🆕｜sd3

[pseudo owl]

You can just input multiple images and they can understand them(you can ask questions or give instructions) They support video understanding and I believe oryx 1.5 and llava onevision support 3d understanding too.

I think MiniCPM v2.6 is the only one supported by ollama so far and I prefer that for vision understanding. They are around 7b-8b in size. I forgot that theres also Qwen2 vl which is pretty great(supports img, multi-img, vids long as 20min).

[toxic bone]

was using 3.1 8b and some extension

[pseudo owl]

Now there are even open models that can natively generate speech output and understand speech.

[toxic bone]

there's a lot of things all iterating along, and will soon converge and change the way we live.

[cursive frigate]

I've been using this one for a little over a month now and its been great. Was just hoping there have been some new stuff to torture my pc with.

https://ollama.com/doreilly/minicpm26_q5_k_m

[mortal mesa]

i dont find myself doing thing with vision models but way back when i did i used Moondream, that not even on the radar anymore? the newer stuff eclipses it?

[toxic bone]

theres a weird thing you can do with flux an sd3. use one of these vision models to describe a long detailed prompt about something that was generated, and then give that back as a prompt and get a very similar image

i used florence large for that a few times

[pseudo owl]

Moondream is still really fast and amazing but its hard to instruct it, it only gives captions or for vqa. You can feed that caption to a normal llm for better instruct capability but its somewhat limitated.

[cursive frigate]

That is how I use it all the time

[toxic bone]

it's a wild feedback loop. black box level magic

[cursive frigate]

This is the workflow I use and the results are pretty good. The base of the workflow came from Aitrepreneur but I changed a lot of it to do more.

I made a section for just feeding the prompt LoRA keywords because AI sucks at keeping that text exactly the way it needs to be to trigger LoRAs and another section where I can tell it some basic instructions for the image and style I want and it also analyses the image and provides a prompt then I have it combine all of that to generate the image. The results have been really good.

If you want the workflow let me know and I will upload an image with it baked in.

[toxic bone]

oh comfyui is not my scene right now. thanks though! i am coming back to dive into comfyui again soon. Factorio space age came out recently though so all the noodling part of my brain is , occupied currently

actually loaded fooocus today for a bit

[cursive frigate]

ComfyUI is great once you start digging in and figuring out how to modify it to do the things you want. I used focus a little bit about a year ago. It was pretty decent.

[toxic bone]

i've used a lot of comfy. i'm no stranger to node graphs. factorio is just oen giant node graph of a game and that brain gear is engaged on that task. once factorio itch is scratched, i'll experiment with it again no doubt. maybe by christmas

i got time off then

[cursive frigate]

Looking forward to see some of your results. They always seem to be good when you put images in the chat.

[bitter hearth]

I've had a theory for about 6 months now that if we can get good enough captioning models then we can fix tiled upscale

cos the reason tiled upscale gets duplicated subjects is that we are using the same prompt across all the tiles

[craggy crest]

you sure about this?

[bitter hearth]

that's the usual explanation yeah

like imagine if it was just a photo of a cat on a rug

when generating the initial image every tile has a cat in it, cos there is only 1 tile

but imagine if we are at 4x upscale, there are 16 tiles, and many don't contain a cat yet the prompt says cat

[mortal mesa]

ive prompted separate tiles, to much work for my use case which is pretty much nothing

[craggy crest]

https://build.nvidia.com/nvidia/consistory

[craggy wave]

Hi there ! Anyone here have a comfyUI workflow for sd 3.5 large with Lora Loader + hire fix feature ? Thanks 🙏

[kind acorn]

Paint a picture of sunset and lone ducks flying, autumn water together in the sky

@dusky thistle Paint a picture of sunset and lone ducks flying, autumn water together in the sky

[bitter hearth]

yeah I have a workflow where I hand prompted 25 tiles it was too much effort

hoping vision models can do that well soon

[bitter hearth]

new MiaoshouAI tagger version https://huggingface.co/MiaoshouAI/Florence-2-large-PromptGen-v2.0

it can give output for T5 and ClipL now

[noble coyote]

President Donald Trump - again - PotUSA number 47!

[craggy crest]

shouldn't that be in #🌶｜off-topic

[noble coyote]

"I think it should be in knead two kneau!" 🥳

[craggy crest]

i so want to make a bread pun, but i'm too tired to think

[noble coyote]

Florence2/Flux img2img

[noble coyote]

Yeah - you're in here all-hours - I've had 7 hours shuteye since our paths crossed!!!

[craggy crest]

it's just now midnight for me

[noble coyote]

Mid-West USA

[craggy crest]

arizona

no daylight savings

[noble coyote]

Beautiful State - I love the series of Route 66 photos of Hackberry Az (by Carol M Highsmith)

Are your clocks set for 'winter time' yet?

[craggy crest]

i love driving on parts of the old route 66

there are still burma shave signs on part of it

[noble coyote]

Its 7 a.m. here in London UK

[craggy crest]

good morning then :)

[noble coyote]

Yes, Route 66 is a wellspring of Americana

[craggy crest]

a lot of it is.

[noble coyote]

"Good Morning Arizona!!!!!!!!!" 🥳

[craggy crest]

there was a guy in here earlier trying to figure out how to generate. he never did, but he donated a very nice prompt "Paint a picture of sunset and lone ducks flying, autumn water together in the sky"

[noble coyote]

I saw CSBW's version - an eagle in place of some ducks!!! 😄

[craggy crest]

i saw that too, i think he added that on the sly

[noble coyote]

Nice, very soothing

[craggy crest]

i wrote that one down, i quite like it

[noble coyote]

Here's a vast prompt I got from another forum ... freely posted ... if you're into B&W?

[muted dove]

Surprisingly nice image for such a bad prompt 😄

The ducks I mean

[craggy crest]

yeah. that was with 3.5 large, no loras. very plain vanilla

[craggy crest]

yeah, but i'd be likely to render each sentence as it's own prompt

[muted dove]

"Lone" ducks, flying "together" 🤦🏻‍♂️

[craggy crest]

too many tokens in that

[craggy crest]

actually, there's a comma after ducks and so: "autumn water together in the sky" i really don't know how autum water can be together, but if it's in the sky, it's normally rain

[noble coyote]

"There are no bad prompts!"
Prompts are rescued by the alchemy of SD.
Prejudice against "short-prompts" should be disparaged 😄

[craggy crest]

an apple spinning in space

[noble coyote]

It produces amazing results - I think the prompt-maker doesn't realise that only about one tenth of that text ever gets used!

[noble coyote]

OK, so you have the most left-field and surreal workflow!!! 😄

I love the way you have the cougar "looking at that spinning apple!"

[craggy crest]

rofl. i gave you a prompt, and then posted an image from the skip layers test i'm doing. they really don't relate to each other

[noble coyote]

🥳

[craggy crest]

told you i was tired ;)

[noble coyote]

It brings out your sense-of-humour

[craggy crest]

here's the first paragraph of that monster prompt you posted

[noble coyote]

It works.

[craggy crest]

that it does. i get fairly silly

[noble coyote]

Silly is good. This room can get a tad doomy at times 😦

[craggy crest]

second paragraph

[craggy crest]

third paragraph

[craggy crest]

5 prompts for the price of one ;)

[untold valley]

ppl need to add token*counter to comfyui cant have over i belive 256 tokens or something like that

[muted dove]

The whole thing

[noble coyote]

There is a URL available (!) which parses your prompt and returns the number of tokens it contains

[craggy crest]

lost the black and white somewhere. love the way it wrote the words out though

[craggy crest]

this is for stable https://sd-tokenizer.rocker.boo/ BUT it is model specific, each model will tokenize differently

[noble coyote]

Yes - the prompt-owner either has a great grasp of language - or he's sleeping with an LLM!

