🆕｜sd3

[bitter hearth]

and cat walks

[hallow lion]

https://huggingface.co/Kwai-Kolors/Kolors

[bitter hearth]

not enough vram to view that link

they are working on next version of sd3

not sure if that can compete with flux

[hallow lion]

Luckily you can usually get away with a bit less than the minimum requirements... Especially with Comfyu..

[bitter hearth]

lot of sd3 team went to create flux

[bitter hearth]

Me in the morning

[bitter hearth]

Sd3 is better than Flux x3

[bitter hearth]

thats how you fail math

[hallow lion]

SD3.5 will kick huge butt. It's always every other model that SD makes that is legendary.

[sacred jewel]

Whatever it is it's scary LOL ...

You will NOT believe the prompt: 😛

tony hawk using trains as skateboards

[bitter hearth]

Flux just seems better at anatomy idk

[torn wharf]

do you guys ever think about the drake equation and how our observations don't match the predicted expectations, giving rise to the fermi paradox? they say there is a great filter that is causing the fermi paradox but we don't know what that filter is. somethings going on cause the math that math says so

[bitter hearth]

Nerd talking about math again

[torn wharf]

you guys ever freak out about the math?

[bitter hearth]

I didn't like it in school

[hallow lion]

SD 1 meh SD 15 amazing SD2 ugh SDxl wow SD3 ugh and the next one will be outstanding

[sacred jewel]

Now look at this seed...

Literally

A Hawk
A Train
A Skateboard

Can't fault it... it is good 😄

[torn wharf]

i programmed my graphing calculator to play snake in school. maybe shoulda paid attention

[bitter hearth]

Tony learned math and this happened

You won't believe it

Click now

[frigid saffron]

Hi, everyone, I have problems on flux

I find Flux is bad at multi-views for background
When I generate a landscape or city or background, I failed to control any viewpoint using like"from distant view/ bird's view/ from blew/top view/", Is this happening to all of you?

[bitter hearth]

algebra is fun but trigonometry gives me headaches

[frigid saffron]

For example, Prompt like: "It's a vibrant Japanese town with traditional Edo period architecture, with narrow streets lined with wooden buildings and tiled roofs. Blooming cherry blossoms against a blue sky add a touch of pink to the scene. At the center of the image is a tall, multi-story pagoda that looms large in the background, evoking a sense of historical grandeur. The image captures the serene atmosphere of spring, with falling flower petals scattered along the streets and large red lanterns hanging from the facades of buildings, making for a charming and picturesque cityscape. The whole town can be seen from a high altitude"

[bitter hearth]

I have a theory for why only the apple was weird
the steps and schedule you had
was giving steps that were too large during a range of the sigmas where the ODE was likely to have a large local discretization error
and the apple had less straight trajectories throughout that sigma range
because the relationship between trajectory straightness and sigmas range can vary based on the composition
notably the apple is a much simpler composition than the others
you could test my theory by running the workflow with ten times as many steps

[frigid saffron]

flux can't give me a high altitude

[bitter hearth]

if my theory is right then running the workflow with ten times as many steps will fix the apple

because it was a case of local discretization error

[bitter hearth]

Lmao

[frigid saffron]

How you control the viewpoint using flux?

[bitter hearth]

when it comes to anatomy flux has it's flaws too, i think the most aesthetic you can get with anatomy is with pony models

Cx

[torn wharf]

true nature of reality

[bitter hearth]

nice big balls

[bitter hearth]

Try putting the view you want at the beginning of the prompt

[bitter hearth]

Finally some good images in this chat

[craggy crest]

you just don't know how to prompt, that's the problem

[bitter hearth]

you are too absorbed with math to understand the basics of prompt

[frigid saffron]

No effect I try it.....like birds view

[craggy crest]

you're also not goign to get a fight started tonight

[bitter hearth]

not looking too, i dont like cat fights

[craggy crest]

@bitter hearth

[bitter hearth]

Now THIS is how you skateboard

[frigid saffron]

@bitter hearth can you change the viewponit using this prompt on flux? #🆕｜sd3 message

[bitter hearth]

Ugly

[bitter hearth]

you talked the same thing about sd3 as long as you could even tho everyone got deformed junk with sd3 and anatomy

[bitter hearth]

flux however is much better but has its flaws

[craggy crest]

go away child. go play with your ponys

[frigid saffron]

#🆕｜sd3 message, I just can't change the different view using this promt on flux

[bitter hearth]

are you democrat, you talk like one

[craggy crest]

drone flyover footage, clouds below

[bitter hearth]

missing logic most of the time, i wonder what kind of math you actually do

probably the kinda math that goes into sd3

[craggy crest]

go away child. go play with your ponys

[bitter hearth]

I feel like flux is just bad at stuff like that, I also had problems trying a simple prompt and gave up

flux isn't as bad

flux can mess up some of the anatomy but not horribly bad

[frigid saffron]

but that's not a general view control prompt i think..

[bitter hearth]

you atleat would have to cherry pick good ones with flux but with sd3 is a surge of mess one after the other

I tried a few times on flux with that prompt with the view from above at the start of the problem and it doesn't do anything

[frigid saffron]

Midjoruny is good at view control

but flux seems no concept

[errant dust]

When people resort to personal attacks it is because they ran out of arguments

[sacred jewel]

I broke it...

prompt: motion blur

[bitter hearth]

yep

[errant dust]

You are the one who resorted to this

[frigid saffron]

also about viewpoint control?

[bitter hearth]

that's the reason im not bothering to respond to him

not me you missed the line of convo then

[craggy crest]

yes it is

[errant dust]

No,,,,, I don't think so

[errant dust]

really not

[bitter hearth]

scroll up then and im not interested to talk about it, shut it pls

[errant dust]

I rest my case

[bitter hearth]

dont insinuate like a fool

[errant dust]

and you just keep burying yourself don't you

[torn wharf]

all these balls and you gonna tell me math isn't real? i unno man. i just dunno

[bitter hearth]

if you keep your mouth shut that wil help

I feel like in the long run the captioning for image model training data needs to be done by a VLM that was made specifically with that task in mind
cos a lot of image models are using something general like CogVLM for the captioning and that's not really ideal

or talk about things that are not nonsense

[errant dust]

Assuredly mister personal attacks

[bitter hearth]

ok time to block you two

[craggy crest]

he's a 12 year old child that's really not old enough to be on discord

[errant dust]

you do know that blocking anyone on Discord is an utter waste, right? You will instead get drowned with messages saying 'Message Blocked'

[brittle nexus]

https://github.com/ataylorm/FluxAIGridComparisons/blob/main/CameraAngles/FullGrid.jpg

[torn wharf]

aw man i was gon let you play with my balls but then you went an made a comment like this. such a shame. you're cut off now

[bitter hearth]

gees hounds

[torn wharf]

no no . deez balls

[errant dust]

@sterile pendant SOmeone posted the combined file Dev+Schnell in Civit, and I must say it is a total winner

[sterile pendant]

alright im back, got side tracked with stuff around the house. the second image has the workflow. it will need a custom node

as you can see, you don't need the second model to load for the sigmas

[bitter hearth]

Models are being stupid for me can't get higher view than this

[craggy crest]

i would agree with that except: here are the full range of heunpp2 tests, all the same prompt, seed, settings, except all the different schedulers

[torn wharf]

so i love that all these models can create really destructive scenes now a days. but i've noticed people and crowds are just business as usual. my proposal, put a traffic control person in the image then it makes sense

[frigid saffron]

That's it, bird's view also no work. I try to think that flux has another special way to control viewpoint.

[craggy crest]

if it's higher, it can't draw that temple correctly

[bitter hearth]

so the test would be to run each of these with 10 times the steps

what that would do is specifically test the hypothesis that the issue is local error from going from continuous to discrete

[frigid saffron]

Here is the same prompt from midjourny that successfully control the viewpoint:

Why flux can't do this

[bitter hearth]

running a very high step count simulates having a perfect schedule at a lower step count
because it a very high step count gives you small steps everywhere

a perfect schedule at a lower step count would have small steps only when needed, and large otherwise

[craggy crest]

midjourney is a different AI - just becsue it understands something one way, doesn't mean other models do. and the next version of midjourney might not either

[bitter hearth]

I removed everything but from above and the temple lmao

[craggy crest]

and that makes sense

[frigid saffron]

.........So original prompt have something ban the view control?

[bitter hearth]

when the step count goes high enough all schedulers trend towards the same output
as discretisation errors trend to zero

[sterile pendant]

assuming they are in the same family of samplers, yeah, usually

[craggy crest]

unfortunately, everyon'es trying to get steps down to 0, or at least no more than 2

[bitter hearth]

0 steps 🤔

[craggy crest]

you know they'd love to

[bitter hearth]

haha yeah they would

if it's ever possible to use the flux algorithm / or some lora to fix hands/fingers with other sd models that would be fantastic

[bitter hearth]

Maybe you need to shorten the prompt so the model has more focus

[sterile pendant]

training datasets and things like supervised finetuning

you need a LOT of data for hands

[bitter hearth]

Millions!

