🏞｜general-with-images

[clever oar]

oh anathomy still bad in new model

extra legs

[grizzled cosmos]

is it possible to inpaint with SD3.5?

[clever oar]

dont know i use it online

[languid pebble]

Ahoi!

[clever oar]

I don’t understand why it’s so difficult to fix anatomical problems because flux doesn’t have this, even sd 1.5 is better

[crisp stream]

I wouldn´t necessarily call it bad in case it (occasionally ?) adds an extra leg only, as long as the shape is fine. Yet from what I´ve seen here so far, as well with the black forest gump on the bench FLux/SD 3.5 comparison, SD 3.5 still doesn´t seem to be doing too well with hands and stuff

[clever oar]

i think SD is better than flux in working with styles, sometimes in details and colors. i sad that this problem exists

[steep sleet]

Early days - we'll know by Christmas as to whether 3.5L is a good SD Model; or just a large bottle of milk!!! 😄

[clever oar]

on the other hand, you can use SD3 where there are no people or living beings🙂

Mojo

you try it?how fast you make one picture

[languid pebble]

Not really fast ...

[clever oar]

one min?

[languid pebble]

Hard to compare cause I upscale my flux picture. 2 Minutes in Energy safe mode maybe

[clever oar]

nice result?

[languid pebble]

I switched back to Flux ... but no complete failure ...

[clever oar]

good you have hope😎

[crisp stream]

Cheating (Flux with locally trained LoRAs):

[languid pebble]

Let's don't talk SD3.5 dead and wait for the finetunes ... base model has never been good

[crisp stream]

prior to SD3 they actually have been, well except for 2.0 🙂

[languid pebble]

But we all have been working with the finetunes ....

If I understood right the base models aren't trained 100% so we can do finetunes without overtraining it ....

[crisp stream]

Well, not really here, while I started using dreamlike photoreal 2.0 for SDXL at some point, the basemodel was already fine, just like with 1.4/1.5, even though finetunes naturally improved the overall output qualities. Yet the base model was a pleasant basis, while I feel with SD3 (/apparently 3.5) and Flux that isn´t the case anymore, reason it´s hard to produce anything out of the boring realism and 3D comic look, because the training material has obviously changed a lot. With Flux for example it looks like they at least for quite an amount trained on a Chinese dataset, with lots of stock footage appearances. Reason I basically went to generating with Cascade again. This merely changed after having started local training for Flux.

Also addressing artists doesn´t really work, neither in Flux nor in SD3, don´t know how it is with SD 3.5, just guessing it isn´t any different than SD3, yet who knows

[crisp stream]

I don´t know, it´s more like what they trained on and less how they trained

[languid pebble]

But think of LoRas. You can overtrain them. I think it's the same with models. So if you want the users to be able to create their finetunes you can't train it 100%

Maybe I'm writing garbage ... it's just how I understood it ....

[crisp stream]

I´m also wondering why SD3 has got 3 different textencoders, making prompting a hassle to work with and best is the textencoder doesn´t even fully grasp the milelong "natural" language prompts, yeah, I know they can be reduced, still, they still struggle then and basically having to translate a prompt into that textencoder language by using an LLM feels counter-intuitive and decouples the artist from the actual prompt engineering to at least some extend. Isn´t there a textencoder being able to understand at least keyword and natural language in one go? Flux luckily "only" got 2, still too much if you ask me. Yes I can combine both, so the prompt will be sent to each of them, then that´s not really the way how its use is being intended and that for, at least how I experienced it here, only partially works well.

[royal monolith]

because they each do different things. you should really prompt each of them seperately

[crisp stream]

yes, that´s the thing I´m talking about, why have 3, again 3 different text encoders to begin with, like said, especially in combination with that natural language-LLM approach, supporting generic results anyway. Natural language includes all types of languages including keywords, at least that´s how we funktion 🙂

[royal monolith]

because they all do different things. and IF you give them each prompts that work to their strengths, you get much better end results than if you just use one of them, or give them all the same prompt

think of three different humans - a mathmatician, an artist, and a chef. put them all together on a project and tell them all to 'bake a cake together' and you'll get a mess. but tell each of them to do specific things for the project, you'll get a much better result

[crisp stream]

yes, that´s why I´m asking if there isn´t a textencoder understanding a broad width of language styles

[royal monolith]

text encoders are models. you coudl train one but at the moment, no, there isn't

and they're just as expensive to train as base models are

[crisp stream]

the model could simply differentiate between the varying input styles and process them in their individual way of understanding.

[royal monolith]

you'd have to train one for that. thre aren't any out there at this moment that are trained to do that

[crisp stream]

then LLM, at least something like ChatGPT seem be smart enough to understand basically any style given or can make sense of it and sort it into its varying parts. Then I´m not too much into the technicalities behind LLM and how they actually can work like a textencoder, it´s simply a thought

[clever oar]

tried to make a moofi avatar😃

[clever oar]

last try🙂

[languid pebble]

Fail 😛

[clever oar]

only moofi know secret how do this😃

[stuck totem]

@late sorrel i cant post images in the other channel but heres an example, and also i think i realized the major problem, the frames are the exact same color as the eyes so 100% the ai got confused, which is unfortunate because the other parts seem to be decent in consistency

[late sorrel]

that looks good man

[late sorrel]

ah ok, now i see what you trying to do

that indeed can be tricky

[stuck totem]

i think next time ill train it without the glasses or just make the glasses a different color or something, idk

[late sorrel]

well yea cause you can change the color of the glasses with a text prompt anyway, just learn the concept first

[stuck totem]

yeah im new to this, either way its exciting though, some generation epochs are pretty decent and not complete trash so its promising at least, just the glasses that need fixing

[late sorrel]

making a lora is like half science, half art lol, you mess with the params in a certain way until it gives you what you are looking for, there is no definite way to train. i mean there are things that could be considered objectively better than others, but overall i would say its still a dance with the params

[stuck totem]

true, no wonder stuff like controlnet and inpainting were developed, ai art is completely unpredictable and you need all those tools to make it do what you want. But its real exciting when you finally get a generation thats coherent, or at least good enough to minorly edit with inpainting, it truly is a work of art

[late sorrel]

yea

lot of fun

[candid river]

This is what I'm trying to do, just way better

[late sorrel]

i see

[candid river]

It butchered my prompt, I tried have chains as like hip and boot accessories

And the person standing etc

[late sorrel]

but wait what are you using to generate? sd3.5?

[candid river]

I'm not using any kind of stable

I'm using a image generator on a site called "perchance"

[late sorrel]

well the first problem with online stuff is that they are very limited in workflows, not to mention if they are limited in censorship as well

[candid river]

There's a bunch of awesome AI generators on there made by people, for anything from fantasy towns, to plants, to minerals, etc

[candid river]

This site allows one to pump out 30 results for free per batch no credits used

And there are no restrictions for content

[late sorrel]

well idk what that website is using under the hood, because if you have that information, then you can technically know how to prompt it correctly along with the params

[candid river]

The one thing I can't do, is use an image as an asset on it

[late sorrel]

like image to image?

[candid river]

Kind of, yeah, but more so using the image to guide the prompt

To get an image

[late sorrel]

yes that is image to image

[candid river]

Oh my bad

[late sorrel]

because the guidance comes from the clip conditioning on the image, it receives info from the image. but its better to guide it with text on top anyway, not just image

[candid river]

Ironically enough, I can actually draw freehand, and have been for over 13 years. I primarily draw on paper and use digital stuff to make it pop more

[late sorrel]

i wish i could draw, i would make some very interesting "manga" :3

[late sorrel]

is that from the freddy game? i never played it, just seen some stuff lol

[candid river]

Yea

I also have a few oil pastel arts of nagatoro hayase

[late sorrel]

you are speaking to me as if i know all of this :3

[candid river]

And a colored pencil of a clicker from the last of us

[late sorrel]

i mean to be fair, i know a lot of characters, but i maybe dont know their names

for example, the danganronpa games, or however you spell that, i know the characters because i seen them, but i dont think i know any of their names lol

[candid river]

https://tenor.com/view/nagatoro-happy-anime-senpai-smug-gif-16186799704885949550

[late sorrel]

huh.. not sure if i know that character

[candid river]

That's nagatoro

[late sorrel]

where from?

[candid river]

Don't toy with me, Miss Nagatoro!

[late sorrel]

yep i dont know that one 😦

[candid river]

I plan on drawing Marin Kitagawa at some point

[late sorrel]

is that a newer anime?

[candid river]

("my dress up darling", it's an anime series about two unlikely friends, a doll making male, and a cosplay gaming and manga geek, female, they make costumes and go to conventions

And do photoshoots and stuff

[late sorrel]

what year did it release?

