompt.

[frail shoal]

i tried now with a new sigma finetune

[terse belfry]

SDXL Vers

[lilac bridge]

Ooooh getting closer to overcoming the infamous laying in grass meme xD

Anyone want to test this workflow out with me? It's working kinda well 😄

[terse belfry]

With higher CFG

[keen palm]

side shot please 🙂

[lilac bridge]

what do you mean by that? laying on her side?

[keen palm]

Yeah, lying straight isn't the issue. It's the other poses that are

[odd basalt]

I’m anti Bing but try Dalle XL it’s uncensored and on hugging face or git hub … way better

[lilac bridge]

oh sure, gimme a sec

[terse belfry]

SD3 versions with same seed?

[odd basalt]

It’s still around just ppl forget how to do it…

[lilac bridge]

First try after adding "side shot" to prompt

[keen palm]

https://discordapp.com/channels/1002292111942635562/1230206273451069540/1253472608108675133

like the second image

[lilac bridge]

ah okay, gimme a sec

[odd basalt]

Yes it does h3ll I still go the long way sometimes and not use AI at all…

[terse belfry]

These are with FlashSD3, almost similar to the ones u generated.

[lilac bridge]

going to cut out the first image cuz it's kinda nsfw, but guess I gotta find a way to put her belly down hmmm

[odd basalt]

Facts that’s why I don’t pay for AI… I draw and only program I pay for is adobe cloud PS is all u need

I only use SD procreate and adobe programs that’s it

And Lightroom for my photography

U can edit them in Lightroom and use camera raw

[terse belfry]

Yes, it's doing really well with that.

[keen palm]

But I wanted a seed that was chosen. The chosen seed number 1477598 😦

[torn wharf]

Sunday Funday! What's new with SD3 today? I saw kohya has dreambooth scripts up an runnin

[lilac bridge]

hmm yeah, getting other poses is way more difficult... Could probably go crazy with the prompt to fool it, but that'd defeat the point of this anyways

[keen palm]

Draw me like one of those french ladies

[lilac bridge]

oooh that gives me ideas, thanks xD

hahaha sofa is OP

[mortal kite]

make the sofa out of grass

[lilac bridge]

LOL

[lilac bridge]

that broke it for some reason xD

[keen palm]

It's the chosen one 😮

[terse belfry]

Ppl will think it's real If u don't say that it's AI 😲 with some VSCO filters will be more realistic.

[keen palm]

Same seed and the 14456634 batch from start of it, i promise..it makes perfect hands

[torn wharf]

consistently creates 3 fingered hands though

[terse belfry]

Uncanny Valley

[lilac bridge]

Awwww so close to greatness

Going multiple ksamplers starting from 512, 768, then 1024 is helping to overcome some "safety" measures haha

Should also mention that the first pass uses this bottom textencode to trick SD3 into making a "safe for work" image, and then throwing the actual "laying in the grass" prompt into step 2 and 3

[keen palm]

Feels like you could use pony for the first sampler then?

[lilac bridge]

my life would be much easier with an rtx 4090, but my limited specs is forcing me to stick to just SD3 to avoid hour long workflows xD

[craggy crest]

avoid the ancestral samplers

well, if that's the effect you want, they are very good for that 🙂

i'm working on a sampler/scheduler comparison chart

i'll share them with you at least

[green shoal]

thanks for this, I was just writing a system prompt for this (not nearly as exhaustive as yours) and was scrolling up for clip_l/clip_g/t5xxl examples to feed in a few-shot style

we're not using negative prompts anymore for SD3, am I right ?

[mortal kite]

God, SD3 just has trouble with some words ugh

[green shoal]

from all comments and educated guesses I've read recently, and confirmed by SAI staff, seems like SD3 was not trained on negative prompts

[bitter hearth]

the negative prompt debate was confusing

[green shoal]

It's crazy how good the vision module of gpt-4o is

[craggy crest]

i never use negative prompts

[mortal kite]

a lot of inpainting and had to do the text myself:

[craggy crest]

try synonyms for those words?

[mortal kite]

I'm sure I could improve it if I ran it through an upscaler it might fix the inpainting mistakes

[craggy crest]

if you increase your steps, it does help the text

[mortal kite]

SD3 sucks at doing "worms eye view" like this

[mortal kite]

oh ..derp I think I was only at 20 steps, my bad

no, it just sucks at the word "Pathetic" no matter what

[green shoal]

it's also very powerful to ask GPT to rewrite the prompt and include some elements it didn't identify all the way or that you need to add to the composition

[craggy crest]

then try: deplorable,
feeble,
heartbreaking,
miserable,
pitiful,
woeful

[green shoal]

so, basically, really verbose and expressive prompts are the key to enable SD3'

[craggy crest]

no

[green shoal]

tbh, I'm really getting more improved images than my puny manual prompts

maybe I'm just bad at prompting SD3 in general..