I did spend a while this week using Realvis Lightning
which is one of the strongest out of the low step models
its actually kind of okay

[sterile pendant]

probably in the tens of millions range, honestly

[frigid saffron]

I'll test it now, insteresting, even put the view control at the begining doesn't work

[bitter hearth]

Then with flux... It's just a bit dumber than sd3

[craggy crest]

you should pay more attention to youtube

[bitter hearth]

Realvis Lightning at 4 steps is good enough for maybe 99% of AI images I have seen

[torn wharf]

see. now it makes sense that people are chill

[craggy crest]

godzilla looks so confused

[bitter hearth]

Even Godzilla stopped

That man is powerful

[sterile pendant]

think about all the locations on the screen hands can be, now think of all the various distances they can be from the camera, all the different ways they can be obscured by things like being in a pocket, then think of all the infinite combinations of positions/angles they can be in, all the potential combinations of individual finger positions, then all the possible textures and skin colors, the list goes on and on

[torn wharf]

i prompted Joe Biden

[frigid saffron]

I thought the successful image is from flux, but from sd3?

[bitter hearth]

UniPC 10 steps with beta scheduler is also not bad most of the time

or maybe 12-15 steps for safety

[bitter hearth]

I mostly use sd3

[craggy crest]

sd3 2b medium is excellent - if you bother to learn how to actually use it

[craggy crest]

we know where torqx's mind is, don't we?

[bitter hearth]

Booba

[craggy crest]

12 year old kid

[bitter hearth]

at low steps PAG goes a long way

PAG is the most impressive node to me

a tiny drop of PAG has fixed images perfectly sometimes

[sterile pendant]

oh and i fucked up and didnt link the image correctly, think i just copy pasted instead of uploading the actual file

this should have the workflow to show it all

[craggy crest]

okay, thanks

[bitter hearth]

makes sense cos discord normally does preserve the comfy data

[craggy crest]

what is pag?

[sterile pendant]

main takeaway is that you dont need to load the second model to get the sigmas when they are just a simple y=mx equation that ranges from 1.0 to 0

[sterile pendant]

perturbed attention guidance

[bitter hearth]

something about white girls

[sterile pendant]

it's basically tangental to self attention guidance

[bitter hearth]

perturbed attention guidance
you can just set it at 0.3-1 on every image TBH

[errant dust]

@sterile pendant https://civitai.com/models/628862/flux1-schnell-merged-with-flux1-dev?modelVersionId=703068

[bitter hearth]

or up it to 1-3 if needed

[craggy crest]

on the snarky subject of Flux not able to do hands

[bitter hearth]

SAG is also great, again at low strength like 0.2

[sterile pendant]

i saw they got PAG working with SD3 on diffusers. doesn't quite work the same way, but they get some interesting results from it

[bitter hearth]

there was a very very recent comfy change
to make PAG and SAG work below 0.1
so it can be used with CFG++

[craggy crest]

okay, how do you use it?

[bitter hearth]

yeah diffusers gets a lot of stuff first

if you just put the PAG node in before ksampler
and set it to 1
that will do it

[craggy crest]

comfy is releasing a new node, did you see it?

[bitter hearth]

no I haven't seen it

[torn wharf]

like no shit the crowd is calm. look at all those traffic control devices out there. just another day in nyc. they got it handled.

just a quick hack if anyone wants to use this for their images.

[craggy crest]

sorry, not comfy, matteo - he posted it last night

[bitter hearth]

ah I think I saw it
a ksampler

[sterile pendant]

if you're using it with sdxl, you just set pag to something like 3.0 and lower whatever CFG you were using by 3. basically, if you were using a cfg of 7, you'd lower it to 4. 4+3=7. you can also use the advanced version of the node and set rescale to like 0.5-0.7(personally, i do it and seem to get better results). there are some pros and cons to pag, but try it out and experiment

[craggy crest]

he said people were goign to hate it

[bitter hearth]

lol

[bitter hearth]

Hes rich

[craggy crest]

cool. i'll play around with it, thanks

[bitter hearth]

I hate the blur

Who invented blur on cameras

3 PAG is strong
that's not always bad
I mostly use more like 0.3

high PAG has a certain "look" to it, which you may or may not like

[craggy crest]

i'ts used to make the actual focal point stand out

[bitter hearth]

Bad, I just want clean images

this one was very high PAG for example

it can start to get pretty cooked

[torn wharf]

do you see how it immediately changes the narrative of the image from "why is everyone chill?" to "oh i see why they're all chill" it's just a little hack to remove distraction from your action images and make them more believeable.

[bitter hearth]

if I remember rightly that image had something else weird in it
either high CADS noise or FreeU with low B and high S

[sterile pendant]

note the two different CFGs on the samplers

[bitter hearth]

FreeU with high S was a major source of funny proportions

you could probably tonemap these a fair bit to remove the PAG+CFG burn

[sterile pendant]

oh they aren't burned, it's just hte prompt

point is, target cfg=pag+cfg. you can also throw in automaticcfg into the mix as well

[torn wharf]

you could totally pair this pag stuff with the TCP hack i described for ULTRA believeable images

[sterile pendant]

same settings, pag on the left, normal on the right

[bitter hearth]

I'm just obsessed with the low CFG look its a personal thing

a bit of ComfyUI_experiments/sampler_tonemap_rescalecfg.py can help

[sterile pendant]

f that, ive been photoshopping for 25 years lol

if i want to make adjustments, i just do it in another DCC

[bitter hearth]

its in a node though

you just hook it up

[sterile pendant]

oh i know, just 1000x easier to adjust things like levels in photoshop and see the image preview change in realtime

[bitter hearth]

ok but in that case
you would still want a rescale cfg node
just not the tonemapped version

rescale cfg does its operation each step

so it can't be replicated in photoshop

its just this:

2 # Apply regular classifier-free guidance.
3 cfg = neg + weight * (pos - neg)
4 # Calculate standard deviations.
5 std_pos = pos.std([1,2,3], keepdim=True)
6 std_cfg = cfg.std([1,2,3], keepdim=True)
7 # Apply guidance rescale with fused operations.
8 factor = std_pos / std_cfg
9 factor = rescale * factor + (1 - rescale)
10 return cfg * factor

so step 3 here is just normal cfg

[sterile pendant]

the middle one is pag+automaticcfg. it's my goto setup for sdxl

[bitter hearth]

photoshop can replicate changes made at later sigmas
like colours or textures or contrast level
but rescale cfg will also affect early sigmas that decide the layout

this lora doesn't seem to make any big difference from what the flux model can already deliver

[sterile pendant]

i usually use the advanced version of the node though and tweak some shit

[bitter hearth]

automatic cfg has some rescaling in anyway if I remember rightly

[sterile pendant]

yeah it does

it scales everything per setp

to ensure good ranges

but your cfg is no longer your cfg basically

[bitter hearth]

the reason I started using thresholding/rescaling is that for some reason the DPM++ paper mentions it a lot
and says it really helps

[sterile pendant]

the dynamic threshold node is a good alternative to automatic cfg as well, but they basically do similar things

[bitter hearth]

yeah its a similar idea

[sterile pendant]

main difference is that you can set your mimic cfg target to like 2, 4, 8, etc

[bitter hearth]

he wrote a new one BTW
this week

skimmed_CFG

[sterile pendant]

ill have to check it out

oh right

i told you about it lol

[bitter hearth]

lol ah ok thanks

I get quite a lot of things from you like the sigma graphs

[sterile pendant]

havent tried it out yet though, looks cool for stylized stuff

[bitter hearth]

and I got FreeU from crystalwizard

I haven't checked if skimmed_cfg works with CFG++

yeah its a shame cos I love FreeU

[sterile pendant]

well i was just saying how they added PAG to SD3 on the diffusers repo, so im sure someone can make a comfy node for it

[bitter hearth]

I can get FreeU effects a bit using CADS
for example this was SD3 with high CADS
and it looks a bit like FreeU with high S1 S2

[sterile pendant]

dont know how well SAG would work with sd3 though

[bitter hearth]

if I remember rightly SAG is basically blurring part of the image and then sending that to the unconditional
should be possible in any model

not flux though

[sterile pendant]

oh i know, but with transformers, you can also do some similar shit

they are layered and blocked as well

https://github.com/huggingface/diffusers/pull/8861 they already have, just needs to be ported to comfy. but again, it operates and behaves differently

[bitter hearth]

there's a new paper also that is meant to be an upgrade from PAG

https://arxiv.org/abs/2408.00760

[sterile pendant]

ah yeah, i saw that one the other day but didnt have time to read it

[bitter hearth]

and this one is meant to be an upgrade from CADS
https://arxiv.org/abs/2407.11555

[torn wharf]

if thats what they want i mean who are we. who are we really?

[sterile pendant]

flux is also really good for img2img like if you denoise at like 80-90%. you can sometimes correct f-ups from sd3 with it

no, wouldnt do that

not losing the 16 ch vae

you could do a really convoluted sd3->flux->sdxl refine->resample at like 30-40% with flux again though

[bitter hearth]

on reddit someone did this order:
flux -> sd3

[sterile pendant]

to gain back the benefits of the 16ch vae

that would work as well

sd3 is hella good at details

[fleet meteor]

Do anyone have a workflow for flux + a lora loader?

