🆕｜sd3

[sacred jewel]

SAI should just release 8B and hope for redemption...

[craggy crest]

1. it's not ready 2. sai doesn't have any need to get into a battle with black forest.

[sacred jewel]

https://tenor.com/view/leonardo-dicaprio-shrug-shrugging-oh-well-smiling-gif-27275540

[rain current]

#🆕｜sd3 message

[craggy crest]

do you honestly think that if it was ready it wouldn't already be released? the community here and their toxic comments are what forced sai into relesaing the unfinished version of SD3 2b - are you trying to make them do that with 8b?

[sacred jewel]

https://tenor.com/view/leonardo-dicaprio-shrug-shrugging-oh-well-smiling-gif-27275540

[silver sluice]

I don't get it, im using the fp8 model, the fp8 dtype, and Im getting this error, is this related to out of memory or is something else wrong?

[sacred jewel]

You pick the right VAE?

[craggy crest]

how much vram do you have?

[torn wharf]

fwiw i use the 16bit weights and setting the fp8_e4 option works still

[dull star]

im using 16bit weights too and fp8_e5

[silver sluice]

yeah i think you're onto something

[sullen moss]

I'm falling in love...

[torn wharf]

kind of love that her hair looks almost like a chemo patient's

[sage burrow]

Flux has its own specific vae as well

[torn wharf]

overwatch?

[silver sluice]

i was getting frustrated bc it wasn't finding vae.sft (bc i wasn't updated to the latest version) so i tried renaming it to ae.safetensors lol anyways after i updated and reset the flow its worked fine

[sacred jewel]

Nah...

Generate an image of a cybernetic mystic in an ancient corridor. The figure has a robotic face with a single glowing red eye and is wearing a hooded cloak with golden ornate patterns. The mystic's armor combines metal and glowing runes, suggesting a blend of magic and technology. Focus on high detail and contrast, with a balance between the fantastical elements and the grounded, realistic textures. The guidance scale should be set high to maintain the intricate details of each element.

[torn wharf]

has anyone loaded the sd3 vae?

[torn wharf]

sweet prompt. llm?

[sage burrow]

Doesn't work, I tried

[torn wharf]

the final sentence was a clue

[torn wharf]

good effort lol. was just curiouus

[sacred jewel]

Not sure... stole it from some website a while back... it was in my prompts repo text file 😄

[gusty gale]

Write this song with cool text on a background that fits the words:

Giant steps are what you take
walking on the moon.
I hope my leg don't break
Walking on the moon.

flux's insane

[sage burrow]

I tried all options lol

[silver sluice]

wow

[torn wharf]

t5 is fun to prompt cause if you throw in irrelevant shit, it has self attention so it can just roll with it or ignore it

[torn wharf]

those are hot lyrics

[silver sluice]

• still getting the error
• workflow points to the vae decode
• i tried redownloading the file once again just to be sure
• i'm using the ae.sft found in this link: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
• i verified it's in the right directory

what could i be doing wrong?

[craggy crest]

per the post that comfy made on github on the issue, you have to update pytorch, did you do that?

[mortal mesa]

i havent used it but any input on the Flux guidance setting

[gusty gale]

defaults to 3.5

[silver sluice]

lol damn how do i not think of these basic things? thank you man, really appreciate it

i ended up just doing a git pull on the repo and not reinstalling dependencies

[craggy crest]

welcome. also, don't rename to safetensors, update pytorch, update comfy with the update script in the comfyUI/update file (not with manager) and make sure your models are in the models/unet folder and the ae.sft file is in the vae folder

[silver sluice]

my dependency chain is very fragile i was trying to avoid nuking it lol

[craggy crest]

i had to update torch, then xformers, then all of the numpy files

[torn wharf]

from orbit. it's the only way to be sure

[silver sluice]

i had to actually resolve conflicts with the git pull for comfy bc i had changed 3 source files on the original comfy backend, I had just added some print statements along the way no actual logic changes

[craggy crest]

and other dependencies. here's the actual issue https://github.com/comfyanonymous/ComfyUI/issues/4179#issuecomment-2267091681

[silver sluice]

wow that's flux?

[torn wharf]

dev

[cunning lintel]

lower is sometimes better, it's not what cfg used to be, it's more of: model knows best from training (3.5 up) vs I really know what i do (1.5-2.5). Has more influence on style than prompt following

[silver sluice]

i think im more impressed how you have all the open comfyui issues indexed in your mind

[mortal mesa]

the answer was update comfyui if you read it lol

[silver sluice]

yeah pytorch2.4 here i come

[torn wharf]

Two men are standing very close together. Standing at a bus stop on the side walk. Their shoulders are touching they're so close. Above one man is a speech bubble that reads "Don't Stand So Close to Me" speaking towards the other man 20steps ddim

[craggy crest]

i opned it, i have updates to it in my email so i have the link

[silver sluice]

how cool would it be to have a browser extension for discord that would allow me to click on the images and in the popup modal, modify it so that it'll try reading the metadata/workflow information for the image to display the prompt and parameters used right under as a caption for the image

[craggy crest]

https://github.com/olaviinha/Discord-Image-Modal-Zoomer would this work?

[silver sluice]

that's just to control the zoom?

[torn wharf]

i just keep xnview open in a tab generally. also it relies on people making sure metadata is included. which since i hate modifying comfy noodles, i haven't done yet on my images. so i'm a problem too.

[silver sluice]

i'm talking about reading the exif data on the image itself and parse the UserComment or however it's stored and display the image gen stats, like now people are posting images and I wonder if they're using SD3 or Flux to generate it, Id love to be able to click into it and see how it was generated lol

[silver sluice]

im guessing xnview is like a website where you can paste the image into and itll do that for you?

[torn wharf]

no i just cehcekd. i got meta data on mine i think. i think that worked

[torn wharf]

free app for windows

[silver sluice]

ive noticed that sometimes if i copy from clipboard it wont copy the metadata but if i drag and drop the file itll always include the exif

[torn wharf]

don't copy the preview. open the full image

maybe.

[silver sluice]

yeah the preview never has the metadata

but even the full image doesnt have the metadata if you sourced it from copying the image to your clipboard rather than dropping the file

[torn wharf]

yeah when i copy and paste the image, it just pastes it as a node.

[sage burrow]

Mine have all been Flux via mage or huggingface today. I don't think those have exif data

[silver sluice]

anyways its a lot of hassle to manually load the full image, copy it into something and then just see, i could probably code that in like less than 2 hours maybe? I'll def tag you if i create it lol

[silver sluice]

yeah i can imagine, stableswarm, comfyui, a111 are the only formats i know store exif data

[torn wharf]

need a bridge extension on comfy. where the save load buttons are. click load from discord and it shows all the images from the channel you're in.

[silver sluice]

or how about a button to load that image as a workflow right next to the Open In a new window link?

[silver sluice]

so it'll say Open New Window | Load Workflow into ComfyUI

[torn wharf]

there are certainly good ux improvements to be made in this field yet. we're not in the ux era yet. this is all still raw programmer UI

[silver sluice]

that would probably be a lot easier to write too, way less work bc its just this simple button that does this thing

[torn wharf]

gets the job done and there have been a lot of ux work yet, but ... yeh

bridign in workflows and managing a library. we gotta get there yet

[silver sluice]

i alright comfy updated here we go...... if this works ill play with it for a bit and then see if i can make that quick browser hack

[acoustic kite]

there's also this one :) flux upscaled with sdxl lightning

[silver sluice]

if only you've seen what ive done lol ive created a crazy management library system for comfyui (thats private not shared)

[torn wharf]

contingency solutions are the best

[silver sluice]

i would code it in TamperMonkey https://chromewebstore.google.com/detail/tampermonkey/dhdgffkkebhmkfjojejmpbldmpobfkfo

[dry wave]

lol, right. The SAI dev team worked for months on SD3, it got never released, they quit, found a company, made an even bigger and better model while in the meantime SAI is still cooking their sd3 model. Do you really think they don't release sd3 because it's unfinished?? They just don't have interest to deliver the model for free to the community. They search for ways to make money

[torn wharf]

when asset managemetn solutions aren't provided, the things people come up with in lieu of it. mmmm. mm mmmso good.

I've seen a lot of personal organziation systems. people are awesome.

[silver sluice]

helmet is way too big right?

[sacred jewel]

Nah, I have a huge head 😛 😛 😛

[mortal mesa]

ComfyUI's UI is changing in 12 days, just so you wont be shocked

[torn wharf]

i've had a couple good images like this like car's have big heads lol.