[candid river]

2022? Maybe

[late sorrel]

ah ok so yea im not that familiar with the newer anime

[candid river]

I drew denji from chainsaw man

[late sorrel]

i barely even watch these days

now i know chainsaw man lol

[candid river]

And I might try drawing power next too

[late sorrel]

damn

[candid river]

Here's my crate of "seasonings" they let me cook

[late sorrel]

it's kinda interesting, cause i remember having fun drawing some stuff as a kid, but i never really continued that path

[candid river]

This is another one I did freehand

[late sorrel]

nice

[candid river]

It's a scanned paper artwork

[royal monolith]

@wispy nest 2d cell-shaded anime image depicting a woman with long flowy blonde hair, wearing bold makeup, in her casual t-shirt and denim pants, navel cutout, indoors, solid red background. dramatic perspective.

[wispy nest]

that done with sd3.5?

cause im getting this w/o artist ref 2d cel-shaded anime , a woman in large white t-shirt, bare legs, at a beach as the sun is setting. the ocean water is deep turquoise with orange hue in the sky.

[royal monolith]

yes. workflow is in it, click the image, open in browser, right click, save as, drop into comfy

i've got no problem using an artist ref, i just tend not to need to usually

[wispy nest]

hang on, my workflow is ok, but i might be missing something with prompting

here is another example which i dont understand ... it should render anime..

a woman in large white t-shirt, bare legs, at a beach as the sun is setting. the ocean water is deep turquoise with orange hue in the sky. Anime style by Katsuhiro Otomo

[royal monolith]

change anime style by katsuhiro otomo to: by artist "Makoto Shinkai" <--- and see what you get

[late sorrel]

nice water :3

[wispy nest]

not too fussed with the anime thingy atm, im liking the overall image aesthetics

[royal monolith]

that AI just be trolling you

[wispy nest]

one thing that stands out for me with sd3.5 is that it can render full body shots lot more fluidly than flux

[royal monolith]

i've been telling people all this time that flux is not what they thought it is. sd3.5 is fluid, it's flexible, and it's almost effortless to use. and it IS trainable. and the base model really doesn't need any loras

[wispy nest]

im using the distilled version of sd3.5.. aka large turbo

[royal monolith]

supercute chibi anime woman in large white t-shirt, bare legs, at a beach as the sun is setting. the ocean water is deep turquoise with orange hue in the sky by artist "Makoto Shinkai"

[royal monolith]

i'm just using 3.5 large directly from the huggingface page

[wispy nest]

i had to struggle quite a bit with flux to render ful body shots, it feels like flux focuses on upper body only how its trained

[royal monolith]

FYI - you can get 3.5 down to 8 steps and it still looks good. cfg 4, steps 8, shift 2

[wispy nest]

yeah that upper body thing is DPO
DPO likes to "zoom in" a lot and fill the frame with the subject

[royal monolith]

you're using turbo. i mean sd3.5 large, not turbo, can still go down to almost no steps

[late sorrel]

im not familiar with that artist, but just wondering, does it really know that artist style? how closely does it match?

[wispy nest]

oh

[royal monolith]

i dunno. i don't use names cause i want the style to match. i use them as concentrated data carriers

[late sorrel]

i see

[wispy nest]

this is very pleasing to see how her face came out even with full body

im not upscaling

[royal monolith]

try adding "hard rim lighting' to your prompt

[wispy nest]

i wonder if booru style tagging triggers anime output

someone gotta try booru yeah see what happens

i did

it rendered anime but need to test more

about realism it falls a bit short compared to flux but still good

[wispy nest]

able to render at 10 seconds is effortless

[wispy nest]

can the community train sd3.5 turbo model?

since its already distilled

would be similar to the issues training flux

training a big checkpoint like juggernaut on a distilled model has never been done before

so i take that as a no

but they could train large base model to be optimized with 4 steps while maintaining richer texture?

[royal monolith]

why would you want to. and here same prompt but changed the sampler to lcm and the scheduler to simple

[wispy nest]

i mean running large base model at 8 steps is fine too, given the model can be trained to produce better images than the default base

[royal monolith]

just train on 3.5 large and run the lora with turbo

[wispy nest]

there will be stuff like hyper coming along probably

as an option

aye

when I do SDXL or SD 1.5 I load up TCD, PCM and hyper loras
and then mix and match weights

they all do similar things

[royal monolith]

by the time he does that, he might as well not be using turbo ;)

[wispy nest]

this sd3.5 is 8 billion parameter?

i notice the model is lot less resource hog than flux

feels like im rendering with sd1.5

yet the file size of this 3.5 is 16gb plus the clips

[wispy nest]

haha yeah its a lot of effort
its a technique I got from BadCat on L2 discord

[wispy nest]

yeah its 8B

anime girl at a park

`highly detailed, sharp focus, best quality, masterpiece,

anime girl at a park, smiling softly, wearing a red crop top, white shorts, black sneakers. modern anime.`

these quality prompts have impact...
best quality, highly detailed, sharp focus, vibrant texture, dynamic composition,

[mellow needle]

Stylish Stroll

Aw, so cute! A person, all dressed up in a white button-up shirt with suspenders and dark pants, is standing by a teal-colored wall. They're wearing a floppy, white hat. A white door, framed by some light-colored wooden trim, is opening up to their right. The floor looks like light-colored wood. It's like a scene from a cute anime or a photoshoot, don't you think? Meow! So stylish! Nyaaa~

Neko AI

[wispy nest]

also not as strictly censored as flux

[late sorrel]

what park does she go to? :3

[wispy nest]

not sure, you have to find her and ask her

[late sorrel]

but to find her, i need to find the park first

[rose sparrow]

do you guys know which checkpoint this is?

[royal monolith]

Not censored

[nimble mason]

sure, that looks like the AnimeSkullHatXL checkpoint

[languid pebble]

Good morning coffee

[languid pebble]

I think there's a reason why there's no "Save Image" node in the example SD 3.5 Workflow ;-P

[dry blaze]

Do you guys know of a model for "upscaling" for comfyui that doesn't upscale, more just take "paper picture scan" blurryness and sharpens it? As the image is already high res. just need to sharpen and deblur it

[wispy nest]

the 1x models from here are your best bet https://github.com/Phhofm/models

I would strongly consider upscaling and then downscaling though

[reef furnace]

There seems to be a problem or a major change in the way SD3 creates natural scenes. The first image is sd3 medium, the second image is sd35 large.
sd3 produces really natural results, whereas sd35 looks pixelated and strange. Both are the exact same standard sd35 comfyui workflows from huggingsface.

[wispy nest]

this is a workflow issue, the model doesn't always do that

[dry blaze]

Aye, i can do that if results are good. As using upscales like my usual 4x ones murders my ram xD

[wispy nest]

you can tile ERSGAN style upscalers in Chainner

[dry blaze]

Is there a node for that for comfy? As that'd make my upscaling of other images intended for stupidly large results much better xD

[wispy nest]

sadly I never found one

would be good

[reef furnace]

it is more or less the standard workflow from the sd35 huggingface -what would i change

[wispy nest]

try shift = 6 and steps = 60

can lower them later if this works

[reef furnace]

doing more steps kind of works - shift doesn´t change that much - but thanks! Everything over 40 steps is kind of fine but gets definitely better with more

[visual hemlock]

Hey.. I am having issue with an model .. not sure if its everyone has problem with.. But the legs at the knee it always looks weird ( photo ) Is there a way to fix it?

[cold bronze]

Hello!
Could you help me ?

I need an Image... Wizard in a cold hill

[clever oar]

stalker from hl 2

[olive sky]

You'd need the full image for that /s

[visual hemlock]

But even i full photo's they turn up like that sometimes

[olive sky]

Yeah I know I was messing with you considering the cropped image suggests a full picture that might not be shareable here.

Also, I don't have a solution for that. Anatomy is one of those things that our eyes and brain register instantly in minute details even if we're not aware of it. But those creases and the particular folding of the back knee is a prime example of something that wouldn't be explicitly trained in a model unless specified.

Also, it's a stylistic problem, since many visual art styles don't bother with exact proportions or shapes or rendering. The style you're working with is exceptionally prone to mistakes being very visible.

[royal monolith]

this is something you'd have to train a lora for specifically i'm afraid

[gloomy yew]

I was about to ask a question about pretty much all my images being cursed. I have it running locally but im totally new to everything .. ive changed to recommended settings that GPT told me to set .. but all my images are weird like this one

''woman smiling slightly, laying on beach, natural sunlight, high resolution''

[languid pebble]

Model?