[bitter hearth]

if you VAE encode and decode a photo
SD3 VAE is better than flux VAE

[sterile pendant]

flux or sd3 depending on what youre doing

if it's humans, flux

if it's art, sd3

[bitter hearth]

best VAE last

[sterile pendant]

both flux and sd3 have amazing vaes

[bitter hearth]

yeah

its probably situational

[sterile pendant]

but the last pass would be a 30-40% resample to fix whatever got broken in an sdxl refinement stage dropping down to 4ch vae

[fleet meteor]

I only saw this one, but they used cli in the example (not comfy ) https://www.reddit.com/r/StableDiffusion/comments/1emrprx/feel_the_difference_between_using_flux_with/?sort=new

[bitter hearth]

also if you have too much noise injections then SD3 3B can't handle it
so for noise injection workflow it has to be flux

[sterile pendant]

i was messing around with a 4step schnell->sd3 workflow earlier and getting some really good results, but then again, i don't make people that often

[bitter hearth]

SD3 is still my favourite TBH I have gone off flux
this is SD3 2B

[bitter hearth]

you can use any lora loader to load up a lora they released altho there is no big difference

SD3 2B (not even the current 8B) is the closest to a photo

8B will win eventually though after fine tuning

@fleet meteor here is a workflow you could try ...

[sterile pendant]

I know I do, but I always add that shit in post if I want it

[craggy crest]

wrong color hoodie

[fleet meteor]

Thank you! I´ll try it 🤗

[bitter hearth]

don't push your narrative on others, everyone can try things the way they like .. you narcisstic freak

[torn wharf]

You dont want to use any of these models for text??

[viral grove]

/Tide play, Green theme, Doll personality traits and accessories according to MBTI personality, Make an INFJtrendy toy plush toy, Cartoon face made of silicone, The body of a plush toy, Short white hair, White goat huzi, Round eyes, Dense black silhouette, Simple nose, Simple mouth, Minimalist facial features, The body is the plush body of the toy koala, Wearing a robe, Lovely, Holds a staff in the shape of lightning, Quiet and mysterious, An inspiring and tireless idealist at the same time, Indicates the MBTI personality type infj, diplomat, Advocates, creating a simple yet captivating style. This sticker perfectly captures personality using vibrant colors and adorable elements, High definition, oc, c4d, 3d, blind box toy style, rococo style, Bauhaus, god rays, High quality --s 400 --niji 6

[tough oriole]

Anyone try the Control net for flux is it safe? How much VRAM added?

[bitter hearth]

its a safetensor file, but doesn't load up on comfyui, tried with every loader, might need a node update from comfy

[torn wharf]

https://github.com/comfyanonymous/ComfyUI/pull/4260 good info about it here

[tough oriole]

Thxs

[torn wharf]

as for vram, i'm not sure. gonna wager a guess. its' 1.5gb and probably in fp16 mode so i'm thinking comfyui might take either another 1.5gb or 750gb.

[bitter hearth]

its the colours mostly
for example look at the red on this hand

SD3 looks more natural to me

[sage burrow]

Ty, I was wondering how tp try that 🙂

[sage burrow]

Perhaps it's supposed to be flux chnell or something (didn't work for me)

[sage burrow]

Not nearly enough for this server!

PS what was your prompt? 😄

[torn wharf]

oh if you wanna use that you need to get that specific pull request merged to your side. it's not merged yet.

[sage burrow]

SD3 is pretty awesome 🙂
I like both though

[torn wharf]

i haven't tried using it yet

xlabs_flux_controlnet would be the branch it's on. i'm unsure how to swap git branches in swarm environment

[sage burrow]

they put out a model sans workflow?! So mean! lol

[torn wharf]

a company called xlabs put it out with inference code and the diffusers team has code supporting it now. comfy is testing their code. thats what i think is going on. there is a workflow for it. you just need the right branch

they said tommorrow is the 1024 x 1024 version

[bitter hearth]

yeah its out in diffusers properly already

generally if you want the latest stuff you need to be able to use both comfy and diffusers

hi-diffusion came to diffusers first for example

[torn wharf]

Guys I just realized something just now and I have to tell people

Barclay from star trek is Murdoch from ateam

That's all thanks for listening

[bitter hearth]

wow I had no idea

[bitter hearth]

that's actually legit

his face looks different lol

[torn wharf]

i saw him in another movie about oppenheimer and i was like .. you know? he kinda looks like murdochhhhooh wait one second

i got it working. used the jump man logo. not much quality above 512 but it works great at this res

just as i thought. the claims i seen about the model not even guiding well are just wrong. as usual. cool.

"it's 12 billion parameters it'll never work it's outlandish!"

[rotund vine]

/style prompt a cat

[torn wharf]

source for the canny guidance. i run it through the filter for the outline but thats about all it is

tommorrow it'll be the 1024 pixel model and they have others on the way. neat

[strange grotto]

does anyone know if cascade has 16ch vae?

or 4ch vae?

[hollow swift]

can you do elepahant and rat

[torn wharf]

cascade's vae is stage a and b as it decompresses the latents through a diffusion network instead. i don't fully get it. stage a i guess is the final leg of that model

[strange grotto]

So it's special? Can't compare it to sdxl or flux?

[torn wharf]

i'm just not sure how it compares. maybe its like a 16 channel one. it looks good. i think it works different though

[hollow swift]

wow crazy man

[torn wharf]

with just the one guidance net theres a ton of potential . it does poses well

if even just with a sillohuette of the shape for the first 15% of steps

@sage burrow all i had to do was open a terminal in the swarmui's backend comfyui folder and type git checkout xlabs_flux_controlnet and i was set. THOUGH, i am going to monitor this and when it merges i'll run git checkout master in the same folder. for some reason i thought it would be harder with the project in a project situation.

[hollow swift]

have you tried lora for flux?

[torn wharf]

theres just the one i saw and i didn't see a point to it.

[strange grotto]

I get the Canny outline, but it doesn't proceed past the Load ControlNet model node.. i get this error:

'NoneType' object has no attribute 'keys'

and i found this

did u convert the canny cnet?

[spice spear]

how to install the canny cnet?

[spice spear]

the one for flux

[bitter hearth]

Do a star fish

That's really funny

[torn wharf]

lol i'm off to sleep soon but i could do a starfish on jumpman logo soon

thats a good one

[torn wharf]

you need ot be on the development branch of comfyui that has the pull request. its only in testing

[torn wharf]

see this message

[dry wave]

4 channel vae

[dry wave]

to be precise: it's a weird hybrid of VQGAN and vae, but it's technically very similar to a vae. It's a convolutional-only decoder (effnet), though, so it might less powerful than the SD VAE

[strange grotto]

got it

[sullen moss]

If someone has already tested the Realistic LoRA and Kenny, please share the workflow.

[acoustic kite]

honestly can't wait

[sullen moss]

https://github.com/lllyasviel/stable-diffusion-webui-forge/

Not stable but...

[hallow lion]

Ah control nets are here.

[errant dust]

So on a slightly different note, has anyone tried Auraform? They released v0.2 just recently

Support for fp16 is out for lesser HW and can run in Comfy

[odd basalt]

I was just about to look that up I was reading the paper on that one

[dry wave]

you mean Auraflow?

never found it good

[errant dust]

Yes, they are using the Dall-E 3 technique, per OPenAi's paper, of massive labeling

[dry wave]

soo many parameters for quality worse than even pixart

[errant dust]

Well, they just released 0.2

as in... 10 days ago

[dry wave]

everyone is doing that

[errant dust]

SD used a different method

[dry wave]

no

SD3 also used synthetic captions

[errant dust]

Yes, they even published a paper on it

Yes, but not DE3's method

[dry wave]

what's Dall-E 3 method?

if I remember their paper correctly they just run an vllm on 90% of the data and keep 10% intact

which is what everyone is more or less doing

[errant dust]

SD3 used their own proprietary method as published in their paper

regardless, Aura Form is in beta and actively being developed. 2nd release in a month

[dry wave]

if you cannot be more precise....

there is no conceptional difference here

[errant dust]

and since it has only been out a month, the talk about 'never finding it any good' seems a bit precipitated

[dry wave]

all modern diffusion models train on synthetic captions. Pixart was I think the first one which did this

[errant dust]

Maybe so, but same concept does not always translate to same results

[dry wave]

right, because results are quite bad for Auraflow xD

[errant dust]

oh boy

[dry wave]

yeah, maybe it will at some point get better. Currently, it's just bad

technically it's not much different to Flux and SD3.

[errant dust]

pot.... kettle

[odd basalt]

This is a example just used it

[odd basalt]

I don’t see a difference

[hollow swift]

different workflow, create different image

[bitter hearth]

Give prompt

[odd basalt]

Of course , A majestic star destroyer dominates the darkening skies, its imposing form silhouetted against a fiery sunset casting deep crimson and purple shadows across its angular hull. Turbulent clouds swirl around it, illuminated by flashes of artillery fire leaving wispy trails in their wake. Distant explosions cast flickering glows on the ship's weathered surfaces, as if foretelling an impending reckoning. The composition is cinematic, emphasizing the star destroyer's colossal scale and menacing demeanor.