[craggy crest]

you just keep telling yourself your assumptions and be happy with your misinformation. go back to your real bosses adn tell them your attempt to destroy SAI has failed again

[torn wharf]

#🆕｜sd3 message

[silver sluice]

as long as they keep v0.4 for the format of the metadata I'll be happy

[torn wharf]

i can't remember what i did with the other

[mortal mesa]

https://github.com/comfyanonymous/ComfyUI/issues/4169

[silver sluice]

its unclear if they're changing underlying formats

[bitter hearth]

I want this automated charger for my phone and stuff

[torn wharf]

just want to go on the record that i have always wholly disagreed with blaming the community for releasing 2b in it's state. i'm not sure of the logic there. yeah maybe people were anxious but.. i mean.. how are they to blame?

rhetorical. it's a dead horse

[mortal mesa]

its ego pumping not reality, the comments

[torn wharf]

logitech's forever mouse. the rolex of mice they say

[mortal mesa]

i dont belong to a "side" others do lol

[sullen moss]

Moarrr

[torn wharf]

thats actually really pretty

[bitter hearth]

IDK how meaningful the LLM that made images that came out is

[bitter hearth]

this thing https://github.com/Alpha-VLLM/Lumina-mGPT

but there are a few other similar projects

[acoustic kite]

3072x1536 image :)

flux of course

[bitter hearth]

it can do over 3000 pixels horizontal?

is this upscale or native?

[acoustic kite]

that's upscaled sorry

[dry wave]

native often looks blurry and low quality

[acoustic kite]

it stops looking good when getting close to 2048x2048 px imo when native

sdxl has major issues with how images look repetitive so using flux for that initial image and sdxl lightning works great

[mortal mesa]

thats that 90 MPH roof A/C

[torn marsh]

impressionist oil painting of a sick rose flower. The flower has dark spots due to disease
Guidance 3.0 and guidance 1.2

[sacred jewel]

Mrs. SporkHand

[bitter hearth]

[granite haven]

This is with SD3? What prompt did you use, if you don’t mind sharing? It seems too good for SD3

[sullen moss]

Flux

[bitter hearth]

this was my experience too

[acoustic kite]

here is a 6k image made with flux and upscaled with sdxl lightning

[torn marsh]

Guidance bellow 2.0 is just...

Impressive I'd say

[sage burrow]

Forget this either or stuff, just use ALL the models! They are free afterall

[torn marsh]

If it's flux could you share the prompt?

[acoustic kite]

thank you you're right, this is the picture above yours with guidance of 2.0 instead of the default 😄

[sacred jewel]

a portrait of an old coal miner in 19th century, beautiful painting with highly detailed face by greg rutkowski and magali villanueve

[torn marsh]

Guidance 1.6 and 2.5

[torn marsh]

So far more guidance = crispy image
Less guidance means more realistic photos (sort of). Also painting style

[icy drift]

This exactly how much menace a cat would exude while slicing tuna, if a cat could slice tuna.

[acoustic kite]

flux is crazy

[mortal mesa]

anyone play with lower steps on flux dev, its painfully slow on my rig, found an ok balance?

[errant dust]

What was the prompt?

[silver sluice]

almost done guys

I found the easiest way to go from discord to view the details (for now) notice how the corner now says Copy to clipboard

[pseudo owl]

flux schnell is better with low steps, try 6ish for good results and close to dev quality. I found 4 is alright but results are considerably lower then dev.

[silver sluice]

@acoustic kite i see you have a complex workflow for that image

[acoustic kite]

yeah 😄

[bitter hearth]

ksampler custom advanced is a nice way to do it

[brittle nexus]

Flux can do different color wings while SD3 fails

[acoustic kite]

setting the guidance scale gives you higher realism and lower gives you more stylized and abstract images, just tried it out

this is with a scale of 10

[bitter hearth]

wait did i have it wrong way

[silver sluice]

https://gist.github.com/dasilva333/b8a5590846515ec379dd0fcae5689b07

[acoustic kite]

maybe i am wrong, my thoughts are based on very few images 😄

[silver sluice]

@torn wharf if you want to try it, this is the best I could come up with in a short time, you get a new link in the modal, you click it, it looks at the full image's metadata, copies it to your clipboard and then you can just hit paste on comfyui to load it up

[torn wharf]

oh like paste the text of the image metadata. right.

[silver sluice]

it'll copy the workflow into your clipboard

so right now it only supports png and it only supports images generated in comfyui

i could make it support png, jpg, all sorts of different formats, and like recreate a workflow if it's missing but this is just like super simple implementation

[silver sluice]

this image is the only image generated in comfyui from all the recent images if you're curious, all the other ones I can't read the metadata for (with the current support available)

[acoustic kite]

yeah feel free to use my images to test on

most of what i upload is directly from comfyui

[acoustic kite]

this as well @silver sluice

should be atleast

another one

but be aware the prompt is quite boring

[silver sluice]

yeah this one works, i don't have realvis checkpoint tho

[acoustic kite]

best one i've found yet before flux came out lol

[sacred jewel]

'''

Create a 3D infographic-style image titled 'Infographic Odyssey: The 3D Future of Resumes.' The design should feature a futuristic CV Resume with multiple layers connected by lines to cut-out boxes containing icons for different CV sections. Use a vibrant and diverse color palette to accentuate each section. The textures should resemble brushed metal, glass, and high-grade plastic, with advanced lighting techniques providing reflections, highlights, and shadows to enhance the 3D effect. Incorporate infographic elements such as charts, graphs, and pictograms to represent skills, experiences, and personal attributes creatively. The overall design should be sleek, modern, and organized, emulating a high-tech interface, while focusing on visual storytelling and innovative presentation over traditional text.
'''

[acoustic kite]

prompt looks ai generated. does anyone know if it's worth manually creating a prompt?

[mortal mesa]

i AI prompt because im not a creative writer

[acoustic kite]

that's fair 😄 i was just curious if it works better than manually prompting

[gusty gale]

after about 20-ish steps it stops making changes to the image it seems

I do 25-30, because in that range it always reaches that point

[mortal mesa]

1920x128

[cunning lintel]

uf, seems in the distilled version a few biases are baked in. Notice this is for a male faery, first is flux dev (never got a male with it in 8 tries), 2nd flux pro

A panoramic view captures a dynamic scene featuring a seductive male faery in a steampunk-inspired cityscape, overrun by wild greenery. The setting showcases open sewers winding through the city, flanked by lush vines and vibrant foliage that envelop the metallic structures. His intricate glass dragonfly wings shimmer with iridescence, casting colorful reflections against the muted brass tones of the urban landscape. He wears a uniquely crafted corset dress made from cyborg circuitry, blending organic shapes with mechanical elements, emphasizing his lithe physique. The level of detail is profound, aiming for an 8k resolution that reveals every aspect of the scene, from the textural contrasts between the natural overgrowth and the steampunk architecture to the interplay of light reflecting off his glass wings. Realistic, dramatic lighting is achieved through advanced ray tracing techniques, giving depth and realism to every object and surface. The composition resembles a movie still reminiscent of Alan Lee's style, enriched with motion hints, such as steam gently rising from the sewers and the faery's wings caught mid-flutter, adding a dynamic quality. Shot as if on a 35mm camera, the image merges a classic film aesthetic with modern fantasy, creating an enchanting atmosphere.

[acoustic kite]

looks like a real image apart from the chair outside lol

[brittle nexus]

Got a male on my first try with dev. (two images, one male, one female)

[cunning lintel]

How?????

any special sampler or something?

I just use api, but it was totally completely impossible, might be some hidden filtering 😵‍💫

[cunning lintel]

cool, i'll have to set it up local. thanks for trying!

[mortal mesa]

nice curves

[brittle nexus]

That zipper is so non-sense that's cool

[brittle nexus]

Flux can do selective color more easily

Actually very good at selective color concept

[sullen moss]

How much is the fish?

[silver sluice]

i finally figured it out. turns out util.py wasn't updating despite having run git pull, super weird but this is my first flux image gen took a whole 407 seconds to generate at 1024x1024. i don't know if 6.7 minutes to make an image is worth it tho lol

[craggy crest]

i had to update pytorch, xformers, and a ton of other stuff by hand, as well as comfy. that came out beautiful

[silver sluice]

any tips you have to increase procesisng speed?

187 seconds at 512x512 not bad i guess

[hallow lion]

handporn

[silver sluice]

can i optimize it further if i change it form t5_xxl_fp16 to t5_xxl_fp8 maybe? lol

[craggy crest]

add more vram?

