🆕｜sd3

[dry wave]

I wished they would have made a 4b version 😦

[frail shoal]

sigma

[frail shoal]

prompt ?

[turbid grotto]

Is this possible to finetune FLux?

[torn wharf]

should be. they're weights. we'll see

[bitter hearth]

I used too much CADS and got a chonky cat on the left

[dusky thistle]

no prob. photoshop was rpolly killing you

[torn wharf]

was gonna work on dataset building but flux happened and a few other things too. tommorrow chore i guess

[cunning lintel]

An atmospheric HDR 8k digital octane render, a living house with a worn, wooden facade, overgrown with vines and moss, walks through the modern city, its foundation roots splayed out like legs, windows blinking like eyes, chimney puffing out smoke like breath, intricate details, rich textures, warm lighting, long shadows, vibrant colors, dynamic movement, surreal atmosphere

[dusky thistle]

yea i had a bunch of coding and reverse engineering stuff planned for today

got literally nothing done

well actually i read half a page of a paper

lol

[torn wharf]

@sacred jewel if you've got windows 11 and a cpu with an integrated gpu, you can go into windows settings and makes ure the ps.exe uses the integrated one

[cunning lintel]

i just wanted to add cascade to my own personal use discord bot, now there's this 🤣

[frail shoal]

sigma

[torn wharf]

tbh i was annoyed about all the flux talk in sd3 but.. i get it now

[bitter hearth]

flux is king yes

[cunning lintel]

it's super off topic, but asking for a 3rd party models channel is prob too much since even cascade is blocked

[frail shoal]

i'm liking this prompt better in sigma

[torn wharf]

they're build from the same parts right? but.. sd3 an flux

[frail shoal]

here is schnell 4 steps

[proven pecan]

SD3

[torn wharf]

internal meetings at stability all day today

mods too busy to bother with us

[turbid grotto]

And we don't need any code to be released?

[frail shoal]

lol

[cunning lintel]

after the first WOW moments, i still kinda like sd3-8b looks better overall i think, it just reacts more to styles which i kinda value more than prompt adherence when the difference is relatively small (but maybe next week flux is figured out how to get more broad styles out of it, still early)

[torn wharf]

diffusers team will figure it out

there's more coming too surely. i think they have people udner nda who have been working on things. there seems to be lots of knowledge about the model that's been seeded around

[dry wave]

code is released

also it's already on diffusers

[torn wharf]

training or inference?

[dry wave]

inference as usual

[torn wharf]

we're talking about training

[frail shoal]

sd3.1 would save me from buying a gpu to run flux

[dry wave]

they usually don't give that out. Its also not of huge interest for the community

[torn wharf]

ok

[dry wave]

we would use different training code anyways with our consumer hardware

[bitter hearth]

can you even train the full flux model on a 4090

nvm can't train for sure, it takes massive amounts of vram

[dry wave]

you can barely load it into vram

[bitter hearth]

was thinking more about things like lora

[dry wave]

you have to quantise it to get it run

[fleet meteor]

My PC is somehow not dying if i use 50gb of my SSDs as swap

Now the vae gets processed in 5 seconds instead of 120

[bitter hearth]

there will be optimizations

i guess

[frail shoal]

sd3

[proven pecan]

DMD2

[hallow lion]

FluxPony

[frail shoal]

dmd2 ?

[proven pecan]

yeah

[frail shoal]

what is it

[proven pecan]

https://huggingface.co/tianweiy/DMD2 the unet

[frail shoal]

new model ?

[proven pecan]

newish

[frail shoal]

schnell

[sterile pendant]

new clip encoder for flux that has guidance* (not 100% sure if it's cfg as we know, havent tested yet) in it https://github.com/comfyanonymous/ComfyUI/commit/a531001cc772305364a319a760fcd5034e28411a

"clip_l": ("STRING", {"multiline": True, "dynamicPrompts": True}), "t5xxl": ("STRING", {"multiline": True, "dynamicPrompts": True}), "guidance": ("FLOAT", {"default": 3.5, "min": 0.0, "max": 100.0, "step": 0.1}),

[undone meteor]

What speeds are people getting with flux schnell? I have a 24 gb 4090 and I’m barely getting -40 s / it. Seems slow. But I’m using the workflow from the example blog. Is that what people are seeing?

[sterile pendant]

use a smaller t5

use the fp8 one from sd3

[fleet meteor]

about 8-9 s / it, but it takes more time loading the models than actually doing the iterations 🥹

[acoustic kite]

comfyui is crashing instantly when using flux and the default workflow that comes with it, does anyone know what to do? RTX 4070TI Super 16gb, 32gb ddr5 ram

there should be enough to make it run

[undone meteor]

Ya it’s a shock seeing it load the model in “lowvram mode”. I have 24gb man. Use them!

[fleet meteor]

Use --lowvram in the launcher, even with a 3090/4090 you gotta use that

[acoustic kite]

thanks :) do you know how i can do that?

[sterile pendant]

the model is massive, it's roughly 23gb for the transformer ALONE

not counting the t5

[acoustic kite]

i did NOT realize that

[turbid grotto]

5-7s/it on rtx 3060 with exmple workflow on flux dev

[acoustic kite]

that's crazy

[sterile pendant]

its a 12b model

[fleet meteor]

Right click the comfyui_run nvidia.bat and click edit, and add this "--highvram"

[dull star]

its him... goku...

[sterile pendant]

also, make sure to update it again, there were some comfy pushes for memory management

[lavish sparrow]

flux vs sd3

[dull star]

hehe

[fleet meteor]

@acoustic kite It should look like this

[torn wharf]

the future of personal ball computing

[acoustic kite]

ohh okay thank you so much, let me see if it still crashes

[sterile pendant]

yeah im testing it now after the memory management update, it's capping out almost all my vram now, before it was only using like 6gb out of 8gb

[torn wharf]

home gpus this gen be like

[prisma rampart]

switched to fp8 flux dev and fp8 T5 and no more low vram or unloading models on 4090

[sterile pendant]

FYI: The guidance on the new CLIPTextEncodeFlux node doesn't seem to do anything yet, or maybe it doesn't work with the schnell model

[torn wharf]

get it?

[hallow lion]

whats the quality drop form 16 to 8 if any?

[torn wharf]

little visual pun because pillows are stuffed full

[prisma rampart]

havent really run any comparisons yet

[acoustic kite]

it made my whole pc crash at first 😂 i think it's doing better now

thank you so much

[torn wharf]

too fun. just toooo fun

[undone meteor]

I may have solved the slow generation. In the load diffusion model node, the example sets the weight dtype to “default”. Setting it to fp8_e5m2 or fo8_e4m3fn gets images down from 200 seconds per image to 3 seconds. No idea what those options are though lol

[fleet meteor]

Lol i hope it works good 😅

[acoustic kite]

it's a bit slow at 4.9s/it but i think the prompt adherence is crazy good lol

nevermind it doesn't like me 😂

[acoustic kite]

interesting, maybe that means i'll finally be able to run it

[lucid swift]

wow this is very cool

[hallow lion]

Can flux images be used commercially or is it just for fun?

[torn wharf]

the gpu we need

[fleet meteor]

Did it work after this? press f5 on comfy if it doesnt load

stolen from 4chan, i gotta try this prompt

[acoustic kite]

it works if i use fp8_e5m2 instead of default, default just crashes

[torn wharf]

diamond is after TI

everything about this card is better than current cards. prove it wrong

[fleet meteor]

power consumption 😁

[torn wharf]

we'll have the ice cube in the ocean by then its all good

[fleet meteor]

Oof

[acoustic kite]

looks like upscaling like you would normally do also works fine

[acoustic kite]

minecraft from wish 😂

[sullen moss]

New toy in our hands😌

[frail shoal]

SD3.1 wip, when ?