[bitter hearth]

I see @odd basalt in chat I must gen one

haha

[odd basalt]

I love it \m/

Great concept art anon ghost ship

[errant dust]

Can you share the workflow? I am getting crash in Comfy, but may be missing nodes

[odd basalt]

I am using my iPad not on pc https://fal.ai/models/fal-ai/aura-sr/playground?ref=blog.fal.ai

And I used Flaux and took it on aura … when I get on pc I can def use comfy and share if it does not crash

[errant dust]

Ok, I had thought it might have been on local PC and was trying to solve, but the given Comfy workflow is missing a lot of nodes. There are no options to specify their VAE or Clip safetensor files

[odd basalt]

Thanx for the update sense u are on pc u can be very right it’s missing things … I’m curious if we need to download them missing nods,,, it’s okay u can try it that way what I gave u sense the other way is not working :*(

[limpid thunder]

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

• Member counts
• 344652 ↗ 344702 ↗ 344792 ↗ 344840 ↗ 344910 ↗ 344965 ↗ 345012
• Action members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 116
• Message members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 66
• Reaction members
• 0 → 0 → 0 → 0 → 0 → 0 ↗ 68
  More details

[acoustic kite]

thought this was kinda funny

[sage burrow]

Pretty sure that's where I was driving in NYC! "recalculating" 🤣

[mortal mesa]

too much grass for NYC

[torn wharf]

i keep see people using sdxl as a refiner workflows. don't they know that they should do a 3rd pass through a superior vae model and end the refining chain with the king of the game?

do people just have bad eyes and not even notice that sdxl washes out details in their images?

[gusty trail]

Most people don't care

[torn wharf]

seems to be that way. they pretend to care about things like quality, but that turns out to just all be theater and dramatic hyperbole.

[torn wharf]

https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Canny-alpha/tree/main 7gb 👀

this one is only 512 resolution too. i wont bother testing it. i had great results from the smaller one

[sage burrow]

More like I haven't figured out how to add them back in at the end. Also, would it ruin the sdxl 🥒?

@bitter hearth

er, not sure if catw4gbgpu is who asked for it, but I just realized, no way you can run it, except maybe glif?

[wispy spade]

is it okay to create sub folder on comfyui like checkpoint> sd1.5, sd3 or Lora>style subfolder, posing etc so i can organizing my model in a better way? it wont create any error right?

[green shoal]

yeah, you can have subfolders in checkpoints/ and they will show up accordingly in the loader node

[wispy spade]

thanks

[sage burrow]

hmmm, that long list of 100 checkpoints is getting out of hand, think I should try this 😄

[green shoal]

at 100 checkpoint files, it will eventually get out of hand even with subfolders

maybe we need a new UX for the loader node

[torn wharf]

i think a low denoise and high step pass throuugh sd3 or flux would maintain most composition elements and pop that 16ch quality too. i haven't done much testing here. i just figure, if you're refining images from SOTA models with late generation VAE's, what's even the point?

[acoustic kite]

😂

[bitter hearth]

yeah this is right

you just want SD3 to refine the fine details

it works great

[sage burrow]

amazing backgrounds for my nsfw! 😄

Does anyone have a workflow I can borrow that brings SD3 back in at the end? 😄

[odd basalt]

Flux + AUra + camera raw

[odd basalt]

Becky oh no what happened … check public work flows that has it at the end of the I don’t think I have one for u :*( I could help u with

[torn wharf]

i haven't started toying with multistage workflows yet on sd3 and flux yet. just a little. should be ez to set one up. there are nice workflows on civit available but many of them refine with the late gen models

[odd basalt]

https://civitai.com/models/623832/flux1dev-img2img-gpt-interrogator-2k-upscaler-workflow

[mortal mesa]

should call that one how to complicate a basic workflow

[odd basalt]

Yes I believe so 😝 not mine

[mortal mesa]

i see that has a negative prompt not sure why, the first thing i do is undo that to save time, so many workflows i see are encoding a negative just to plug the hole in ksampler

like so, to not have to process a negative that i wont use, time saver on my old rig

[bitter hearth]

conditioningzeroout and an empty prompt are not neccesarily the same

[craggy crest]

not the same at all - i like putting just a . or a space in the neg prompt

[mortal mesa]

ya just skipping the encoding negative prompt, was my point, less time

[torn wharf]

i thought negative prompting just piggy backs the cfg process and it's cfg which is causing the extra generation time

[mortal mesa]

basically

so i dont care if zero out or encoding a blank prompt are different, i get the same result faster at 1 CFG

[craggy crest]

you can also do some interesting tweaking by dropping a generic ascii charcter like a . or a , into negs though

[torn wharf]

The new movie was killin me the other night

[bitter hearth]

looks for balls

[errant dust]

I read something quite curious and wondered whether others had seen this:
Flux comes in two variants:

Timestep-distilled (black-forest-labs/FLUX.1-schnell)
Guidance-distilled (black-forest-labs/FLUX.1-dev)
Both checkpoints have slightly difference usage which we detail below.

Timestep-distilled
max_sequence_length cannot be more than 256.
guidance_scale needs to be 0.

in other words Schnell should be run with a guidance of zero

https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux

[torn wharf]

time particles count as balls.

[odd basalt]

Flux dev is better Schell suxx

[errant dust]

I use Dev-Schnell

[odd basalt]

I use Dev and pro

[errant dust]

I run from my PC

[odd basalt]

Pro on iPad and dev on pc

[errant dust]

well, Dev-Schnell is exceptionally good and fast

[odd basalt]

From the iPad

[odd basalt]

Yess

Pc

As long as I can do my horror I love it 😍 inspired by true events

[errant dust]

Blind date?

[torn wharf]

hehehe its a full moon out

[odd basalt]

Actually I never thought of that when I made it 😂 I was mashing up scream with pretty girl and piglet from Pooh turning hott girls into killers

[bitter hearth]

Happy little fish

[sage burrow]

Flux Canny, but apparently I haven't gotten the settings right. I haven't looked into it yet, I just chanbed the checkpoint in Crystalwizardd's apple workflow

[icy drift]

I'm been having a very hard time getting 4K renders out of FluxDev. 😦

I want to replace my desktop background, because it's a great new model.
But with multiple upscale steps and a low denoise on the final render, I get all sorts of artifacts.

The model's physics are amazing, for sure. Those reflections are flawless as far as I can see.
For comparison, here's my current desktop background, which is a Cascade render. The reflections are obviously wrong, because Cascade isn't very smart, and the details aren't perfect. The flowers are all just impressionist mushy messes when you zoom in.

But it's still overall better than FluxDev. Anyone know how to get it to render well at 4K? 😕

[rain current]

It works very differently from usual. If there are artifacts, you need to increase the denoise, and if it generates an HDR effect, you need to increase the max_shift. It is complicated to reach a good point for all types of images...

[icy drift]

I'm not sure though... Higher denoise was giving me a grid / vertical bar pattern.
Haven't experimented with shift at all, need to close and update to get the node.

[rain current]

Well, if that happens, increase the max_shift even more, even to 2..

[acoustic kite]

why are my images noisy? could it be a wrong setting somewhere?

i'm using sdxl_lightning so the steps aren't too low

[rain current]

But I have given up... when it seems like I have achieved a good breakpoint with another type of image, it is a disaster.... I am done with dev and the upscale with sdxl...

[mortal mesa]

what sizes people generate at, i was doing what i was used to doing 1MP resolutions, but that was a 2MP, i forget this can do 2MP out of the box

[bitter hearth]

you can't expect Flux Dev to be as good at high resolutions as a model with Würstchen architecture like Cascade or UItrapixel
its just not going to be anywhere near as easy

your best bet would be waiting to see if this model gets support by the Hi-diffusion team
as that is currently the top method for high res diffusion outside of Würstchen

otherwise, until then some form of block by block scheduled latent downscale could help, i.e. a manual version of deep shrink

[dull star]

man I can't wait for boring reality lora for flux, but the realism lora is nice for now

[bitter hearth]

oh there's a lora?

[dull star]

https://huggingface.co/comfyanonymous/flux_RealismLora_converted_comfyui/tree/main

before lora, after lora

no more strong (+ overexposed) plastic skin

[bitter hearth]

its a lot better yeah

its still too saturated and plastic for me but not as bad

[dull star]

boring reality will take things next level

[bitter hearth]

sadly not neccesarily

in order for a neural network to be really good for finetuning it has to be a little bit under trained

in Flux's case it seems to be overtrained a little to bake in an "aesthetic" or midjourney look

this can't necessarily be removed by fine tuning that easily

[errant dust]

mno — Today at 4:24 PM
Hi, let's skip the community chat this week, and let us continue to work hard on shipping the best text to image model ever existed, called Ideogram 2.0. It'll be in beta testing very soon.

So new Ideogram is on the way

[bitter hearth]

I kinda just ignore the closed source models TBH

[brittle nexus]

2x upsscale flux

[odd basalt]

Flux + aura

Well least we all get mj for free through SD … making up for the SD3 XL model that they want us all to forget … there back up model that was suppose to be sd3 makes up for it all… the boat switch worked sd3 sunk like the titanic

Sarcasm^

[sacred jewel]

How are folks plugging in the LoRA? I just tried the reality LoRA from Comfy but not sure if I have it connected properly...