[silver sluice]

lol if only i could, im on a laptop

[craggy crest]

oh lord. i would never try to use that. just make sure you don't have anything else running on it at all i guess.

[hallow lion]

flux flex

this channel could be renamed "whatevers the latest best model this week"

[silver sluice]

lol pretty much, before flux it was pixart and auraflow

[silver sluice]

I'm going to try this text encode and benchmark it against the fp16 one to see if there's a performance difference https://huggingface.co/ckpt/stable-diffusion-3-medium/blob/3b998e720927a2f7bf10b7b17c540942d2214b7c/text_encoders/t5xxl_fp8_e4m3fn.safetensors

[craggy crest]

okay. let us know what your results are.

[silver sluice]

Setup

32GB RAM/8GB VRAM RTX 4070
latent image size 512x512 px

clip name/weight d_type
t5_xxl_fp8_e4m3fn/fp8_e4m3fn = 234, 235 seconds
t5_xxl_fp16/fp8_e4m3fn = 232, 261 seconds
t5_xxl_fp8_e4m3fn/default = 452 seconds

I tested it twice to be sure

[craggy crest]

not bad gen times on those

[silver sluice]

i appreciate how it's at least supported and runs unlike AuraFlow that requires a minimum of 16gb and won't run otherwise, I much rather have something that runs slow over something that doesn't work at all

[craggy crest]

agreed.

[brittle nexus]

The bottom of the food looks real, the top looks like a painting

[sterile pendant]

Visually, 8 vs 16 doesn't really make any difference in the long run. You could run an A/B test with thousands of samples and it would result in a 50:50 +/- a percent or two.

Just use the q8 versions, it's a waste of system resources if you don't have the room or ram/vram

[hallow lion]

Eraserhead cyberpunk.

[acoustic kite]

loading flux with fp16 takes up 56gb ram, how is this possible 🤣

[craggy crest]

you must have a lot of swap space

[acoustic kite]

i have 64gb, it took up like 62gb out of everything i had

[craggy crest]

well - fp16 is floating point to 16 places... what happens if you step down to fp8

[acoustic kite]

it becomes manageable to run without my pc freaking out, but i wanted to see if there was any difference in quality 😄

[craggy crest]

there is, yes, but if you're not going to be putting it on the big screen, or printing billboards, probably not all that noticible

[craggy crest]

from the discussion earlier: 1344x768 seems to give the best, sharpest, images

[acoustic kite]

interesting! thanks for telling me

i'm gonna upload 2 images, if you want you can try to guess which is fp16 and fp8 @craggy crest

[craggy crest]

the one on the left is fp16

[acoustic kite]

yeah

it's kinda crazy

i definitely see the difference

[craggy crest]

the differnces are subtle, but they are there. look at the veins of the leaf, the sun streaks, etc.

but do they matter enough for the exctra time and system resources?

[acoustic kite]

192 seconds vs 325 seconds 😄

i think it's worth it

[craggy crest]

go for it then. to me, it's not worth it unless i'm doing something to put on on an AMC theater screen

[acoustic kite]

haha that's fair, to be honest it's a lot of time spent on nothing if it's just for fun

[craggy crest]

sure. but in my opinion, i don't need to pay the electric bill that using that sort of system resources on a regular basis is going to cause.

[sacred jewel]

My new iPhone case

[brittle nexus]

Flow as Cascade knows "Kapow", not SD3

[brittle nexus]

Amazing kapows by the way

[hallow lion]

Our daily dose of ballz.

[sage burrow]

I hope the right one is 16?

[sage burrow]

I actually clickedo on them now, the left one seems to have more detail, but I like the composition os the right one better

[bitter hearth]

We need a hero

Randomly I want to test if the ai knows what a cybertruck is

Flux, bad model

[hallow lion]

This looks better than a cybertruck tho.

[sacred jewel]

A few ballz

[bitter hearth]

That car goes from like a cool 7/10 look to a -5/99 on some angles

[sacred jewel]

CFG 1.5

[sacred jewel]

CFG 50, 10 steps, DEV

[sterile pendant]

It doesn't really work that way. An analogy would be fp8 saying something like "a green leaf in a verdant forest" vs fp16 saying "a green leaf within a lush forest." Basically, it's just subtle differences in the outputs of the tencs like if you use a temperature of 0.4 with an llm and get slightly different responses every time. So you can almost think of it like a microseed

If you're talking about the t5 being fp8 or 16

[craggy crest]

snort. you want to go back and re-read what he said? and maybe see what I said?

i mean, actualy READ the text this time rather than assume and go on a technical rant

[sterile pendant]

Well last I remembered, we were talking about t5 precision, I didn't read much more after that. Just replying where my chat left off at before scrolling down more

[craggy crest]

okay well all he asking was which image was which.

and he was asking about fp16 and fp32 specifically

not the encoders

[sterile pendant]

Yeah he was the guy that I told to use fp8 t5, so I figured he was talking about that

[craggy crest]

he'd moved on to making two images and wanting to see if people could guess which was which

[sterile pendant]

Yeah, so based on the scope of my context window(I only have a 4096 token context window, with limited rope or sliding window options), I figured he was still talking about the same stuff.

But I'll add it to my RAG

For future reference

[craggy crest]

i think he decided to stick with fp8

[sage burrow]

ROFL

[torn wharf]

https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md lol could i spend money to train a flux balls? lololm maybe maybe

[craggy crest]

do it 🙂

[hallow lion]

Balls are only one step away from other round shaped things.

😁

[craggy crest]

like... pumpkins?

[hallow lion]

Flux. Now featuring small horses.

[hallow lion]

In a day or so every major AI youtuber did a sd is dead flux is king video.

Dang.

[errant dust]

How is that surprising?

[craggy crest]

in 48 hours the entire internet went from arguing to laser focused on one word

[errant dust]

Every time a new model comes out it is the same too. New model of SD/Flux/whatever kills the original.

[craggy crest]

yeah - not like this

with the right video model, this is the end of hollywood

[errant dust]

lol

[craggy crest]

not following the posts on twitter, are you?

[errant dust]

Let me put it this way: I keep having to reactivate my Twitter/X account because of lack of use

[craggy crest]

the actual movies that have been posted in the last month are almost better than what hollywood puts out, and what's been posted since flux released is better

[errant dust]

Well, Hollywood as a whole puts out a LOT of stuff, much of it tripe. But not all. and the good stuff floats to the top and the rest either disappears or is soon found on your favorite streaming service

you still have to write a good story produce a good film, act it, light it, film it, etc etc

[craggy crest]

yup. what's been posted since flux released is better

you don't have to 'act it, light it, film it, anything it'

[errant dust]

Huh, Feaurue films in two days

ok

lol

[craggy crest]

you generate it, animate it, run it through the lipsync programs that are now out there, including now talking animals with live portrait. and yeah, 5 to 10 minute film in a couple hours

you should get back on twitter and watch the flood

[errant dust]

I was never 'on' Twitter. I have an account

as to staring at social media all day. No, defintiely not

[hallow lion]

Yeah...

[hallow lion]

Good ones endure tho, some kinda ceiling will be reached eventually too - how good can these models get, consistency and later on video is the next big thing

until images are good but no consistency is achievable this is just going to be stock footage/product placement

hollywood won;t be threatened

[low stone]

it's not dead. they'll release new versions and it'll be great and people will finetune etc. then black forest will release open weights text to video. THEN sd will be dead. :p

[hallow lion]

looool

The new SORA killer.

Flux Kino!

they better call it flux kino

[low stone]

realistically if kling can do it, no reason why others here can't as well.

[hallow lion]

totally

[low stone]

but seriously, i was playing with sd3m again today. too much wrong with it to keep using it, but it's crazy fast compared to the other new stuff out there. if it fixes the issues I would definitely use it.

[hallow lion]

it does one thing well - backgrounds

i replace my backgrounds with sd3 generated ones and it always kills it

and yes the speed is good too

Emad got some things right... yeah yeah he aint around anymore well hes the sai CEO of my heart forever 😛

XD

[craggy crest]

https://x.com/recatm/status/1819348949972476019

XiQiao 西乔 (@recatm) on X

SD3 team has shared some generated results from the WIP training SD3.1 (not done training yet). comparing them with Robin's Flux 3.1 Pro using the same prompts. The 3.1 model is almost on par with Flux Pro in terms of quality, but it's nearly 5 times smaller, with only 2.5b

Twitter · Aug 2, 2024, 12:26 PM

[low stone]

well now he just said that the black forest guys (ex-SAI people) couldn't have done the same thing while at SAI. not sure what that implies.