[fleet meteor]

Im trying to get it, almost there, only the hud remains

[acoustic kite]

looks great so far

i'm about to make a 2048x2048 image

but it takes such a long time 😂

[grand wyvern]

R u guys trying to make Minecraft in sd😂

[acoustic kite]

no i'm making gta 6 before it releases haha

there we go, i guess it sort of works lol

[lucid swift]

what romt did you use?

[dreamy sundial]

yeah sd3 8b wins by a mile

[fleet meteor]

what is that?

[lavish sparrow]

xD

[lucid swift]

the letter on my keybord is broken

is missing

[fleet meteor]

Wait i stole that image, idk the prompt they used or the settings

[lucid swift]

oh

[dreamy sundial]

it's true

[hallow lion]

Aw we cant put negative prompts

not that it matters at this point a lot :))))

[acoustic kite]

to go around this i'm upscaling it and then i can use a positive and negative prompt

but it doesn't really work, only if you don't want a certain style

[bitter hearth]

maybe perp neg works?

[acoustic kite]

this one? will test it rn 😄

left 1024x1024, right 2048x2048 :)

[hallow lion]

very nice

[acoustic kite]

ty

[fleet meteor]

almost XD

[acoustic kite]

xD

looks great tho

[acoustic kite]

@bitter hearth sadly this one didn't work :(

this is what came out of it when trying to add a negative prompt

[acoustic kite]

this looks sick man

[hallow lion]

instant VHS lora

[acoustic kite]

2048x2048 image.

sampler: euler
scheduler: beta
steps: 30
prompt: gta 6 gameplay footage, letters on tyres saying "michelin", license plate saying "GTA 6", realistic, background characters

sampler: euler
scheduler: beta
steps: 16
positive prompt in upscaler: gta 6 gameplay footage, letters on tyres saying "michelin", license plate saying "GTA 6", realistic, background characters
negative prompt in upscaler: anime, cartoon, deformed body parts, misspelled words, watermark

[compact forge]

is SD3-Universal an improvement over SD3?

[grand wyvern]

is beta scheduler better than normal scheduler

or whats the difference

[acoustic kite]

i'm not sure what the improvement is but there is definitely an improvement imo

it's just overall looking better

[grand wyvern]

ill try it

[acoustic kite]

okay 😄

[low stone]

i tried flux but...

[torn wharf]

flux is the new crysis engine

[vapid radish]

Make sure you switch down to the fp8 weights and Text encoder, went from 10 mins to 25 seconds on my 3090 after that.

[acoustic kite]

512x512 test at 12 samples

[acoustic kite]

thank you this was the only thing that worked, my pc shat itself trying to use fp16 😭

[prisma rampart]

1424x2048, nice to see flux holding up really well at higher resolutions out of the box.

[hallow lion]

this is what alarm clocks really look like

[bitter hearth]

thats perp neg yeah

not sure if it works with SD3

[acoustic kite]

i pinged you a bit later, it sadly didn't work :(

[sage burrow]

I made a bunch (woth mj) using reference images which were screenshots from the leaked video. It worked really well!

[acoustic kite]

@bitter hearth this was the result

[acoustic kite]

thanks for the tip!!

i tried using controlnet with flux and it didn't like that either, hopefully in the future though

[hallow lion]

yeah i didnt even try... but i sure hope well get controlnets for it and ip adapters. imagine flux with sdxl community support

one off images are good enough - next consistency and next how to make them move, once we get 5-10 seconds of actual consistent img to video. we can make movies

[sage burrow]

This was 7 months ago, probably better now and SD also now:
https://www.deviantart.com/uniquecharacters93/gallery?q=Gta

[dusky thistle]

you know what's weird? vram use goes way up with the most recent update of comfyui for this shit

i reverted to an older commit and manually added the new clip text encode node and now i can generate at 2048x1024 without oom

very easy fix

git reset --hard 2f88d19

[bitter hearth]

why is flux able to handle more resolution?

[dusky thistle]

won't wreck any models, data, outputs, etc, will just un-update

with the current repo i can't gen past 1536x768

now i'm back to 2048x1024

literally twice the pixels with the same vram

[torn wharf]

smart science

[prisma rampart]

it doesnt seem to do the weird distortions on the edges like sd3 when you try to gen huge images, it just does some screen door/pixelation effect when you push resolution too far

maybe flux uses a different way of doing positional embeddings compared to sd3 or just conditioned on higher res images

[bitter hearth]

part of the reason I want to know is
if it shrinks or compresses the latent like hi-diffusion, deep shrink, cascade or ultrapixel
that compression has some side effects

I'm not saying latent compression should not be done, it seems like a good solution, but you do take a hit to the composition

[craggy crest]

go read the paper

[bitter hearth]

ok thanks I should just read the paper yeah

I do hope they managed to do it
without compressed latent like the other high res solutions

but if there's some compression that's fine I will still use it

[craggy crest]

in case you don't have the main link https://blackforestlabs.ai/announcing-black-forest-labs/

[bitter hearth]

can it balls

[bitter hearth]

thanks

somehow made a breakthrough
and R2D2 got his long legs out
not sure if its the seed, the sampler or CADS that did it
never seen SD3 do this before

[acoustic kite]

thx!! where can i write this?

sorry i meant to ask if this was with comfyui?

[dusky thistle]

you gotta pull up a command line and get into your comfyui folder

[acoustic kite]

ohh okay i just typed it in the default location lol

[dusky thistle]

yeah that's to roll back to the version from earlier today that seems to be better with vram

[bitter hearth]

on L2 discord
someone used the C++ version of SDXL, with a library that uses the VRAM as a storage system on linux, so they could bake the model into the VRAM
it seemed cool

[dusky thistle]

nice, got a wf version?

[acoustic kite]

this was in the comfyui folder

sorry i'm probably missing something 😆

[dusky thistle]

wait what?

[dusky thistle]

oh maybe you're on the portable

[bitter hearth]

I will go find

[acoustic kite]

ohhh yeah i am

[bitter hearth]

ok

[torn wharf]

balls

[torn wharf]

the technical paper comes later thats just an announcement.

[craggy crest]

did you happen to read it?

[torn wharf]

yeah. they give links to all the research they used to build it. papers for other projects. then say theirs will come soon

go check

[craggy crest]

and a lot of technical information.

[torn wharf]

guess i have a higher bar for what constitutes a paper

[craggy crest]

for what he wants, all he has to do is what emad told someone else earlier today 'go read the SD3 paper"

[acoustic kite]

Cyberpunk Venetian carnival, holographic masks, bioluminescent canals, floating neon gondolas, crystal skyscrapers in the background, steampunk clockwork pigeons, quantum aurora borealis in the sky, fractal fireworks, hyper-detailed fabric textures on futuristic Renaissance costumes, zero-gravity confetti, time-distortion ripples in the air, AI-powered street performers, 8K resolution, photorealistic rendering

[tough viper]

Does sd3 curently work in auto?

[lucid swift]

yes but these images are not sd3

[torn wharf]

yeah but its innefficient and if you intend to use t5 enc with it, use comfy

[craggy crest]

i.e. - not very well

[hallow lion]

cute and disturbing at the same time

[hallow lion]

Classic case of noodle denial.

Everyone goes through that phase before giving in.

[torn wharf]

i lean towards a111 where the memory isn't a problem. but t5 breaks the bank and needs to be cached. last i tried it wouldn't do it that way

the man who took over reforge is working on the core of it, to enable sd3 support from upstream to bypass the unet patcher part of forge, last i looked

[tough oriole]

Flux can do the Pixar horrible movie titles meme.

[craggy crest]

powerful new earth shaking tool is released. Internet uses it to make memes

[mortal mesa]

flux1-dev-fp8.safetensors 11.9 GB on HF

[low stone]

logo Typography "I Love Samplers" with design elements and artistic flair, background is various noise patterns mixed with dj mixing boards

it does good typography as well

[mortal mesa]

i was getting errors stuffing it with alot of text, but ya its pretty good

[patent magnet]

The Hellstar Remina prompt attempts. The first image is the original by Junji Ito. The second is another original, Remina eating a planet. The others are by Flux Pro. I got it first time. I couldn't do this with SD.