[uncut river]

damn you flux-schnell ! got it right in about 2 or 3 images - 1 prompt tune

[errant dust]

https://civitai.com/models/628862/flux1-schnell-merged-with-flux1-dev?modelVersionId=703068

Same schpeed, better results

[fleet meteor]

👀 👀 👀

[fleet meteor]

Amazing, I wonder if there´s a comfyui node for it

[torn wharf]

ideogram . meh. flux pro is the king if you want to pay for a service

[bitter hearth]

nah its still midjourney

if midjourney was open source I would barely use anything else

[brittle nexus]

You don't need a new one

[errant dust]

No, it really varies. MJ is amazing at many things, but not all. For example with text it is just crap, and some styles of art too. Its understanding of complex prompts is also not the best, even if it can often produce the most amazing visuals. However, if you want uncensored productions of artist styles or known celebrities, it has no peers. Ex: give me an illustration in the style of Frank Frazetta. It will, like nothing you ever saw. Ideogram was king of text and has fantastic art too. I say 'was' because I haven't tested Flux enough to really compare. Flux has serious limitations so far in terms of pure creative illustration-style art. A lot of it seems by the letter, with few extra details and touches. For a lot of things it is amazing so don't get me wrong. None of them are universally king though. Not even Flux.

Here is an example of Midjourney asked to make a fantasy image in the style of Frank Frazetta. For this type of request, it is peerless, it is true. just don't ask it to spell anything:

Here is the sort of artistic text and variety in fonts and presentation in which Ideogram until now has had no rivals:

Dall-E 3 is king IMHO in terms of general illustration style art. Others can get it hit or miss, but it is the most consistent. And it will imitate the style of known artists too. Censored you say? Only the version in ChatGPT/OpenAI. Microsoft's Copilot has no such issues:

Assuming it recognizes them of course.

The last one I can't seem to get Flux to do anything similar no matter how hard I try. Ideogram is okish, and MJ can too, but Flux seems to resist any attempt. Might simply need some major prompt tweaks and adjustments on my part, but it hasn't been for lack of trying

And dont think I am dissing Flux. It is absolutely amazing, but it is clearly better at some things than others. It has a huge focus on photorealism, but illustration type art? Not so much so far.

[fleet meteor]

is this dalle3?

[errant dust]

Yes

Through Copilot

[fleet meteor]

I´ll try achieving something like that on flux rn

[errant dust]

You can use it 15 prompts a day for free, so 60 images. More if you don't mind waiting a bit for more. Ideogram gives 10 a day.

I have subscribed to them all, and used SD as well since 1.4

TO show Dall-E 3 can do artist styles, here is the same image but requested in the style of Keith Haring.

This is a picture by Keith Haring to illustrate the style:

[torn wharf]

if a company made an ai accelerator that looked like this, woudl it sell?

[errant dust]

Would be nice if the AI did not insist on blurring that as if depth of field meant anything here

[torn wharf]

its dramatic that way. makes you think about how deep the symbolism is

[errant dust]

lol

[fleet meteor]

Do you have the prompt? (i just wrote something )

[errant dust]

Or reassures me that AI is still dumb

[fleet meteor]

Yeah I hate dof, idk why most AIs do the same thing blurrying the background

[bitter hearth]

can anyone get skim cfg node to work

[errant dust]

Dall-E's biggest fault is its absurdly small history, so no. But Ideogram has no such limitations and I found one I may have used. Might not be exact.

"A vibrant and whimsical comic of a tree with numerous cats of various colors, sizes, and expressions perched on its branches. The cats exhibit a range of emotions and designs, from smiling faces to curious expressions in a caricature style. Some cats are perched alone, while others are grouped together. The overall mood of the image is playful and enchanting, and accentuated by the dark outlines in the drawing."

For the record, Ideogram was not great at this either.

[fleet meteor]

Trying it rn , its so slow 😅

[errant dust]

You don't want to know how slow it is on my laptop

[fleet meteor]

Better but not that good

[errant dust]

Yeah, Ideogram was like that, though sometimes a bit better.

Like I said, they all have their strengths and weaknesses

[errant dust]

I mean some things they all do similarly well, like those godawful boring Japanese cartoon styles. But others are very hit and miss and some are better than others.

[torn wharf]

just put blur in the negative. ez

[young blade]

did you negate flux? if so, how 🙂

[torn wharf]

hooked up cfg guider and fed it a negative prompt

doom guy comes out better when i negative doctor victor and halo

[young blade]

got it, the default 8 cfg is obviously too high lol, you find a good average? 1-3 seems ok

[torn wharf]

i hover in that range. seems to do something. more data still needed kinda though

[torn wharf]

its still off but its better

[errant dust]

Well, may be a matter of prompting to some degree as I am getting closer. But you really need to spell out everything. For example, if you don't tell it the tree has a lot of leaves, it gives you:

So I added "dense foliage" and "LOTS of cats":

Not quite as creative or interesting as DE3's images, but still waaaaay better than the initial ones.

Will try Flux Dev, and after playing more with the prompts. See what I can learn from it. (BTW, I always leave prompt and workflow in images I share here)

[bitter hearth]

how long did it take on what hardware?

[fleet meteor]

yaaay!

[torn wharf]

if you use an LLM to really elaborate your prompt concept, it'll go hard too. that's often what the walled services are doing.

[sterile pendant]

PSA: the latest https://github.com/comfyanonymous/ComfyUI/commit/413322645e713bdda69836620a97d4c9ca66b230 for comfyui cut my inference times in half. Used to get like 16 sec/it with flux, now I'm getting like 8 sec/it (i only have 8gb vram)

I see why, it's actually sometimes using all 100% of my CPU alongside it. Normally, it will only use the e-cores (usage sits around 40%). Was noticing that sometimes it would be 8 or 9 sec/it, then other times, like 12 sec/it.

[errant dust]

To the credit of Ideogram, they offer the option of an LLM enriched prompt or not, and you can even have it apply a different LLM prompt for several or all of the four sample it produces per prompt.

It will show you what the LLM version was in each case as well.

[dry wave]

I have a 3090 with 24gb. With batch size 1 I usually have 5 seconds per step. So definitely slower than SDXL, but still doable. Problem is rather to find good parameter settings first. My validation images always look as if the method blows up, but then it always turns out it is still undertrained

[icy drift]

Testing negative CFG now. Thanks @torn wharf for hint.
FluxDev FP8, 1024x1024, euler/beta, 16 steps, fixed seed
same prompt to clip_l / t5: `mountain landscape, grass, lake, sky, birds flying, soft white clouds, mountains reflected in the rippling water, wildflowers in the grass,

cinematic masterpiece, dynamic lighting and shadows, extremely highly detailed, landscape photography, desktop background wallpaper`
Baseline result

With negative prompt red, crimson, carmine, ruby (Holding all other parameters constant throughout tests. Will only change negative prompt.
Obvious absence of red, so far negative CFG appears strongly effective. Reason for blurriness unknown, will experiment after CFG tests.

With negative prompt blue, azure, cobalt, marine, navy, ultramarine Definite effect, less obvious accuracy.

With negative prompt green, jade, emerald. Again less obvious than red, but definite absence of green from original on fixed seed. No observed blur means probably side effect of seed, not of negative.

With negative prompt color, rainbow, saturated, vivid, lush, colorful, multicolor, intense, rich. Obvious reduction of color from baseline.

[restive notch]

laugh girl walking to the moon

[icy drift]

Positive guidance 3.5, negative guidance 9.0, CFG 3.0

Positive 3.5, negative 9.0, CFG 9.0

Positive 3.5, Negative 9.0, CFG 1.0 (holding CFG 1.0 from now on)

[dry wave]

maybe run cfg only for the first 90% timesteps?

[icy drift]

No idea how to do that. I think I would need two sampler nodes and some kind of time split node?

[dry wave]

and I would use same guidance for both

[dry wave]

that's the easiest way to do that I guess

[icy drift]

I'm experimenting with the effect of varying guidance.

Positive 9.0. Negative 9.0. Desaturation maybe less pronounced.

Positive 9.0. Negative 1.0. No change.

Positive 1.0. Negative 1.0.

Again, Positive 3.5, Negative 3.5. (Repeat baseline.)

Upweighted negative prompt (color, rainbow, saturated, vivid, lush, colorful, multicolor, intense, rich:5.0). No change.

Removed negative prompt. No change.

Restored negative prompt color, rainbow, saturated, vivid, lush, colorful, multicolor, intense, rich (not upweighted), CFG 2.0. Definite desaturation effect has returned.

All else constant, raised to CFG 3.0. Desaturation effect intensified.

Upweighted negative prompt to 5.0. Desaturation diminished (opposite expected effect).

Restored negative prompt to unweighted state. Raised negative guidance to 9.0.
Desaturation effect increased again.

Tests complete! 😊 🎉

Negative prompting for FluxDev is now 100% usable.

[dusky thistle]

used random attention guidance for this

no upscale

[icy drift]

Negative CFG workflow. Will also post on Civitai I guess? Never posted there before.

On Civitai. (If I did this right.)
https://civitai.com/models/633586?modelVersionId=708340

Prompt painting of a woman warrior by Frank Frazetta, moon maiden, oil painting, concept art illustration

Same seed, with negative NSFW, nude, naked, bare skin, exposed skin

😊

[dry wave]

that's not bad. I never got flux into good Frazetta paintings

they always look too photorealistig and not vintage enough

[hallow lion]

Wow you have your beer brand.

[sage burrow]

Ty , finallly! 😄

[hallow lion]

holy molly!

[bitter hearth]

He's the guy who gives powers to the main character and leaves without a trace

[signal shuttle]

Flux

[hallow lion]

The 5th element.