[craggy crest]

he's sleeping with an llm - however, at least for flux and SD3.X - since the images in the dataset were captioned with CogVLM - getting an LLM to create prompts for them works perfectly

you just have to remember to tell it how many tokens

[craggy crest]

because everyone should have beer with their breakfast

[muted dove]

Ferrari inspired armour design

[noble coyote]

When I visited Hungary many moons ago - so many people breakfasted on beer!!!

[noble coyote]

YMCA Fireman!!!

[craggy crest]

usually you put it IN the pancakes, however

[noble coyote]

The prompt consisted of "the white witch eats pancakes and drinks beer with the red knight!"

A secondary prompt dropped the beer and pancakes, and substitued "does the laundry"

[muted dove]

It's really struggling with monochrome!

[craggy crest]

prompt: a group of giggling girls chasing geese

[noble coyote]

^..^<

[craggy crest]

@-->--->---

[noble coyote]

Hi, you look familiar?!

[craggy crest]

<%%%%|==========>

[noble coyote]

Sword

[craggy crest]

--~~~=:>[XXXXXXXXX]>

[noble coyote]

<*))))))))))><

[craggy crest]

:)

@@@@:|

@@@@:)

that works better

[noble coyote]

Indigenous beefcake

[craggy crest]

black and white portrait photo of an elderly native american in the 1800s

[untold valley]

Oh damn, Trump won.

[noble coyote]

PotUSA 47

[muted dove]

What a disaster and a terrible image of the American population. They were just starting to gain some credibility back with the rest of the world. I fear it'll be a worse shit-show than last time too.

[craggy crest]

prompt: diana ross and the supreme sandwich

[dusky thistle]

so far beyond fucked up i don't even know what to say

[craggy crest]

can we not use this channel for politics? there's an #🌶｜off-topic channel for stuff that's not Stable diffusion

[dusky thistle]

it will be much, much worse

[noble coyote]

The USA is entering a period of Unstable Diffusion

[muted dove]

Weighing up her options after finding out who the next prez will be.

[dusky thistle]

absolute nightmare tbh

[noble coyote]

The evil cabal of Farage Musk and PotUSA

GM all y'all

Welcome to a new dawn

[noble coyote]

Your excellent w/f can take up to 15 minutes per image on my 8Gb VRAM PC 🥳

[craggy crest]

there's a 12 step program for that

[dusky thistle]

thse take about 50 sec on mine

[noble coyote]

Great!

[craggy crest]

they don't take any time at all on mine.

[noble coyote]

I wonder how my (desired) 5090 will do?

[craggy crest]

(cause i can't run 'em)

[noble coyote]

🙂

[craggy crest]

waiting for him to get his node into manager

[noble coyote]

I just updated RES4LYF

[dusky thistle]

did you try installing from the link using the manager

[noble coyote]

Is installing via Manager any worse or better than a git pull?

[craggy crest]

not yet, no. i've been slogging through one skip layer after the next for 48 hours now

[craggy crest]

yes

[noble coyote]

How?

[craggy crest]

you mean git clone

[noble coyote]

Git pull as in update?

[craggy crest]

but you hve to have it before you can update

[noble coyote]

So a Manager-inuced install wins hands down over a git clone? How?

[dusky thistle]

not sure

[noble coyote]

OK

[dusky thistle]

but he doesn't have pywt installed in the right spot

i guess manager handles that

[craggy crest]

it doesn't

[dusky thistle]

you can also just comment out the import pywt btw

[noble coyote]

In truth, it probably is no difference

[dusky thistle]

it just won't let you use the wavelets noise type then

but it's one of 17

[craggy crest]

there is, actually

[noble coyote]

But what?

[craggy crest]

sometimes manager's install works great. and othertimes it barfs and git clone is what you wind up with

[noble coyote]

What is the essential difference between a git clone install/Manager install?

[craggy crest]

when one doesn't work, the other usually does?

[noble coyote]

OK, but even if both work - which one is better?

[untold valley]

you do it yourself vs a program does it for you

[craggy crest]

neither, both, doesn't really correlate

[untold valley]

same thing end result

[craggy crest]

"does not compute"

[untold valley]

anyone been able to recreate 1.5 skin textures with 3.5 yet? like somehow removing the "stylized" format that it outputs. its nice, eye pleasing but sometimes you want that generic non-retouched looking image

[dusky thistle]

it can do better than 1.5 skin textures

tbh, sdxl can too

[untold valley]

need like prompt or keywords, ive been struggling w it

[muted dove]

Maybe it's down to the workflow and not just the prompt. You shouldn't need to prompt for skin detail.

[untold valley]

yeah maybe its user error

[muted dove]

The prompt I typed for this was just a face

This was with freckles, but the refiner went overboard with it 😄

[dusky thistle]

a close up amateur cell phone photo of a woman smiling in her messy apartment

[untold valley]

maybe its just that we are genning in a higher res but you can kinda see how they have some sort of plasticity,

not my images from *photoreal civitai page, but it just feels different, resolution may be the key. but sdxl,3,3.5 dont have this type of airbrush, high make up feeling. idk maybe im just seeing things and need sleep.

[dusky thistle]

those do look a lil airbrushed yea

[muted dove]

This looks perfectly normal to me

[muted dove]

That's a noisy mess 🤷🏻‍♂️

[dusky thistle]

that's the point

amateur cell phone photo should look like this

[muted dove]

Why? It doesn't look good. Not even like a poor amateur photo.

[dusky thistle]

disagree

[muted dove]

It's ok as small images on Discord, but full size they're bad.

[dusky thistle]

one shot quick generations with no refinement using sd35M

the faces you posted above look pretty plastic

it's more convincing on the low end quality side

most camera photos are amazingly crap

[muted dove]

I think yours are too far in the low end, so they don't look good or realistic. I'm not being confrontational, just my own opinion 🙂

[dusky thistle]

search around for phone resolution photos on google

without any other context

what i posted is better than most of em

they're blurry, hazy, full of artifacts

[muted dove]

Nobody has phones that bad nowadays though 😄

[untold valley]

ok awesome at least im not losing my mind and you kinda sorta are getting it that left one leans in the right direction, but u see the plasticity thing on the rights forehead.

[dusky thistle]

most photos look worse than that

and maybe theoretically the camera phones are good but they're usually not used that well

glare, dirty lenses, blur from the hand shaking, poor lighting, etc

half the time you can barely make out the structure of the iris

without the cell phone part

sd35M

[muted dove]

Isn't that just bright light reflecting off moist skin, or do you mean the texture in that area?

[untold valley]

texture

[untold valley]

kinda sorta zeroing in this and the left hand image of glaxy pink tank top shirts going the right way

[muted dove]

That could be down to a combination of using Flux as a refiner and the sampler/scheduler choice.

[muted dove]

Skin is textured, I think it looks acceptable in this one. 🤷🏻‍♂️

Anyone got a pin?

[untold valley]

what prompt did you use for image on the left if you are pry to share? think that was the closest

[muted dove]

They were all the same prompt.

[untold valley]

so just "a face"

[muted dove]

This one was...

a young tanned __nationality__ woman with fair skin and a petite physique. She has a round face with a soft, natural makeup look, and her hair is styled in a casual, side-swept bob. She is wearing a fitted, sleeveless, pink ribbed tank top that accentuates her small to medium-sized breasts. Her attire includes thin, transparent straps that add a modern, minimalist touch. The woman is wearing round, gold-rimmed glasses that frame her face, giving her a studious appearance. Her expression is happy, with a smile, and she is looking directly at the camera with a confident stance, one arm raised to adjust her hair.
In the background, a dark street scene

[muted dove]

@untold valley I do use an LLM in my workflow, so that prompt isn't necessarily what is used for the end result.

[untold valley]

baby shark do do dooo dooo baby shark! that hats awesome.

[untold valley]

thank you both for helping, i have some ideas i may experiment with. they aesthetic may yet be there. a prompt that big will not make it consistent but ill dig thru and see if there are any key tokens there

[muted dove]

baby shark

[dusky thistle]

doododoo

[untold valley]

yeah we got close! the idea is there. thanks!