[hallow lion]

probably cencorship BS

they were forced to tamper with the training material

[low stone]

true

[craggy crest]

he still owns the majority of the shares

[low stone]

fal.ai has a censorship slider on their site and api. amazing stuff, giving you the choice of what you want. it's the way it should be for all of these apis.

it puts the choice on the user.

and the liability

[craggy crest]

y'all should go read the post i just linked

[bitter hearth]

Are you telling me to willingly visit Twitter

what has this chat become

[craggy crest]

just to read that article

[bitter hearth]

He said 2 weeks

That's what I got from it

[dry wave]

the SAI twitter post is just stupid. Besides the SD 3.1 images look subjectively worse, it doesn't matter. I can make stunning photos of dogs with sdxl or even sd 1.5, too. What matters is that sd 3 is not able to make a human correctly, in particular if he holds something in his hands. As soon as the object you generate is a bit more complex sd3 generates garbage. So when they want to teaser their new sd 3.1 model they should showcase prompts that current pre-flux models cannot handle

[hallow lion]

Give us consistency Emad. That's what the masses crave.

That and electrolytes.

[noble coyote]

"I just did a Flux/SD3 merge - and now, NASA's Cold Fusion Laboratory just got in contact!!!" 🥳

(There's a joak there somewhere?!)

[uncut river]

merge, cold fusion, yes

[noble coyote]

Flux.Schnell

[noble coyote]

Flux has a small foible: it can make hands and feet look ginormous - anatomically correct - but oversize!

If you search at fineartamerica - you can pick up some fine prompt ideas! Flux.Schnell

[uncut river]

sd3 is also known for big hands, or females with male hands

[noble coyote]

Georgette Washington?

[uncut river]

I was asking for a Dutch flag, but sd3 got it wrong

[noble coyote]

Russian Flag and Dutch Flag are almost the same

[uncut river]

yeah, and they also stole some words from us, for ships and marine stuff

[noble coyote]

Tsar of Russia was the first cousin of Queen Victoria (UK).

He visited London one day, and he took interest in the railways.

[uncut river]

royals, it's almost all family. money and power, keep it in the family

[noble coyote]

He wanted to take them back to Russia

Victoria told him that if he went to Vauxhall - he could see as much of the railway as it was humanly possible to see.

When he went back home - and built the railway - Vaksal was the new Russian word for the Ticket Hall!

Flux.Schnell

[uncut river]

Is there a flag lora out there for sd3 ... ?

[noble coyote]

Looks like a flag from Puerto Rico

[uncut river]

im still trying the dutch flag, red-white-blue

maybe flux does better

[noble coyote]

https://civitai.com/models/314768/all-asia-flags

[uncut river]

flux-schnell

yes, it got the flag right. But the males have gone away ...

dutch masterpiece painting, historical gorgeous painting, hard contrast, beautiful realistic lighting, white female with 2 black people around her, tight victorian dress, (group of people holding the dutch flag:0.7), red-white-blue flag

modern style

[noble coyote]

Dutch Blue is sky blue - if you're going to use dark blue then you're inviting comparisons with France, USA and UK 😄

[uncut river]

the old painting has faded colors, it's old ...

[noble coyote]

Flux.Schnell - almost surreal prompt = bewitching beauty peruvian arpillera rococopunk cubism dada style of victo ngai henry rousseau Vladimir kush peregrine heathcoate catrin welz-stein Andrea kurti Andrea kowch remedios varo Oxana dobrovolska. A cat washing the dishes. A stylishly dressed woman in a large hat and high heels sits on a modern chair in a minimalistic, futuristic space with geometric designs. Beside her, an elegant handbag and a cocktail add to the sophisticated ambiance. "Simplicity is the ultimate sophistication" - Leonardo Da Vinci.

[uncut river]

you like surreal stuff?

[noble coyote]

Indeed I do 🥳

[uncut river]

maybe i should try

I tried photoreal, cartoon, anime, 3d renders and stuff

[noble coyote]

Flux is seriously good at artistic material

[uncut river]

any suggestion for colors and lighting in the prompt, to go with surreal?

[noble coyote]

Try fauvism for colours, or neon glow

Voronoi is great for geometric shapes; and don't forget fibonacci too

Zentangle gives a nice effect, as does cubism or dadaism

[uncut river]

This is what SD3 makes of it

[noble coyote]

Looks very Renoir-ish

Nice!!!

Use words like retro, futurist, retrofuturist, rococo, rococopunk, vintage etc

[uncut river]

Might be a bit overkill on the fractal stuff:

surreal open room landscape, 3d fractal object, fractal style, fauvism colours, neon glow, 8k, sharp focus, detailed, retro, futurist, retrofuturist, rococo, rococopunk, vintage

Flux-schnell

[noble coyote]

Jello

[uncut river]

sd3 with similar prompt

[noble coyote]

SD3 has a native 'wash' feel to its saturated colours - which is refreshing

[uncut river]

seems ai is going through the same thing as televisions did many years ago, going from subtle color to full-over-saturation! 😄

[noble coyote]

A cat washes the dishes

[strange grotto]

forge

he is updating

[noble coyote]

Good

[strange grotto]

maybe dev fp16 on 8g gpu i think

[noble coyote]

I have SD.Next, Fooocus, SwarmUI, ComfyUI and AUTOMATIC1111 - its great to pick a UI and then get stuck in!!!

I've never tried Forge - what are its outstanding features?

[strange grotto]

Gets you faster than comfyui on the webui interface

[noble coyote]

I'm running fp16 and dev on 8Gb RTX2070, 64Gb RAM - 20 iterations - 5 minutes/image

But Schnell at 4 iterations is 100 seconds/image

[strange grotto]

I'm guessing he must be sure to increase flux dev speed far beyond comfyui

He's an unusual character, but his work never disappoints

[noble coyote]

Dev is slower than Schnell; but I'm not sure that the quality of the dev output is worth the extra wait?

[strange grotto]

I think it's worth it.

If dev becomes mainstream in the future, then finetune or lora will also be centered around dev, and the first thing that needs to be solved now is to run it well.

[noble coyote]

Black Forest you mean?

[strange grotto]

[noble coyote]

ILLyasviel is Black Forest?

Eau keigh

[strange grotto]

He's not.

[noble coyote]

Oh?

Schnell

[dusky minnow]

A beautiful young woman with medium-length, shiny, healthy blonde hair styled professionally in a premium salon manner. She has symmetrical facial features with an oval face, high cheekbones, a neat nose, and balanced eyes and lips. Her hair is smooth, possibly with light waves or soft curls, meticulously arranged. Her skin is clear, even-toned, with a natural glow. The overall look is elegant and sophisticated, highlighting the quality of the hair styling.

[rain current]

Schnell is great and very fast with 4 steps, but it makes more mistakes and is less understandable with the prompt. Schnell:

[noble coyote]

Schnell - prompt via @dusky minnow = A beautiful young woman with medium-length, shiny, healthy blonde hair styled professionally in a premium salon manner. She has symmetrical facial features with an oval face, high cheekbones, a neat nose, and balanced eyes and lips. Her hair is smooth, possibly with light waves or soft curls, meticulously arranged. Her skin is clear, even-toned, with a natural glow. The overall look is elegant and sophisticated, highlighting the quality of the hair styling.

[uncut river]

with same low weight anime style added, SD3

[noble coyote]

Quality prompt!

Added a smidgen of neon

[hallow lion]

i rarely see subjects lit from below in these AIs

the lighting is always flat

[noble coyote]

Here is the prompt for you to try it yourself ... microscopic ::1 closeup ::1 backlit jewel tiffany ::3 intricate needlepoint closeup murmuration ::3 underwater lighting ::2

Neon-Infused Expressionism, Desert Mirage: Surreal desert landscape, Souza's artistic style, neon lights and ethereal glows, blending reality with fantasy, iridescent hues

A beautiful young woman with medium-length, shiny, healthy blonde hair styled professionally in a premium salon manner. She has symmetrical facial features with an oval face, high cheekbones, a neat nose, and balanced eyes and lips. Her hair is smooth, possibly with light waves or soft curls, meticulously arranged. Her skin is clear, even-toned, with a natural glow. The overall look is elegant and sophisticated, highlighting the quality of the hair styling.

[sullen moss]

What I also like about FLUX is the unique new faces. I'm tired of seeing the same "1girl" face everywhere.