[stoic turtle]

Lmao

Shows up again at the end of the story smiling

[stoic turtle]

I think yer camera ain’t straight

[sage burrow]

Flux has expanded a bit

[hot dawn]

Muad'Dead

[acoustic kite]

TIL the max resolution recommended is 2MP, double what is recommended for sdxl which is 1MP

this is what 2mp looks like. the guidance scale is set to 4, i might lower it to 2 and try again

[errant dust]

So for any interested, you can generate a few images a day with Flux Schnell at Nightcafé. Bear in mind that it does have some clear restrictions or lack of control of things such as number of steps used or resolution (1k default), though not ratio. It is also extremely fast. https://creator.nightcafe.studio/flux-image-generator

Note, it SAYS Pro, but I am in doubt as to whether this is Pro or only their Pro subscription plan.

EDIT: turns out only subscribers can use Pro. The free images are Schnell.

[acoustic kite]

wait was it pro that was supposed to be slightly better than dev?

[errant dust]

Yes, that is correct, and after inquiring I found that the free generations are all Schnell. Only subscribers can use Pro.

I edited my messages above to avoid any misdirection

I'll probably plunk some small money down on one of these guys to test Pro and compare. Likely one that allows me maximum settings control

[bitter hearth]

@bitter hearth https://www.instagram.com/reel/C6l6cnwLRb_/?igsh=NnU0dW8wemF0dmQw

rosariomdra

me hizo doler el gato 😡

Likes

529797

Instagram

[bitter hearth]

She's incredibly impressed by the 2mp image

[bitter hearth]

[hallow lion]

She is drunk.

[bitter hearth]

[muted dove]

"Spaced" more like

[acoustic kite]

xD

[errant dust]

Power to Fal.AI. I had credits left from when I plunked some money down for SD3 8B but apparently did not use it all. Can use for Flux Pro.

[acoustic kite]

oops sorry for posting images back to back. didn't mean to :)

[errant dust]

So it turns out you DO get 5 free Flux Pro images with Nightcafe

Their site is hopelessly confusing in some things, but this is now certain.

[acoustic kite]

that's amazing!

i'll def check it out if that's the case then

also i found out that you just have to copy paste your prompt several times if it doesn't give you what you want

a 5x copy/pasted prompt will give you the result all of the time lol

[errant dust]

Be warned their default settings, even for a free account is to send you email notifications about EVERYTHING. So be sure to enter the setting to turn that crap off. That said, free Flux Pro is free Flux Pro, and the settings change is a one-time nuisance

They have an option to fix the seed, but it is entirely unclear to me if the seed matches Comfy's in any way. I suspect not.

When comparing models or prompts I like to fix the seed to be sure I am really seeing an apples to apples comparison, and not lucky RNG vs unlucky

[torn wharf]

free flux pro? more like.. flux SLOW. amirite?

[silver sluice]

hey @jolly swan, hope you're doing well. I look forward to your progress on v6.9 if you have any. Me and my friends we're wondering now that the dust has settled on SD3 per-se, on a few things

• Are you still considering making an SD3 Pony model after 6.9's release?
• Have you considered switching to Flux instead for the next generation Pony model?
• (Sorry I have to ask) Do you have a timeline for when you'll expect to see some results for 6.9?

[mortal mesa]

nah dont fall for it

[icy drift]

Have whoever you're arguing with blocked so your posts are all I see. 😛 I did a bunch of tests on negative prompts after you mentioned the node you were using, and it definitely works! I posted a workflow for it earlier in the thread with instructions on what works / doesn't. 🙂

[mortal mesa]

yikes, you attack and fight small children

[torn wharf]

just crystal wizard saying everybody is thinking about me and laughing fondly of their memories of my antics when i'm not around. I'm still not entirely sold on that one particular node. i'm not sure where the balance of guidance and cfg fits.

I seen that reforge / forge prooject is getting flux support soon. i can't wait. messing around with noods for basic generations is cumbersome

never thought a lame joke like "more like flux SLOW" would set someone off that hard

[foggy cloak]

Where does it say flux is coming to reforge?

[torn wharf]

there's an experimental branch right now

its not working when i checked

[foggy cloak]

From panchovix? Might have to switch to experimental branch to try it out

[torn wharf]

and lly has been adding stuff to upstream too

i haven't had success getting it to work but they're doing stuff

[mortal mesa]

people have broken psyche, alot of people

[torn wharf]

https://tenor.com/view/ben-kenobi-hes-me-gif-23727931

[mortal mesa]

TBF there are a ton of reasons

[torn wharf]

whenever i hear things like "nobody likes you!" from one person, that just makes me think that all the people who don't like me are afraid to tell me so. so i intimidate them and that makes me feel good about myself.

logic

[foggy cloak]

Never seen you before, but I just remember crystalwizard going crazy glazing SD3 when it came out 😂

[bitter hearth]

balls

[torn wharf]

as i was setting up to train the ball lora on the one trainer experimental branch, he was arguing with me that it couldn't be done and it was impossible to train a lora on sd3 . pretty sure he insisted i give up

i think that was the earliest impression of him i can remember

[bitter hearth]

Nothing is imballsible

[torn wharf]

i know i dont know everything. i think i'm dumb in this field. i also recognize that many self proclaimed experts are serial bullshitters. Thats why i like to just get down and dirty and start throwing balls at the wall to see what happens

can only get so far with book smarts. like dr jones used to say, field work is important

bad doxxing skills

[mortal mesa]

yell him he watches FOX like you di dthat other time, thatll get em

[icy drift]

I experimented to figure it out, and the results are in my workflow as notes.
CFG 1 = negative prompt does nothing.
Higher CFG = more powerful negative prompt
CFG > 3 = artifacts and buggy output

Guidance on negative prompt: increase for slightly more powerful negative prompt. High guidance will cause artifacts.

Best way to boost negative prompt: Add more words, synonyms, and related terms.

I locked the seed and all other gen parameters, and for different sets of prompts with / without negative I proved that negative was working perfectly.

[torn wharf]

OH good data! taht's a lot more structured than my tests. thanks.

[mortal mesa]

i did some feelings based tests and went back to FP16 T5xxl

[torn wharf]

feels based tests are my main sail

[mortal mesa]

i dont know how else to say it lol, it does indeed feel better haha

and im not going to test

[torn wharf]

i'm sure that fp8 t5 diminishes the quality. i've felt it. but i don't have the data to demonstrate it to others.

[sage burrow]

The comic in question (tbh I'm only reposting the actual image it's really awesome)

I sent to icycold (and crystalwizard also but accidentally bc i mixed my windiws up at first)

I think too many people look down on those with any psychological stuff going on.

I wasn't making fun of icycold.

I just woke up, so apologies if I'm not making sense yet; I also didn't scroll back very far for context.

[torn wharf]

lol it's the usual situation. I expected it to be honest, just not from crystal wizard. mention last week in passing that i had some mental health issues i've been working on , focus issues mostly, and now it's blowing up in my face. everyone laughs about me and shares mental health memes about me. smh. what a catastrophe

[bitter hearth]

I just got here, have no idea what's going on, nothing makes sense in that image too

[bitter hearth]

Does flux not have negative prompts

[mortal mesa]

technically, no, just Pro

[bitter hearth]

[mortal mesa]

also technically, people added them anyways

[sage burrow]

Flux scnell. It's super fast tho. A smart way to attract new users to their packages I figure. Bc once there are flux loras and checkpoints I can use id get pro plus cause my computer is slow?af.

[mortal mesa]

i just do the slow

i mean you do like art business stuff maybe i think, i just play, so experiences may vary

[sage burrow]

If you mean the on my own computer slow, for schnell it's 5 mins compared with 30 seconds. My lack of patience satanic $30 p/m isn't bad 😄

[sage burrow]

I've got 2 workflows sp fat that add neg prompts!
Np waiting 20 mins for my 8gb vram to make one image 😭😭😭

[mortal mesa]

yup i dont blame you, im just not a buyer of AI yet..

[torn wharf]

I like the idea of them releasing a special distill that's not good for images, but we can use to train Loras that work on the distilled models.

Would that work?

[torn wharf]

I want to say the same but I use my new GPU more for this than gaming

[sage burrow]

Scroll up. Though warning there's Distracting stuff on the way such as training info released for schnell!

[sage burrow]

Glif us freeeeee 😄

[bitter hearth]

my poor thing cries on games

[torn wharf]

I just need to remind people that Barclay from TNG is also Murdoch from the a-team.

Like, what else haven't we discovered?

[sage burrow]

So which rented gpu site do you use? I want to use one to make some loras 😄

[bitter hearth]

is it gonna be easy to train or something

Im dumb

Lalita is the name of an extremely acidic soda thing over here, few like it

I like it

Your name always makes that come to mind

[sage burrow]

I'm kinda hoping for one of those pre-made diffusers training workfliws or something lol

I trained a checkpoint before, learned everydreamtrainer setup, and it worked! Was sdxl tho. But it took me a few days to figure it out. That would cost a lot on a rented gpu!

[bitter hearth]

[sage burrow]

I forgot to title that mental health cartoon, supposed to he titled "if we treated physical ailments how ee treat mental health ailments" or something like that.

[torn wharf]

I was watching a-team AND TNG while they were both broadcasting new episodes and I never connected those two characters.