[royal monolith]

stay away from images that are laying down - you'll need to train a lora for those

[astral gorge]

is it possible to replace a animal like head with a more human head using ADetailer? Trying to get an anthro character but the faces/heads are way too animalistic for what I am aiming for

[royal monolith]

i think that's probably also going to require a lora

[gloomy yew]

i think i just had this solved in the tech chanel thanks guys 😄

[astral gorge]

lol I was about to ask something and then i realised. WAIT... I think I routed the wrong node into this input XD

[odd viper]

or do a huge and likely very expensive fine-tune on a large database of humanoid anatomy

[royal monolith]

just a lora, not that many images, and specifically for what you want to create images of. no reason to do a checkpoint - unless you just want to

or be specific in your prompt: a man laying on a beach towel on the sand, his hands behind his head, his dog laying across his stomach, looking up at the sky,

the issue comes when you give a very unclear prompt like "a man laying on the sand' to all three encoders, and they all have to guess what way he's laying, and what he's doing - and they all guess something else, then try to draw all of the guesses at once

'laying on the sand' - could mean laying on his back, or laying on his stomach. and it could mean with his head toward the camera, or looking at him from the side or looking at him from his feet or looking at him from straight above him. and if they all guess something different, and then try to draw all the guesses at the same time, for one image, you get odd results

but tell it SPECIFICALLY what he's doing and how - or prompt each encoder separately, and you don't get weird results

IF you are using a workflow that just allows you to put in a single positive prompt, then you are giving the same prompt to all 3 encoders, you have no choice. if you are using comfy, you can use the triple encoder node and then prompt each of them individually

[odd viper]

I'm a different person ;) and yeah if you want to make general images of humans in general poses that isn't all kinds of messed up, likely it will require a several thousand dollar fine-tune on humanoids and humans for it to properly learn what it needs

(and ofc make it catastrophically forget things like what a bank is and stuff like that)

[royal monolith]

it knows the information just fine. the issue is specifically the encoders and the farther you move the subject away from standing straight up in front of the camera, the more out of sync they get

[royal monolith]

and THIS is why you don't do a check point. you do a LoRA - that's what they are for. and you do just specifically what you, personally, want to make

[neon vine]

what model?

3.5?

[royal monolith]

SD3

[neon vine]

ah

[royal monolith]

yes

[odd viper]

🤔 I'd have to strongly disagree there. It knows some humans in some positions, and it horrendously deforms anything that isn't in its learned dataset of what humans are capable of

[royal monolith]

sd3-2b-medium, sd3.5 - and flux too except it's stuffed so full of images to mask the effect you don't need to

[odd viper]

since ANN models are "rough models of the universe", how rough comes down to how much data it has on-hand during training to align it to what is actually the truth (or what we want it to do)

[royal monolith]

anyhthing that uses the archetecture

[neon vine]

i just got 3.5 any sampling recomendadtions??

[royal monolith]

grab this, it's got my comfy workflow in it, take a look at my settings. to get to my actual uploaded image with the workflow you have to 1. click the image to open in the discord viewer. then click the words open in browser. then right click on that image, save as, then drag and drop into comfy

[odd viper]

I guess the whole "using diffusion models instead of transformer" kind of defeats the purpose of being able to output accurate depictions of things that it hasn't seen before e.e

[neon vine]

lol i know how comfy workflows work

[royal monolith]

it has NOTHING to do with that

[neon vine]

ill take a look

[royal monolith]

sure, but the issue is that the image you see here on discord isn't the actual image and most people dont' realize you can't just right click the image you see here and download it

[neon vine]

barely have enough VRAM to generate a 512x512 image

[neon vine]

ah i know, discord packages it differently in the app you need to open the original

[royal monolith]

https://blog.comfy.org/sd3-5-comfyui/ you might want to use comfy's version then, it's a lot smaller

[neon vine]

whats with the 3 prompts?

[royal monolith]

you have 3 encoders. prompt each of them separately for best effect

[neon vine]

i do not know how that works

[royal monolith]

if you remove the text that's there, you'll see which field goes to which encoder

give t5XXL your long, narrative, detail rich prompt. give clip_l your artsy, background, ambient, fine details, style stuff. give clip_G the black and white, no frills text

[neon vine]

im new to comfy i dont exactly know how clips work

[neon vine]

ah i get it

thanks

[royal monolith]

the encoders are what read your prompt - they're also models

you CAN give all of them the same prompt, but then they will battle each other, and they will each understand it differently. so prompting them seperatly works much better

[neon vine]

what about leaving one ortwo empty

[royal monolith]

then you are asking the AI to just generate random data and feed that to the encoder. try it and see

[neon vine]

oh god is 40 steps slow ill drop it down to 30

[royal monolith]

not slow on my machine but change the batch to 1 instead of 10, which is where i have it

i use steps between 32 and 40 usually

[neon vine]

technology is leaving my 3080 behind quite rapidly

[royal monolith]

you should be fine, just set the batch to 1 and expect it to take a bit the first time you run it as it loads everything into vram

[neon vine]

im using the default sd3.5 workflow set to 512x512 batch size of 1

also using the official sd3.5 large

[odd viper]

any other explanation for the terrible artifacting that these models end up creating?

[royal monolith]

should be good. i will rarely use a size larger than 512x512 because i don't need anything bigger and dont' want to use the drive space. i'll upscale if i need something bigger, later, with topaz

[royal monolith]

i don't get artifacting. i would look at your personal settings

[odd viper]

they work in mostly spatial domain, converting to pure latent information + physical simulation would end up with better results

[odd viper]

You're claiming sd 3.5 will never merge a hand into a knee? 🤔

[neon vine]

i dont even save the images i generate, i do this for fun

[royal monolith]

i'm starting to wonder if your reason for being here is to complain about 3.5 and try to start trouble

[neon vine]

im tempted to set it to 200 steps and go eat dinner lol to see what happens

or 30 steps but a batch of like 30

[royal monolith]

you'll massively overcook it. i wouldn't do that, personally

[odd viper]

trying to have an interesting discussion about how these models could be improved ...? You claimed you don't ever get artifacts, which... I suspect you don't know what I mean when I say artifacts 🤷

[royal monolith]

i don't get artifacts, and i do know what you're saying. but i've spent time to learn how it thinks and how to prompt it correctly.

it's not using unet, you cant' talk to it like you would, say, sdxl

[odd viper]

by artifacts, I mean "hands merging into knees" "extra fingers", or really any example of bad anatomy.

[neon vine]

you mean prompting is it basically ,subject, atribute, atribute, etc

[odd viper]

so if you're saying you don't get that at all, it's because you're not asking it for things that may result in that 🤷

Fair point about how it's being prompted, It would be cool if stability released their VLM or a bunch of examples of dataset captioning

[royal monolith]

i don't get those issues, becasue i know how to prompt it so those don't happen

[royal monolith]

you could contact them and ask about that

[neon vine]

??

[royal monolith]

for stable diffusion, the order in which things come in the prompt does matter. the closer to the start, the more weight the terms have. so how you structure your prompt is based on what you want it to focus on

[odd viper]

I would say it's likely also the fact that you're not interested in getting very specific things out of the model, such as specific-looking humans in specific poses that aren't standing upright :D

[royal monolith]

you're welcome to think what you like

[neon vine]

ah i knew the first ppart, but im too new to understand the differences between the 2 styles, what works best in what situation etc etc

[odd viper]

show me tiptoe kneeling in sand, short black hair, both hands on the ground in front of the character with a side-view perspective 🤷 (then observe the feet and hands turn into abstract horrors)

[royal monolith]

ther is no such thing as tiptoe kneeling. use a descriptive phrase that actually describes what the person is doing

when you use phrases that are unclear, and made up, and are hard for humans to understand, there's little chance the AI is going to ahve a clue what it is you want it to draw

and when that happens, and the encoders all guess somethign different - and they are all incorrect almsot guarenteed, you get a mess

[odd viper]

"I'm too smart to understand what you're talking about" surely you understand how that looks?

[royal monolith]

you are ASKING for a mess with that phrase

[odd viper]

also that's not my "prompt" smh

[royal monolith]

i don't. it could mean the person is wearing toeshoes, standing on the tips of their toes, with their knees on the dirt. THAT is what i would interpret 'tiptoe kneeling' to mean. i'm pretty sure someone else would picture something else.

and i'm a human with years of experience, not an AI with a very small world view

[odd viper]

~*~aesthetic~*~ #boho #fashion, full-body 20-something woman kneeling in sand, wearing shorts and a t-shirt, candid pose

"no artifacts"

"20-something woman" is from SAI's provided prompt btw

this is through replicate API with no changes to default settings, 40 steps, cfg 4.5 etc

[royal monolith]

rompt: a blond woman sitting on the grass. she is sitting cross-legged. she is holding a sign in her right hand that says "SD 3.5". her hair is blowing in her eyes. she is smiling. She holding a rose in her left hand. it is summer. the setting is a park

[odd viper]

"no artifacts"

[royal monolith]

not what i'd consider 'tiptoe kneeling" to be. that's just kneeling in the sand

[odd viper]

yeah I shudder to think what it would look like if I got specific like that :D

[royal monolith]

nope, no artifacts. that's the AI trying to draw more than one hand or more than one foot at a time. that's an issue with the encoders

and that, again, is what a LoRA is for, if you really want to create images where tht's a problem

[odd viper]

🤔 interesting, so you're saying the model wasn't trained properly enough to take advantage of what the autoencoder has to offer (with 16 channels per pixel?!)