[mortal kite]

she needs to say "Pathetic" , it's a meme

[torn wharf]

fetishes are weird. why do people insist on revealing them to the world? can't we go back to a time when people would keep that stuff to themselves?

probably a steep ask

[mortal kite]

this is the correct meme but SD3 can't draw this

[craggy crest]

that's fine, I'm just saying there's no 'key'.

[mortal kite]

skill issue?

[upper snow]

how is this not the correct meme?

[gaunt tiger]

this is the correct format 100%

[mortal kite]

I want to try the regional prompting in Stable swarm but I have no idea how to use it. The prompt itself indicates you can include keywords but I am not seeing any effect yet

[mortal kite]

They have enhanced the meme with woman looking down over her breasts at you, it's funnier

[torn wharf]

no boobs is why. theyr'e really important to some

[mortal kite]

because it's funnier

its like she's peeking at you instead, and barely even can stand looking at you

[torn wharf]

seems like reverse reasoning to justify a homelander attitude about it

https://tenor.com/view/boys-the-milk-gif-18445775

[mortal kite]

please don't talk about that disgusting show

created by hedonistic disgusting people

[bitter hearth]

hey chat

[torn wharf]

The Boys is hedonistic for parodying modern society and satirising superhero tropes, but Pony XL is fine. Weird

[bitter hearth]

pony makes elf girls, good model

draugrs looking good on unreal engine 5

put some long hair on them

Can you gradient hair lmao, very colorful and stuff

try with 80s glam rock wig

[keen palm]

Rolling bones?

[mortal mica]

Sorry for poor quality but recently my dad has been sending me photos of fake products generated by AI (obvious to me not to him) hah. So I decided to show him that it's pretty easy so I made water and beach sand flavored lays using sd3!

[bitter hearth]

xD

[craggy crest]

try clown hair

[upper snow]

if anyone's using the ODE solver node we just made changes to it to fix the broken progress bars and the overall problem that none of the fixed step solvers worked at all this entire time. it'll also work properly on non-flow (non-SD3/Lumina) models.

[bitter hearth]

wtf is that book of a prompt

LLM gen?

i figured, aint no way i'm writing a 500 word prompt

[cinder ether]

Here are some images I did with SD3:

[hallow lion]

Cigarette ad.

[bitter hearth]

would be cooler if there was a tiny snail on it

[torn wharf]

now try making them laying on grass

[cinder ether]

Thank you.

[cinder ether]

I tried to add a snail on the leaves and I think it did a good job:

I think the second one is great too.

[lilac bridge]

Oh dang, SD3 is actually going to be my new dnd item generator. img2img is highly recommended for gloves though

[torn wharf]

knew it. Can't even draw a leaf on grass. Dead model

[twilit cobalt]

Goliaths. I wanted to create guideline compositions based on MOSFET/GPU design schematics, but there's no controlnet support in ComfyUI for SD3?

[torn wharf]

only diffusers controlnets are available

[cinder ether]

I tried making leaves on grass, but I struggled a bit with the prompt. Here what it generated:

[torn wharf]

lol it's all good 'm just memeing. They're high quality crispy images all around

[craggy crest]

pick the right sampler/scheduler pair, get a normal looking hand. pick the wrong sampler/scheduler pair, however ...

[past flame]

My friend that hand on the left ain't normal unless this apple is extremely poisonous to touch

[hallow lion]

Reagan.

[mortal kite]

Heheh

[lilac bridge]

Swords! 😄

[split bramble]

that's Tifa...

[mortal kite]

trying to get action based lines.

ROFL

[vast condor]

sounds perverted

why is the morning person green

[mortal kite]

they are zombie they didn't have coffee yet

[torn wharf]

suggesting it should be censored? aren't we past that since it's been revealed that's from pretraining?

[vast condor]

that was somewhat tongue in cheek given your comments earlier about fetishes

[torn wharf]

its fun to mock the problem but lets be accurate about what the problem is. misinformation is poison

[mortal kite]

ummm. Kind of?

[craggy crest]

tell that to reddit

[torn wharf]

can't tell anything to reddit

[mortal kite]

LOL

[digital thorn]

Do you send it to SDXL or is it trained on Gaga?

[terse belfry]

Face changed with Retake

Generated without Lora

This is straight out of Sdxl

[digital thorn]

Ok, and with only sd3?

[terse belfry]

Opendalle fine tuned SDXL

[digital thorn]

Ok, but what is made with SD3?

[terse belfry]

Oh wait... I thought I'm posting in SDXL channel 😂

Lol. My bad...