[untold valley]

a lot of workflows now incorporate LLM's keep trying to see examples but its crazy how much you can tell a ai wrote it. we have come full circle, using ai to generate ai prompts to generate ai art.

[untold valley]

@dusky thistle @muted dove instagram photo is helping

[muted dove]

an amateur iphone photo of a face

[untold valley]

awesome

iphone, instagram jfc just need to think like a white girl out to get her venti mocha latte from starbucks

[untold valley]

i could tell you used "sharp image" in your prompt lol

[muted dove]

This is raw sd3.5 output

[untold valley]

that's a really clean crisp gen

[muted dove]

I don't like the neck

[dusky thistle]

held up surprisngly well considering how close up it is... sd35 is a lil better than expected there

normally that just turns into nonsense

they're pretty coherent models for sure

[untold valley]

regarless of anything sd3.5m is a great model

[untold valley]

yeah

[muted dove]

I'm using SD3.5L

[untold valley]

have you tried medium?

[muted dove]

Yes

[dusky thistle]

i bounce back and forth all the time

[muted dove]

Me too, they're all just different models

[dusky thistle]

everything recently above is medium

yeah def

no doubt the training set was different

[muted dove]

I don't like how SD3.5 left the face blurry, but refining with Flux makes freckles look like the pox.

[dusky thistle]

flux tends to make skin look plastic

it's better to have something be noisy and shitty looking than too clean

very refined sets off our BS detectors

[muted dove]

Sometimes I'd agree, but Flux does tidy up a lot of the "mess". It fixes hands and fingers for example.

[untold valley]

i will say based on that comparison the textures on materials like brick etc stand out, the skin texture not so much but the bricks/cement 🤌

[muted dove]

...and the watch detail

...eyes, hair... 😄

[untold valley]

kinda reminds me of like 2.0 good at "landscapes" horrible everything else

except better at many things

minus the humans

thing that bugs me ab 3.5 is it loves loves the 3/4 shot

doesnt like to zoom out

[dusky thistle]

yeah i'm referring more to textures, details

jpeg artifacts are good for fooling the eye obv

the smudge patterns phones give too

[bold fossil]

Underwater world with colorful fish, coral reefs, and sunken ship, illuminated by natural light filtering through water, in a hyper-realistic style

[limpid thunder]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[dusky thistle]

Here is the image you requested.

[muted dove]

Love the window/water detail in this (raw sd3.5L)

[untold valley]

could I ask which one of these appeals/ looks the best according to wte criteria you want?

[dusky thistle]

ask dinogator

[muted dove]

I don't like snow on the face and the eyes are worst in the 1st one. So it's a difficult choice.

[untold valley]

I can’t make up my mind either

[muted dove]

Not that sort of "pool"! 😄

[ocean mural]

make image of an alien galactic council

[noble coyote]

Thwaark!!!

[dusky thistle]

forget if i said this last night, but i got the deis samplers working!

it's kinda interesting - the third order multistep samplers seem to struggle a tiny bit with SD3.5, maybe because we're using CFG

DEIS_3M is basically what comfyui uses when you select "DEIS", i'm finding DEIS_2m is much better

[craggy crest]

how about releaseing a clownsampler node that's just your versions of the samplers and schedulers that can replace the ksampler node

[dusky thistle]

def a good idea

i was thinking of having a ksampler select and a "klownsharksampler" (who knows what i'd call it, but yeah)

equivalents to that

gotta work out some good presets first... generated like 9k sd35 images now to that end

but yea def agree it would be good to have a simple interface for easy entry

SD3.5 Medium

living rooms are something diffusion models really struggled with before

[craggy crest]

Kitty!

[dusky thistle]

it's great with animals

[old walrus]

medieval empress sitting on her throne, dark fantasy, digital art illustration

[bitter hearth]

oh no

I opened the workflow on the kitty image

[craggy crest]

when your cat has an affair with a rat

[craggy crest]

what'd ya find?

[bitter hearth]

chaos cloud

[craggy crest]

??

[noble coyote]

£10 cash

[craggy crest]

that was interesting

[bitter hearth]

just a lot of nodes

in a big cloud

[marble sand]

lol

[noble coyote]

Flux and Silhuflowart2 LoRA

[pseudo owl]

Tried will smith eating spaghetti with Mochi-1 from genmo website(apache2.0 open model)
I mean not bad but not perfect.

[sacred jewel]

TESTOTHER is my preference

[untold valley]

I appreciate the feedback thank you 😊

[bitter hearth]

found a gguf of flan T5

https://huggingface.co/dumb-dev/flan-t5-xxl-gguf

its better than normal T5

its similar idea to common thing where people replace Clip-L with this https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14

don't rly need a gguf for Clip its so tiny already lol

[craggy crest]

he ate the fork!

[sacred jewel]

Cool LoRA thanks for heads up 😄

[noble coyote]

You're welcome!

[rapid pivot]

This prompt

The nightmares ive seen...

[pseudo owl]

lol, yeah its a massive increase in quality compared to previous models

[sullen moss]

https://blackforestlabs.ai/flux-1-1-ultra/

[pseudo owl]

Cool but sadly closed source, images do look great but are they even going to release something open now?

[bitter hearth]

the flux pro update does look amazing

2048x2048 native, and they made a realism mode

[sullen moss]

Yep

This is really crazy

[craggy crest]

pictures or it didn't happen ;)

[cedar axle]

SD3.5L

[cedar axle]

really fun texture on this one 😍

[craggy crest]

A dark, Gothic castle floating on a mist-covered, rocky island in the sky | towering spires stonework, eerie, isolated atmosphere | surrounded by thick fog, fading into a bleak, overcast sky | minimalist composition with high contrast | surreal, haunting ambiance

[errant dust]

I must have missed the news. What update? NVM, found it. Well, I am sending a en email complaining and demanding a refund for my Flux Dev

[craggy crest]

waste of time most likely

[errant dust]

Not a problem. I am a serial email complainer

(curiously got me two jobs at companies I was complaining to. True story.)

Maybe three, if I include writing for a newspaper back in 90s. Not sure that counts, and was not by email.

[dusky thistle]

the great north american sandfish

[bitter hearth]

FLUX1.1 [pro] – ultra mode: This option enables image generation at four times the resolution of standard FLUX1.1 [pro], without sacrificing prompt adherence. Unlike many high-resolution models that experience significant slowdowns at higher resolutions, our performance benchmarks show sustained fast generation times—over 2.5x faster than comparable high-resolution offerings. This model is available at a competitive price of $0.06 per image.that's the thing, their price for one single image is around the same as renting an RTX 3060 12GB for 30 minutes

[untold valley]

I have a big complaint.... this looks so good on discord but then you click iand the rocks/waterfalls so blocky. water still looks nice tho. but still bait n switch lol

[dusky thistle]

sounds like you need a new monitor lol

https://www.bestbuy.com/site/lg-signature-97-class-m3-series-oled-evo-4k-uhd-smart-webos-tv-with-wireless-connectivity-2023/6550242.p?skuId=6550242&utm_source=feed&ref=212&loc=TVsGeneralPREMIUM&gad_source=1&gclid=CjwKCAiAxKy5BhBbEiwAYiW--0Pi08Qp59SXEkGZO22ZdCy7eptwAaKnv23aIJKF51hDGn_thJbXeBoC56UQAvD_BwE&gclsrc=aw.ds this will help you see that image more clearly

[untold valley]

lolololol the blocks

i have two screens

one is like a 14 or 17" POS, my main is a nice 27" IPS one

[dusky thistle]

damn

i'm rockin dual ultrawide

got a 57" dual 4k and a 49"

[untold valley]

we are not all millionaires

[dusky thistle]

me either lol

i'm just a big believer in selling passenger doors off cars to fund more hardware purchase

[noble coyote]

Flux and Silhuflowart2 LoRA

[craggy crest]

did you actually open the original image? or just look at discord's compressed version in their image viewer?