Prompt is:

a giant hellish planet with a giant eye looking surprised at the viewer and a deformed mouth with a tentacle tongue licking the planet, horror by Junji Ito

And slight variations of this prompt, and tests with different intervals.

[bitter hearth]

Very cool!! i love junji ito's artwork

[patent magnet]

He is a master of creative horror. I love the fact that in this story the whole apocaliptic ordeal (Remina getting to Earth) starts right in the first few pages.

[rain current]

I am testing Flux and it is working really well

It performs very well with anatomy

[supple gulch]

When release SD3.1

[hallow lion]

Deja vu....

[hallow lion]

incredible isnt it?

i tried my ideogram promts on flux and even on schnell they came out way better

[hallow lion]

I'm not gonna go there...

[rain current]

I am testing with Dev; I haven't tried with Schnell yet

[bitter hearth]

2 weeks

[mortal mesa]

Soon™

i caught myself almost prompting for screenshit, whew

[bitter hearth]

[frozen summit]

#✨｜sdxl happy Cartoon head portrait

[signal shuttle]

SD3 can't compare with flux

[bitter hearth]

Meencraf!

[muted dove]

Flux is looking promising!
a glass ball half full of red liquid is to the left of a glass square that contains blue liquid. To the right of them is a green glass bottle

[frigid mango]

try to put persons in positions.

[undone timber]

an Apple

what？

[signal shuttle]

flux does burning cars better then SD3

[wispy spade]

helo guys can anyone help me how to fix this problem?

[hallow lion]

holy molly these really look real, its harder and harder to spot ai

[signal shuttle]

Does the clip names match the ones you downloaded?

[hallow lion]

looking for melted hands and deformed numbers or text wotn help soon!

😄

[wispy spade]

i dont think so

i only downloading the check point from hugging face

[signal shuttle]

thats the issue, if your downloading the checkpoint with the clips then you don't need a triple Clip loader

just the checkpoint node

[lunar ingot]

hello every one

[wispy spade]

i suppose checkpoint from hugging face already have clip inside right? should i remove the triple clip loader?

i try sd3 frist time with workflow provided on that library

[lunar ingot]

🍉

🍉

🍉

[wispy spade]

hi

[signal shuttle]

If you downloaded the SD3_medium_incl_clips_T5xxlfp8 or SD3_medium_incl_clips then you don't need a triplecliploader

[wispy spade]

i have this SD3_medium_incl_clips

[signal shuttle]

how much Vram and ram does your system have?, since T5xxl is pretty important to get a good image out of SD3

[wispy spade]

i just testing with sd3 meidum incl

i can run it but another problem come up

[wispy spade]

32gb ram and 8g vram mobile gpu rtx4070

[signal shuttle]

Is this the latest version of comfyui?

[wispy spade]

yes brother i havnt running comfyui for months i just updating all nodes and comfyui itslef while ago

i dont really understand such technical issue so i dont know where to start looing for this problem

this is how it look like

[signal shuttle]

Try clicking on sd3_medium in the load checkpoint node and try switching to SD3_medium_incl_clips

[wispy spade]

omg i click wrong checkpoint

hahaha

thinking it was SD3_medium_incl_clips

yes it running already

thank you very much mate

[signal shuttle]

yep thats the issue, the cliptextencode node couldn't tokenize your prompt since there wasn't any clips

[wispy spade]

yes

@dreamy egret by the way how good is this SD3_medium_incl_clips_T5xxlfp8?

sigh even 8gb not enough

[signal shuttle]

You pinged the wrong guy but SD3_medium_incl_clips_T5xxlfp8 is way better since it has T5xxl in it

[wispy spade]

😅 how much vram i need?

[signal shuttle]

I think at least 12? but i might be wrong

since it been a while since i ran SD3

[wispy spade]

well i least we got 50s family the end of the year or next year Q1

it will bring older series prices down

[frail shoal]

i'm gonna try to run it in a laptop with 6gb vram and 16gb ram

the flux i mean

[signal shuttle]

Yeah i hope so but considering the RTX 3090 is still expensive, i dont think it will

[signal shuttle]

You can run it, but it will take you over 6 minutes per image

[wispy spade]

i think it can if only SD3_medium_incl_clips

[frail shoal]

will see

[wispy spade]

TI and super is the way to go

[frail shoal]

no sd3 runs fine and very fast with it

[signal shuttle]

the minimum vram requirement for Flux Schnell and Flux dev is at least 16gbs of vram

and the recommended is 24gbs of vram

[frail shoal]

sure but with paging to ssd and ram might do it in fp8

[signal shuttle]

i am not saying that you can't run it, you can but it will take you a long time to generate one image

[frail shoal]

i know, but i want to know how much

[signal shuttle]

someone said with 8gbs of vram it took them five minutes per image

[frail shoal]

it would be way faster if we can use align your steps or something else for it

[wispy spade]

what is this flux something?

[frail shoal]

also i'm running xformers which is more optimized for shitty gpu

[signal shuttle]

Flux is a new opensource SOTA text to image model that is 12B parameters made by black forest labs, it beats SD3 by a long shot

[wispy spade]

what the .... it sound so awesome but why i never heard of them?

[signal shuttle]

since it was just released yesterday

[wispy spade]

what you even want to make with 12B para

[signal shuttle]

and it came out of nowhere, they didn't tease the model

[wispy spade]

i mean it a big thing and never hear any news regarding that before hahaha unless they want to be a dark horses

[signal shuttle]

nobody heard of or even knew who black forest labs were before Flux's release

[wispy spade]

free for use?

[signal shuttle]

Flux has three versions, a pro version which is closed source and only useable through a API and Flux Dev version which is free and opensource and a small Flux schnell model which is also free and opensource

they have a playground to use flux pro for free (20 images per day for the pro version) https://fal.ai/models/fal-ai/flux-pro

[wispy spade]

i see what is the community reaction it sus for such hot potato come out of no where

Flux Dev version can run locally?

[signal shuttle]

both the dev and the schnell versions yes

[signal shuttle]

and the ironic thing about black forest labs is that the guys behind it is the same guys behind Stable diffusion (People who left stability ai)

[wispy spade]

that nice gonna try it later

[wispy spade]

im glad that the one working on it have real abilities

[dry wave]

the open models are bad at paintings - or we haven't found the right prompt for them

besides that the model is great

it gets hands right literally every time xD

it even can make people holding swords or axes correctly

[wispy spade]

cant ask that much for something that is free, but yeah it still have a lot of room to growth since it was a new model

[dry wave]

and it can make women lying on grass 😛

[hallow lion]

I can run it with 12.

[wispy spade]

why grass?

[hallow lion]

You're new here huh?

[wispy spade]

how big is the model sizes?

[hallow lion]

23GB!

[dry wave]

24gb....

you have to quantize it to run it 😦

[wispy spade]

i know sd3 have problem with that but you can test it with women lying every where right?

[dry wave]

I'm not sure if it's possible to train loras for such a huge model on consumer hardware

[dry wave]

it was ja joke. I just wanted to say Flux can do human anatomy very well

[hallow lion]

I wodner when flux gets a slot on civitai

[dry wave]

it sometimes struggles with feets, but everything else seems perfect

[wispy spade]

hopefully they realease smaller model thought not all people that run such big model locally for any serious work

[hallow lion]

yes it can the hands of flux are magnificent

[dry wave]

yes, I hope that, too...