[noble coyote]

Agreed, it is quite refreshing

[rain current]

Terribly good... (these are with Schnell)

[noble coyote]

Gorgeous

[uncut river]

flux-schnell is too smooth

real photo of a 21yo woman wearing a sweater, golden sunset, round nose, rectangular face, Armenian female, dark hair, face portrait, happy eyes, realistic grainy photo quality, image noise, jpeg artifacts

even with addition of "realistic grainy photo quality, image noise, jpeg artifacts" it still looks polished like a photoshopped marketing image

how to get that grainy real photo quality?

[sullen moss]

Yes, I agree with that. Perhaps we just haven't found the right tokens yet.

My super combination for DALL-E also doesn't work.

Very grainy direct flash photograph taken in 2019, 35mm iso 200 Cinestill 800T dust and scratches light leak slow shutter speed motion blur

[rain current]

You can do a second pass, or upscale

[frail shoal]

why don't you run fp8 dev. It's basically almost the same but faster

[acoustic kite]

let me try this prompt with flux dev fp16 and upscaled 4x with sdxl lightning

[uncut river]

wtf?! replicate.com finds this prompt NOT sfw :

Very grainy direct flash photograph taken in 2019, 35mm iso 200 Cinestill 800T dust and scratches style, 26yo lady wearing street outfit, pants, top, tight fit, long hair, city background

[sullen moss]

I think it's not bad, even the tiny hairs on the arms are visible.

[acoustic kite]

probably the "tight fit"

[uncut river]

ah, yes that could be it

some generated images using same prompt with sd3 seem pretty harmless

oh, and the grainy photo style works better with sd3 compared to flux. at least we found 1 thing that it's better at! 😄

[acoustic kite]

now it just needs to get anatomy right 😄

[rain current]

🤔

[uncut river]

those claws ...

[sullen moss]

I'm still amazed at the quality of the skin texture in the default generation.

[acoustic kite]

are you comparing flux schnell to sd3 8b?

[uncut river]

yes I think so, sd3MediumInclClips

[acoustic kite]

try flux dev vs that

[uncut river]

I see, flux dev is indeed a bit less smooth and polished (or burned)

but still not quite the grainy photo quality sd3 can do (more easily perhaps)

[sullen moss]

I think this can be fixed with a LoRA

[acoustic kite]

this was the best i could do 🤷‍♂️

it isn't really grainy

[hallow lion]

Especially with SD15 thats a problem 😄

[acoustic kite]

that is a DRAMATIC improvement in variety compared to anything else

i'm upscaling it now

upscaled 4x with sdxl lightning

another upscaler, upscaled 4x with sdxl lightning again

[sullen moss]

https://civitai.com/models/621563/fillluxspoopy?modelVersionId=694866

[rain current]

boring snapchat photo circa 2015, {prompt}

[proven pecan]

I got a lot of them, they're fun

[icy drift]

Hmm. Okay, but this is obviously hearthstone, you know? I wonder how I can get it to give me another design.
a trading card from a trading card game. in the top left corner is an image of a fire icon. at the center is the card title that says "KNIGHT". in the bottom left corner is an image of a shield with the number "4". in the bottom right corner is an image of a sword with the number "7". the card art in the center is a knight with a sword. below the card art is a box with the description text that says "The brave knight will not be defeated."

[noble coyote]

But it's only 152 bytes?

[rain current]

well, it has to generate images really fast

[noble coyote]

Is it for real?

[rain current]

https://huggingface.co/lividtm/FillLUX_Spoopy/tree/main

[sage burrow]

SD3

[noble coyote]

Was this done with Flux.Dev or Flux.Schnell at all?

[icy drift]

Flux.Dev Checking how higher guidance affects its ability to follow the prompt.

[noble coyote]

OK, the spelling is poor when I use Schnell

[sullen moss]

I don't understand what the catch is either.

[remote holly]

Flux schnell is too good for liminal spaces

[noble coyote]

Actually it is a link to a huggingface 21.8Gb d/load

[remote holly]

144p, vhs filter, low resolution, 2000s photo, jpg artifacts, analog image, shadows, suburban garden at night, house across the street, neighborhood, trees, starry brown night sky, darkness, night, low light

[noble coyote]

Schnell

[hallow lion]

He will not le defeated!

Maybe he is French.

[uncut river]

144p, vhs filter, low resolution, 2000s photo, jpg artifacts, analog image, shadows, factory at night, industrial area, neighborhood, trees, starry brown night sky, darkness, night, low light, empty industrial buildings

[sullen moss]

Time to grab a snack.

[noble coyote]

Dev not better 😦

[uncut river]

surreal painting of goldfish-sharp, (sunny, blue sky in the background, fractal style sharp clouds, contrasting styles, contrasting colors:1.18), historical 1920 noir style, (one green alien holding a blue box, blue box:1.1), (flying orange fish:1.28), (orange shark:1.05), (many big evil goldfish:1.1), (vicious orange shark, sharp teeth:1.2), (blue roses, field of blue roses, blue roses everywhere:0.9), (massive war, tanks, explosions, armored sharks:1.2)

sd3

[noble coyote]

@uncut river and @icy drift prompt mish-mash in Schnell

I love the tanks rolling over roses!!! 🥳

[sage burrow]

Flux pro via glif

[icy drift]

Very high guidance (30) hurts prompt following, it doesn't improve it.

[remote holly]

👀 amazing

[noble coyote]

Goldfish turns the Knight into an alien!!!

Schnell

[noble coyote]

Nothing bottom left nor bottom right

[icy drift]

It just makes up random text. Low guidance sticks to the text you give it.

[noble coyote]

Is Flux Pro based on the 12B Dev model?

[icy drift]

Other way around. Flux.Dev was distilled from the Pro model down to a lower parameter count.

[sage burrow]

I have nooo idea?

[noble coyote]

How many B is the Pro model at all?

[noble coyote]

I'm finding that Schnell is adequate and satisfactory for most of my output

[sage burrow]

It was really difficult previously to talk models into flying anthros

[turbid grotto]

lots of people are saing the just replaced clip

[noble coyote]

It is probably faster for me to setup and SDXL Base/Refiner to get the last drop of detail out of it. Dev would take a deal longer to do the same thing

[noble coyote]

Is that a recommend - or a fail - for this merge?

[sage burrow]

Feel free to share that wirkflow 😄

[icy drift]

Oh apparently they're all 12b parameters. The difference is how many steps they're trained to run at.
They're going to publish a technical report soon, so we'll have more details then.

[sage burrow]

No commercial use, is flux going to try to sue me?

[icy drift]

You can't host Flux.Dev on a server and sell people generation access. That's forbidden.
You can use the images generated for anything you want, including commercial.
You can also create derivative models all day long, as long as you're distributing them for free and not hosting them for image gen services.

[sage burrow]

I mean for selling my images?

Though they'd have to get in line after Pokemon

[icy drift]

I mean... You can try to sell AI gen images if you want. Eventually every device out there will have image-gen powers superior to Flux.Dev, so get that cash while you can.

[noble coyote]

Hope it works? - Flux-Schnell+SDXL

[noble coyote]

I've sold about £50 worth of AI Art!

[sage burrow]

With selling ai images, quality is mostly irrelevant (unfortunately)

[icy drift]

Found it. Also, I misremembered this. No training competition to Dev using its output. (But training anything that doesn't compete with it is fine? This is iffy.)

[sage burrow]

I've sold..... $1900 usd worth.
Don't bother, it doesn't pay min wage)

[noble coyote]

In 16 years I have sold mebbe $750 worth ...

Pays my subs on the sites I use 😄

[icy drift]

I've been doing art commissions since before AI, and I've used AI in my art a little since it became available, but it's just still not good enough. (Flux.Dev might be a game changer once it has controlnet though. We'll see.)

[noble coyote]

So merging does not count as training at all?

I just d/loaded FiiLUX_Spoopy

[icy drift]

Merging across models is nonsense IMO. You're just merging the CLIP, right? Black Forest didn't make these CLIP models. The license doesn't apply.

[noble coyote]

I'm gonna try FiiLUX_Spoopy ...

[sage burrow]

Now trad art, that at least pays min wage! (Or I make sure it does lol)

Combing trad and ai would be fun. I've used sike if ny trad art as reference images.... makes them look like masterpieces lol

[noble coyote]

Here's one which panders to my surreal taste (not Spoopy - yet!)

[icy drift]

I made a painting app that I use personally. Haven't updated the public repo in ages though. https://github.com/QuintessentialForms/ParrotLUX

[sage burrow]

You need to charge more for your trad art 🙂

[noble coyote]

Agreed 😄

[sage burrow]

What is spoopy? Cluffnires version

[noble coyote]

https://civitai.com/models/621563/fillluxspoopy?modelVersionId=694866

[sage burrow]

Ooh android w pen pressure:D

[turbid grotto]

it even seems to not be a merge, just replaced clip, I see no reason to redownload same 20gb

[noble coyote]

I forgot to ask - 4 iterations of Spoopy if Schnell-based? Or 20 iterations+ if Dev-based?