[digital thorn]

No, you've been posting several times in SD3 channel lol. Not the first time we've seen you posting Gaga's here 😄

[terse belfry]

The others I posted was with SD3

Just face changed with Retake app

[digital thorn]

Ah, okey. I was thinking we had our first female celebrity trained in sd3 for a while

[bitter hearth]

wow the joker is cool

what was workflow like? the colors on the left are good

[craggy crest]

far more normal than the hand on the right 😉

[frigid saffron]

What's the right sampler/scheduler pair?

[frigid saffron]

Is there a lora training method for SD3?

[mortal kite]

I think people are attempting it, but I don't believe it is working yet

or if it will ever work..

[frigid saffron]

Hope it works, important for the future development for SD3

[craggy crest]

prompt: a green haired boy eating a blue apple - t5XXl encoder only, no negative prompt

prompt: a green haired boy eating a blue apple - clip_g encoder only, no negative prompt

prompt: a green haired boy eating a blue apple - clip_l encoder only, no negative prompt

prompt: a green haired boy eating a blue apple - all 3 encoders at the same time, same prompt, no negative prompt

[past flame]

Skinwalkers 💀

[craggy crest]

and just how far out into lala land clip_l went

[bitter hearth]

Is some chucky doll cursed movie

[mortal kite]

SDXL could never draw a speaker to save its life

[bitter hearth]

Huh, what was it spatting out?

[odd basalt]

Lizzie Borden

[mortal kite]

just weird random shapes, I could never get it to draw a raw speaker

like it would draw speaker "boxes" if the cone was't visible, it could never draw an actual speaker CONE

[sage burrow]

The ones which are physical impossible, even with juice, are extremely popular for some reason. I figure the reason that people like exagerrated aspects of the physical body is that it's just something that one never sees in person. Sort of like superhero comics, not realistic proportions at all.

[mortal kite]

like this : this is all SD3. I could never get SDXL to do this:

[mortal kite]

SD3 is also very good at isometric view it appears

[bitter hearth]

[limpid drum]

Oil paintings make everything more classy, including chocolate dipped bacon ...

[craggy crest]

Wendy's bacon shake...

[mortal kite]

i like the baconator but let's be real that thing will put you to sleep

[sage burrow]

Hmm, I did it all with just one prompt, but now I wonder if I can talk about each panel set seperately

[mortal kite]

in Stable Swarm it kind of looks like you might be able to make image layers, that would be useful for adding subjects in the foreground if I could figure it out

[dusky thistle]

love this

[sage burrow]

the top 2 always look amazing, then the bottom two look like the artist got board/ran out of time

[brittle nexus]

I think the SD3 is horrible for any kind of angle. How are you doing low angle?

[mortal kite]

"a comic illustration: a tall muscular brunette woman standing in a forest. The camera is looking up at her from ground level in a worm eye view. "

and then just keep genning til it gets it right...

[brittle nexus]

I will try. Ty

[sage burrow]

ordering some at the bar 😄

[sage burrow]

So, character consistency, how is that done is SD3? I'm guessing the usual, describe the character in detail, the colour of their clothing, and perhaps give them a name?

[sage burrow]

I get perfect hydras with SD1.5, but not with SD3 <pout>, they are still awesome though

[torn wharf]

https://tenor.com/view/breaking-bad-hector-salamanca-sad-emotional-gif-17812145

[upper snow]

make it do loss

[faint breach]

the day that CAD lost me forever as a reader. i'll always remember it. "Wtf was that comic? Out!" I think i was so upset i stopped reading Penny Arcade too

I've come back to PA a few times though. They're good guys. They've changed though.

[raven fern]

that almost looks like something from an anime 😮

[low stone]

I've discovered anime art. With lykon's sdxl checkpoint with the same stuff in it as refiner.

[low stone]

sd3 2b

[raven fern]

very wide pic :3

[raven fern]

just like people generating big booba, very popular.. "for some reason"

[raven fern]

aww that's a cute one

[craggy crest]

thanks 🙂

[raven fern]

and what did you conclude overall?

from what i am looking at personally, i like the results from steps 15 and up along with cfg 2 to like 5

[raven fern]

nice

[sage burrow]

Loss??

[upper snow]

|  ||
|| |_

[raven fern]

assuming by PA you mean PixArt, i saw some news on reddit, they are working on the next model with Nvidia? 😮

[coral sable]

Stability also works with Nvidia, and AMD

[raven fern]

yea i saw similar conclusions from other people too

[raven fern]

i did some tests recently with like TCG cards that have a lot of text, it seems higher steps helps to get better more coherent numbers and text on the cards, even tho it's still gibberish lol, so it really depends on the type of thing you want to generate

[craggy crest]

you can easily overcook an image.

the more steps you have, the more passes the AI will make, and the more details it'll put in. and that gets to point of no return very rapidly

modified HEAVILY by the sampler you are using (and other values)

[coral sable]

xy plot is best to determinate optimal settings

cfg is too high drop it to 3.5 at least

[craggy crest]

DMing you

[coral sable]

his GF arrived

[raven fern]

she can kill me any day

but nice colors tho

[noble coyote]

"Ollama-up baby!" The problem with a lot of Generative AI - it has all the answers - just that too few know how to formulate the questions!!! Get gud skilz ! ! !