[untold valley]

open in browser

[dusky thistle]

looks like sandstone cliffs to me

got stuff like that in my area

the water gets into the sandstone and separates the layers through freeze thaw cycles over the years

[robust echo]

An old 1960s vintage photograph album depicting a parked car in front of a house --ar 3:2

[dusky thistle]

Here is the image you requested.

[untold valley]

Clownshark has became the Ai.

[robust echo]

a Chinese man standing before a blue lake,ride a red mountain bike, wear a sunglass, on the plate refelcting the lake

[dusky thistle]

Here is the image you requested.

[muted dove]

Damn you Scotty, you got the teleport coordinates wrong again!

[lone eagle]

We lie down together amid the cacti. The succulent leaves closest to the ground are a marbled grey, as if turned to stone, and we become absorbed by them, feeling our way around their rounded contours with our fingers. As we gaze up, following their odd tear-shaped forms bundled together, the sprawling double cypress tree – two trunks locked in an embrace – claims our attention with its swaying branches splitting into ever finer branches and twigs ending, here and there, in clusters of cones. Initially, we can’t really tell if it’s the wind that’s causing the canopy to stir and sway that way. It forms a dark shifting frame that we enter and get lost in as one does in a forest.

[noble coyote]

Flux Turbo LoRA w/f with 2 x KSamplers, 2 x Upscale and Sharpen

[bitter hearth]

the 8 step Flux Turbo LoRA gave me about as good an image as 1200 steps of Flux Dev

its so amazing

something about these new models means they can compress like that

[rapid pivot]

I remember this prompt lmao

[sacred jewel]

Yeah, revisiting old prompts with new LoRAs 😉

[signal shuttle]

Finally just started to train a 3.5 medium lora, its going pretty well, so far i noticed that it learns styles better then flux

[craggy crest]

if you're trying to generate that, and that's a really lousy prompt, you need to do so in the #artisan-faq channels

[pseudo owl]

Which one? The one from alimama?

[lunar canopy]

@noble coyote mind accepting friend request?

[bitter hearth]

yeah the alimama one

the hyper one is also very strong its not bad either

the bigger your final resolution, the less harm they do

so they do especially well for tiled upscale

[pseudo owl]

Yikes it’s amazing at prompt following. Even auraflow, flux dev can’t get this right most of the time

It nailed everything

Prompt: A white cat on top of a blue dog sitting on a brown couch in a living room. Behind them is a window with 4 cow pictures, one in each corner. Outside the window is outer space and a ufo.

[bitter hearth]

often acceleration loras can improve various aspects

people often view them as always being worse

but TCD is substantially better than regular SDXL, for example

[noble coyote]

Flux Turbo (8-step with LoRA) AliMaMa

[noble coyote]

'Ullo?

[pseudo owl]

Seems pretty impressive so far, both didn’t get it 100% right, I think flux dev might win at this one tho.
Prompt: A man holding a sign that says “This is the grand contest of the teacher or the student. Will 8 steps win or 30 steps?”

[bitter hearth]

oh if you want accurate text that is probably a pretty big exception

[untold valley]

Ooooo Torcello big trouble.

[bitter hearth]

I think using DiTs for text will not be the way to go soon

Omnigen-style models are much better for text

the huge downside of Omnigen models is they can't add noise to correct for past mistakes like we do with diffusion

[craggy crest]

not sure that's a problem

[bitter hearth]

not sure

it may well not be a problem

we're using stuff like DPM++ 2SA, Euler A and Clownshark stuff on the Rect Flow models but you're not really "supposed" to

[craggy crest]

you know what "you're not supposed to" really means?

[bitter hearth]

on Comfy server someone was saying, the big API providers like Fal aren't doing that

lol

[craggy crest]

not doing what?

[bitter hearth]

using fancy stochastic solvers
they are likely using DPM++ 2M or heun or euler or something like that

[craggy crest]

for which model?

[bitter hearth]

oh I mean for Flux and SD3/SD3.5

[craggy crest]

i know what FAL is using for SD3.5 - i gave them the settings.

not sure about flux

[bitter hearth]

on comfy server someone speculated that FAL have made a hand-written INT8 kernel
their Flux Dev endpoint is a lot cheaper and faster than their competition

[craggy crest]

they might have

[bitter hearth]

but yeah the point I was making was the research tends to just want simple samplers for these modern Rect Flow models
compared to diffusion where SDE/ancestral was favoured

[craggy crest]

for 3.5, they're using: sampler: euler, scheduler: simple

cause that's what I told them to use

[bitter hearth]

oh thanks that's really helpful

so we've finally reached the point where the trajectories are straight enough to use euler

it was gonna happen eventually

if your line is straight then euler is optimal

[craggy crest]

did you, or did you not, praise a euler image the other day?

[bitter hearth]

haha

its swings and roundabouts

[craggy crest]

;)

[bitter hearth]

diffusion models tend to have curvy trajectories, is the issue

[craggy crest]

plotting out the path through an image is worse than plotting the trajectory to hit the moon with a rocket from earth

[bitter hearth]

its a pretty rough method yeah

[craggy crest]

"getting better all the time"

[bitter hearth]

GANs and VAEs don't do it they just jump in one go

apparently SD 1.5 VAE was almost a GAN anyway, there's a lot of overlap

I still find it funny that with diffusion we run the models backwards

[craggy crest]

computers don't care which way is up

[bitter hearth]

yeah they seem quite chill about it

[craggy crest]

as long as you don't drop them off the second floor balcony or something, they're happy with whatever

[icy coral]

I don't know where the beef with the author of the P-word model originated from, but I was wondering if the team behind Illustrious decided to train a model on 3.5 would they be in a better position when it comes to obtaining a license?

[bitter hearth]

fairly sure the situation was that Stability AI were busy and were behind in terms of dealing with the licenses

rather than them specifically denying a license

[pseudo owl]

p-word is pony?

[rapid pivot]

If that's it

I call stupid

lmao

[signal shuttle]

cagliostrolabs (Creators of AnimagineXL) announced that they had plans for a 3.5 fine tune. so far they are the only only ones who publicly announced that they plan on fine tuning 3.5. i believe that they will be the first to publish a full anime fine tune for 3.5 but i may be wrong, we will have to wait and see

[icy coral]

Someone should probably get in contact with him then, because it looks like he's pretty convinced he's being targeted somehow

[untold valley]

@icy coral and everyone I guess, you need to think about optics with Pony models we are 100% talking about heavy NSFW and SAI or any company likely wants to steer 1000000000ft away from it. This is because in doing so you lose potential for future funding. It’s all about a business and money and lastly public perception.

What SAI needs is a adept PR person for retail.

[signal shuttle]

Man flux ultra is so good, like what

[icy coral]

I'm guessing the answer is no then

[untold valley]

It’s generating nice images but can you run it in your own pc like sd3.5M?

[untold valley]

No what?

[signal shuttle]

No, but its fun to use while waiting for a major fine tune of SD 3.5

[bitter hearth]

flux ultra is incredible quality but closed source models are absurdly expensive
for the price of one flux ultra image you can rent a 3060 for 30 minutes

[signal shuttle]

MAN I REALLY WANT A OPENSOURCE MODEL THAT CAN DO SICK STUFF LIKE THIS

[pseudo owl]

prompt? I'm sure open models can do this with a bit of extra settings.