Just goes to show how we can get stuck in this domains.

[sage burrow]

On a special server with a giant gpu....

[torn wharf]

I also like the idea of them releasing an undistilled 5b version . Ideas are fun

[sage burrow]

Where do you live?

[bitter hearth]

The worst place

[mortal mesa]

chicago?

[errant dust]

not sure what you mean. Takes about 5 seconds to generate.

[torn wharf]

This lame joke has really set some off. It was a lame joke

[bitter hearth]

||bruhzil||

[mortal mesa]

it was a joke for the americans, no power there

[errant dust]

Look, on this laptop, by comparison, a single Dev image takes a minimum 20 MINUTES. so....

[torn wharf]

Wait until people figure out they give free flux gens too. Right now it's 5s.

Also, it's not a hill I'm gonna die on. Read it aloud for instance.

[errant dust]

I have credits in Fal anyhow. The real interest was/is to see how Pro compares in quality and output.

[bitter hearth]

idk what I'm doing

[torn wharf]

when the prompt was for something entirely different it seems to make something cool anyways?

[bitter hearth]

No I did prompt for an overgrown ruins with portraits...

[torn wharf]

oh it got it!

[sage burrow]

USA?

[mortal mesa]

ide like to say i re-remembered if you select a bunch of nodes you can combine them to a mega node

[bitter hearth]

Answer is above

[mortal mesa]

so not the worst place by far

[sage burrow]

Even with schnell?

[bitter hearth]

Lmao

[torn wharf]

first thing i noticed is it tries to make the magic kingdom castle accurately tall instead of perspective trix like in real life

[sage burrow]

What's wrong woth Brazil? Good food, amazing festivals, nice beaches...

[bitter hearth]

Shrek is the way

[bitter hearth]

No gpus

[sage burrow]

Amazon.com? But they prob cost equivalent to 6 months revt for you guys 😦

[errant dust]

No, absolutely not, thank goodness. And I get better than Schnell quality with the special Dev-Schnell combo suggested by Comfy

[sage burrow]

Glif or mage 😄

[dull star]

dev schnell combo suggested by comfy?

is that the merge?

[errant dust]

yes

[dull star]

and comfy recommends it? damn

[errant dust]

a special merge

not a random one

https://huggingface.co/drbaph/FLUX.1-schnell-dev-merged-fp8-4step/tree/main

[bitter hearth]

Comfy secretly put fox girls in the merge

[mortal mesa]

i think the recomendation was in the method of combining

[bitter hearth]

Rob a bank angle

[mortal mesa]

i use that one with 6 steps

and fp16 t5

[torn wharf]

sd3 adn flux and all these SOTA models. I can't wait until this tech converges with the meme spheres and gets things weird af online. i feel like we're too slowly getting to the next gen of online content and things need to ramp up. we canmake relevant new memes for any situation now. it's a memevolution.

[errant dust]

Exactly and this file I linked has it pre-made, with Clip and VAE embedded

"Flux.1-Dev double_blocks (MM-DiT) onto Flux.1-Schnell, then quantized to fp8 giving you a higher quality model that still runs in just 4 steps"

[mortal mesa]

i feel like consumers are going to get edged out on local inference

[dull star]

how does this compare to dev and schnell, and can it perform even better at higher steps (it only says 4-8)

[torn wharf]

wont be for long since the weights are out and hardware just keeps coming out.

[mortal mesa]

true

[dull star]

like if I want something super complex that might only work with dev at >25 steps

[torn wharf]

only a matter of time before a smart person figures out how to run fast inference on cpus with clever instruction set tricks and we have access to the main system memory

or intel is almost certainly designing instruction sets specifcally for inference

[errant dust]

Yes, consider that the number one driver for tech updates in consumer PCs is not AI, but games.

[torn wharf]

I know jim keller, the guy who architected the current amd64 ecosystem of instruction sets, he's working on dedicated hardware for this stuff. his chips are kind of old gen tech now though and i haven't seen new stuff from greyskull labs for a bit

[errant dust]

And in the cause of image AI aficionados, it is a match made in heaven

[torn wharf]

i dont know him. i know that hes' working on stuff

[torn wharf]

games need local model inference to be fast and efficient enough for the game to run. only a matter of time for that.

It's basically what all these dlss and fsr techs are. local models of sorts. local language modesl are the next big one to crack

[mortal mesa]

TPU for all

[torn wharf]

i don't think cloud inference services like chatgpt will ever work well for games outside of chatbots in games

[dull star]

ggml implementation of stable diffusion and flux would be interesting

there is already one for stable diffusion 1.5, 2.1 and sdxl I think

oh even sd3 daaamn

https://github.com/leejet/stable-diffusion.cpp

[torn wharf]

live services for gaming is an ongoing problem. they're often not needed and are essentialy just DRM

[torn wharf]

yeah there's lots of smart scientists excited to get inference on cpus. all that ddr just sitting there going ddurrrrrrr

[dull star]

but this repo also allows you to partially use your gpu like with cuda, vulkan and rocm

like llama.cpp and whisper.cpp

[bitter hearth]

ddurrrr...

[torn wharf]

only a matter of time till a smart developer puts diffusion upscaling into a game engine, so that all the high fidelity textures are generated at run time using lower quality scafolding assets

[mortal mesa]

i think Microsoft and NVidia have that developed already

gonna get many game remakes down the line

[torn wharf]

theres a ton of controlled art direction that can occur with model. ip-adapter style transfers and very elaborate prompts would get enough consistency for game textures i have no doubt. just no one's made a product showcasing that yet. usually takes about 5 years for a game to go from initial ground breaking to a released product.

[torn wharf]

yeah tools are there for sure. nvidia has all those remastering tools available.

[torn wharf]

the GTA trilogy remake was done with "remastering" tools too, though that was a poorly applied effort imo

remember when we first started seeing GAN upscalers? people were remastering final fantasy 7 prerendered backgrounds with them. like 2002ish?

[errant dust]

Games don't really do model inference, unless you are thinking specialized NNs such as chess or go, but regardless the tech benefits both, so if gamers push Nvidia to make faster cards, and that happens to also be exactly what helps Ai model inference, then it is a win-win. No surprise Nvidia grew so valuable in today's AI boom.

[torn wharf]

games already are today doing inference with systems like raytracing, frame generation, upscaling, all in real time. that's because of trained models.

instruct models have a lot of potential for game master style design. a game master is traditionally what separates table top rpgs from video game rpgs. that's just one design goal that ML models have a lot of potential for.

nvidia is still a gaming tech company at their core. they didn't design these new chips for openAI 's level of business at first. it was primarily to make video games go harder and they knew that chasing that would put them in a very strategic position for the future.

the one AI i've been watching them train that's so cool is a player agent. it can learn to play any games with any control schemes. they're doing a ton of practical research in that field. so instead of dippy dumb AI team mates, you can have one that teaches you some shit

[errant dust]

The ray tracing you see in games such as Cybrpunk 2077 and others has nothing to do with neural network inference.

[bitter hearth]

How can you guys type so much

[errant dust]

Nor the massive polygon output in Unreal 5

[bitter hearth]

I'm lazy to even read

[mortal mesa]

2 very fast fingers

[torn wharf]

https://youtu.be/5LL6z1Ganbw](https://youtu.be/5LL6z1Ganbw?t=120 this is just one aspect of that research. they taught it to play minecraft by watching streams. it can do text to minecraft! "build a castle" and then it does!

[errant dust]

There are papers out to use neural networks to improve ray tracing, but they have not made it into games yet

[torn wharf]

coffee. also typing so much that it's as natural as speaking

while i don't believe nvidia has published a paper on how RTX is achieving real time ray tracing, i'm fairly certain they use a machine learned model

[errant dust]

But as I said, NN inference does exist in games, as shown with AlphaGo first, but is now well embeded in other mind games such as chess and shogi

[errant dust]

Ray tracing is not a secret

[torn wharf]

the cyberpunk raytracing doesn't use nvidia's tech at all

imo crytek's crysis remake does the most accurate real time ray traicng

[errant dust]

I have trained many chess-based neural networks from scratch

[torn wharf]

DLSS "Deep Learned Super Sampling"

[errant dust]

using established tools mind you, but data and structures my own

[torn wharf]

with programming i got as far as remaking tetris from scratch in a console text engine. also i made a template language for html forms that was over engineered and then jquery came out and i said fuckit.

never got much further with code

[errant dust]

I'm a chess player at my core, not a programmer, so my interest was in leading the training of the NN through specific steps and methods from the perspective of a chess player

[torn wharf]

hmm. at my core.. what am i? break dancer with a knee injury. that'll do

[errant dust]

I had help with the code from a friend who is a coder for Amazon

[bitter hearth]

[bitter hearth]

Making things breakdance is very funny with ai

[errant dust]

I take it this frustrated your Olympic gold dreams

[torn wharf]

oooh yeah tragic memory just came iin. thoguht i was clever and named my library "phorms" because it was a php situation and i thought "wow thats gotta be the best library name anyone has ever come up with" and i was really into it. i remember feeling so proud of myself. but everyone who did forms in php had the same idea

[sacred jewel]

They have Skillz and Balls

[torn wharf]

so long ago that break dancing at the summer olympics was un heard of

[sacred jewel]

I got as far as printing ASCII table using embedded BASIC in the 1980s. 🤷‍♂️

[torn wharf]

you dealt with ascii in the 80s. remember an app called "The Draw"

[sage burrow]

My level of coding is: built websites (in html in notepad) before dreamweaver etc. was even out. Make tons of glifs
ROFL

[torn wharf]

olympic ball dancing event

[bitter hearth]

10/10

People just standing there looking at them not do anything for 10 minutes

[torn wharf]

i put "resting still" in the negative

[torn wharf]

tbf that is a lot of olympics events as the athletes prep for the next event

[acoustic kite]

where is this on the realism scale

here's another one :)

[rain current]

Not only realism, it seems that the images have soul 😅

[acoustic kite]

awww :)
also this is insane

[torn wharf]

the height chart is a dead give away. the measure lines are very none standard and dont persist properly behind her. her left ear the one hidden behind hair, wacky postioning for her head tilt. the glasses are sitting way higher than they should be. goofy hair line. weird red smudging under the frames but otherwise perfect skin everywhere else. obviously the text is weird but that could just be an in joke that no one gets. but the one thats super uncanny that most people wouldn't notice i bet. light seems to be coming from the top left and top right at the same time. the shadows are all over the place.