[hallow lion]

no more

[signal shuttle]

I living my life over here with two RTX 3090s

[wispy spade]

just get 5070

sooon

[signal shuttle]

nah i plan on getting the RTX 5090

[dry wave]

nah, I guess things will develop as they did with llms

there will be bigger and bigger t2i models

[wispy spade]

that good to but for common people 70s is a good middle ground

[dry wave]

and then somebody shows that a 6b model can do better than a 70b model with right training

[signal shuttle]

with the amount of companies producing open source text to image models, the scene is looking more and more like the early days of the Opensource LLM scene

[dry wave]

yes

which is a good thing

[wispy spade]

race to the moon. more player competing more better for us

[signal shuttle]

Like you have pixart and HunyuanDiT and auraflow and now you have black forest labs with their flux models, SAI has real competition now, they need to step up their game

and now black forest labs is even making a open source SOTA text to video model

[hallow lion]

dont forget Kolors

[dry wave]

I mean, SAI is a company, what counts (besides money) is the people behind it

[frail shoal]

168s for 4 step with schnell. I have a 6gb vram mobile gpu and 16gb of vram

[dry wave]

who is left on sai? All good people left

[wispy spade]

speaking about video model, hvnt really payng attention for this area. which model is the greatest right now?

[hallow lion]

video is not there yet

lol

come back in a year

[signal shuttle]

you have Kling and gen 3 and luma

[hallow lion]

its still to AI-y

but getting there

[signal shuttle]

My two favorite models are Kling and Gen 3

[hallow lion]

its impressive

but not there yet

[signal shuttle]

I find it best to do Image to video using Kling

[hallow lion]

yeah

but its not 100% still

[wispy spade]

that awesome something good doesnt come that easy, just looking at her finger precision 2-3mnts waiting time is worths it. as long the model very accurate with the prompt given you dont really need trying that much for good result

[hallow lion]

until we cant get at least 5--10 seconds of real consistent images its useless for anythign other than pschedellic imagery or stock footage

[wispy spade]

i will try them later

[frail shoal]

yes those were the first two generation with euler. Trying other samplers now, dpmpp 2m is my favorite so i think i can get better looking images with it

[wispy spade]

both of them run locally?

[frail shoal]

did you guys test the schedulers ?

[signal shuttle]

nah both are closed source

[wispy spade]

wait what webui you use to run flux?

[frail shoal]

right now in comfyui, will test later on swarmui

[hallow lion]

https://huggingface.co/spaces/black-forest-labs/FLUX.1-schnell

schnell here

[wispy spade]

thanks

[frail shoal]

168s is not that bad considering the quality the model outputs with only 4 steps

[signal shuttle]

For some reason schnell does paintings better then pro or Dev

[wispy spade]

come back 1 months after sd3 release and now learning it wasnt a hot potato any longer hahaha

[wispy spade]

agree

[frail shoal]

it does

[hallow lion]

i guess pro and dev are more for realism...

[signal shuttle]

which is weird, why does a small model outperform its bigger counterpart on something so specific and simple?

[dry wave]

😭
I only tried dev so far. i don't want to waste 50gb memory for both

[dry wave]

it's not smaller

[frail shoal]

fp8 variants are here

[hallow lion]

i downloaded both but only tried shnell

[frail shoal]

it's distilled more, who knows

[dry wave]

both dev and schnell are distilled model. Maybe schnell had more paintings in it's training data

[wispy spade]

wait this is not the model repo

[bitter hearth]

Now sd3 became flux channel

[hallow lion]

some youtuber did a live flux session and dev is muc nicer than shnell

[frail shoal]

does dev do good images with low steps like 4-8 ?

[hallow lion]

ah yes.... 😄

[dry wave]

dev needs around 10-20 steps I would say

[signal shuttle]

just like how the subreddit was taken over by Flux

[dry wave]

schnell is the fast one

[bitter hearth]

Reddit puke

[dry wave]

schnell=German word for fast 😉

[wispy spade]

schnell why i cant find this model i want to download

[signal shuttle]

here is the huggingface repo https://huggingface.co/black-forest-labs/FLUX.1-schnell

[frail shoal]

dpmpp 2m sampler on 4 steps. Not bad.

[wispy spade]

do i need all those 3 file?

[frail shoal]

a stunning artist depiction of a sexy woman made of the cosmos. Her body is made stars, nebulae and she is playing the guitar

fucked up on the prompt, but okay for a first test

the hands, OMG

[wispy spade]

why her upper body not ethereal with cosmos?

[signal shuttle]

I think you need the two files that end with .sft

[frail shoal]

4 steps dude, with a shitty prompt. Need to make a better prompt

[wispy spade]

but dam that guitar look good

[signal shuttle]

flux dev absolutely rocks, this is 40 steps

[wispy spade]

that trump look handsome

@signal shuttle wait are all flux model dev schell all 24gb in size?

[hallow lion]

its so amazing that like out of nowhere a model like this comes. its mind boggling how good it is. if it does get finetunes and controlnets its like perfect. i really am impressed

[signal shuttle]

yes i think so

I really want to download the schnell version but i cant be bothered to download it with my amazing 500Kbs internet speed

[frail shoal]

I'm affraid all other models are dead when compared to this. Perfect anatomy most of the time

[wispy spade]

have you try it with dept and canny controlnet? can it run

[hallow lion]

you could hear genuine amazement in Nerdy Rodents voice lol

[frail shoal]

download the fp8 version it's only 11. something gb

[signal shuttle]

Nerdy Rodent did a fantastic video on it

[hallow lion]

i havent tried it with any contorlnets or loras or even any workflows yet

[wispy spade]

dam even this feel slow for me

[signal shuttle]

that will take me overnight to download

[hallow lion]

yes

[wispy spade]

where ?

[signal shuttle]

your lucky it only takes you half an hour, it took me the entire night to download Flux dev

[frail shoal]

this model is the perfect gift, it isn't even Christmas

[hallow lion]

ikr!

[frail shoal]

dev or schell ?

[wispy spade]

i wont try overnight download it always went shit midway

[frail shoal]

https://huggingface.co/models?search=flux

[wispy spade]

schell 8fp that you talking about

[frail shoal]

https://huggingface.co/maximsobolev275/flux-fp8-schnell

[crude lily]

Can it work for flux with 3060 ti 8g ram

[frail shoal]

it works with mine 2060 6gb vram and 16gb ram

[wispy spade]

just slow buddy

but go for it

wash your feet with flux right now

[hallow lion]

i get 90 seconds 1024/1024 on 3060 rtx 12gb vram and 32 gb ram

[signal shuttle]

Flux dev gets me 50 seconds on 1024/1024 with 30 steps on a RTX 3090 with 128 gbs of ram

[frail shoal]

when the model loading is done and first image is genned. The other images take 160s each to gen in 4 steps

[crude lily]

Pretty insane after saw the example with flux

[wispy spade]

right anyone willing to share some workflow for flux schnell that i can try?

[frail shoal]

the initial time with model loading and first image takes 300s

[frail shoal]

comfyui has one in examples

[wispy spade]

have you try intricate design image with flux?

landsacpe

[frail shoal]

just tested it, will try other prompts after. Now i'm testing settings

[signal shuttle]

How long do you think it will take until the mods come in here and tell us we can't talk about Flux anymore

[wispy spade]

aerial view, never ending futuristic wall strecthing into horizon, intricate design, highly detailed

[wispy spade]

much better then dead channel

[hallow lion]

make a flux channel then! its your destiny

[frail shoal]

i hope they come and tell us don't talk about flux anymore because we have brought you sd3.1 and enjoy.

[hallow lion]

you can replace cascade channel with flux

[bitter hearth]

You can't talk about models other than sd3 in the sd3 channel

👍

Did it first

[hallow lion]

That and balls.

[bitter hearth]

balls are important

I need more

[wispy spade]

what is this

[bitter hearth]

My profile

[hallow lion]

lmao wtf

[wispy spade]

oh you write it by yourself right

[hallow lion]

:))))

[wispy spade]

🤣

[hallow lion]

looool

[bitter hearth]

exposing me

[wispy spade]

nah just clarification

just in case you are a wanted

[hallow lion]

he talks about balls everywhere so he gets banned

[wispy spade]

well... his taste is heavy

[wispy spade]

that ball is shiny

[sullen moss]

white cat in black hat on red dog in blue shoes

[bitter hearth]

Correct aspect I'm dumb

[wispy spade]

add black and blue spot for the cat and tiger spot for the dog

[hallow lion]

I must amdit these are some fine balls.