[noble coyote]

Ollama indeed! 😄

Try Jan.ai - free LLM generator for prompts and more. Ollama, ChatGPT (via Quality of Life nodes), Anthropic's Claude 3.5 Sonnet etc.

"Get Gud Skilz!!! ™

[sterile pendant]

The model acts like an ancestral or SDE sampler where adding or removing steps will cause the image to change. SD3 usually hits a hard diminishing returns around 35-40 steps

Which is why the suggested default step count is 28

If you want to get micro variations in an image, add in some , , , , in the prompt one by one

But leave the steps and everything else the same

on a side note, comfyui has a new sampler type that seems to work really well with SD3. it's the eulercfg++ but you have to use the alternative mode option of it. obviously, use it with sgm_uniform sigmas

it's based on https://arxiv.org/abs/2406.08070

you use it like this, obviously, finish hooking up the other stuff needed

[dusky thistle]

yeah this is pretty damn good

[sterile pendant]

showing a side by side of eulercfg++ vs euler, everything else the same

just using florence-2 for the prompt off an old generation because i'm lazy

[dusky thistle]

yeah i'm finding some really nice improvements myself

[sterile pendant]

i'm going to keep tinkering with it, but out of the dozen or so side-by-sides i've done so far, it's won every time for me

[dusky thistle]

yep same here

sometimes resoundingly, including vs dpmpp_2m

[sterile pendant]

yeah i should throw dpm++ 2m in as well to compare all three

[warm sonnet]

How to get SamplerEulerCFG++? I had upgrade to the lasted comfyui but can not find it

[sterile pendant]

scroll up a handful of posts, i showed an image with the setup

you have to use it with the custom sampler

[dusky thistle]

dpmpp_2m vs eulercfg++ with my weakass sd3 lora

stil ljust can't get the loras to do much of anything

[sterile pendant]

dpm++2m on the right

[dusky thistle]

dpmpp_2m, cfg++, euler

[sterile pendant]

basically the same as euler

[dusky thistle]

huge improvement there. it's a shit image, and suddenly great

[sterile pendant]

yeah ive seen a few where it's drastically different than the other samplers

like even shifted the art style

[dusky thistle]

yeah

very cool... what i like the most is how much it improved the coherence

got two major issues with sd3 myself: coherence, and how the hell do we train it?

[sterile pendant]

dunno, i've mostly been working with pixart sigma lately and doing refinements with sdxl

[dusky thistle]

i wonder how many tens of thousands of watermarks are in their training set

left is base, right is my most recent test lora

i manually shop out all watermarks in my training sets, i don't leave any bullshit in them

[sterile pendant]

if it's art related, almost every single image would likely have either a watermark or signature. have you ever seen a painting that didn't have a signature on it in the bottom right or left corner usually? by definition, the fact that the models try to add them is more correct than if it didn't

but i feel you, we dont want them lol

[dusky thistle]

yeah, just never used a finetune or a base that had them so consistently in the image

[sterile pendant]

they probably didnt do as much random cropping and whatnot and actually trained with the full images

[dusky thistle]

mostly cascade myself^

yeah, prolly

[sterile pendant]

here's one where it changed the style up, but in this case, nails the prompt far better

i mean look at those black and orange flames, glorious

yeah i think this eulercfg++(alternative mode) is going to be my new go-to

[dusky thistle]

Yeah agreed

[dusky thistle]

By far the biggest thing is the improvement in coherence. There's a lot less separated object crap going on. Anatomy is still horrible but in general there's a substantial improvement

[sterile pendant]

i kinda wanna try it out with sdxl, since i can test both modes of it there

[dusky thistle]

Oh, yeah

[sterile pendant]

id have to look at the code to see what all it's doing differently between regular and alternative, but obviously the alternative mode isn't doing w/e ancestral-like stuff in it

[noble coyote]

I've installed the CFGpp node but I cannot find SamplerEulerCFG++ at all?

[sterile pendant]

it's part of stock comfyui now, you have to just update comfyui

[sterile pendant]

looks like the regular mode is broken, i've tried it with every other sigma type as well. cfg is only 4.5. oh it works fine if cfg is set to 1 though

this is with sdxl

[limpid drum]

Hmm... I did three minutes ago... not there. Will try again

[sterile pendant]

cfg1.0

[craggy crest]

did you do an update all?