[craggy crest]

@bitter hearth https://www.reddit.com/r/StableDiffusion/comments/1gm1kxm/ever_look_at_real_photos_and_see_ai_artefacts_in/

[signal shuttle]

my prompt was made with Claude "Vibrant anime key visual poster, "The Adventures of a Clumsy Female Knight!" prominently displayed in stylized Japanese text, featuring a cheerful young woman with messy blonde hair in oversized, ill-fitting armor stumbling forward, sword awkwardly held aloft, against a backdrop of a whimsical medieval castle and lush fantasy landscape, dynamic action lines emphasizing her clumsy movement, soft pastel color palette with pops of bright accents, highly detailed character design in the style of Studio Ghibli meets "KonoSuba", expressive eyes and exaggerated facial features conveying both determination and embarrassment"

[craggy crest]

standard anime lora can do that. all the base models can do that with the right prompt

[signal shuttle]

Can they also do the text correctly?

[craggy crest]

it's non-english characters, so it'll look correct but not be readable.

but we had this discussion yesterday

[signal shuttle]

Look above the non-english letters and you will see english words

Its small and in black

[pseudo owl]

the english sentence is not hard to do at all, its the detail which can be done with a lot of upscaling and refining.

[craggy crest]

sd3.5 base, no loras

[signal shuttle]

Large?

[craggy crest]

yes

prompt: a blond haired anime cartoon knight, happy, wild hair. across the top of the image is written the text "the adventures of a clumsy female knight"

[bitter hearth]

lol yeah looks like a bad AI generation

[craggy crest]

and yet, it's a photo - just goes to show it's really not possible to tell what is ai generated and what isn't, any more

[bitter hearth]

yeah essentially not possible anymore

the impressive thing about Flux Pro Ultra is just the size without tiling, done within 10 seconds

it doesn't even need to be a new technology though, hand-optimised pipeline on Nvidia H200s, plus more finetuning and distillation shenanigans can go far

[mortal mesa]

digital cameras(cell phones) don't necessarily take true pictures like film

[bitter hearth]

yeah film was nicer IMO

[craggy crest]

course, there's the issue of the actual brains that are processing the data coming in through the eyes of the human that's viewing whatever it is. not everyone sees the same thing in the same way. re: that blue dress on the internet a couple of years ago

if you want to get really technical, you're not seeing the actual objects at all, just the light that is bouncing off them

[river sleet]

You aren't even seeing that. You're seeing vague shapes and a tiny circle in the middle of your vision that's actually clear and sharp. Everything else is a hallucination made up by your brain.

[craggy crest]

yeah, the light is activating sensors and hopefully your brain will decode the data, and extrapolate in the same way everyone else's does. but what if it doesn't?

[bitter hearth]

nice, really sharp and good colours

[mortal mesa]

the illusive leapordtigerpuma

[craggy crest]

he was Dr. Moreau's kitty

[mortal mesa]

hah nice

[craggy crest]

he's self-polinating

[pseudo owl]

Some interesting prompts with mochi

[rapid pivot]

Trump, drugs and one interesting prompt

[rapid pivot]

Twitter after the x accident

[craggy crest]

@dusky thistle @bitter hearth https://youtu.be/EvmML-aFRsQ?si=Cw1f-PXzOSSzg7tn

[bitter hearth]

the canvas looks amazing

[craggy crest]

yeah. so many comfy nodes I want to go download and fill my drive up with

[bitter hearth]

I've essentially switched to diffusers/pytorch only at this point

I did enjoy comfy a lot though

[craggy crest]

gonna miss out on all the toys ...

you know anyone that's made any 3.5 medium finetunes?

[halcyon yarrow]

has anyone here tried mochi? I just tried it and it takes me 12 minutes with 8GB vram to render 31 frames or 2 seconds @ 15fps lol

[bitter hearth]

seems too early for 3.5 medium finetunes but there's been some progress in koyha, ST communities etc apparently

people are split between more models now, and that's probably gonna get worse as time goes on

[rapid pivot]

It's good and bad

[craggy crest]

it's pretty promising

[halcyon yarrow]

same reason i don't want BFL to release a flux-dev 1.1 it's just going to split the community further and invalidate all our work

14.75 minutes to render 73 frames aka 4.8 seconds

i'm going to keep increasing my frame count to find my max

[craggy crest]

pretty sure the companies don't care about whether the community gets split or previous work invalidated. they're going to keep researching and releasing. and the community isn't going to care whether others are training, they're all going to do their own thing - as always

[halcyon yarrow]

i can totally wait 15 minutes for a 5 second video clip if it's worthy and high quality

[halcyon yarrow]

i agree, feelngs won't stop progress basically lol

and progress won't dictate communty sentiment

[craggy crest]

yup

@bitter hearth

[halcyon yarrow]

looks like something clown would make, good stuff

[craggy crest]

skip layer tests for 3.5 medium. workflow is in each image if you want to play around with it

[halcyon yarrow]

ive seen that in a few youtube videos and a few like "how tos" in civitai it's liike the trendy new thing to try, just picking at the model's layer to see how output changes given the same input, interesting stuff indeed

[craggy crest]

SLG - skip layer guidance - was released with 3.5 as a way of tuning the image and fixing some of the issues it might have.

[mortal mesa]

sooo this is pretty wild [MOVIE-SHOTS] In an enchanting tale of nature's wonders, [SCENE-1] shows <Sophie> observing butterflies in a sunlit meadow, her expression one of awe and delight, [SCENE-2] transitioning to <Sophie> sketching the butterflies in her notebook, her brow furrowed in concentration, [SCENE-3] wrapping up with her lying back in the grass, gazing at the sky with a contented smile, surrounded by nature's beauty.

[craggy crest]

you do need to make sure you're on the latest level of comfyUI though or you won't have the new node and the new scheduler

[mortal mesa]

https://huggingface.co/ali-vilab/In-Context-LoRA

[bitter hearth]

I don't think the community splitting is neccesarily bad

back with SD 1.5 and SDXL it was more needed to rely on this big ecosystem of fine tunes

but the models come out of the box stronger now, and loras can be done in 2 minutes on Fal

[bitter hearth]

really cool paper, didn't realise Flux could be trained as an edit model so fast

[halcyon yarrow]

just saying imagine if SDXL 1.5 came out or SDXL 2.0, I don't think we can compare SD 1.5 to SDXL since it's a complete archicture change, I'm talking about a company doing an incremental model release, it would suck to have to keep around 2 loras that amount to the same thing for 2 similar models becase they're not cross compatible

that's why i'm glad flux 1.1 was liimited to just pro so it doesn't fragment the community pool

[bitter hearth]

I don't think its that big a deal anymore

people can mostly just use the base models now

and recreate loras themselves if needed as it can be done so fast

[halcyon yarrow]

I guess storage is cheap enough where it's not that big of a deal but still sucks in other ways like having to keep 2 sets of vaes for each release. i just wish when flux does the next release it's a major release with significant changes that make sense to split the community for, i for once am not okay with incremental updates to major models

for SD3 vs SD3.5 it makes sense, SD3 was dead anyways and SD3.5 kicked it back to life, I'd consider 3.5 a worthy incremental change where they just messed with the training data and made everyone happy lol

[mortal mesa]

im trying them its unexpected to me

The four-panel image showcases a playful bubble font in a vibrant pop-art style. [TOP-LEFT] displays "Pop Candy" in bright pink with a polka dot background; [TOP-RIGHT] shows "Sweet Treat" in purple, surrounded by candy illustrations; [BOTTOM-LEFT] has "Yum!" in a mix of bright colors; [BOTTOM-RIGHT] shows "Delicious" against a striped background, perfect for fun, kid-friendly products.

[bitter hearth]

I wouldn't store stuff necessarily just download fresh when you need it

[halcyon yarrow]

lol i'm like the hoarders of lora my collectioni is at 1.7TB 👼

[bitter hearth]

I get what you are saying, it would be convenient for there to just be one big model with everything on it

1.7TB wow

okay yeah your perspective makes a lot more sense then

[halcyon yarrow]

can you try the film one? I'd try it buy my GPU is busy rendering a video

I'm going to try this prompt as-is with that film-storyboard one

[MOVIE-SHOTS] In a vibrant festival, [SCENE-1] we find <Leo>, a shy boy, standing at the edge of a bustling carnival, eyes wide with awe at the colorful rides and laughter, [SCENE-2] transitioning to him reluctantly trying a daring game, his friends cheering him on, [SCENE-3] culminating in a triumphant moment as he wins a giant stuffed bear, his face beaming with pride as he holds it up for all to see.