[noble coyote]

Just using it now - if it amuses me I will make space for the 20Gb 😄

[sage burrow]

how do merges even work? I did some merges with comfy and it basically just made them whatever the frist checkpoint I put in. It didn't really merge them. Or is comfy merge nota real merge?
I'd like to merge Flux and SDXL or 1.5

[noble coyote]

As I can now see - it'll be 20+ iterations as is a Dev merge

[turbid grotto]

You can't merge different architectures

[sage burrow]

But, but, but, I want a Flux/Redwater merge <cries>

[noble coyote]

OK, imho - Spoopy is no real increase on Schnell ...

If you drop the noise down below 0.7 then u get this greyed-out sketchy look

[turbid grotto]

I did check what redwater is

[sage burrow]

Flux doesn't do very good(it's terrible) nsfw lololol

[turbid grotto]

but better than other bases

[sage burrow]

Did all the other base models have no nsfw whatsoever as well? I wasn't around yet for the sdxl release, but I've never been able to even get sfw to work with base sdxl

[icy drift]

I got one image that did correct female anatomy down there yesterday, and one that did correct male anatomy when I was checking if it was censored. There was some in the dataset obviously, so fine-tuning shouldn't take too much resources.
All I really want / need is controlnet though. I don't have any NSFW commissions or personal projects.

[sage burrow]

Most of the SD3 and Flux models out so far don't seem to add anything that SD3 and Flux can't do with good prompting

[noble coyote]

Flux.Spoopy

No better than Schnell - bye bye 20Gb

[sage burrow]

I've gotten lots of nsfw from both sd3 and Flux, but it's pretty terrible nsfw lol

[sage burrow]

Flux straight up version? 🙂

[icy drift]

I'm also not really sure what people mean when they say NSFW. I'm just talking about nudity tbh, and I've gotten it by accident, without specifically prompting for it.
(I had to prompt "robes shirt tank-top tube-top sportsbra" to try to get Flux.Dev to cover up the top half on my cards yesterday, and it was still only like 25% successful.)

[hallow lion]

hey since u mastered that ball lora you can make an image with ëmad dropped the ball with sd3"theme somehow 😄

[noble coyote]

Yes, Spoopy checkpoint, no SDXL

[sage burrow]

Try adding "sfw" to your prompt, it helps.
But also, the private parts it sometimes shows are really terribly "drawn"

[sage burrow]

I mean, side by side comparison please 🙂

[noble coyote]

Wait up ..

[proven pecan]

Schnell

[prisma rampart]

I wouldnt expect much out of it, its just regular flux with a swapped clipL from some sdxl/1.5 finetune, at best it will act like a really shitty lora without any unet training.

some proper finetunes/loras will show up soon im sure

[lunar rivet]

schnell

[frail shoal]

liking schnell more than dev. Follows prompts better and styles.

it just has a bit worse anatomy

but nothing too problematic

[noble coyote]

Spoopy and Schnell are significantly different when using the same starting parameters! Spoopy is the first picture - and the Schnell version second.

[sage burrow]

I was using Dev, downloading Shnell now... HF is sooooo slloooooowwwwww!!!

[sage burrow]

Darn, it doesn't work with Dev 😦 DLing schnell now..

[acoustic kite]

hey guys you should switch to fp8_e5m2 instead of the other fp8. it's almost like fp16 but as fast as fp8 e4m3fn

[noble coyote]

I'm using e4m3fn with no problems. What would e5m2 bring which is better/any different?

[sage burrow]

That's the one I've been using apparently. Perhaps that's why my computer hasn't crashed lol

[noble coyote]

I'm to finding no difference between fp8 and fp16

Then again, all I'm using is Schnell - as Dev takes a lot longer ...

[lunar rivet]

all the fp8 checkpoints are somehow slower for me than fp16 so just went with upgrading ram instead

[noble coyote]

Schnell+SDXL "is as good and a lot faster" than Dev!

[uncut river]

Even for flux-dev, combining prompt for low grainy old photo quality WITH woman or maybe person in general, always gives nice smooth polished images.

[noble coyote]

You won't need it with Dev - Dev is crisp and sharp on its own. SDXL just makes the 4 iterations of Schnell as sharp as the 20+ of Dev - but much quicker!

SDXL's been working with Dev on my PC?!

Schnell

[frail shoal]

how many steps are you using on sdxl ? Also why don't you use 8 steps instead for schnell ?

it will be way sharper

[sage burrow]

Flux with 8gb GPU. Dev = 10 mins per image; Schnell is 5.5 mins per image

[noble coyote]

Schnell is OK, softer than Dev. I may try 8 steps. As it took me 10 seconds to hook-up to SDXL, I guess I never thought of doubling the iterations first!

[frail shoal]

post your 8 steps comparison

[frail shoal]

also you are losing the 16ch vae by passing sdxl after

[noble coyote]

2 steps Base, denoise 0.05; 8 steps Refiner, 0.05 denoise. SDXL result is less contrasty - typical flaw with SDXL is a certain lack of contrast

[noble coyote]

Schnell 4 steps

[frail shoal]

depends on the model, there are contrasty models out there

[noble coyote]

Yes, but there was also a LoRA released alongside SDXL which added contrast back in - OffsetExample

[noble coyote]

All I need to deal with now is the tendency to oversaturate in Schnell ...

[sage burrow]

Your workflow crashes comfy no matter what I try, darnit! 😦

[noble coyote]

You're prolly too little RAM already 😄

I have 64Gb

Sorry to disappoint!

[acoustic kite]

3360x1920 image, 60 seconds
flux schnell @ 8 steps upscaled 2x with sdxl lightning @ 4 steps

[noble coyote]

Do as suggested - make Schnell use 8 miterations, instead of the 4

[sage burrow]

but even straight up Dev works just fine (at long as I have other stuff to work on for 10 mins lol), but with your workflow, faster about 30 seconds it crashes, whether I use Schnell or Dev 😦

[noble coyote]

That's weird, and no crashes to report here!

[sage burrow]

I cant usually run other porple

[noble coyote]

Porple Crayon?

[sage burrow]

s workflows for whatever reason. I can always run the main ones, but not others

[acoustic kite]

advanced system settings > performance > settings > advanced > virtual memory > change > set it to something higher > restart pc

it runs out of memory

very quickly

[noble coyote]

Your PC/Mac is limited. Work within its limits 🙂

[sage burrow]

LOL, but, but Dev works lol

[frail shoal]

as you can see schnell has good quality on 8 steps. And knows artist styles

[noble coyote]

Coochy coo!

[frail shoal]

fp8_e5m2 has indeed more quality

[sage burrow]

What else woud you call that ? Im going to ask Gemini (nothing like that on win 11)

[icy drift]

Anyone using ComfyUI know how to get the "feed" / preview of images to show up? I just found out about it when I used Comfy on my laptop for the first time. It shows up there, but it doesn't show up on my desktop. What gives?

[acoustic kite]

i'm doing this on win 11 😄

[noble coyote]

Send me link please?

[acoustic kite]

it's right here

[acoustic kite]

that's what i've seen as well

closer to fp16

yet as cheap as fp8 something someting else

i forgot what it's called rn

[noble coyote]

You mean on the KSampler? Use Preview Method (via Manager) TAESD

[acoustic kite]

don't ask me what this is, i don't know 😂

[frail shoal]

it's just the weight type

[icy drift]

No at the bottom of the screen a box appears showing all of your generated history as thumbnails, and you can click a thumbnail to pull up the generated image, and use arrow keys to browse history, just like in A1111.

[noble coyote]

I didn't d/load it yet ... ?

[frail shoal]

don't know what you mean

[noble coyote]

OK, that would be ... Default UI - History (via manager)

[noble coyote]

I don't have fp8-e5m2 on my desktop

[lunar rivet]

you can use fp8 with fp16 models too, they should be cast automatically if you start with --fp8_e5m2-unet --fp8_e5m2-text-enc

[frail shoal]

yes you do

[noble coyote]

OK, let me see where I've stored it?