[limpid drum]

Yes.... that's why I said I did 🙂

[limpid drum]

Something is goofy with my symlinks. maybe that's it

FETCH DATA from: C:\Applications\StableDiffusion\ComfyUI\ComfyUI\custom_nodes\ComfyUI-Manager\extension-node-map.json [DONE] JSON file does not exist. Update ComfyUI Cmd('git') failed due to: exit code(1) cmdline: git stash stderr: 'error: 'models/checkpoints/put_checkpoints_here' is beyond a symbolic link fatal: Unable to process path models/checkpoints/put_checkpoints_here Cannot save the current worktree state'

[craggy crest]

oh. yeah.

maybe use full paths instead

[limpid drum]

The JSON error might have to do with the standalone Windows version? Who knows, will try later.

[limpid drum]

Figured it out, the put_checkpoints_here file was missing from the symlink 🙄

All good, have access to the CFG++ now to play with and still somehow mess up the art 🙂

[sterile pendant]

yeah im thinking the cfg range with this eulercfg++ node is mapped weird or something. images all the way down at 1.0 cfg come out completely normal and if i bring it up even to 3-5, it starts to behave like you've got your cfg set to 8 or something(in terms of almost burning the image with too much contrast)

maybe this will be like cascade where you use 1.0/1.1 for cfg

either way, i'll have to play with it more to find the sweetspots for it

[scenic shadow]

the link you posted earlier did say it was better at lower cfgs ... though it did say it is supposed to be better across "all scales". just now getting a chance to catch up on things

has there been any good updates in the past few days other then this cfg++?

[limpid drum]

Are you using the regular or alternate version? This is CFG 8 on "alternate", regular is bad off the bat

[maiden monolith]

/help

[sterile pendant]

Oh from the actual paper "The resulting process, which we call CFG++, works with a small guidance scale, typically λ∈[0.0,1.0]" so yeah it likes low cfg. Not positive, but I think comfyui's cfg 1.0 is actually zero, so 1-2 might be the right range for this sampler

If not, then the range would just be 0<x<=1

[limpid drum]

CFG 3 is not bad, trying 2.0 next. Or, going to bed because I'd have to be up in less than five hours

[sterile pendant]

So maybe ranges in the 1-3 are like regular samplers in the 2-8 range or something. As you get closer to the cap, it starts to burn

[limpid drum]

I think so, CFG 8 was burned bad, but oddly worked well on my old workflow (where CFG 4 was blocky)

[sterile pendant]

The fun of trial and error experimentation

https://tenor.com/view/scientist-science-im-a-scientist-gif-14090115

[placid swallow]

it's not an anime girl but eulercfg++ def seems to follow better

[placid swallow]

of course after i say it's not an anime girl it makes anime girls

[noble coyote]

Wow!!! This CFG++ is really making a difference - REALLY!!! Incredible!!! Just drop the CFG from 8 to about 2.5 or 3 before you start!!!

1st effort CFG++, 2nd effort no CFG++

[sterile pendant]

testing from 0-7 cfg with cfg++

0.25 cfg increments, starts with zero

[noble coyote]

Nice!

[craggy crest]

@lavish osprey making sure you don't miss this discussion

[noble coyote]

And the text has never been crisper!!! Still garbled, but so much classier!! 😄

[sterile pendant]

ill make some more test animations like this but with a tighter range now that i know what to expect. like from 1-3 but with smaller increments

[noble coyote]

CFG++

[noble coyote]

Look! Readable text - not just the odd time - almost every time!!!

[noble coyote]

CFG++

[warm sonnet]

seem CFG++ not work with SD3 Tensorrt, tried but not luck

[scenic shadow]

that is utterly horrifying

[noble coyote]

Get friendly with an LLM! "Get Gud Skilz!" ™ SD3 including CFG++

You may be doing overkill if you implement CFG++ with a ReFINER. CFG++ really improves text!!!

[sterile pendant]

here's another cfg++(alternative) example with SD3

it definitely looks like it's 1-2 for the "normal" range

though conveniently, if you're trying to do more posterized illustrations, with less shading, cranking it up higher actually helps quite a bit, without ruining the whole image

i'll do another one in a realistic style

[noble coyote]

CFG++

CFG++ definitely increases the rate of keepers when it comes to text!

[frigid mango]

These are a great help

[sterile pendant]

no problem, figure if im doing it for myself, might as well share some results

On a side note, I was thoroughly pleased to see there was a text overlay node in one of the common packs I'd had installed. Makes it easy to overlay the concatenated cfg: and whatever the selected value for the cfg is for the frame

[muted dove]

Close to a person sitting down! I had to blur what was going on in the background because his trousers were pulled down 😄

[sterile pendant]

"a portrait of a young woman with a dramatic makeup look. She has long blonde hair styled in loose curls and is wearing a large flower crown on her head. The crown is made up of pink and orange flowers and green leaves. The woman is also wearing a gold necklace and earrings. She is looking directly at the camera with a serious expression on her face. The background is black, making the woman and her crown stand out. The overall mood of the image is dark and mysterious."

it really is looking like the sweetspot for cfg++ and sd3 is around 1-3

well more around 2-3, but it depends on what you're going for

i'll do a couple more before the morning chaos begins aka the kids get up

[signal shuttle]

Amazing work RX808, CFG++ really does work super well with SD3

[sterile pendant]

thanks, and also just an fyi, it's a completely default workflow, nothing special, just model and prompts straight into the samplercustom, 30 steps

[frigid mango]

also with cfg++?