[bitter hearth]

Im out of dad jokes about balls

[wispy spade]

no worries take rest and come back again tomorrow with more balls... cough i mean with more vigor

[remote holly]

Can flux schenell generate images with 2000's aesthetics ?

[hallow lion]

AESTHETICS.

[wispy spade]

never try, have you try it already?

[hallow lion]

hey

[bitter hearth]

Go and touch the ball to see the truth

[hallow lion]

can you make the Death Star? its basically a giant ball XD

[remote holly]

Heastheticks

[hallow lion]

hehe

[bitter hearth]

I have no clue if sd3 knows that, or if I even know that

[hallow lion]

No ship that small has a cloaking device!

[frail shoal]

best samplers for flux ?

[bitter hearth]

@hallow lion you exploded

(flux) does it know what a death star is?

[hallow lion]

😏

[remote holly]

I just tested the distiled model , i am suprised he understand correctly the style

[wispy spade]

how far? 100%?

can you share

[remote holly]

Maybye 70%

[wispy spade]

lemme see and judge

[remote holly]

prompt :144p, vhs filter, image noise, low resolution, jpg artifact, , nostalgia, photo from the 2000s, a dark living room, shadows, lamps, sofas, overturned chairs

for a distiled model is too good

[wispy spade]

this is flux?

[remote holly]

yess

[wispy spade]

last prompt get ingnored

[remote holly]

the distiled model has less prompt understanding ?

[wispy spade]

pretty good

[remote holly]

i love this one

144p, vhs filter, image noise, low resolution, jpg artifact, , nostalgia, 1990s photo, a field next to the entrance, houses, streets, electricity pylons, tall grass, starry brown night sky, red moon, black clouds

how many gb i nedd for flux.1 dev model ?

😮

i will take one month to install it lol

[signal shuttle]

16gbs of vram with the fp8 version and at least 32 gbs of ram to run it somewhat fast and not having to wait 3 minutes per image

[remote holly]

i have 12gb vram 😔

[dry wave]

hm, I currently try the SCHNELL version

[hallow lion]

more than enough, how much ram?

[dry wave]

the results look so much worse than on DEV. I wonder if I do something wrong...

[remote holly]

16gb

[hallow lion]

i have 32 and comfy maxes it out

maybe youll max out too but last longer, you shoudl still try

[hallow lion]

Dev looks way better than shnell

So you are not doing anything wrong i think

[dry wave]

hm, okay, then I can delete it xD It's not worth the speedness

[remote holly]

i will try it

[signal shuttle]

Schnell is more for like artistic styles, dev trys to inject realism into every prompt

[frail shoal]

what is the lowest step to get something good in dev ?

[wispy spade]

what ist the problem brother?

[dry wave]

I always use 20 steps, but I think you can go much lower

[signal shuttle]

People recommend 20, in my case i use 40

[frail shoal]

so not worth it for me

[dry wave]

with 3090 I get an image in 20-30 seconds - which is totally fine for me

if I would have to wait over a minute I would agree

[signal shuttle]

I love using Flux dev

[dry wave]

what's the prompt?

[wispy spade]

do you have any solution brother? @signal shuttle

[dry wave]

I don't use comfyui for flux but I guess you have to download the T5 and CLIP text encoders

[frail shoal]

what about 12 steps with dev ? Can someone try it ? I want to see if it's worth downloading for me

[signal shuttle]

"Field next to entrance with houses and streets, nostalgic 1990s VHS camcorder footage, extreme low resolution 144p, heavy image noise, visible VHS tracking lines, severe JPEG compression artifacts, desaturated colors, grainy texture, blown-out highlights, electricity pylons silhouetted against starry sepia-toned night sky, oversaturated red moon, ominous black clouds, tall grass in foreground, lo-fi aesthetic, CRT TV scan lines, time-worn visual degradation, analog video glitches, muted palette, soft focus, light leaks, chromatic aberration" it didn't follow my entire prompt but it gave me something good so eh

[dry wave]

yeah, my feeling is that if you give it long prompts it starts to miss a lot of things

[signal shuttle]

The names in the node don't match with names of the files you downloaded

[wispy spade]

but there is no other files i can download on hugging face page for schnell'

[hallow lion]

I hear 20

[dry wave]

T5 and CLIP-L are the same you use for SD3

or PixArt, or SD 1.5

[signal shuttle]

did you missplace the files then?, because that error comes up when the file either is not named correctly or doesn't exist

[wispy spade]

i download this one i place on vae and another on checkpoint

[wispy spade]

i have t5

[frail shoal]

the other on unet folder

[wispy spade]

which one?

[frail shoal]

the big model

[signal shuttle]

did you download Clip_l and T5xxl?

[signal shuttle]

Wait is Flux not a DiT model but a Unet one?

[wispy spade]

here i place on check point folder

[wispy spade]

i dont know about clip_l that you refering to

[wispy spade]

wait so not on checkpoint folder?

[frail shoal]

it's a dit, just comfy that wants to keep everything on one node

[frail shoal]

no

[wispy spade]

this good?

[signal shuttle]

you need to download Clip_L

[wispy spade]

where to find this model brother?

[hallow lion]

clip goes into clip vae into vae and models into unet

[wispy spade]

t5 go to?

[wispy spade]

but there is no clip vae folder

[sullen moss]

https://www.reddit.com/r/StableDiffusion/comments/1ehqr4r/you_can_run_flux_on_12gb_vram/?share_id=DXXXEjafcpxYm4ZBc83GU&utm_content=1&utm_medium=android_app&utm_name=androidcss&utm_source=share&utm_term=1

[wispy spade]

i got it

[signal shuttle]

Is there a difference between the flux and sd3 clip_l encoders?

[signal shuttle]

to be safe just use this https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

[frail shoal]

yes

[wispy spade]

where should i place this?

[wispy spade]

this right?

[frail shoal]

that is not the correct one

[signal shuttle]

yes the clip_l one

[frail shoal]

you need t5 separate and clip l separate

[signal shuttle]

Download clip_l.safetensors and one of T5 Encoders: t5xxl_fp16.safetensors or t5xxl_fp8_e4m3fn.safetensors. Both are going into \models\clip

[wispy spade]

last one?

[frail shoal]

yeah if you don't have enough vram like me, or take the bigger model if you do

[signal shuttle]

Fp8 if you have low vram and low ram fp16 if you have a beefy computer

[wispy spade]

i see fp8 then i use laptop

[sullen moss]

This excitement around FLUX is amazing. Everyone is pleasantly surprised. I think most people will forget about SD3 2B now.

[dry wave]

what is a beefy computer? fp16 would be already 24gb. You would need two gpus to do something with it

[wispy spade]

12B and 2B

[signal shuttle]

well i have two rtx 3090s with a 128gbs of ram so i dont have a issue running it

[frail shoal]

i won't if they release a decent sd3.1

[dry wave]

I know xD I just say: with a normal setup you cannot run it without 8b

[sullen moss]

Could you please advise which LLM model is best to use with Comfy for prompt enhancement?

[signal shuttle]

open source or closed source?

[sullen moss]

open source

[frail shoal]

llama 3.1 ?

quantize it to 4bit and it should be decent

[sullen moss]

I understand that it's possible to switch the LLM to the processor?

[frail shoal]

yes

[signal shuttle]

Using gguf models with llama cpp or kobold cpp will always use the cpu

[sullen moss]

Thanks

[frail shoal]

you can use it with ollama

[signal shuttle]

Ollama uses the gpu doesn't it?

[frail shoal]

i think you can do both, but i thought he wanted to have llama 3 on comfyui

so ollama i'm afraid

i don't know if there is another way to have it inside comfyui

[sullen moss]

It's still interesting to know how many parameters DALL-E 3 has.