CivitAI would have a field day if someone reposts those LORAs on there

this is exactly what CivitAI loves the 4-panel story board, film board, side by side concepts and it's not available there yet

[mortal mesa]

the first one was film lora

[halcyon yarrow]

oh i see thanks for the heads up, you did a good job maintaining the same syntax they asked for but i feel Sophie iisn't consistent in every frame

[mortal mesa]

portrait-illustration.safetensors This two-panel image presents a transformation from a realistic portrait to a playful illustration, capturing both detail and artistic flair; [LEFT] the photograph shows a woman standing in a bustling marketplace, wearing a wide-brimmed hat, a flowing bohemian dress, and a leather crossbody bag; [RIGHT] the illustration panel exaggerates her accessories and features, with the bohemian dress depicted in vibrant patterns and bold colors, while the background is simplified into abstract market stalls, giving the scene an animated and lively feel.

[halcyon yarrow]

crossbody bag you say? oh you must mean crosseyed hag lol

85 frames seems to be my max @ 24 minutes for essentially 5.6 seconds at 15fps

[untold valley]

finetunes taking forever fr, way too slow. lol/s but the higher we go in params the slower things will go.

need that nsfw fix from time to time....

[untold valley]

I think im happy w faces.

[muted dove]

https://tenor.com/view/indiana-jones-indiana-balls-sisif-sisifus-rolling-gif-17011737

[dusky thistle]

all SD3.5M

my mona lisa

[pseudo owl]

This is pretty amazing for quantization of models like flux/sd3.5. It uses less vram then bnb4bit while being similar quality to 8bit. It’s also far faster then bnb4bit. Works well on smaller models like pixart too.

https://github.com/mit-han-lab/nunchaku

You can also apply Lora’s on it while you can’t on bnb4bit

[turbid grotto]

can it already be implemented in comfy?

[craggy crest]

that's why you create LoRAs instead of full finetuned checkpoints

[craggy crest]

gravity fails

[untold valley]

and fight the actual model with concepts it does not inherently knows a semblance of, and have to have stacks on stacks of loras, nah. There is a need for better finetunes, like all Sai models that have been released to the public have mostly been designed for. Training Lora's for hundreds of characters and styles and everything in between vs training 1 model, seems theres a clear winner.

[craggy crest]

the reason for a LoRA -low rank adaptation - is so that you can update the model with new or revised information and not spend hours or weeks, and hundred or thounds of dollars doing a retrain on the entire model

a full fine tuned check point is going to be vastly more expensive and take a whole lot longer time

[untold valley]

loras are great fill a void/niche/gaps but its nicer when the model inherently understands fundamentals. SD3.5 is a superb base. a finetune would be wonderful to train other loras on.

And then SDXL went on to be wildly successful. Im hoping the same for 3.5

[bitter hearth]

I feel like the quote is kinda backwards

for a base model is good to be very underfit

so that fine tunes have more room to change the model

[untold valley]

thats what quote saying.

[bitter hearth]

is it?

I read it as saying the opposite

[limpid thunder]

Last 7 days <Nov 01 2024> → <Nov 07 2024>

• Member counts
• 345955 ↗ 345969 ↘ 345955 ↘ 345930 ↗ 345948 ↗ 345967 ↗ 345976
• Action members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 85
• Message members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 58
• Reaction members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 43
  More details

[untold valley]

saying you need good base model, not too over trained so community finetunes.

[dusky thistle]

Yeah I think it's just phrased a little oddly

[bitter hearth]

maybe, not sure
I never really like trying to work out what past ambiguous quotes are saying

anyway if he means it should be underfit I agree

[dusky thistle]

Yeah I think that's what he means

That it won't look great cuz it's designed to be a little undercut

Underfit

[bitter hearth]

Flux was somewhat helpful in that it taught the population that if the model is overfit its hard to get rid of the Flux Chin

[untold valley]

lololol just yesterday I saw someone getting mad on the butt chin from Flux, ROFL

[bitter hearth]

lol yeah

some of the realism loras are good though

sadly I didn't save the workflow but I found one that works much better than the others

this same debate about fit happened in LLM world when Phi models came

cos Phi models are underfit so they are bad on release

but for example Omnigen is based on Phi, they are really useful

[untold valley]

time to make your own lol, sometimes i feel its more fun messing with settings than actually gening stuff

[bitter hearth]

oh yeah this was my own workflow

I spent $10 on L40s making it

then forgot to save it lol

all I have is this screenshot

[untold valley]

should be enought info there to remake it no?

[bitter hearth]

almost yeah
sadly the nodes to put the settings up top were wired up a bit wrong so it is missing some
but most are there

I feel very lucky I put that label up top and the graphs on the right

I would have nothing otherwise

[hallow lion]

Hey Fartoo.

[bitter hearth]

yeah 🙂

base Flux can't do him well but with realism lora Flux can

[untold valley]

have you tried 3.5 see how it does with him?

[hallow lion]

r2D2 frollic in grass

https://tenor.com/view/padme-anakin-smile-love-gif-17180376

[bitter hearth]

not yet I think I made galaxy bottle with 3.5

[hallow lion]

[hallow lion]

Usual Hogwarts report.

[hallow lion]

Clownsharks images are like a near death experience

couple of days ago I dreamt of nuclear blasts.

https://tenor.com/view/fifth-element-milla-jovovich-leeloo-bada-boom-boom-gif-9053853716468720447

[craggy crest]

yes, well, you said "finetunes taking forever fr, way too slow." and my response was telling you why. they are going to take a long time, because you're trying to update the entire model, not just creating a small set of information that'll be used with the model at runtime. even with highpowered gpus it's going to take quite a while to do this

[fleet meteor]

Nice gen!

[untold valley]

I ordered a pizza and all i got was the pepperoni.

[runic tusk]

I ordered a frappuccino, where's my fuckin' frappuccino.

[craggy crest]

should be some cheese in your fridge

[noble coyote]

Here's my photo! Sign me up!!! 😄

SD3.5 L Turbo

[mortal mesa]

The CogVideoX1.5-5B series model supports 10-second videos and higher resolutions. The CogVideoX1.5-5B-I2V variant supports any resolution for video generation.

[untold valley]

@lunar canopy sry ping but dont know if server has mods, this is a scam.

Actually am not sorry ping

[mortal mesa]

you come here from the SAI webpage then get fished

like this

[untold valley]

Oh that guy spam posting everywhere

Where the mods

[mortal mesa]

abandoned ship

[hallow lion]

Titanic is sinking.

[cunning lintel]

This discord is so wholesome it needs no mods 🤣
Sucks btw, this used to be a nice place, now i still hope to find some tidbits of interesting talk/images here, but quite often i think "ignore, ignore, don't get dragged in"

[hallow lion]

lol

[craggy crest]

with SLG - skipping specific layers to adjust the details, and without slg

[craggy crest]

just ping @lunar canopy if there are issues you need help with

[bitter hearth]

AI image generation communities are very toxic compared to literally any other area of machine learning

[craggy crest]

fulll of emotionally charged creative people - which is usually a recipe for toxic

[bitter hearth]

yes plus the NSFW community

I never see drama in the more dry areas of ML like communities for graph neural networks or time series analysis models

its always in image or chatbot communities

[mortal mesa]

they need help not me

[sterile pendant]

Yeah because they are a hotbed for people on the spectrum. Lots of binary logic/rigid thinking.