[muted dove]

Yes

[green shoal]

where are you hooking up the SamplerCFG++ node in your workflow?

[sterile pendant]

i showed a screenshot of it earlier

#🆕｜sd3 message @green shoal

there, just hook everything else up you need

"a portrait of a young woman with long pink hair. She is wearing a futuristic outfit with a black and silver design on her face and neck. The outfit appears to be made of metal and has multiple layers of armor. The woman has a serious expression on her lips and her eyes are looking directly at the camera. Her hair is styled in loose waves and cascades over her shoulders. The background is black, making the woman and her outfit stand out. The overall mood of the image is dark and edgy."

[green shoal]

thanks! no modelsampling to adjust the shift ?

[sterile pendant]

you can pipe that in if you want, the default for ksamplers is 3.0 anyways

[noble coyote]

#🆕｜sd3 message

[sterile pendant]

everyone's favorite, let's see if cfg++ helps it at all lol:

[digital thorn]

All you have to do is to update comfyui to get the node?

[sterile pendant]

ja

[digital thorn]

okey, ty 🙂

[sterile pendant]

and if youre going to use it, id recommend using the alternative mode and keeping your cfg between 1-3 or 4, 2 is looking like a pretty decent sweetspot

[muted dove]

🤣

[sterile pendant]

looking like a biblically accurate depiction of an angel, well seraphim

[digital thorn]

2 in 1

[sterile pendant]

god i need a new gpu for this pc. this 2080 is showing its age. we have a 7900xt downstairs, but well you know how SD is with AMD, it probably has the same it/s as this card does lol

Alright last one for the morning:
"a digital illustration of a futuristic landscape with a red planet in the background. The planet is a bright red color and appears to be glowing with a bright orange glow. The sky is filled with stars and planets, and the horizon is visible in the distance. In the foreground, there is a rocky landscape with several tall buildings and structures scattered across it. The buildings appear to be made of stone and have multiple towers and spires. The structures are of different sizes and shapes, and they are arranged in a way that creates a sense of depth and dimension. The landscape is barren and rocky, with patches of grass and shrubs scattered throughout. The overall mood is dramatic and awe-inspiring."

[digital thorn]

pcm lora doesn't work with the sampler unfortunately

[cobalt moon]

holy sh

[sterile pendant]

did you try it with both modes?

[cobalt moon]

kohya ss supported SD3 finetuning now ( full model )

[digital thorn]

Yep

[green shoal]

anybody got any decent results with SamplerEulerCFG++ regular ? it's either blurry blobs or pixelated burns for me

[sterile pendant]

it seems like you can't go above 1.0 with it without it blowing up

[cobalt moon]

oh yeah I gonna do my captioning real quick ( although Kohya did say that it is buggy )

[sterile pendant]

oh and that was with sdxl, with SD3, it doesn't work i don't think

it's probably ancestral/sde based with regular mode, which don't work with sd3

[muted dove]

A little cropping got rid of the feet above her arms and the stumps at the end of legs 😄

[digital thorn]

Nvm, it's something weird after I updated comfyui...........I shoulnd't have updated 😦

Before and after comfy updated...great

[sterile pendant]

they might have fixed something under the hood with loras that an addon was using a work around for

[digital thorn]

I've got these versions:
Python version: 3.11.8
pytorch version: 2.3.1+cu121

Error. No naistyles.csv found. Put your naistyles.csv in the custom_nodes/ComfyUI_NAI-mod/CSV directory of ComfyUI. Then press "Refresh".
comfy_extras.chainner_models is deprecated and has been replaced by the spandrel library.

Same errors before I updated comfyui, but different outputs and output was working before updating it.

[sterile pendant]

yo... wtf kinda data they got in this dataset... i accidentally typo'd 0.1 instead of 1.0 for cfg and got this

looks like some kind of science experiment documentation or something with all the timestamping at the bottom

[cobalt moon]

unexpectly good compared to SDXL?

although no one really use 0.1 CFG

[sterile pendant]

right, but going below 1.0 cfg almost always results in some weird feverdream shit, but this is just nightmare fuel

looks like having a stroke visually

[digital thorn]

What python and pytorch versions do you have?

[cobalt moon]

Python 3.11

SD recommended 3.10 though for compatibility

[digital thorn]

Ok, then I know python isn't the issue. What about pytorch?