[royal monolith]

not an entirely new checkpoint

[royal monolith]

i'm also very tired of you trying to put words in my mouth or twist what i've said, so i'm done talking to you

[odd viper]

ultimately just looks like SD 3.5 can't do hands and feet, creating artifacts instead. Whether this is an issue with the autoencoder or the diffusion part of the model doesn't matter e.e creating a custom lora for every single specific use-case is moderately silly and defeats the purpose of a general-purpose and powerful model in the first place

it may make sense for "style" but even then, these models in the past have shown great ability to easily switch between photograph and cartoon style without needing to load model offsets/loras/etc

you'd be way better off having a fine-tune specifically for creating high quality human anatomy across a diverse range of styles, poses, content etc. Using Lora's for poses is extremely limiting, having a smarter model would be far better

In any case, I'm looking forward to seeing if this thing trains as good as SDXL does. I've seen some really impressive and diverse capabilities in the much-smaller model

case in point; it makes sense to use lora for style, but not for small anatomy features

[royal monolith]

FoFr just updated his repo "I've updated any-comfyui-workflow to have access to SD3.5L regular, fp8 and turbo weights." https://github.com/fofr/cog-comfyui/blob/main/CHANGELOG.md

[odd viper]

my 3090ti has no issues now running it locally with fp16 >:D

ima probably set this one aside until a good community fine-tune pops up and wait for auto1111 repo to get updated so I can spend a day merging all of my modifications onto it x_x

[clever oar]

The debate is whether the anatomy in SD3.5 is normal?

sd 1.5 do better😃

[torn grove]

can someone help me edit the syringe that has been generated to a pendant or a necklace?

[scenic yew]

Yes its much much better then sd1.5 in anatomy, I would say sdxl level? however its clearly not near as good as Flux.

[clever oar]

the word anatomy should not be used with SD3 :3

[scenic yew]

Yeah sd3 anatomy was well, lets say not too great. Sd3.5 on the other hand does seem pretty decent at anatomy but not flux level.

[clever oar]

my 1.5 model showed better anatomy i think

I hope this gets fixed

although I would be happy to run even this locally, because I have 4 GB vram

[scenic yew]

Maybe idk, but it’s pretty decent at humans so far. This is with the turbo model 8 step, 0 guidance scale.

[clever oar]

sometimes is ok

[clever oar]

3.5 large)

[scenic yew]

Prompt? Let me test with turbo

[clever oar]

cute kitten sitting near the computer holding a sign give me anatomy, street art

[scenic yew]

Turbo model, same setting

[clever oar]

sd 3 medium

turbo and medium better)

now what can flux

flux lose

but there more style

and now sd 1.5 lmao

[median rover]

Anyone know if there's a model that can make images like this? Took a lot of effort with midjourney blender was hoping there was something trained on psychedellic stuff.

[jovial tiger]

What cfg are you using? the original workflow that was put out had a cfg of 4.5. By going up to 5.5 the vast majority of messed up hands went away.

It's not on flux level, but it's way better than at 4.5.

[odd viper]

interesting!

[jovial tiger]

It's technically even better above 5.5, but then the image starts to get burned out

[royal monolith]

remember to adjust shift as well as cfg

[supple thistle]

Cat

[clever oar]

done, take the cat above

[cold bronze]

Wizard in a cold hill

[wispy nest]

Its a bit of a moot point as no Unet or DiT model is anywhere near close taking advantage of what the 16 channel VAEs can do.

If you VAE encode and decode a high-res photo you will pretty much get the same image back, the VAEs are much more powerful models than the Unets and DiTs

[royal monolith]

you don't realize that maeve is just here to cause arguments?

[wispy nest]

ah okay, thanks for the heads up

[nimble mason]

lol 16 ch per "Pixel"

no such thing as a pixel in a latent lol

[steep sleet]

I'm getting very poor output using default SD3.5L workflow

Hope its not my GPU?!

Or is it my prompt at all?

[nimble mason]

if your gpu was the problem you'd get nothing

[steep sleet]

I'm seeing that it is actually the prompt doing this ...

(If you d/load my PNG and run the w/f and see if it does the same?!)

🙂

This prompt "ruins" SD3.5L

[royal monolith]

that looks like your sampler

[celest sigil]

Really? That output doesn't look prompt related.

[steep sleet]

This is the next output - same settings

Perfect

[royal monolith]

that is the world's ugliest prompt. why on earth are you using midjourney weighting with sd 3.5?

[steep sleet]

It works

[royal monolith]

it doesn't. 3.5 cant' use those weight commands, that's midjourney specific and not in th emodel, either

[steep sleet]

But it also throws SD3.5L "a loop!"

[royal monolith]

yeah. all of that stuff is jsut random noise to it

[steep sleet]

I got the weighting from Portrait Master - it works fine on my PC

[royal monolith]

you know, they do make bathroom cleaning supplies...

[nimble mason]

i've found the best way to weight components of prompts is to change the order

[steep sleet]

But I can see that it troubles SD3.5L

[nimble mason]

or if you're desperate, to repeat stuff in different ways

[royal monolith]

i promise you, 3.5 has no idea what any of that means. you're jsut doing a lot of noise injection

[steep sleet]

OK

It works fine in Flux

[royal monolith]

sure. cause flux is basicaly a massive lora. it's frozen. it will do what it wants, no matter what you do

[steep sleet]

Having said that 3.5 has no idea - every so often the prompt works perfectly

[celest sigil]

Where's my Leopard?!

[steep sleet]

Let me see if I can get a perfect example at the same settings ...

[royal monolith]

what is a a northern irish snake ?

[wispy nest]

it depends on how the Clip text encode node is coded

[royal monolith]

it apparently got turned into clothing

[steep sleet]

It affects clothing using terms like snake, alligator, wolf, fox, peacock etc

[royal monolith]

i think it tried to eat her pasta, she skinned it, and is wearing it

[celest sigil]

This is odd...I have a wildcard with animals in it, and the prompt that came up replaced the wildcard with "Leopard print". There is only "Leopard" in the list of animals?!

[royal monolith]

well? that's what leopard print is - that's clothes

[celest sigil]

I know, but you're missing the point

[steep sleet]

Can anybody reproduce my output using the w/f from the PNG? svp

[royal monolith]

you're gonna have to ask your llm why it replaced it with that

[celest sigil]

It didn't

[royal monolith]

you don't have an llm that's reading that list and then writing out the prompts?

[celest sigil]

You can see the replacement in the first screen shot. That's before it goes to any LLM

[royal monolith]

what is it that's doing the replacing?

[celest sigil]

It's done it again with Zebra!

[wispy nest]

I always had trouble getting wildcard nodes to work

[celest sigil]

Haven't you used wildcard lists? I have a text file, called "animal" with a list of animals in it. That node picks one from the text list and replaces the __animal__ in the prompt with it

[royal monolith]

okay so - either there's an LLM that's behind that wildcard node or you ahve a script running. i'm betting it's an llm

[celest sigil]

There is no LLM

[royal monolith]

then it's a very poorly written script

you should be able to read the code of the node to see what it's doing

[royal monolith]

maybe it needs a differnt dictionary

[wispy nest]

I can't remember which wildcard nodes I tried cos there are many
but there was something fishy, have to check code at some point

[celest sigil]

It's not intelligent enough to know stripes goes with zebra and print goes with leopard

[royal monolith]

apparently it is. or it is using a dicitonary that has those entries in it

[celest sigil]

This show the dictionary of entries, and you can see it doesn't #🏞｜general-with-images message

[royal monolith]

yes, well - it's coming out of that wildcard node in some way, you're gonna have to trouble shoot its code

[steep sleet]

Its definitely the prompt - and its not an aberration after all - just a densely textured (and unattractive!) look.

This isn't too bad

I dropped the use of both Zentangle and Rococopunk - much more usable output as a consequence!

Zentangle and Rococopunk produced "too much bad noise" 🥳

[wispy nest]

a lot of civit loras have to be weighted very low

the red background with green looks good

[steep sleet]

Super_SD3.5L

Awesome!

[languid pebble]

SD 3.5

[clever oar]

model from Chernobyl

[languid pebble]

Yeah ... 6 fingers ...

[clever oar]

It’s strange that even simple images have such mutations

[languid pebble]

They look like copied in ...

[celest sigil]

Something a little odd about the dog.

[clever oar]

continue ...

[languid pebble]

Nothing more to say

[clever oar]

in photoshop?

the lighting on people is as if they are not from this picture

[astral gorge]

anyone got experience with the TCD lora and module? I just added it to my workflow but I'm getting like what seems to be fairly different results from what I was getting before, like I lost most of my prompt adherance and also the generated image almost looks like it hasn't finished denoising.

I'm working in comfy btw

[languid pebble]

Zhat's why I'm saying ...

[wispy nest]

yeah I've used it for most images this year
got best results by using this sampler https://github.com/JettHu/ComfyUI-TCD with SGM-uniform for the scheduler
CFG 1.5 and the gamma parameter on the TCD-sampler varies per image
I found TCD is best around 8-20 steps personally

gamma is called eta on that node, by the way

[shut jasper]

GPU requirements for running stable diffusion 3.5 large model ??