[craggy crest]

scientsts tend to only get emotional if some other scientist argues with their research

and then they get very emotional

[mortal mesa]

that is %1000 wrong

[craggy crest]

you're not a scientist so how would you know

[mortal mesa]

i have a brain

[craggy crest]

you should know better than to give people openings

[mortal mesa]

so you should cry when you do bad science to get people to leave you alone right

[craggy crest]

you are, apparently, in need of coffee this morning

[bitter hearth]

I forgot how your feud with crystal started
was it over the SD3 medium release?

[mortal mesa]

you said one of the craziest things ive heard involving science

[mortal mesa]

no feud, they say crazy ass stuff, wrong often, opinionated to the point of attacking people with diffent opinions, literally the only person ive seen warned, once of the few ive seen attack people. Its like the passive agressive help bot you never wanted. a social media "reputation farmer" it seems, shit is weird, most blocked person

society is wrong i suppose

[craggy crest]

maybe, but your normally zinging rebuttals are lacking this morning

[mortal mesa]

you broke sd3 by demending it came out as a huge member of the community

the clips you have to talk to the clips

[craggy crest]

and he dives down a radom rabbit hole

[mortal mesa]

nah i just get annoyed

i dont expect anything of anyone here

[craggy crest]

how did you go from talking about scientsts to SD3?

[mortal mesa]

you could read, its right here

[craggy crest]

you went right from talking about rabid scientists yelling at each other directly to SD3 being broken, without even a transistion

[mortal mesa]

i dont hate you i just often dont like you

[craggy crest]

i don't hate you or dislike you, but this morning you are really confusing

[mortal mesa]

maybe put it into a LLM to explain

[hallow lion]

🍿 popcorn?

[craggy crest]

naw, i'm just gonna send you some coffee

[mortal mesa]

ty

[hallow lion]

maybe wizard is just misunderstood...

[craggy crest]

i thought you blocked me, mr. coin

[hallow lion]

im willing to give a second chance to everyone

and its miss

[bitter hearth]

I find it hard to criticise people for being wrong when a lot of the advice I gave about image models turned out to be wrong as I learned more about them

there's a lot of stuff I said earlier in the year which I now think is cringe because I understand the math better

[hallow lion]

Right, we're all friends here!

Models come and go

[mortal mesa]

you all ruined sd3

[hallow lion]

[untold valley]

Hello

Did I miss participating in the drama 🎭

[craggy crest]

no one's stopping you from using somethign other than sd3

[hallow lion]

nah

[craggy crest]

wasn't much. everyone's asleep

[mortal mesa]

no one asked

[hallow lion]

very mild popcorn moment

[untold valley]

Wait why are Kagi and Crystal arguing? Don’t y’all share same opinions?

[craggy crest]

we're not?

jsut banter

[mortal mesa]

ya over crying scientists lol

[craggy crest]

i actually respect Kagi's opinion quite a bit. but we do lock horns at times.

[mortal mesa]

im a mere mortal

[bitter hearth]

fairly sure it started over the SD3 release

[craggy crest]

eh, we've been tradeing barbs longer than that i think

[bitter hearth]

ah okay

I find it hard to keep track of the drama

[craggy crest]

you can't complain of being bored here ;)

[untold valley]

Idk but crystal likes gaslighting the community over sd3 failure and it seems kagi also thinks it
Good thing 3.5 was success imo mostly cuz they removed some training restrictions.

[craggy crest]

no, i do not. i don't gaslight, ever, at all.

[untold valley]

[craggy crest]

you're more than welcome to go back through several months of posts and read what the commuinty was saying between march of this year and when SD3 releaed - and read the extreme toxic whining that was constantly going on

they got what they asked for and demanded

[untold valley]

Ok we can go back to posting gens.

[craggy crest]

sure.

[untold valley]

[hallow lion]

[untold valley]

@craggy crest figured out to make texture better. Add node to add grain to hide it

[craggy crest]

that looks really good

[lunar canopy]

nicest fall gens get to be server banner

or...er winter?

[hallow lion]

Ye it does, still amazes me how we can get stuff like this by typing words.

[craggy crest]

i had to read that twice. first time i thought it said nicest fall gens get banned

[lunar canopy]

that'd be fun

[untold valley]

Winter so like igloos and penguins

[mortal mesa]

i dont lora much but i saw a 2gb flux and 900mb sd3.5 lora, is this a new normal

[hallow lion]

lol 2gb flux?

[mortal mesa]

ya its kinda nice too

[untold valley]

Good lord

[bitter hearth]

I like adding grain yeah

[mortal mesa]

https://civitai.com/models/796382/ultrarealistic-lora-project

[bitter hearth]

there's people doing Lycoris Lokrs rather than Loras for Flux also

there's some on Civit if you want to try them

[lunar canopy]

banned....

[untold valley]

Ty

[bitter hearth]

it happens quite often that someone comes and puts their prompt in public like its midjourney, and the prompt is either super-NSFW or crazy

[hallow lion]

Good riddance

[bitter hearth]

midjourney requiring discord probably cost them hundreds of millions of dollars
most bizarre business error

they kept saying it is too much work to make a backend which makes no sense

it would have cost them way less than the cost of training the model

[craggy crest]

why would it have cost them anything?

[bitter hearth]

what I was thinking was server hosting cost plus the wages of a few developers

[craggy crest]

you don't pay discord a cent to use their servers. what server hosting?

[bitter hearth]

I mean if Midjourney had made an external website on day one

it would have been better, but they would have had to pay server costs and dev costs

but I think it would have paid off massively because so many people avoid midjourney because you had to use the discord

[craggy crest]

david wasn't trying to make money, still isnt'

[bitter hearth]

oh I see, is it one of those companies that is aiming for goals other than money?

wasn't aware of that

[craggy crest]

he's a researcher. midjourney is still a research project

[bitter hearth]

ah okay yeah that makes sense

[noble coyote]

I left you some pictures 🙂

[halcyon yarrow]

is this generated liike that, stiched together or using that new lora Kagi showed off yesterday?

@mortal mesa thanks for showing off that in-context lora project that thing is really cool man

[craggy crest]

generated. one shot. sd3.5

[halcyon yarrow]

really cool, did you speciify for 5 frames or did it just choose 5 on it's own?

[craggy crest]

prompt was: happyness: colorful autumn trees by artist "Shaun Tan", by artist "Mab Graves", by artist "Rien Poortvliet"

[halcyon yarrow]

ok that answers a lot

[craggy crest]

i didn't, but i used several names, and stable likes to create sections when you do that

[halcyon yarrow]

were you messing with the clip skip layers thing when you made it?

[mortal mesa]

ya its cool stuff, the loras are attempting more consistency

[craggy crest]

nope. no slg turned on. that's medium anyway and this is sd3.5 large

[halcyon yarrow]

there's also some effects ones in there, all good stuff

@mortal mesa I had a feeling the CivitAI community was going to love those loras so I reposted it there and iit's already gotten almost 100 downloads in less than 5 hours https://civitai.com/models/929592/creative-effects-and-design-lora-pack-in-context-lora

[craggy crest]

@bitter hearth no SLG vrs skiping layers 8,10,18,23,22 at scale 1

[bitter hearth]

wow it makes it brighter

I thought it would only help structure but it goes beyond that

[mortal mesa]

wild

[untold valley]

It afecta everything it’s directly messing with how the model creates images outa noise

[bitter hearth]

yeah makes sense

[craggy crest]

vastly. it does a lot more than just adjust struction.

[bitter hearth]

this means you've gotta start liking negatives though, cos this works via the negative 😂

[craggy crest]

you have no idea just how right you are

but with this you can be exact. with negatives, it's hard to be this exact

not only do you skip layers, you adjust how much, and you have other values. you can pin point

[bitter hearth]

PAG and SAG work on the negative too
SAG blurs the subject in the negative and PAG scrambles it

[craggy crest]

this isn't using negative prompts

[bitter hearth]

yeah I mean it uses the negative prediction

SLG drops layers when making the negative prediction but keeps them when making the positive

dropping the layers makes the image worse, which is okay as CFG pushes us away from the negative

its very weird

[craggy crest]

um - sort of but not quite

[bitter hearth]

ah maybe I misunderstood it, haven't looked in detail at the code yet

[craggy crest]

I can't say any more than that. at some point, development might post something.