[cobalt moon]

I think most popular SD application ( ComfyUI, A1111 ) used the latest? But when downloading Pytorch it usually attached and embed to GPU's CUDA

downloading through pypi ( pip install torch ) won't do that if I wont wrong

and also 12.1 and 11.8

[scenic shadow]

i have the same versions, prob isnt a version problem with those 2

[digital thorn]

Okey, it's just weird that first version of comfyui works great for me, but when I update it fucks up the generation :/

[cobalt moon]

you can launch it?

[digital thorn]

yes

[scenic shadow]

workflow?

might be a custom node acting up?

[digital thorn]

It's not any issues with workflow. not even a basic ksampler works

[scenic shadow]

when was the last time you hard reinstalled comfy?

might be worth doing a fresh install in another folder and double checking

[cobalt moon]

time to fuck my captioning of my entire SD3 dataset

[digital thorn]

today, since I got the same issue when I updated comfyui the other day

So it worked when comfyui was download from portable without update, but stops working (bad gens) when I update it.

[scenic shadow]

I know it got updated for the new CFG++ eular sampler, might of had under the hood updates that changed things ... your def beyond by ability to help though sorry ...

have you played with settings to see if you can get the new "bad" gens to line up closer to what your expecting?

I will say I'm running my comfy directly out of a stable swarm install

[scenic shadow]

but the error points to the custom node

quick google search later https://github.com/comfyanonymous/ComfyUI/issues/3463

looks liek its your universal styler custom node

[digital thorn]

Right now I can only compare to my last generation which has larger workflow. But I think it still looks different, Would you be able to see if there's a difference? (workflow included)

[scenic shadow]

id see above first.

[digital thorn]

Shouldn't be the issue since it was already that way before comfyui updated though. But I'll try to fix it again

[scenic shadow]

one is a file that was made on the 14th one is a generation from today with the same workflow ... its prob a custom node

[digital thorn]

I just notice that it's the pcm_stoschastic lora that doesn't work for me in the new comfyui update

[scenic shadow]

ah im still running without loras atm so that make sense.

[sterile pendant]

digging the non-euclidean geometry in this one lol

[scenic shadow]

WTB a lora or hopefully a finetune model that knows how a bow is used

[bitter hearth]

Want To Buy ? Cx

bows are cool why they never work

[scenic shadow]

yeah, sorry old gamer slang

did some more training of my gpt

went with elvish ranger in a vaporwave style

good outputs but the bows are killing me

[bitter hearth]

I play poe, that's the only meaning for WTB!

[desert garnet]

in another game it also means willing to blow

[scenic shadow]

think it might of been the same on craigslist before the personals section was removed

[low stone]

Who needs sd3 when you have cascade

Grotesque Pikachu creature, taking a selfie, burning football stadium, sky filled with alien ships, eerie green lighting, chaotic crowd fleeing, otherworldly creatures descending, intense smoke everywhere, neon electric bolts, unsettling facial expression

[scenic shadow]

3/3 runs ... Im kinda proud of my gpt atm.

[sacred jewel]

Or what it looks like 🤭

[sacred jewel]

QUick question for the model: Where the heck is all that water for the waterfall coming from? LOLOL

[sacred jewel]

" [my gpt] " ??

Do tell 😲

[scenic shadow]

I've trained a gpt to output a T5, G, and L encoder prompt

[silver sluice]

is it a single prompt or did you require a aseries of conversations to train it to do that?

[scenic shadow]

did a session of training in the builder to get it tuned

[scenic shadow]

Its a full custom gpt built in the my gpts section of ChatGPT

[silver sluice]

is it shareable?

[scenic shadow]

this should be a link
https://chatgpt.com/g/g-SrU9OKDp7-prompt-research-assistant

[scenic shadow]

its not perfect, but its a good starting point.

if you dont use a seperate g or l prompt you can always just use the T5XLL prompt as the main.

[sacred jewel]

JESUS CHRIST you're my hero... been looking for this... where can I steal this golden nugget?

[silver sluice]

cool dude thanks for sharing ill play with it in a bit, i was talking to @craggy crest the other day and he was demonstrating how much better output is when you split the prompts

[scenic shadow]

posted a link

[sacred jewel]

See it...

[silver sluice]

check out that replied comment you can see the depth and color range is greatly improved when you let each encoder focus on it's own thing

[scenic shadow]

yeah ... i got the idea to do it from him. then used his base foundation of what each encoder wants to start with

[silver sluice]

oh wow it all comes back around, good stuff thx again for sharing

[scenic shadow]

well the idea that im too lazy to come up with 3 diffrent prompts at least XD

[bitter hearth]

how were you able to do that with free gpt?

[scenic shadow]

i pay

[bitter hearth]

so you are paying for gpt to let everyone use for free?