8/10 realism because it would probably fool 80% of people.

[acoustic kite]

woah ty

[vast condor]

are those real words?

[torn wharf]

this one i would only suspect because of the power button position on the bottom of the device

[acoustic kite]

yeah it seems i got a bit unlucky there, maybe some photoshop can fix that

[torn wharf]

large crowds doing active stuff it works really good in these new models. but lookit the feet. lots of feet doing lots of stuff and it's confusing the patterns. Faces on that could at least be explained by photography artifacts like to slow a shutter speed or something. the feet tho.

i think thers still ways to spot fake images but, it really really is getting harder

[rain current]

They are not examples of perfection, they have flaws, and the models are not perfect...

[torn wharf]

it really is getting so hard to tell tho

[icy drift]

Flux is the closest to perfect any open model has ever been, but it still took a TON of manual work to get my monster cards ready to print. Every image has mistakes only an AI would make, as soon as you start looking close.

[torn wharf]

you could sometimes get soul out of older models if you used an open pose controlnet on a photo with soul. it's in the latent space it's just hard to evoke out

[torn wharf]

you actually producing those into a physical set? nice to hear. excited to see the fruits of that

[icy drift]

Is this "soul" like your "aesthetic" thing? 😛 Doubt I'll get it. I like data.

[icy drift]

Yes! Sent the prototype off to print Wednesday morning. Working on layout V2 for when it arrives. The borders were just unusable in the end. Need to composite for the revised design, and then I think they'll be really good looking.

[torn wharf]

to understand soul there is but one bible of a film to ingest. May i be so inclined to even provide a link. There's probably more that people would argue about, but i think they get it here. https://archive.org/details/the-blues-brothers-1980-4k

if you're on the fence, might i suggest that it stars Carrie Fisher

[icy drift]

Summary? TLDR;? 🍒

[torn wharf]

They get the band back together

[icy drift]

Explain "soul" in images in a way a computer could understand it? 🤔

[sage burrow]

Someone needs a giant pool stick 😉

[torn wharf]

https://tenor.com/view/dancing-ayye-hype-excited-dance-gif-17458303

[icy drift]

(I also am a philosopher and believe in the metaphysical soul, so alternatively explain in a way a philospher / logician / linguist would understand it.)

[torn wharf]

i believe that if there is a soul, we can engineer it.

metaphysics is just physics that doesn't have an applied science to it yet

[icy drift]

But what do you mean by "soul" in images???

[torn wharf]

https://tenor.com/view/tank-of-gas-blue-brothers-hit-it-gif-18357795782957779468

[mortal mesa]

believable actions/poses/expressions/body language

[torn wharf]

https://tenor.com/view/blues-brothers-ray-charles-dance-blues-music-gif-7215971032810803123

[icy drift]

I assume you would add "diverse" to that list? RevAnimated could do believable poses... Like 6 of them. (And I don't mean diverse in the nonsense angry way.)

[torn wharf]

also maybe i might show the computer some alex grey paintings or a tool music video

[sage burrow]

you know that famous painting of Venus in the clam shell? That's one that google images chose. I can't post it on discord tho

[icy drift]

Google chose it for something? (I know the image.)

[sage burrow]

I did a search for rennaisance painting of the soul

[icy drift]

Doubt that's what we mean here, but Google's opinion is noted. 🙂

[sage burrow]

I was hoping for one of those fancy ones where there is a translucent version of the person half out of their body....

[sage burrow]

I don't see why that one wouldn't count 🙂

[icy drift]

Because I think he's talking about body language and expression. Venus is very painter-pose, since the model had to stand like that for days on end. Cameras are different.

[torn wharf]

if the birth of venus isn't SFW, i don't wanna work there

boss is like "JENKINS?! YOU GETTING OFF TO RENAISSANCE ART AGAIN?!" its like.. no dude thats a weird accusation

[icy drift]

Imagine working at a museum, can you have that as your desktop? 😛

[torn wharf]

just rip the band aid off and post it

[sage burrow]

Who's talking about? Apparently I mised the original post or something

[icy drift]

I feel the same way... But @torn wharf was looking for images with "soul", and I wanted to know what he meant.

[mortal mesa]

its a 6th sense

[icy drift]

Balance?

[mortal mesa]

cant say it but can see it

[torn wharf]

i dont think it was a portrait of someone posing. it has all the elements of like a constructed person. like they planned it out her pose and face and expression and eye lines and all that. like how we have a style like anime or pixar style. it's a design language.

[icy drift]

Wait now what are we talking about? Venus? Constructive drawing is used to draw models.

[torn wharf]

knowing the italians probably the position of her toes have extreme symbolism too

[icy drift]

Yeah he could definitely make changes like that. He would work from multiple sketches.

[uncut river]

I believe in soul as in spirit, with a certain focus and mindset, in the style of, alive, purposeful

[torn wharf]

i gave all those blues brothers pics to represent soul. and also alex grey

[icy drift]

Well those were representations of people, who have souls.

[uncut river]

not only people have souls

[torn wharf]

whatever LSD idoes to neurons, thats what we need to encode into these networks. if computers are going to find soul, that'll either do it or just make things interesting. either way, fun.

disclaimer: i have no idea what lsd does i've only seen it in movies

[icy drift]

But can you tell if a specific image has soul?

[uncut river]

shrooms work better for the visual, I heard someone tell to some other

[sage burrow]

Alex Grey created a lot of paintings along those themes, in the 80's

[torn wharf]

like a soul classifier?

[icy drift]

See this clearly has no soul IMO.

[uncut river]

If you need a scan device to exaclty measure the amount of soul in something, you'll probably dont get it

[torn wharf]

contextualization could easily give that image some soul. but it does parodoxially feel stiff to me. and that's interesting. it compels me to examine it closer

[mortal mesa]

the scientologists got that thing

[sage burrow]

dips my gpu in lsd

[torn wharf]

they've got a lot of things

[uncut river]

careful, dont touch with your fingertips

[torn wharf]

i always say handle it like a hockey card. but i guess thats dated and most kids would say pokemon card now

[mortal mesa]

like its a Michael Jordan rookie card

[torn wharf]

my heart aches so hard when i see people planting their sweaty sebum bum thumbs all over circuit boards

[icy drift]

Like kids grabbing DVDs back in the day. 😛

[torn wharf]

every gpu be like

seeing it out of the case is actually making me really nervous

[mortal mesa]

anything could happen a sudden gust

Heres a horror story for you, i pulled out like a half a cup of matted down dust out of my video card when i started with AI, its much quieter now

[torn wharf]

yikes

soilent dust is ||people||

[mortal mesa]

freaked me out, was well hidden behind the casing

[icy drift]

Guy says Differential Diffusion inpainting works with Flux? 🤔 Really great results. I'm at work. Anyone want to test it?
https://youtu.be/Qiltbxvhr_A?si=UpIN8In_7df77WiT

[torn wharf]

lol invoke was telling poeple that flux inpainting was impossible

[icy drift]

I'll test it later. Might be tomorrow though.

[uncut river]

I think sd15 has some soul

[torn wharf]

i was thinking "if soul were anywhere you'd find it in people" but then i remembered this monster

prompted it by name with a description of it. i think it knows giant cedars but not that one in particular. its close! it vibes.

[bitter hearth]

did you use simpletuner or diffusors?

i wana train a lora on flux on my 4090

kinda waiting for kohya_ss script

[dry wave]

yes

SimpleTuner is fine. Not more complicated than kohya_ss

it supports quantisation and all kind of loras out of the box as it's using diffusers

I could imagine it takes some time until kohya_ss supports stuff like quantisation, dunno

[torn wharf]

yeh its not getting that one specific tree but it does get creative with trees. hehheheheh hole hehee

[rain current]

six fingers... welcome back

[jolly swan]

SD3 is not competitive with other options like AuraFlow or Flux so there is not much reason to build anything on top of it (or even use it, honestly). Flux is good but overall not as exciting as AuraFlow which lacks in terms of aesthetics but I think I may be able to fill that gap. tldr: I am getting ready to train the first smaller finetunes first to see how it goes but I think the main training run is getting close.