[wispy spade]

no matter how many it have with that kind of restriction i wont even bother looking at dall e 3

[signal shuttle]

i love flux

[dry wave]

in contrast to SD the text does not look like photoshopped

[wispy spade]

what the hell almost 10lit

[dry wave]

what gpu?

[wispy spade]

mobile gpu 4070

does it dont need negative prompt for flux?

[dry wave]

it does not have negative prompts

it does not have classifier free guidance at all

[wispy spade]

it dont take much time generaing but take a lot of time loading the model to gpu xd

[sullen moss]

I hope that soon we will see tunings and upgrades for FLUX.

[signal shuttle]

I hope this doesn't turn out to be too expensive or too restrictive to train, i just hope this doesn't turn out like deepfloyd's "IF" model

[wispy spade]

flux good with long prompt?

why it keep making small child for me

[signal shuttle]

did you use schnell or dev?

[wispy spade]

schnell

[signal shuttle]

What prompt did you try to use that kept generating a small child for you?

[wispy spade]

Image is a digital illustration of a young woman with a fair complexion and short, dark hair. She is dressed in a medieval-inspired outfit consisting of a white blouse with a high collar and a black ribbon tie, a dark green cape, and a high-waisted, knee-length black skirt with a lace-up front. She also wears black tights and white shoes. The character has a delicate and refined appearance, with large, expressive eyes and a calm expression. The overall style is reminiscent of fantasy or historical-themed artwork, with a focus on detailed clothing and a soft color palette.

just testing with civitai prompt

3 attempt all result is a child half her size

even when i add young adult it same

[signal shuttle]

give me a second, let me try that prompt

[wispy spade]

good luck

[signal shuttle]

i get this when using your prompt using dev

[wispy spade]

3 attempt on my side is half body shot

cant understand why it become small child thought

one from civitai with same prompt with flux

[signal shuttle]

i think it is because you used the term "Young woman" in your prompt, here is the image when i removed young

[wispy spade]

does flux have some filter?

[dreamy sundial]

love the aethetics of this prompt (sd3 8b)

[signal shuttle]

no it doesn't, i made some wild stuff with it so no

[signal shuttle]

i think it is just because your using schnell with a low step count

[dreamy sundial]

flux pro for comparison, meh

[wispy spade]

low step huh, i will try it with different step

[wispy spade]

how many step for this? i just 7 last time

[signal shuttle]

but the image is more coherent with flux pro then SD3 8B

[dreamy sundial]

hell nah

[signal shuttle]

who in the hell holds a sword with a stick?

[signal shuttle]

just use 28 for a small test

[wispy spade]

this is good

[dreamy sundial]

i loved the aesthetics, it can deffinetly hold the sword properly look (not cherrypicked)

[dreamy sundial]

thx

flux...

[signal shuttle]

can SD3 get a map of north America in the background right without it being a mess?

[dreamy sundial]

let me try, prompt pls?

[dreamy sundial]

you have to remember sd3 8b is still an old undertrained model

[signal shuttle]

a military emblem showing GLIF FORCE 1, empty white background, clean and uncluttered, military emblem, clear outlines and borders, do not show anything but the emblem against isolated white background

[dreamy sundial]

thx

[dreamy sundial]

first output, it's what you asked

[signal shuttle]

here is two of my favourite plane images made with flex dev

[icy drift]

The lines are so solid and consistent even with perspective. 😊

[signal shuttle]

take a look at my image and your image and tell me which is more coherent?

the text looks worse on SD3

[dreamy sundial]

mine follows the prompt better, yours added thinks you didn't ask for and sd3 is still an undertrained model

[signal shuttle]

followed the prompt better?, it gave me what i asked from it, a military emblem with Glif force 1 written on it, it just made the image look cooler

[icy drift]

I posted some tests including prompts in the SDXL channel. Testing for physics. I'd love to see how SD3 handles them. I was just comparing SDXL.
(I was running FluxDev fp8 on a 4090, 20 seconds per image. Not great, but fast enough to experiment.)

[dreamy sundial]

it added so much you didn't ask for

[dreamy sundial]

alr let me test some of these

[signal shuttle]

It add that much because i suck at prompting, if i had given it a better prompt with much more instructions then it would've given me exactly that, and i was using 1.5 guidance becuase i wanted to test out how low i can get

[dreamy sundial]

the problem is not your prompt it added unnecessary things

[signal shuttle]

of course it add unnecessary things, try using SD3 on 1.5 guidance scale and tell me it doesn't add unnecessary things

i was surprised it even gave me the text

[dreamy sundial]

@icy drift it struggles with mirrors, at least it got the text right

@icy drift can't do the dice, it does 5's not 6's lol, interesting

@icy drift

[wispy spade]

@signal shuttle right to left 28 step > 12 step

same prompt just removing word young

all child sigh

[icy drift]

Yeah you can see some reflection in the marble below it, "NIIKBOK" or something.

[dreamy sundial]

lol yeah

i still prefer sd3 8b for aesthetics, it's almost the same as mj 6 or 6.1 (you'll need an insane prompt tho)

sd3 8b

[icy drift]

The girl-reflection here is what I've come to expect from all models until FluxDev. It's the same girl with the same outfit, and they're standing side-by-side. But the lighting is flipped. Leftmost girl has light coming from right, and so does her twin. FluxDev isn't exactly photo-real, but it really is trying to do a sort of copy-paste flip. It surprised me.

[lunar rivet]

has anybody tried quantizing the flux transformer with quanto?
been running for almost an hour here now, not sure if stuck or churning

[icy drift]

I got a quant from HF, didn't try myself.

[dreamy sundial]

yes that's pretty good

[dry wave]

yes

[dry wave]

takes only 1-2 minutes for me

but you need high ram

[lunar rivet]

uhhh really? is that in qint8 or 4?

[dreamy sundial]

sd3 8b

[lunar rivet]

can you link it? I only saw the normal fp8 quant, I don't think that's done with quanto

[icy drift]

I don't even know what quanto is. Asking the wrong guy.

[lunar rivet]

okay nevermind, I'm guessing you mean the model from kijai then?

[signal shuttle]

Isn't quanto for llms?

[dreamy sundial]

sd3 8b

[turbid grotto]

guys, what do you think, is tensorrt for Flux possible?

[lunar rivet]

not only, hf made a blog post about quantizing diffusion models yesterday

[turbid grotto]

cause it is literaly free speedup

[signal shuttle]

huh, i need to read that

[dreamy sundial]

sd3 8b

[signal shuttle]

what was the prompt for this one?

[dreamy sundial]

A photo of an orange cat diving into a pool at the Paris Olympic games, with "Paris" written on one side and "2024" in a vibrant blue color with some black elements, Olympic rings taken with a Nikon d850 with a wide-angle lens, f/2.6 aperture, high shutter speed. The competition was filled with excitement as spectators watched from their seats. In front of them stood three high podiums, each decorated to represent different sports events like swimming or gymnastics. This moment captured a tense but joyous atmosphere.

even mj struggles with this one

another one

[lusty oyster]

Is SD3 better than DALL .E?

[dreamy sundial]

depends on which things

[dry wave]

SD3 is trash

[signal shuttle]

the closed source 8b? maybe, the opensource very bad 2B? absolutely not

[lusty oyster]

Show me some examples

Of good images

Also is two minutes long for gen time?

[turbid grotto]

flux first try

My hardware does not allow to cherrypick

or my patience

[sullen moss]

There's no point in arguing about it. It's pretty clear to me that FLUX is miles ahead of SD3 in every way.

[lusty oyster]

The Olympics is degenerate

And lame

YOU VILL WATCH ZE MOCKERY OF THE LAST SUPPER

[wispy spade]

hahahahaha

[dreamy sundial]

nice

sd3 8b

[signal shuttle]

Can SD3 do this? (Made with flux)

[dreamy sundial]

prompt?