[sacred jewel]

Yup... so do I ... both UI and API... worth it for the power and ease of use.

(disclaimer, I use it a lot for my day to day job, not just the hobby)

[scenic shadow]

no i pay for gpt plus which means I get to play in the builder and have higher message limits.

[mortal kite]

Comfy and Stable Swarm, etc, IMO really are too far behind in Inpainting powers, Fooocus is just so much better at it. I think Comfy has a Fooocus inpainting node? Without such powerful inpaint, honestly things are a bit more useless

[bitter hearth]

damn, that's amazing ... thanks for that im gonna play with it

[sacred jewel]

Why are you now happy with Foooocus? why Use Comfy if you like Foooocus?

[mortal kite]

Foooocus always inpaints and makes things look like they belong in the scene...Stable swarm default inpaint still you get blur and some seams

[mortal kite]

Um because Foooocus doesn't support SD3 yet so I have to use inferior products

[silver sluice]

have you played with claude 3.5 btw? as a paid gpt member i feel kind of offended now that there's something better (as far as coding) than openai's system, like it's remarkably better too not just slightly better

[signal shuttle]

Comfyui is in no way inferior to Foooocus, Comfyui is really good if you knew how to use it, Foooocus is more friendly to less tech-savvy people

[mortal kite]

it IS inferior with inpainting

[fast dew]

Should make this into a Yu-Gi-Oh card

[scenic shadow]

i pay for gpt becuase i use it to help run a dnd campaign. thought about playing around with it soon but its a pain in the ass to flip all of the campaigns information that has been made generated in one ecosystem to the other

i know becuase i went from claude to chatgpt when 4 came out.

[signal shuttle]

Claude 3.5 is really good at understanding different concepts, i pay for it and use it to generate prompts, and its far better at understanding my instructions

[sacred jewel]

Results of a test...

The one with the taxicabs was from a set of prompts from your trained GPT
The other one is from my own asking the general GTP-4o to write me separate CLIP-G, -L and T5 prompts

[bitter hearth]

cody does it cost you for ppl to be using that gpt prompt?

or are there caps on how many ppl can use?

[scenic shadow]

no, they made using 4o free for everyone it just has a tight usage limit

[bitter hearth]

ahh sweet thanks man

ppl are gonna use it heavily

[mortal kite]

I mean how can this even be useful? Inpainting a soda. Stable Swarm..you always get fuzzies

[scenic shadow]

yeah the robot one in its suggested prompts kinda jank for me. the main thing is that you can back and forth with it to get it better

[silver sluice]

i use it more for ad-hoc programming like just random stuff I need to code rather than relying on it for an entire built up system but I could understand how you're already deeply entrenched into one system and for your use case it doesn't make sense to switch back and forth, for me i have no issue switching back and forth, whoever gives me the best answer wins personally

[mortal kite]

if you did this in Fooocus you'd never even tell. And it's not the settings, I tried them all, its special magic that Fooocus does

[mortal kite]

almost like it does a AI content fill after the inpainting to make everything seamless

[scenic shadow]

if i inpaint i always go back to sdxl then come back to sd3 to upscale

[sacred jewel]

BTW correction: I meant "not" not "now" :;

[mortal kite]

that might work - because the issue I was having is if I go back to Fooocus just to inpaint I would lose detail right, but i guess if you come back to sd3 and upscale you could get it back

[sacred jewel]

Are you using the official ComfyUI example workflow for inpainting or you built your own?

[scenic shadow]

honestly i really like invoke for inpainting. havn't used it in a bit and im sure fooocus is just as good of a program. its just a nice all in one package that has a clean ui and a really nice universal canvas

[mortal kite]

This wasn't comfy is is Stable Swarm (which uses Comfy backend)

[sacred jewel]

Also keep in mind, there are no specific inpainting models yet for SD3... that I know of... and Differential Diffusion I have not tried yet if it works with SD3 as it does for SDXL

So, in ComfyUI, you are using a proper inpainting model for SDXL?

ChatGPT is not loading for me right now... anyone else? My sessions to it are just spinning

[mortal kite]

I heard that Comfy does have a Fooocus inpainting style node, though. But I don't know if it would support SD3 or not

[sacred jewel]

I love how it JUST GIVES YOU the meat 😄 No fluffy explanations ... lol. Trying it now. GPT is working here again

[frigid mango]

awesome thank you! are the generated prompts safed so a possible databaase could be made?

[sacred jewel]

do you or anyone know how to reference a trained GPT like yours or others if using the ChatGPT API?

[scenic shadow]

wanted to do a side by side of mine then yours

same seed

think we are both pretty much on the same path

mines not super trained for it but lets see

[sacred jewel]

Source image and result from SD3

[mortal kite]

what look like if you just put all in one prompt