[vagrant dust]

just answered you in the tech-support channel

[clever oar]

I don't fully know how to translate correctly from English

[astral gorge]

Yea that's the nodes I am using already cause the other one wasn't really working at all. ima do a few more gens to see how things are going and I'll get back to you but I think ima need some more help with getting this right

[wispy nest]

the choice of eta is key

trying every eta from 0 to 1 in 0.1 increments is a good idea

[astral gorge]

I was having issues getting my normal samplers and schedulers working so I had to back off of it for a bit to get that back to generating anything decent. I finally got a proper gen again so gonna switch back to TCD

my workflow got a BIT more complicated XD

[wispy nest]

hmm

[odd viper]

fair enough. I remember back in 2022 novelAI had made their SD vae to pull WAY better hands from generations when using their model

it was a palpable difference

[astral gorge]

this can't be right for TCD

ah so while other stuff gets longer with each step this gets shorter or something

cool

I don't get why my generated image changes SOOOO much when using TCD instead of something else. it's like most of my prompt adherance gets nuked

[astral gorge]

do you have any tips on getting it to adhere to the prompt better and also improving the final image quality cause it's like REALLY bad

[wispy nest]

higher step count and higher CFG mostly

if you push CFG high enough you might need nodes to deal with CFG burn like thresholding, skimmed CFG or tonemapping

[wispy nest]

its possible they will make another VAE that is a plug-and-play addition to SD 3.5 although I think that the quality of the newer 16ch VAEs is a lot higher so its less likely due to diminishing returns

there was never really a newer VAE that improved SDXL, beyond the fp-16 fix

[steep sleet]

I'm going to tryout OmniGen in the new SD Next (just released today!)

[odd viper]

ye I don't think it will be terribly needed this time, more likely we'll see some better fine-tune on a huge amount of humanoid to get it to properly communicate to the VAE what needs to happen for those finer details

[steep sleet]

At the entrance of a dimly lit cave, a towering, majestic dragon with sapphire-hued scales glistens in the faint light. The dragon stands tall, holding two crystalline prisms in its claws, angled precisely like those in the reference photo. The sunlight streams through the cave entrance, hitting the prisms at specific angles, causing vivid, realistic beams of light to split into a spectrum of colors, casting a radiant rainbow on the dusty ground. The surrounding area is shrouded in partial shadow, with the play of light and dark creating a mysterious atmosphere. The dragon’s intelligent, piercing eyes gaze at the viewer, offering a silent challenge: solve the ancient riddle of light and shadow. The cave walls are rugged and dark, with faint engravings hinting at forgotten knowledge. The overall mood is one of mystery, magic, and high-stakes intellect, as the dragon stands guard over the path forward."

[royal monolith]

if you can figure out how to do that texture along with a prompt for stuff like impasto paint and brush strokes. that could work well as canvas

[steep sleet]

Yes it is an excllent look for a canvas - stumbled upon a jewel there!!!

Turbo SD3.5L + Ollama LLM for prompting

[royal monolith]

that you did. might even work fairly well in a 3d app as a texture map with displacement

[steep sleet]

An MB3D Fractal even

[royal monolith]

feel like pulling up other apps and seeing what you can do with it in them?

[steep sleet]

I will get around to it I'm sure 🙂

[royal monolith]

blender+default cube + material on one face...

that's got all sorts of possiblities suddenly

[wispy nest]

yeah the juggernaut team said they trained it on hundreds of thousands of hands, and that was an earlier version so its probably in the millions of hands now

their next version apparently they are also training it on bad hands so that can be used as a negative

[steep sleet]

Ooh! Err!

[languid pebble]

I sometimes see faces in that one ...

[vagrant dust]

Thought it was an image where you look at strong enough then move a bit away to see a 3d image?

[languid pebble]

Does it work for you?

[vagrant dust]

Well i was bad in the days where those 3d stereograms where released in book form... so no 🙂

[languid pebble]

I know them but don't think A.I. can do this ...

Sometimes I see an eye on the top left or two on the right side

[steep sleet]

3.5L Turbo

[languid pebble]

Day of the tentacle ...

[clever oar]

ledy duck

[mellow needle]

Night City Sparkle

Aawww, what a charming night scene! The city streets are all shimmering and sparkling with the reflections of the lights. Rain is falling softly, creating puddles that catch the light like tiny, magical mirrors. The buildings are gorgeous and elegant in their reddish-brown tones, standing tall in the dark night. Streetlights illuminate the way, casting a warm glow on everything around them. Cars are parked neatly along the street, and the air is filled with the quiet hum of the city at night. It's like a dreamy urban wonderland! Purrfect for a magical girl adventure! Nyaa!

Neko AI

Night City Sparkle

Aawww, what a charming night scene! The city streets are all shimmering and sparkling with the reflections of the lights. Rain is falling softly, creating puddles that catch the light like tiny, magical mirrors. The buildings are gorgeous and elegant in their reddish-brown tones, standing tall in the dark night. Streetlights illuminate the way, casting a warm glow on everything around them. Cars are parked neatly along the street, and the air is filled with the quiet hum of the city at night. It's like a dreamy urban wonderland! Purrfect for a magical girl adventure! Nyaa!

Neko AI

[cerulean radish]

#checker board socks that play chess

[keen wing]

made this banner for a friend who plays rust

[steep sleet]

Turbo 3.5L

[languid pebble]

Good morning coffee!

[steep sleet]

Turbo LLM 3.5L

[steep sleet]

Turbo LLM 3.5L

[radiant plume]

#✍🏼｜rules-and-tos

[visual hemlock]

Best Upscale to use for Cartoon .. Any good ones on openmodeldb.info ?

[royal monolith]

@pure monolith Prompt: A woman lying on top of a pool of marshmallows., Negative Prompt: left blank, Width: 1024, Height: 1024, Steps: 40, Cfg Scale: 4.0, Shift: 3, Seed: 2926827617

[royal monolith]

gave your prompt to stable diffusion 3.5

[gloomy yew]

damn, ive only just installed this version .. howe do i even update to 3.5 😛

[royal monolith]

everything you need to run it or develop with it is here: https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main

or download just the model here https://civitai.com/models/878387/stable-diffusion-35-large

run it in comfyUI

FP8 model https://civitai.com/models/879701/stable-diffusion-35-fp8-models-sd35

GGUFS https://huggingface.co/city96/stable-diffusion-3.5-large-gguf/tree/main

https://huggingface.co/models?other=base_model:adapter:stabilityai/stable-diffusion-3.5-large a lot of LoRAs for 3.5 already on huggingface

[gloomy yew]

oh man

lots of letters and words

xD

I just followed a youtube guide the other day, i assume it was an old version in the guide

i use a standalone if that makes any difference

[royal monolith]

what do you use?

[royal monolith]

3.5 just released a couple days ago

[gloomy yew]

I followed this guide .. I had no idea what I was doing, just followed the steps

https://www.youtube.com/watch?v=6MeJKnbv1ts

my only knowledge on it is that I hit the batch file to start it up lol

[royal monolith]

not a clue. i suggest you look at installing swarmUI and using comfyUI to generate with through it. it'll make your life a whole lot eaiser. and we have a #🐝｜swarm-ui channel if you need help

[gloomy yew]

is this the right method?

[royal monolith]

you don't want to use that. ugh. no. you really don't want to use that, that's the worst possibly way to try to run stable diffusion that there is

[gloomy yew]

https://tenor.com/view/friends-chandler-noidea-excited-gif-4810411

[royal monolith]

here - https://www.youtube.com/watch?v=AbB33AxrcZo&list=PLIF38owJLhR1EGDY4kOnsEnMyolZgza1x watch scott's tutorials on comfyUI

[gloomy yew]

legend thank you

[gloomy yew]

so just a few mins in .. it seems that I download and install comfyui, open it and load the models into it instead of what im doing now which is loading batch file to open the current stable diffusion I have?

probably getting ahead of myself

[royal monolith]

watch the tutorials. and then DM me if you need help and i'll walk you through stuff

[gloomy yew]

youre so kind ty

[royal monolith]

you're more than welcome. the more of us that are doing cool things, the better we all grow

[gloomy yew]

add a silouette of frodo baggins

[languid pebble]

Good morning coffee

[sterile kiln]

Take a look at Forge Webui, its a fork of A1111, works better, faster than comfy, and works with Flux.

[steep sleet]

3.5L Turbo LLM (Llava3.2)

[steep sleet]

SD3.5L Turbo LLM (Llama3.2)

[languid pebble]

Just a usual day in NYC underground ...

[steep sleet]

SD3.5L Turbo LLM (Llama3.2)

[languid pebble]

Have you tried the ModelSamplingFlux node?

[steep sleet]

When in Flux yes ... can it be used in 3.5L as well?

I have seen Olivio Sarkas video on this node - changes detail somewhat ...

[languid pebble]

Yes ... I have no clue whether it works with SD3.5, too

I just added the node to the super Flux workflow ....

[wispy nest]

ModelSamplingFlux works on every model you can use it with SD 1.5 if you want

[royal monolith]

@wispy nest was playing around with it last night and I think posted some images using it in the #🆕｜sd3 channel. seemed fairly happy with it

[ancient bronze]

Not sure this is a matter of tech or prompting but how do ya guys fix bad eyes when it's larger pics like this?