[icy drift]

Help? I can't find "Default UI - History" or "manager"

[frail shoal]

it's not a separate file, it's just how the weights are handled

[acoustic kite]

just press the dropdown

[noble coyote]

OK, I was looking in the wrong place

Found it! 🙂

Was using Default

[sage burrow]

Tyvm 🙂
200,000 mb should be enough
It could mess things up, but dying to try it 😄

[hollow swift]

A picturesque natural decor with a Provençal house of olive trees, a lavender field and lush greenery. In the foreground, a 10-year-old girl with blond brown hair with a braid in front of blue eyes, wearing a pink dress, next to her, a little boy with blond hair has a green t-shirt with the drawing of a rooster and white shorts, behind them, an elderly man wearing glasses, blue eyes, wearing a white Panama hat, pocket shorts and a dark green T-shirt marked "Papy" the three make a barbecue while smiling, a blue sky and a late afternoon sun the atmosphere is warm and serene

crazy prompt but still got it

[acoustic kite]

just make sure you have 200GB free storage space lol

i set it to 64 so i have 32gb ram + the extra 64gb so it doesn't crash

it used 76gb at its peak

[sage burrow]

It said I have 210, prob should have gone with 180 heh

[acoustic kite]

haha maybe

[noble coyote]

When you use Flux to make as much artistic output as I do, changing parameters simply makes "different art" - whether that is better art is moot! 😄

I guess if my output was closer to photographic realism, then parameter changes would be crucial

Schnell and fp8_e5m2

[sage burrow]

I prefer the 1st one

[acoustic kite]

haha schnell doesn't quite know how to do text as dev does

but it's fun that it's doing it regardless

[frail shoal]

it does it better for me. dev forgets the text when the prompt is long

[noble coyote]

If you do too many iterations, Dev blows out garbage text!!! 😄

[acoustic kite]

when the resolution is closer to 1mp in total, text seems to pop up more frequently

hey but it added a watermark 😂

[sage burrow]

So dev online and Schell on own system serens like a good idea

[noble coyote]

It seems to add tex early; then decides that more detail is needed, so the text is overpainted

[frail shoal]

somebody should change this channel to flux lol

[torn wharf]

Locally devvvvvv

Maaaachinne

The state of sd3 is in flux

[noble coyote]

That we discuss Flux in here more than SD3 shows just how disappointed the SD3 Community really is?! 😄

It didn't say Papy!!!

[acoustic kite]

do weights like ((this)) and [[this]] work with flux? 😄

[noble coyote]

Not tried - give it a go

[sage burrow]

Schnell plus sdxl = 6 mins

[uncut river]

how much to add

lol

[acoustic kite]

ok i'll try wait a sec

they do nothing

[sage burrow]

Neither does:2.1

[icy drift]

Monster Card Series V2
17 - Minotaur

[acoustic kite]

the sloth is still having a stroke

[frail shoal]

remember when they said skill issue with sd3 ? Nobody is complaining about flux.

[noble coyote]

It's nice that Flux is inclusive of our Disabled Sloth Community! 😄

[sage burrow]

do prompt permutations work with Flux? (prob not, but I can't test it with this brand new obsolete computer)

[acoustic kite]

not much to complain about

ok all text seems to work every time at 16 steps with schnell as opposed to 8 steps

everything works all the time at 16 steps

[noble coyote]

Getting anatomy right with a tool which is pre-programmed "not to do anatomy" is never a skills issue: it is a bottle-job!!!

[icy drift]

2-shot / second try. Different prompts to clip_I and T5, more detailed prompt to T5 containing everything relevant about pose and expression.

Monster Card Series V2
18 - Gryphon

[noble coyote]

😄

[icy drift]

No way that text was generated by Flux. It's too ugly.

[noble coyote]

PhotoShop and SD3 😄

[signal shuttle]

Flux dev, i love it so much

[noble coyote]

Woman lying on grass remind you of anything?!

[icy drift]

Have you tried it in Flux? I'm sure someone has. I guarantee it will work just fine.

[noble coyote]

I changed the orientation from Portrait to Landscape; and all of a sudden, the text has showed-up!!!

[noble coyote]

Flux is not good with a lot of text ... but I guess I could make several takes - and cut and paste them together?!

[sage burrow]

Flux

[noble coyote]

Two Right Arms - or a hell of a contortionist!!! 😄

Love the tank guarding the lavender

[torn wharf]

even when you get good text from flux , it's not perfect.

[noble coyote]

Nice try!

[icy drift]

That doesn't fit my experience at all? Flux is the best text model I've ever used by far.

[noble coyote]

OK, so you're saying the sweet-spot (using Schnell) is +/- 16 iterations?

[icy drift]

Monster Card Series V2
19 - Banshee (Actually nvm failed hands retrying)

[noble coyote]

Beowulf? Hrothgar?

[signal shuttle]

I saw the word banshee and the first thing i thinked of is halo

[icy drift]

I'm not sure where I first heard of banshees. They're an Irish myth monster. The LLM I used yesterday to list monsters mentioned Banshee.
Also just realized her hands are messed up here. I'm still just thinking about text.

[uncut river]

trying out the Copax lora for SD3

[icy drift]

Fixed.

Monster Card Series V2
19 - Banshee

[torn wharf]

in terraria i think they're the women that look like they're lost then you go to help them annd they fuck you up

no thats different mb

[acoustic kite]

schnell is also pretty good

2 different guns staying in the same style at the same seed

[torn wharf]

i was thinking of the terraria nymph.

[sage burrow]

flux

[torn wharf]

feet for hands

[acoustic kite]

schnell or dev

[sage burrow]

Via Mage, so Pro I think

[acoustic kite]

ohh that's interesting

did you choose how many steps?

[torn wharf]

it gives a strong effort but longer sets of lyrics it doesn't manage as well

[craggy crest]

mage only has schnell right now

[icy drift]

Have you compared it to other models? I've never tried DallE models or whatever Google has.

Monster Card Series V2
20 - Gargoyle

[icy drift]

She just asks ship captains to hand over their good with big puppy-dog eyes. She's never been defeated.

[torn wharf]

i've done lots of text comparisons. dalle through copilot barely manages long blocks like that at all. i haven't played with mj yet, apparantly the new is really good.

google i think blocks their ai services in canada

[icy drift]

Okay guys I need more monsters again.
Gonna try my LLM, but I think I'm just running out of monsters that exist in general (with appearances suitable to this card collection).

[uncut river]

Have you tried a snake with many spider legs?

[sage burrow]

Nope, I just set it to auto, bc I

m too cheap to pay for the proper plan. Also, it worked extremely well for everything else

[torn wharf]

cheap people make the world go round

[icy drift]

What monster is that? But also the body has to be a chibi anime girl, so specific body features that can't be mapped to humanoid are out.

[uncut river]

oh, sorry i missed that context.

ill give it a try

[icy drift]

I'm doing classic fantasy monsters as chibi anime girl mascots.

[sage burrow]

SD3 seems so fast now!

[uncut river]

hm, not really

flux-schnell, online

anime style snake with many spider legs, centipede, anime monster, horror anime, colorful, high contrast, dark realistic lighting, black shadows, body of a cartoon chibi girl

[icy drift]

So far: Vampire, Werewolf, Slime, Dragon, Mummy, Harpy, Mermaid, Ghost, Dryad, Medusa, Kraken, Fairy, Golem, Unicorn, Yeti, Siren, Minotaur, Gryphon, Banshee, Gargoyle, and I'm doing Goblin now.

[acoustic kite]

sd 1.5 512x512 seems very very fast 😂

for me it's like 40 times faster

[noble coyote]

Banshee of Inisherin - starring Brendan Gleeson and Colin Farrell

[icy drift]

Yeah we keep getting more quality and then more speed and then more quality and back and forth. SDXL Lightning has been my go-to model until this week.

[noble coyote]

Hrothgar is a good monster

[sage burrow]

I did side by side comparisons with it vs the huggingface Dev yesterday, and most often I couldn't decice which I liked better, hmmmm

[noble coyote]

I'm not sure why Schnell and Dev are so close in output - but Dev can be extended and developed I'm supposing?

So Dev (if you like) is in 'alpha'

[icy drift]

Yeah the hardest I tried was a Kraken. It took absurd quantities of rerolling and trying different prompts. Flux just wants to put the girl and the monster side-by-side instead of fusing them.

[noble coyote]

Girl as Kraken; Kraken as Girl

[torn wharf]

people be like "THERE WILL BE NO FINETUNES" and the license explicitely states finetunes of dev are allowed and the description of the model is so that developers can build on it and it's name is called [dev] short for devellopment.

The insidious rumor spreading is so insane. and we can clearly see it rooted with invoke ai on this particular point. They're patient zero in this misinformation pandemic

[icy drift]

You are welcome to try. It's obviously not impossible, just takes forever to get a good one.