[signal shuttle]

Minecraft screenshot of a battleship, pixelated 3D render, isometric view, blocky aesthetic, voxel-based graphics, retro gaming nostalgia, low-resolution textures, cubic architecture, digital sandbox style, vibrant primary colors, orthographic projection, 8-bit inspired design, procedurally generated seascape, minimalist game art, simplified naval warfare representation,

[dreamy sundial]

thx

[frail shoal]

btw flux works with the aligned scheduler

[signal shuttle]

schnell?

[frail shoal]

yes i have not tried the dev yet, but tell me if you do

[midnight oasis]

/credits

[signal shuttle]

Would you look at that camenduru made a IMG2IMG workflow for Flux comfyui

https://www.reddit.com/r/StableDiffusion/comments/1ei7ffl/flux_image_to_image_comfyui/

[dreamy sundial]

best thing i could get, i think your prompt needs a bit of working bc it's contradictory

[bitter hearth]

flux is the goat 💪

[signal shuttle]

Well doesn't that mean that flux understood the prompt better?

[dreamy sundial]

you specified for a minecraft isometric pixelated style, all of thesse things contradict eachother let me try repromting it

here's one with the isometric style btw

[icy drift]

Can we encode the image with the clip model instead and concat the conditioning? That way we can talk to the model about the image instead of just generating a variant.
You convert the image to tokens in the prompt. I know we've done it for other models before.

[dreamy sundial]

@signal shuttle reprompting it can do it much better

[signal shuttle]

That doesn't look like a ingame screenshot from Minecraft

[dreamy sundial]

• flux didn't get the naval warfare ship right

[dreamy sundial]

neither yours

[dreamy sundial]

it followed the prompt better tho

now this looks more like a vanilla minecraft screenshot

looks like sildurs shaders lol

[signal shuttle]

I have been using Flux Dev let me just switch to Flux pro and see how it compares

[dreamy sundial]

ok

[signal shuttle]

(Unrelated image but i made it using Flux dev)

[signal shuttle]

can you give me the new prompt you made?

[icy drift]

Flux dev couldn't do trading cards at all for me. That's just dataset though. Loras are perfect for missing data, they just can't improve the models basic IQ very much at all.
It might not have UI stuff or game stuff. Try some more prompts.

[dreamy sundial]

Minecraft screenshot of a battleship, 3D render, blocky aesthetic, voxel-based graphics, cubic architecture, digital sandbox style, vibrant primary colors, orthographic projection, procedurally generated seascape, minimalist game art, simplified naval warfare representation

[dry wave]

I really hope we will be able to train Loras on 24gb 😢

[icy drift]

Uhm... Hmm.

[dreamy sundial]

this is flux pro, pretty good

[icy drift]

Are there wave shaders for MC? That would be cool.

[dreamy sundial]

yes there's a wave mod, i think it's on the physics mod

[signal shuttle]

flux pro

[dreamy sundial]

damn these shaders look good look at those clouds

[signal shuttle]

3.5 guidance with 40 steps

[dreamy sundial]

another flux pro one

[icy drift]

I have 6 new physics problems to test when I get home. So exciting!
I haven't had to think up new tests since we got SDXL. Nothing ever passed my testset until Flux. 😊

[sterile pendant]

"This lets you adjust the guidance on the dev model which is a parameter
that is passed to the diffusion model."
https://github.com/comfyanonymous/ComfyUI/commit/eca962c6dae395cab1258456529030880c188734

[dry wave]

I currently use it in diffusers only. Did anyone tried if inpainting works in comfyui? Might be a reason to swicth

[dusky thistle]

it does

works surprisingly well too vs sdxl

[dry wave]

nice 🤩

[dusky thistle]

yup

[dry wave]

now, please make a few controlnets and ipadapters black-forest-lab! ;D

[dusky thistle]

i did some tests with 100% denoise and didn't end up with visible edges or anything like that

yeah i think ipadapter could be huge here

12b is obviously too much for us to train style loras on consumer hardware

if we could use the style ipadapter stuff here, it'd fill one of the biggest gaps this model has (artist styles)

[dry wave]

yeah. Although I still have hopes that training loras on the quantized model might be possible... We will see

[craggy crest]

yes

[acoustic kite]

text you didn't specify tends to be hilarious, this is no different 😂

enal teepee weack axy leack 😂

[sacred jewel]

How to use it?

It's different than the defualt guider node... no model input. I assume a different workflow?

[craggy crest]

you don't use it. it's added support for flux. you update comfy, and then if you use flux in comfy, it knows what to do with it

[brittle nexus]

Herrison Fiat?

[sacred jewel]

Hmmm... I am lost then 😦

[sacred jewel]

😄 😄 😄

It's a stolen prompt 😛

[craggy crest]

if you'e trying to run flux in comfy, be aware that it's HUGE - you're going to need more than 16g vram

[brittle nexus]

Nah. I'm running with 10gb. It just slow

[acoustic kite]

i am THIS CLOSE to losing any potential future job 😂

[remote holly]

Is possible to run flux.1 dev localy with 12gb vram and 32 ram ?

[brittle nexus]

Yes

[remote holly]

Lessgoo

[acoustic kite]

i get 1.19s/it with 16gb vram and 32gb ddr5 ram for reference

[mortal mesa]

whoa

[acoustic kite]

fp8 e4m3fn

[brittle nexus]

It's a whole different thing

[craggy crest]

i mean if you want any sort of reasonable generation time

[craggy crest]

how long do you want to wait for each generation?

[sacred jewel]

Me? you may have been tryign to reply to another 😄 .... I have a 4090 and 64GB RAM...

That being said, I did still have small issues with Photoshop running at the same time and also a Google Chrome tab eating up enough VRAM to put me over the edge. Ffux definitely likes a very empty 34GB VRAM 😛

[acoustic kite]

we're all cooked 😭

looks great

[craggy crest]

i was, yeah, discord must have jumped the text scroll and i got the wrong message to reply to - and it does. that's the current conversation topic on Latent Vision 2 discord

[brittle nexus]

It's relatively faster than SD3 since i don't need to generate 10 imagens to get 9 shit and 1 half-shit

[craggy crest]

if you're still having to do 10 images and only getting a couple that are good while using SD3, you don't have a clue what you're doing. i do at the most 4 and get exactly what I want, usualy in all 4

and they're high quality, and excellent images

[brittle nexus]

Yes. Whatever. I won't fall for your trolling

[acoustic kite]

woman laying on grass doing absolutely nothing but looking like she has blown up...

[craggy crest]

i'm not trolling you, you just don't know how to use SD3 correctly

[brittle nexus]

Generate something like this

[craggy crest]

i've only gone into great detail on this discord about exactly why it's 1. not just females, and 2. why that happens. so did lykon. you can search for it.

[craggy crest]

if you consider that good, you should probalby just stick with using sd 1.5

[turbid grotto]

I found out that there is not reason to use fp8 weight dtype, it only slowdowns generation and reduces quality on rtx 3060 at lest

[acoustic kite]

correct number of fingers, grabbing an object, belt isn't doing anything weird. it's good

[turbid grotto]

btw it is good at 512px

[acoustic kite]

if i use fp16 it crashes

rtx 4070ti super 16gb vram, 32gb ddr5 ram

[brittle nexus]

Can't do, right?

[acoustic kite]

absolutely dies and crashes

[turbid grotto]

Shouldn't be like that, here is mine:
rtx 3060 12gb vra, 32gb 3200mhz ram,
flux.1 dev, 1024px, 20steps
100s generation, 60s text encoding
weight dtype - default
fp8 version of T5 encoder

[acoustic kite]

like this? 😄

[raven fern]

gib 😮

[turbid grotto]

yea

[acoustic kite]

thanks i'll try it - i fear the worst though, normally it crashes instantly with no error code

[turbid grotto]

do you have a pagefile? because it maxes all my 32gb, pagefile could be a savior here)

I remember comfy completely colapsing when I had 16gb without pagefile on sdxl workflows

[acoustic kite]

that might help - what is it? 😄 it also maxes out my 32gb

as you see here it just crashes instantly

[raven fern]

weight dtype to fp8, did you try?