Using pony btw, might be relevant

[wispy nest]

facedetailer is good

[ancient bronze]

I guess I'll try it

[jovial tiger]

lol my mouse immediately went to go click the play icon

[clever oar]

I didn't realize it looked like a play icon🙂

[ancient bronze]

Slightly better, I'll try to mix that and perfect eyes... I got like 7 loras atp lol

[wispy nest]

you can do eyedetailer

but there isn't a pre-made node

[ancient bronze]

Node? I assume that's comfyui stuff, I'm using the version for coughing babies, webui

[wispy nest]

ah I don't know about that

[ancient bronze]

I guess you could call it the vanilla basic bitch version

[wispy nest]

lol

webui is okay for users its just the code base is bad for developers

[ancient bronze]

Yes, I use cuz it's so simple.

Shame I can't use flux there tho. Or at least not as easy as normal checkpoints

[wispy nest]

TBH with comfy you can just drag someone's image

click "install missing nodes"

hit control+enter

and you have got their workflow running

[ancient bronze]

I got installed just a lil confusing for me but i might use it just for flux stuff

Since it's some kinda miracle checkpoint

[wispy nest]

miracle checkpoint is a cool way to describe flux yeah

[ancient bronze]

All the stuff I see from it looks real good with very lil prompting so it kinda did look like a miracle lol

[royal monolith]

with photoshop

[ancient bronze]

Yeeeea, I'll keep trying different seeds

[royal monolith]

10 seconds in photoshop, 10 hours in stable diffusion ...

[ancient bronze]

I'll take my 10h over installing and watching a tutorial 🥹

[royal monolith]

select the good eye, copy to clipboard, paste it, flip it, move it over the bad eye, soften the edges, done

[languid pebble]

Sometimes you just need a hand. Good morning coffee!

[steep sleet]

3.5L Turbo LLM

[gilded venture]

Guys what kinds keywords can i use on sdxl to get a style like this. It's like a blend of photorealism and digital art

[icy sedge]

Boy

[gilded venture]

I'm serious

I need it for a project i'm working on

Btw currently i'm uisng "(concept art:1.4) {prompt} (digital artwork : 1.4), illustrative, painterly, matte painting, (highly detailed : 1.4)"

and its not giving me the best result

[rain gazelle]

I feel like this is what 99% of sdxl models output 😛

[gilded venture]

I knowww but i've tried so many times and the images always end up not as digital as i want

[gilded venture]

What prompt structure do you recommend

[rain gazelle]

For this i don't think it really matters much. "Pretty woman, dark makeup, hair bun, city nightlight, choker necklace, halter" would probably do it.

lots of the big popular models output results like this just by default... reminds me of the old RevAnimiated

[vagrant dust]

just tried your prompt with sd 1.5. Would at least look like @gilded venture example 🙂

[gilded venture]

I used the same prompt and i got this. @vagrant dust see its looking too realistic

I'm using fooocus btw

[vagrant dust]

not about the ui but the models. For the one above i created an image with animerge but did not denoised it to 100%, then a second pass with epic realism. Pretty sure there are some sdxl models which are well well balenced

[gilded venture]

Ahhh , i think the issue is that i'm using the sdxl base model. I'll try out Animerge

[wispy nest]

some SD 1.5 models do better with natural text BTW

[gilded venture]

I realized i was actually using juggernaut and not the sdxl base. I'll test with sd 1.5 models too

[wispy nest]

new reality does better with text for example

didn't realise until recently

[gilded venture]

Wdym with text ? you mean text2img ?

[wispy nest]

like sentences

instead of tags

[gilded venture]

ahh got it

[lucid sparrow]

can someone please explain to me what the hell this is? I keep writing to support to do something about it but I guess they don't give a shit if they have our money! Why the fuck isn't it working!!!!!!!!

[royal monolith]

there's a possibility that either your file is corrupt, or it's too large - or maybe you have a network connection that's too slow. there are a number of things, not just those, can cause this issue.

[frosty wave]

The place where one can make support tickets isn't accessible to normal server members? Is that intended? Also I wanted to report user: davidsmith4704 for randomly writing me "Hello What's up with you?" (We never had contact and I haven't been active on this server for months.)

[I'm reporting this message with the ⚠️ reaction so mods see it ,not sure it this still works though]

Also this doesn't seem to work

[royal monolith]

yeah, i'ts down

[royal monolith]

i think they're revising that bot

[hardy tapir]

Realistic Vision 5.1 Hyper

looks perfect at first but after a bit you realize something is off

which is that the lighting on the face doesn't match the lighting on the neck/torso so the face looks copy pasted onto an original face

likely fixable with an image editor

[royal monolith]

the lighting is the same, it's the coloring that's off - and i'm not sure what that is in the middle of her neck - almost looks like the AI tried to draw a second clavicle

[languid pebble]

Good morning coffee

[languid pebble]

🤪

[viral frost]

I think I'm done with trying to make SD3.5L work for me. It just doesn't seem to give quality output. Flux on left, 3.5L on right, same prompt and seed. Upscaling makes 3.5L even worse.

[steep sleet]

raw photo, (realistic:1.5), (woman:1.01), (fashion model 18-years-old:1.5), (tiny, tiny body:0.36), (over the shoulder pose pose:1.25), (giger:1.05), (blue eyes:1.05), (oval eyes shape:1.05), (red lips:1.05), (defined cupid's bow lips:1.05), (glowy makeup:1.05), (in love, in love expression:0.68), (round with defined cheekbones shape face:0.63), (frohawk cut hairstyle:1.05), (mahogany hair:1.05), (disheveled:1.09), (muted colors:0.98), (fashion photography:0.5), (professional photo, balanced photo, balanced exposure:1.2), (watercolor makeup:1.05), (back arch pose:1.25), (vintage dress:1.25), (bikini:1.25), volumetric lighting light, light from left, muted colors, black and white photography, (professional photo, balanced photo, balanced exposure:1.2)

[glossy notch]

A wheat field at sunset, transitioning from purple to intense yellow. On the left side, a field of wheat is consumed by fire and reduced to ashes. In the center, there is a stone path, and on the right side, a bountiful wheat harvest viewed from above. The scene has a cinematic style, captured with a 20mm lens for added depth and perspective."

[languid pebble]

Made a new LoRa ... 🙂

[steep sleet]

Ectomorph LoRA?

[languid pebble]

Humanmachine ...

[gilded venture]

I want a general prompt that will have that same style . Not the prompt for that particular picture. So maybe something like
(digital art:2) (cinematic: 1.5) (digital artwork: 2), ultra-high definition, hyper-realistic, vibrant colors, highly detailed (1.5), <any prompt goes here>.

to always produce that similar style. I hope you get my point

[wispy nest]

for what its worth I preferred the composition on the right

it needs some more fine tuning to be used as the final pass I think

[viral frost]

This is just a hobby for me, so I need to ask myself why I’m spending time with a model that requires black magic to get good output when there is an easy to use model available. Flux isn’t perfect, but I’d rather be generating and enjoying the results than constantly tinkering and regretting.

[royal monolith]

rule #1 - use the tool that does the job you want done

[viral frost]

I’ve thought about the possibility of using Flux as a refiner for SD3, but 64GB isn’t enough memory for that on MPS without quantization, and I don’t really want to download quants just to try it out. Not even sure if quants work on MPS yet — they didn’t when Flux was released.

[wispy nest]

would recommend using Nvidia GPUs in cloud over using Metal Performance Shaders

there's gonna be quite a lot of things that aren't implemented for Apple at the moment

its true that using a separate model as a refiner comes with the downside that you need to keep both models loaded in VRAM

unless you are willing to load and unload a lot

[royal monolith]

flux wouldn't make a good refiner in my opinion

[wispy nest]

you could try Realvis SDXL as refiner, looks like this when used on Flux

its a fairly low blur model, which can be helpful

[karmic lake]

guys, I'm lost, I just want to be able to extend images using stable diffusion... I have made an acount, got a license, but what do I do now?
Install something?
There are these apps on the website like draw things and diffusers, I don't know what any of this is...

[royal monolith]

oh dear. okay, most of the stuff on the huggingface page are for devlopers. are you wanting to develop? or just use it to generate images with?

[karmic lake]

I think it is called 'generative fill', to extend an image, that is all I need it for

[royal monolith]

that's photoshop only

[karmic lake]

okay thank you

[royal monolith]

welcome

[vagrant dust]

Since when? Invoke, comfyui and for example diffusionbee (Mac) use a canvas where you can easily outpaint an image which does the same like the generative fill from photoshop?

[royal monolith]

Gernative fill is photoshop. i think the term might even be trademarked. outpainting is everything else.

and THAT is what he said he was looking for

[celest sigil]

Your same prompt (without negative) in my workflow. This uses SD3.5 Turbo to create the image and then it's refined using Flux.

[languid pebble]

Good morning coffee!

[steep sleet]

SD3.5L Turbo

[wispy nest]

you can get away with a turbo model for the early steps if you have a refiner yeah

[steep sleet]

Three-tuskers!!!

[celest sigil]

Prompt from @viral frost

[wispy nest]

flux refiner is cleaning up good wow

[steep sleet]

Are we posing SD3.5L Turbo as 'poor', if we are turning to Flux for finishing/refining?