[torn wharf]

do invoke just think no one will fact check them?

[torn wharf]

really like this one.

[icy drift]

Monster Card Series V2
21 - Goblin

[craggy crest]

https://x.com/yangzhizheng1/status/1819969554388496709

PrimerYang (@yangzhizheng1) on X

I have found how to train FLUX.1 lora，It seems like need 48G VRAM at least，here is a guidelines for training.
https://t.co/8sUpYY1eDM @bdsqlsz @araminta_k @AIWarper @ZHOZHO672070 @camenduru @cubiq

Twitter · Aug 4, 2024, 5:32 AM

@torn wharf here's what you're waiting on

[torn wharf]

like, we know invoke ai's motiv. they're building the open model initiative and are invested in models with no commercial licensing. BFL are a danger to that business strategy. So they're spreading misinformation about Flux accordingly.

But what i don't get is why they're spreading such dumb misinformation with such blatant lying. "Flux can't inpaint" "Flux can't finetune" "Flux can't run on home pcs"

[craggy crest]

why does anyone spread misinformation? but there's the training guide

[sage burrow]

48G VRAM at least lol

Rental I guess 😄

[craggy crest]

for some people, yes.

[torn wharf]

#🆕｜sd3 message linked it last night adn you replied

[icy drift]

😱 I don't think Flux can make one-eyed cyclops anime chibi girls! What do I do?...

[craggy crest]

i guess you don't make one-eyed cyclops anime chibi girls!

[brittle nexus]

Apparently doesn't know FSM

[mortal mesa]

flux tech was TAKEN from SAI

thats what soneone said here

[craggy crest]

flux tech BELONGS to robin, who developed it when he was SD3 lead at sai, and he took it with him when he left

SAI just has permission to keep using it

[torn wharf]

former employees of SAI made it, but it's based on their research that's published openly.

[mortal mesa]

oh i know, other dont

[dull star]

holy crap why didn't I think of these memes

[noble coyote]

If all the research at SAI was "open source" - then whoever has the tech 'owns' the tech?

[mortal mesa]

when he was doing his doctorate, public German money paid for it

[noble coyote]

No centralised owner?

[craggy crest]

SAI is mostly an incubator, and has been from the start. the people there were all originaly Eluther, as is Robin. his tech belongs to him. this might change under the new CEO, who knows, but SAI doesn't own any of the tech at this point.

[noble coyote]

Schnell - a pensive lady!

[mortal mesa]

ya see that isnt correct either

[noble coyote]

Emad left to start a "decentralised structure" (based on BlockChain) where no one individual exerts - or can exert - overall control

[torn wharf]

no permission needed. it's a published paper, not a patent

[noble coyote]

One-eyed Chibi ... more disabled than cyclops

[icy drift]

LOL really though, can anyone find a one-eyed prompt?

[noble coyote]

Leela from Futurama?

[icy drift]

Can't decide if it's cute or creepy...

[torn wharf]

after emad went full cryptobro, i entirely checked out. i couldn't care less about his new endeavor and am glad he left stability so as to not corrupt it with that criminal bs. Same category as Roger Ver is now. Convicted felon that is living life off a ponzi scheme. crypto on paper is awesome looking, but then in real world it's just ponzi crap because of corrupt actors.

[uncut river]

oh, that was cute. here comes creepy

[dry wave]

this discussion leads to nothing. Latent diffusion was developed at University Munich where also the source code for sd 1-2 was written - which was used by sai.

In general we all benefit a lot from open research from Google, Meta, Microsoft and OpenAI

Stability itself did not contributed that much

[mortal mesa]

and cascade

[noble coyote]

Sort of one-eyed

[craggy crest]

still, i was clarifying the situation. they have 'permission' by the fact that it's public and open source, and they don't own it

[craggy crest]

crypto is not criminal and you might want to get used to using it. the entire financial world is moving to block chain and away from traditional monitary systems

[mortal mesa]

gaslight

[craggy crest]

stability contributes nothing. it's an incubator. a company that exists to make it possible for the creatives to create and research and develop.

[mortal mesa]

heres a crypto lesson, dont leave it in an exchange, you wouldnt buy a collectible at a store and leave it there

[uncut river]

this aint a crypto channel, please go somewhere else with those lessons

[mortal mesa]

please dont pretend mod, its against the rules

[uncut river]

im not a mod, it was just a question. I dont think anyone in here wants or needs to know anything about crypto. And if they do, they sure can look it up somewhere else.

[mortal mesa]

maybe off topic would be a better fit for you

[noble coyote]

Emad is using Crypto technology because it guarantees democracy - no one person can own, totally control or singly develop - a product or company

[uncut river]

are you now talking to yourself? I mean, crypto is off topic, I mostly just post images or jump into ai related discussions

[icy drift]

Personally, I don't come here to talk about corporate governance, politics, religion, or financial systems.
I'm mostly interested in AI image generation and getting help with / sharing whatever image generation stuff I'm doing at the moment.

Monster Card Series V2
22 - Cyclops

[noble coyote]

Emad is not involved in Currency - he too understands that Crypto Currency is just downright bad

Blockchain guarantees democracy

[craggy crest]

he's using blockchain because not only does it guarentee that, it's also jsut about unhackable

[noble coyote]

Blockchain guarantees a decentralised working structure

[mortal mesa]

not without electricity

[craggy crest]

you're not making pretty pictures on the computer with out that, either.

[noble coyote]

OK, so keep putting coins in the meter and keep that electricity flowing

Some Art Deco themed Flux.Schnell

[dry wave]

blockchain 🥱

[brittle nexus]

A guy appeared talking so much nonsense that he made me agree with crystalwizard. What a day!

[sage burrow]

Inpaint?

[torn wharf]

lol not really. not really at all. private networks that are built with merkle trees , but they're still operating on the same cobalt cores.

no bank is going to swap their foundational infrastructure to a crypto network that is rife with unregulated abuse.

Blockchain is just a data structure. a novel form of a merkle tree. it's not what crypto networks are. it's just a part of it.

[icy drift]

Already solved it.

[noble coyote]

Tah daaah!

[uncut river]

huh ?!

how you ?

[noble coyote]

Leela as Little Bo-Peep!

[craggy crest]

someone needs to actually research blockchain and the banks are moving toward it, yes.

[icy drift]

Cyclops had sheep. Vaguely remember from Latin class.

[noble coyote]

Block chain and Tors (onion skins)

Amo amas amat

Villa Villae Villam Villas

[icy drift]

Sum Es Est Sumus Estis Sunt

[brittle nexus]

Confusing Blockchain with Crypto is like confusing wifi with internet

[mortal mesa]

get em nordel

[noble coyote]

Crypto is built on Blockchain ... but Blockchain is a substrate for a lot more than Crypto

Flux.Schnell - Elegance

[torn wharf]

block chain is not crypto. it's just an immuteable datastructure.

Banks like to reverse transactions typically. they'll be keeping a private network not ethereum

[craggy crest]

i never said it was. and the entire financial strcture of the world is, yes, moving to blockchain AND are switching to crypto. get used to it

[torn wharf]

we're long past the blockchain buzz where it was the magic pill every developer wanted. now that pill is AI

[craggy crest]

you're obsessed. go read the trianing guide for flux and see if you can train a balls lora for it

[icy drift]

Having problems with Centaur now too... I think the odder the creature the worse Flux handles it. It's sort of like it traded flexibility for accuracy.

[noble coyote]

As I said, Blockchain is just a structure or a platform over which Crypto is built. Blockcain performs this role for a lot more than just Crypto ... but Crypto gives Blockchain a bad name

[mortal mesa]

its all high level math that looks like magic

[icy drift]

I have this guy blocked, so I don't know what you're arguing about, but I recommend it.

[torn wharf]

he was the one who told me i couldnn't train sd3. then i made balls. he's been passively aggresively pretending to be supportive since.

[noble coyote]

SDXL was bad at Centaurs. Try a Satyr (Mr Tumness) as well!

A Minotaur

[icy drift]

A satyr is a half-goat right? 🤔 That should be doable and adorable. LLM kept mentioning it and I kept skipping.

[noble coyote]

Yes, there was one in The Lion, The Witch and The Wardrobe (Mr Tumness)

[torn wharf]

with sd15 we'd just do 1024x512 images for centaurs. why'd they break that functionallity?! ||/s||

[noble coyote]

Gryphon, Basilisk, Gargoyle

[hollow swift]

the thinker

[sage burrow]

Pony solved that ROFL

[noble coyote]

Tee hee