[craggy crest]

not going to bother with YOUR trolling. go back to 1.5, which is what you're good at.

[mortal mesa]

you are mean and usually wrong also

[acoustic kite]

that's what i did before and it ran fine but cds said this: "#🆕｜sd3 message" was better

[brittle nexus]

You're not good even at trolling

[raven fern]

yea but you have to make it better for you...

[sacred jewel]

CFG 15

[turbid grotto]

check how to add pagefile, it might really help

[sacred jewel]

CFG 7

[craggy crest]

i wasn't trolling. i stated facts. you, on the other hand, really need to learn how to use the program before you complain it can't do what it very obviously can do

[dry wave]

dude, just shut up

[tough oriole]

Anyone know how many images Flux trained on?

[acoustic kite]

thanks, do you know what i should search for in order to find it? i'm not quite sure where to add it or what it should look like - thanks 🙏

[mortal mesa]

listen you cant use Flux with 16GB, your gonna need more, one of your facts, mreanwhile people using it with 8GB

[turbid grotto]

yea but it is possible with default on 32gb and 12gb vram, so no quality decrease and time increase

[brittle nexus]

If you can do a good hand holding a sword with SD3, why don't you just do?

[raven fern]

oh wait i just noticed now, you also using upscale on top of that, lol

[acoustic kite]

facts without evidence are opinions

you haven't really proved anything

[craggy crest]

becuase, child, i don't have to. go back and play with your toys

[craggy crest]

and you stay out of this

[sacred jewel]

https://tenor.com/view/trump-gif-26016517

[raven fern]

hmm il try tonight, im still testing combinations

[dry wave]

I mean, they are the same guys who made SD3, so I think training sets might be similar

[craggy crest]

this channel is packed with a massive number of images all doing exactly what he says SD3 cant' do, and are proof enough that he doesn't know what he's doing. just scroll through it if you want that proof

no one's going to play your little game and go run off and make you proof you will then diss and complain about

[mortal mesa]

people dont emulate you because they dont want to have outputs like you...

i know i dont

[bitter hearth]

That face on the left lmao

[turbid grotto]

check something like "how to set paging file windows 10"

[raven fern]

did you try flux? :3

[bitter hearth]

Only a bit

[karmic gale]

can't wait for 8b sd3

[acoustic kite]

oh okay ty 😄

[hollow apex]

Anyone know how much vram it takes for flux?

[turbid grotto]

all

[hollow apex]

As in not needing a work around

[dry wave]

16gb at least. Some people say they can do it with less

but it will take much time then

[craggy crest]

the discussion on L2 discord is that you want more than 16g vram if you don't want to wait quite a while for each gen. it's huge

[hollow apex]

Perfect, 4080 should work then

[hollow apex]

Ah shit

[turbid grotto]

yea full quality dev version runs on 3060 for 100s and text encoding is about 60s on my poor r5 5600

[hollow apex]

Here I was proud of my fast sdxl gens 😦

[hollow apex]

Is it worth it over sdxl?

How uncensored is it

[raven fern]

wait... i actually didnt try,... do you need the t5? what if you just use the clip_l, is it gonna work?

[dry wave]

it's much better than SDXL in text understanding and anatomy

regarding "anatomy" (or in general: structureness of the image) it's the best model I saw so far

like hands are almost perfect most of the time. Hands holding stuff works at least in majority of the cases

[sterile pendant]

put it between the clip text encode and the basic guider.

[dry wave]

background looks clean and not messy

[hollow apex]

Serves stability right, hope black forest doesn't do the same shit this company did

[signal shuttle]

there's people on twitter saying 3.1 will be better then Flux, and sharing images of SD 3.1 wip model

[hollow apex]

How are they?

[signal shuttle]

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

[bitter hearth]

Twitter

[dry wave]

what the model is not good at is paintings, or styles in general

[turbid grotto]

well will see in 2 weeks

2.1 weeks

[dry wave]

like it can do anime and comics, but if you ask for specific styles or artists it is usually not good in that

[hollow apex]

Yeah hope we get some good fine tunes

[turbid grotto]

is there any info about finetuning Flux?

[hollow apex]

Schnell is allowed I think

[turbid grotto]

I am worrying that it is not possible

[turbid grotto]

not dev?

[dry wave]

the typical SAI bullshit as usual

[signal shuttle]

Both Dev and Schnell are finetunable, its just with Dev that you can't sell said finetune

[dry wave]

like none of the images is good for assessing model performance. Probably all of them are cherry picked anyways

[acoustic kite]

schnell at 512x512 is manageable, i get 2.7it/s with 16gb vram and 32gb ram

[raven fern]

good one

[dry wave]

we can make extremely good looking images with SDXL or even with SD 1.5 that look as good as Flux or Midjourney images. But as soon as we ask for something simple as a knight holding a sword these models fail miserably

[signal shuttle]

or use said finetune to train a model that is competitive with Flux

[dry wave]

so if anyone want to convince that his model is good he should show images that are challenging for existing models

[icy drift]

Uncensored tests I did yesterday. I don't consider these NSFW, they're the tamest stuff possible. But it's the first model in ages that I consider uncensored. (And this is plenty for anything I want out of a model.) Obviously I added the red blocks. It was the only anatomically perfect David I have ever seen.

[hollow apex]

Does civit count as selling?

[raven fern]

woman lying on grass is gonna be the first test from now on LOL

[bitter hearth]

try gymnast doing a headstand, usually upside down stuff confuses the models

[raven fern]

@icy drift did you try tcg cards 😮

[dry wave]

problem with Flux is its so huge it will be difficult to finetune... I mean, in theory it should be still possible if you train a lora only on the last k layers of the model or similar. But yeah, I guess some people will have to implement really efficient finetuning scripts for that to make that happen

[icy drift]

Yes not in the dataset at all as far as I can tell. It can do the art, but "TCG" and related terms do nothing. Loras needed.

[bitter hearth]

perhaps the next GPU gen will feature 128gb vram and none of these concerns will hold true

[frail shoal]

pictures of dogs and landscapes were not the reason sai got backlash. Show us humans in different poses holding swords and other things but of course also, a woman laying on grass.

[hollow apex]

Btw what size is sd3

[raven fern]

ah yes 128gb vram, so we can finally run those 400B llm models LOL

[bitter hearth]

i feel u

[pseudo owl]

2b(just the mmdit part), the text encoder is pretty massive tho thats like 3b i think

[bitter hearth]

Llama 3.1 405B, the largest model in Meta's Llama 3.1 family, is approximately 854GB in size when using full 16-bit precision

[raven fern]

damn son

[pseudo owl]

just have to use 1 bit quantization(:

[bitter hearth]

use 0 bit quant

any model fits

[pseudo owl]

yes all problems solved

[bitter hearth]

Black hole happens

[pseudo owl]

run flux, 405b on cpu

[raven fern]

pc gonna explode

[turbid grotto]

llama 400b as text encoder 💀

[bitter hearth]

All that to write "Imagine a serene scene"

[turbid grotto]

oh is this possible to quantize T5 even more? It takes too long on my cpu

[bitter hearth]

flux is pretty okay at doing headstands

[raven fern]

generation so slow, you get one word per year

[dry wave]

first try, not cherry picked

[bitter hearth]

whats that game where you have to put hand and feet on different colors in a group

[sterile pendant]

The diminishing returns though lol....

[raven fern]

twister?

[bitter hearth]

yes!

that foot tho

[raven fern]

nice as..thetics

[pseudo owl]

yes 4bit, you can even remove it if you are using sd3 but quality will suffer(slightly). You can remove it with flux as well but quality will horribly drop i think.

[raven fern]

hahaha 4gb vram.. are you ever gonna upgrade? 😦

[bitter hearth]

twister in flux