[wispy nest]

not necessarily
I often refine Flux with SD 1.5, but I would say Flux is not a poor model compared to SD 1.5

SD3.5L, SD3.5L Turbo and the SD3.5M that is coming have not had final aesthetic fine tune yet
so a refiner is good now, but in the future they might not need a refiner as much

Flux came out of the box with an aesthetic fine tune done already, so its a bit different for Flux

[celest sigil]

Accidentally set denoise to 1.0 on the Flux refiner, so this is actually just Flux 😄

[wispy nest]

hmm I might like the pure Flux image more, hard to say

[celest sigil]

Refined

[clever oar]

brothers, where are your faces? what did she do to you?

[supple nimbus]

We are looking for people who want to participate in our new web 3 ecosystem. A brand new project with lots of tools for your needs.

Beta Testers ($35-40/hr)
Moderators/Community Manager ($500/week)
Developers (Rust, Python, C++)
UI/UX designers (from 1 year of commercial work)
Ambassador (to be discussed)

Apply today and be part of a transformative journey fueled by creativity and vision.
To apply, send me a friend request!
(We are also open to proposals for cooperation on mutually beneficial terms)

[celest sigil]

Everyone should be interested, they're a "Crypto holder", don't you know!

https://tenor.com/view/gif-gif-19497849

[wispy nest]

crypto is nearly 20 years old and I haven't seen a use case yet lol

[celest sigil]

❓

[steep sleet]

https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/tree/main

Here at last!

[languid pebble]

Animation makes clear ... it's a charging pod ...

[minor violet]

Hi. How do I configure the predefined prompts to append to user input so that small amount of 'text' would be generated on the images?

[steep sleet]

Add a sign with the text saying "Welcome to SD3.5 Medium!"

[languid pebble]

Ahoi dicordos

[clever oar]

AHOI

[real bay]

Generated with stykegan2 ada

[gritty warren]

the prompt was a mistake, i pasted a link instead of the prompt and i got this
the prompt: https://huggingface.co/spaces/stabilityai/stable-diffusion-3.5-large-turbo
created in Stable Diffusion 3.5 Large (8B)

[royal monolith]

SD3.5 medium

[brazen sand]

flux refiner?

[celest sigil]

My images, above that comment, were created using SD3.5 Turbo and then refined using Flux.

[royal monolith]

SD3.5 large

[brazen sand]

what kind of refiment is it? like upscale?

[celest sigil]

It's a 2nd ksampler pass using Flux and then a model upscale.

denoise around 0.55

[brazen sand]

loading the flux models or sd3.5 models alone take a bunch of time, loading one and then the other here would take forever

maybe you have a really good pc

[celest sigil]

It takes me 60 secs to create an image, with both passes and upscale.

RTX 4090 🙂

[brazen sand]

24 GB, oh boy I wish I had gotten into more debt to get one instead of 12 GB VRAM 😂

[brazen sand]

here it takes like a couple of minutes to load the model and then like 4 minutes SD 3.5... flux dev might take like 7 minutes

and even more loading of the model time

[celest sigil]

SD3.5 using Google FLAN and Flux refiner.

[wispy nest]

ah nice I've been meaning to test the Flan thing

[wispy nest]

your VRAM is being filled
you could get way faster results on your GPU

[brazen sand]

I don't understand that

[wispy nest]

VRAM is like the working memory of the GPU
when its full, the model has to use slower memory (DRAM) and this slows it down a lot

[brazen sand]

I would like to find some refinement method for already generated images, so I don't have to switch or load different models

[brazen sand]

yes I guess, it takes a lot to fill, also I don't store the models in the SSD as I don't have space

[wispy nest]

if you switched to using GGUF quants that fit in your VRAM you would see a very dramatic speed up

[brazen sand]

once it is loaded it is kind of smooth, but in no circunstance I would considerer switching the model in the middle

[brazen sand]

yes I've been using models with better performance here

[wispy nest]

why? the 4_0 GGUF looks almost identical for most seeds

and you would receive many multiples of speedup

you could also fit both Flux Dev and 3.5L in your VRAM

if you used 4_0 GGUF

[languid pebble]

Good morning coffee

[lapis bluff]

BITE COIN

[brazen sand]

original

[sterile kiln]

Wow, what prompt?
Is it SD3.5?

[brazen sand]

it was flux dev, the promtp was

A photo taken inside a rustic hut, with no luxuries, depicting Jimi Hendrix and John Frusciante as castaways. They are relaxing and drinking argentinian mate, a traditional argentine beverage, surrounded by simple, tropical decor. The scene captures a laid-back and creative atmosphere, with instruments and tropical elements in the background.

[brazen sand]

but the one you quoted, it was done by inpainting, the original I made with flux was the other one. In the original, one is drinking "something", and they are not quite jimi hendrix nor john frusciante, I just inpaint it until I produced that

[visual hemlock]

I was sitting and doing prompts from different shows and they all became so fucking tall and I did multiple ones and they all became the same.. Eventually I look at my photo size and it was 832 x 2048 😄

[royal monolith]

@nimble mason prompt: Abandoned, century-old lighthouse on rocky coastline at dusk, with crumbling stone walls, rusty lantern room, and overgrown vegetation. Waves crash against weathered foundation. Flickering sunset light casts eerie shadows." sampler: dpmpp_2s_ancestral_cfg_pp scheduler: linear_quadradic

[nimble mason]

oh, that sampler might be fucked lol

a lot of that stuff stopped working with sd35, including the euler_cfgpp in comfy

SamplerRK is definitely working though

[royal monolith]

it's something ;)

[nimble mason]

the "dpmpp_2s" in the dropdown in samplerRK is a dpmpp_2s_ancestral implementation

[royal monolith]

well... um... yeah, because of several reasons but more work with it than i expected. i'm half way through this chart i'm building. but that was unexpected

this is the entire set for that sampler

[nimble mason]

lol

[royal monolith]

i've got distribution on all but that one

[nimble mason]

yeah i probably need to go do some rescaling with the math or something

[royal monolith]

that might be good. cause, well ... yes, it's not working all that well ;)

this is another one

but euler_ancestral is nice

[nimble mason]

oh yeah my cfgpp ones are actually not for RF, i just looked

[royal monolith]

i'm just using the default that's in comfy as of the latest update for 3.5 medium

[nimble mason]

haven't updated em... hopefully never will as the goal is to just get all this shit rolled into the same core architecture

gotcha

[royal monolith]

someone stuck lms right in the middle of the dpm's

[nimble mason]

recommend trying this for euler ancestral

[royal monolith]

later. right now i'm putting a chart together to hand out to people that'll just be using the build comfyUI comes with. then i have every intention of playing around with what you have on your repo

[nimble mason]

gotcha, makes sense

[royal monolith]

but you might want to look at wht's going on with that sampler anyway

[nimble mason]

the euler ancestral in comfy?

that should be pretty much equivalent to this:

[wispy nest]

comfy adapted euler ancestral for ret flow recently

if i remember rightly from the commit history

[royal monolith]

#🏞｜general-with-images message

[nimble mason]

yeah it's not adapted for RF

[royal monolith]

becuse i should have at least gotten a distribution for that scheduler, not solid red

[nimble mason]

that was back in the pre-SD3 days

afaik

very very work has been done by pretty much anyone afaik to get noisy sampling working

[royal monolith]

all the other samplers i'm at least getting a noise distribution, though, for that scheduler. just not that one

[wispy nest]

I don't know why but it feels like the entire academic field wants these models to be one step of euler

[royal monolith]

they probably do. scientists get in ruts and don't like to get out of them

[nimble mason]

ultimately im' gonna delete all of those and just replace them with functions that pass the call to the same core sampler code

lol, yeah, i mean... here's how i see it

either you get it to the point where it's one euler step, or there's benefits to better sampling methods

[royal monolith]

yann posted something really dumb on twitter the other day, to which i responded with 'proof that scientsts need to get out of the lab and live in the real world'

[nimble mason]

test of linear quadratic, res_2s with one implicit step

hard noise eta 0.5

same as that first one but with beta

[royal monolith]

that's not bad at all

i like the first one best

beta's too saturated

[nimble mason]

agreed there

another linear quad

[royal monolith]

that one's good too

skip layer 2

[nimble mason]

agreed

[wispy nest]

I thought using linear_quadratic on image models instead of video models was a joke
but its working kinda well 🤔

[nimble mason]

main thing i'm looking out for here: the grainy gritty look that most mmdit models seem to be plagued by for whatever reason

[royal monolith]

never assume ;)

[nimble mason]

another linear quad

gonna gen these prev two with beta too

[royal monolith]

maybe because they're really LLMs not image models?