[bitter hearth]

maybe they could post
... an Arxiv paper 😄

[craggy crest]

you know - as many things as have shown up on arXiv in the last 2 years that were junk, i think it's better everyone move their papers to huggingface

[bitter hearth]

there are some bad arxiv papers out there yeah

[craggy crest]

there are more than 'some'

huggingface is really doing a nice job of currating papers now

(we're back to carrying on a single conversion in mutiple channels ;) )

[bitter hearth]

AI moves so fast that

it might be too fast for curation

[craggy crest]

huggingface doesn't seem to ahve any issues

[bitter hearth]

not sure

[mortal mesa]

speaking of papers https://hanlab.mit.edu/projects/svdquant

not SVD as in SAI SVD

[pseudo owl]

Yeah that quant is pretty amazing, similar quality to 8bit while being much faster then bnb4bit and using less vram. It can also work with loras, no need to requant(unlike bnb4bit)

They have a space to compare,
Flux.1 Schnell bf16 vs Flux.1 Schnell SVD quant
prompt: A man holding a sign that says “Is 4bit quant better or the full bf16 model?”

[bitter hearth]

 We will replace unsafe prompts with a default prompt: "A peaceful world."```LMAO

[mortal mesa]

the demo loads the Gemma-2B model as a safety checker by default. To disable this feature, use --no-safety-checker

[bitter hearth]

bare in mind the biggest speedups are gonna be with the FP4 version not the Int4 one

and the FP4 version doesn't look as good as the Int4 one

especially with RTX 5090 which will have native FP4 matmul acceleration

[noble coyote]

Wheeeeeeeeee https://x.com/StabilityAI/status/1854989620154712474

Stability AI (@StabilityAI) on X

Straight from the Stable Diffusion Discord, @JackTorcello1 captures a raw, surrealistic creation using SD3.5 Large Turbo ⚡(1/3)

Twitter · Nov 8, 2024, 8:49 PM

[craggy crest]

YAY!

[noble coyote]

[craggy crest]

kitty!

[untold valley]

GGs

[halcyon yarrow]

@short thicket have you been working on any new models? i'm waitinig for the booru-free release

[halcyon yarrow]

cool share, thanks for the link i just went through it but a few points

• where can I download the svd quant version of flux dev?
• I'm thinking it only runs through scripts right now there isn't a comfyui node to load these models?
• i can't find the script or program that will allow me to take any safetensors or gguf file and quantsize it using their technique

[craggy crest]

without SLG

vrs With SLG

[mortal mesa]

links to code and files they offer are here, no for comfy as far as i know https://huggingface.co/mit-han-lab/svdquant-models

[halcyon yarrow]

yeah i looked thhrogh the code, i didn't see what script or command would take an existing and apply their quantzation to it, i realize its the deepcompressor library but it's unclear how to actually use it

see I think specifically the process to quantsize a model is this 3 step thing

https://github.com/mit-han-lab/deepcompressor/tree/main/examples/diffusion

where you can omit the benchmark portion on step 3 if needed but it's like at what point during any of those steps do you input the safetensors file? doesn't make sense to me

[halcyon yarrow]

@mortal mesa using that in-context lora lol

[mortal mesa]

ide buy it lol

[halcyon yarrow]

yeah the concept of some paper chips in a plain brown paper bag kinda intrigues me too

STOIQ did a terrible job with the same prompt horrible text horrible design, I'd avoid those chips bc the bag looks so flat and like there's nothing in there

I noticed for the couple-generation lora it doesn't really seem to abide by rules set forth https://civitai.com/models/929592?modelVersionId=1040555 @mortal mesa check out how like more than half of the images I made it didn't make a clean split it just ignored the prompt and put them together

[untold valley]

giving large more a chance but damn its like 6x time longer gen time, but some results so different.

Medium Left, Large Right

damn

[amber crow]

good

[short thicket]

Yes, but I'm in the very early stages. I just started building an image dataset for fine tuning. I took some time off to touch grass after the time it took to prepare Magic, Matrix and V1.

[halcyon yarrow]

took tiime off to touch grass ini other words to smoke some weed? lol

[short thicket]

https://tenor.com/view/wink-eye-wink-gif-12352742718877841322

I needed to take some time to "brain storm"

my goal is to build a small dataset of about 10,000 images to start.

[halcyon yarrow]

while you collect, caption and otherwisie prep that dataset wouldn't iit be easier to just remove fluuxbooru and leave that cooking in the meantime? so then you can compare that with and w/o your dataset?

[short thicket]

I'm gonna keep fluxbooru in.

[halcyon yarrow]

it's your call but it doesnt hurt to do A/B testing w and w/o it just to see how it's affecting image generations, currently not happy with how it's performing even at 61 steps check it out.... damn nvm i can't show you, I instinctively delete bad generations, needless to stay it's very sloppy and incoherent even with 61 steps and 3.5 cfg on what I'd consider a super complex prompt, meanwhile flux dev destill manages to do remarkably better

[short thicket]

It will get better in time. I'm basically just doing this in my free time as a hobby. There is always room to improve things. But after putting out 3 models in 2 weeks, I'm gonna take some time to chill. It would be just as easy for you to get the models and merge them how you want.

[craggy crest]

just a thought, but a lot of people like those booru tags...

[halcyon yarrow]

makes sense, has to stay fun if it's as a hobby, dont wanna burn out on too much 'fun' lol anyways yeah youre right it's just merging not training, just never messed with that field of stuff

[halcyon yarrow]

the booru tags are cool my disdain is the trainnig methodology the model creator used ended up making things worse not better imo, i think the training dataset was overly ambiitious and he didn't throw enough compute at it so it feels half baked but that's just my opinion from using it, maybe i'm using the model witih the wrong settings still

[craggy crest]

that might very well be the case. a lot of people don't like waiting, they can't stand the idea that something like training a model might take weeks, and want it done in 30 minutes or less

[short thicket]

its super simple. here's a very basic lora merging workflow.

[halcyon yarrow]

well to be fair he threw a bunch of H100s at it but I think his dataset was larger than his compute, the ratio very well likely could've been off, and one more thing the creator specifically mentioned in one of my comment threads it was not trained with booru tags, in other words all the images processed used natural language VLM captions rather than the booru tags they were found to be part of

[short thicket]

yeah he did mention that it was undertrained. That was the entire point of leaving it in. I want to continue fine tuning on it.

[halcyon yarrow]

so you give it based dev, the fine tune and the lora and it generates the merged model, good to know

[craggy crest]

not sure i'd just go with 50 on all the blocks...

[short thicket]

yeah. you can add multple loras in at the same time as well. Or add a model that isn't the base dev

[halcyon yarrow]

its one thing to be undertrained, but he didn't even train it with the booru tags, dude got a lot of grief from that in the comments throughout hugging face and civitai about it

using the in-context lora suggested by @mortal mesa using the STOIQ model on this one

[short thicket]

I said it was very basic lol.

[halcyon yarrow]

lol yikes that thing looks nitense

[craggy crest]

will be interesting to see how it comes out

[craggy crest]

never seen the inside of model before

[short thicket]

what's intense is having to put in the percentages for each block 1 by 1

[craggy crest]

i still might not just set the same weight on every block, however

[halcyon yarrow]

yeah i could't stand that, i would have chatgpt fix that for me

[craggy crest]

chatGPT would make a mess

[halcyon yarrow]

just give it the code and ask it to make you an input at the top that affects all the other ones

[halcyon yarrow]

100% disagree you just gotta know how to use it

[craggy crest]

not even. ChatGPT has no idea what it's doing