[nimble mason]

the thing that's interesting is the fact these schedules work period

yeah not sure if it's arch or training

tehre's certainly more synthetic data around now

[royal monolith]

skip layer 2 and see what you get

[royal monolith]

and image is an image. if you clean up the anomalies before you use it, it doesn't matter if it's synthetic or not

[nimble mason]

yeah, but i mean... more stuff that's got weird artifacts that aren't typical for photos etc

[royal monolith]

if you're talking about flux - that would be the midjourney scrape it's stuffed full of

[nimble mason]

just a random guses at one possible reason for the appearance of new artifact patterns

i've seen it with sd3 and flux

it was at its worst with the sd3 beta

and it tends to appear for certain prompts, which is one reason i suspect training sets being a culprit

when these images come out clean, they're gorgeous

[royal monolith]

SD3-2b-meduim wasn't supposed to be a release at all. and flux IS sd3, remember

[nimble mason]

yea that's why i said beta

[royal monolith]

there's a lot of 'beta's ;)

that's even a fish

[nimble mason]

but yea that is one type of artifact i'm watching out for, and something your comparisons had less of in the linear quad outputs

i've noticed other models doing this shit too now now that i think about it

[royal monolith]

like i said - skip layer 2

and see what you get

[nimble mason]

and different versions of it in certain ways... when flux spits out the fake toy story world look, it's especially prevalent on texturse like dark pavement

dots, dots, dots, fake detail

noise passing as detail

think most ppl think it looks great but... it does not, it's fake bs detail

[wispy nest]

I thought it was positional embeds
but then Sana came out with literally zero positional embeds and still has it

[royal monolith]

it's the core architecture

[nimble mason]

beta with the same seeds

linear quad is def better

i see some subtle haloing with the beta results, it's possible this is somewhat tangentally related to cfg now that i think about it

[wispy nest]

I think maybe its an inductive bias due to the fact that DiTs have to use patchwise embeds

I really wanna test the flan thing

[royal monolith]

did you try skipping layer 2 yet?

[wispy nest]

someone on civit added flan t5 to sd 3.5 and flux

he said its way better but then didn't put comparison images

[royal monolith]

what's flan t5?

[nimble mason]

nope, need to get that added, don't have it on here

i'm not doing general updates with my comfy environment

[royal monolith]

i did some images with just clip_l and clip_g with prompts and a . for t5xxl - and they came out really nice

[nimble mason]

too much customized stuff

[wispy nest]

flan t5 is a fine tune of t5 made by google

[royal monolith]

take it out of the mix entirely

[royal monolith]

it deals with the tiny details. skipping layer 2 has a tendency to get rid of those dots

[nimble mason]

SkipLayerGuidanceSD3?

[wispy nest]

oh nice this could help with lowering the side effects of too much noise injection

[nimble mason]

the problems with injecting too much noise come from the sigma schedule not getting adjusted to scale for it

[royal monolith]

yes.

[nimble mason]

RF is real sensitive to any perturbations to the variance level

[royal monolith]

also - ddim_uniform warms your image up - has a lot of red shift

[nimble mason]

anytime you add noise, you have to be sure the next step goes farther down that it normally would, to get you back to the right noise level

[royal monolith]

colorado river rapids!

[wispy nest]

ah is this the variance exploding versus variance preserving thing

[royal monolith]

did you not see the avatar i made you earlier?

[wispy nest]

no I didn't see

[royal monolith]

sgm_uniform vrs ddim_uniform

[nimble mason]

i might have to go make another RAG (random attention guidance) node and hope it's as interesting as it was with cascade

[royal monolith]

ddim on the left - lots of red

[nimble mason]

cuz i just love that name

SLG did nothing

[royal monolith]

that one isn't plagued with dots, either

[nimble mason]

is it supposed to have a very visible effect or be really subtle?

[wispy nest]

haha nice

[royal monolith]

it's subtle. skiping layer two on the left. skip turned off on the right. the very fine details

[nimble mason]

gotcha, good to know

scale, start end percent?

[royal monolith]

workflow is in the images

[nimble mason]

says no WF for some reason

just curious if you're letting it run through the entire generation or stopping at 0.15

[royal monolith]

click the image to open in viewer. click the open in browser text. right click, save

i'm using dango's default settings. but i have something else loaded and can't load it right now to check what they are

[nimble mason]

yea i can view others like my own above, idk what's up

no biggie

[royal monolith]

beta's noise distribution leaves a little bit to be desired

[royal monolith]

i'll check later then

[nimble mason]

what sampler is that with? what're you testing here exactly

obv torched lol

[royal monolith]

dpmpp_sde/beta

[nimble mason]

oh, yeah, that thing doesn't have any of the noise scaling corrected for RF

that would explain it

[wispy nest]

need to test Instaflow model

[nimble mason]

what's that

[wispy nest]

its like a ret flow version of SD 1.5

[royal monolith]

not as bad as dpmpp_sde/exponential

[nimble mason]

new?

[wispy nest]

no its old

but it would show something

if it gets flux grid or not

[nimble mason]

oh yeah that's def worth a look

[wispy nest]

one thing with ret flow is the flow paths aren't meant to cross

I wonder if this is the issue

because Pixart Sigma is a DiT and we never heard about Flux grid back then in the Pixart Sigma days

[royal monolith]

how would you test?

[wispy nest]

one thing I've been meaning to do for ages is write a python library that analyses and visualises various group statistics for the diffusion model trajectories
I originally wanted to do it to think about CFG burn but it would be helpful for this sort of thing also

[royal monolith]

i think it would be too. can you beat @nimble mason to writing it?

the difference between dpmpp_sde_gpu/normal and dpmpp_sde/normal

[wispy nest]

this sort of chart is great

they plug a toy 1-D example into the model equation to get these

its showing the CFG in the early steps destroying the image diversity

and then limiting it brings the diversity back

[royal monolith]

https://x.com/PyTorch/status/1851703101771714984

PyTorch (@PyTorch) on X

Learn the inner workings of Triton, the hardware agnostic language for GPU programming and powering TorchInductor: https://t.co/A6JTVjdRXW

Twitter · Oct 30, 2024, 7:09 PM

[royal monolith]

@viscid mural #🏞｜general-with-images message

[wispy nest]

I don't think I need speed bad enough to write Triton for it

[royal monolith]

it's calling you ...

[wispy nest]

there's a nice 3D one from GITS paper too

[royal monolith]

now THAT is a chart i like

[wispy nest]

that's how they made GITS scheduler they just took the average of these

[royal monolith]

i want to see what's going on in more than jsut 2D

i really want a visualizer that charts, in vector space, how a token in a prompt moves to it's final destination

[wispy nest]

3rd dimension adds a lot yeah

[royal monolith]

in 3D cause, yeah

[wispy nest]

I always forget the name of this paper, will find it tomorrow
its the one about delaying the negative
it shows maps of individual token vectors on top of the score function of the model
its the thing you are looking for

could be re-visualised in a nicer way probably

[royal monolith]

bookmarks are a thing ...

so much for assuing that the _gpu on the end of the sampler just means it'll use your GPU if you ahve one

[wispy nest]

lol

its just where the noise is made

its weird how much difference it makes

[royal monolith]

it really does. i'm really surprised how different the distribution is turning out

[rain gazelle]

Their recent share on monodromies was pretty interesting. Will be interesting to see that shift.

chuckles

😎 sometimes it feels like that in here

[royal monolith]

in what way? you have a link to that?

[celest sigil]

https://tenor.com/view/dont-cross-the-streams-egon-ghostbusters-bill-murray-gif-14907508

First thing I thought of when I read this 😄

https://tenor.com/view/ghost-busters-holding-on-keep-going-gif-9935479

[languid pebble]

Good morning coffee

[languid pebble]

Homer J. S. having a midlife crisis 😄

[celest sigil]

Apparently, the street gets narrower with age

[languid pebble]

Cool idea!

[celest sigil]

Quite a few more over in #🆕｜sd3 🙂
I got the idea from a Flux lora, but I did all of mine without the lora.

[rain gazelle]

it was a joke, just gibberish 😛

[sterile kiln]

Looks like Brigitte Macron

[glossy sun]

a teenager`s head in mini style

[royal monolith]

it was just a really, skinny, tank

[languid pebble]

Cause of the wig?

[neon vine]

this unironically looks like near where i live

i guess its that unmistakable ex commie eastern europe aesthetic

[steep sleet]

Andy Warhol Jagger

[steep sleet]

Flux and Ollama

[orchid nexus]

bonjour, puis-je avoir une assistance en Français?

[royal monolith]

Je peux essayer de vous aider, mais je dois utiliser un traducteur, si vous êtes d'accord ?

[orchid nexus]

merci, pas de soucis
je suis novice, je cherche à savoir que faire pour créer une image

[royal monolith]

pour créer une image sur ce discord, vous devez utiliser le canal artisan. Vous pouvez lire les informations à ce sujet dans ce canal #artisan-faq Cependant, il est en anglais, donc je vous suggère de copier le texte et de le coller dans ce traducteur. https://www.deepl.com/en/translator

[orchid nexus]

merci je vais essayer

[steep sleet]

Tres bien

[languid pebble]

The red planet

[little dagger]

I’ve been trying to get out paint to work for hours now and this is the best I get using poor man’s off paint. They really need a new video update of what needs to be done with modern out paint

[clever oar]

where mask??😃

[royal monolith]

he's a rooster, they're obstinate

[clever oar]

he has no right to break such rules!