🆕｜sd3

[torn wharf]

lighter note... new flux lora looks like a huge success https://civitai.com/models/655633/the-beavis-and-butthead-loogie-lora

[bitter hearth]

lol

that was a hard video to make

[torn wharf]

Good song

[sage burrow]

That's actually good! Flux surprises me more and more each day lol

my forge produces nothing but grey boxes

think I like comfy better

[torn wharf]

forge did a lot of ground breaking work recently but i don't think it was ready for an actual release in this state.

[sage burrow]

my sd 1.5 turn out just fine, as you guessed I was trying with flux

[torn wharf]

we're in the #flux channel so i just assumed

[sage burrow]

😄

so has anyone used dreambooth with Flux or SD3? They aren't showing up in the dropdown list

even though I have refreshed and restarted plenty

[craggy crest]

dreambooth belongs to who?

[sage burrow]

Google says Google.... 😦

[dry wave]

we had such discussions, but I haven't said that unets are transformers, I said that the sdxl unet consists of transformers

[noble coyote]

My NF4 Flux Weights are in Checkpoint folder

[dry wave]

I clarified this because people often claim that sdxl has a convolutional unet architecture while pixart and flux have transformer architectures. But sdxl is also a transformer architecture. These terms are not exclusive.

because comfyui is a hacked piece of code. Many namings are off. The t5 text encoder for example is a subclass of CLIPModel which doesn't make much sense

they implemented comfy for sd 1.5 and hacked support for other models in there later - often without changing names

[torn wharf]

T5 loading into clip seems weird af too

[dry wave]

the code is extremely messy - but to be fair: this technology is evolving super fast. Often there is just no time to write clean code if you want to support all new stuff as early as possible

[torn wharf]

I think it's in business leader's best interests to keep the technical terminology obtuse and innacurate. Making sure the community is not a well informed population seems like a Machiavellian approach. it's hard not to believe it's intentional.

[noble coyote]

Flux.Dev and Flux.Schnell are in my unet folder; Flux.NF4 is in my checkpoint folder

[torn wharf]

yeah but at the same time comfy dev / anon publically calls out other devs for having sloppy messy code too

[dry wave]

or: it's a software project of a single guy who is too busy for Machiavellian approaches as he is too busy refactoring his code base

[torn wharf]

it's a business with investors now

[dry wave]

yes, I agree and I found that always very strange behavior

[bitter hearth]

flux sits nicely in my forge where every other checkpoint goes, models\stablediffusion... But the T5xxl goes in the Vae folder lol bizarre https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1

[torn wharf]

t5 in vae?! this is getting preposterous.

[sage burrow]

Perhaps controlnet would work to bring elements into flush which aren't part of the original training?

[torn wharf]

so much terminology confusion. so many avoidable situations. hard to beleive it's just because they haven't gotten around to a refactor

[noble coyote]

running this now but t5 is in clip

4 bit quantisation data not implemented - oops

[bitter hearth]

put the T5 in the Vae folder

along with the flux vae

[rain current]

But that input is for selecting VAE, text encoder,... it's not that the T5 is going to the VAE folder....

[rain current]

With what data has this 'HYPER TRAINED' been trained?

[bitter hearth]

more info here and more images to view the difference , Hope you all like it https://www.shakker.ai/modelinfo/bbe99aa7e86540a082f1e18b3af9131c?from=personal_page

[craggy crest]

so why would google care if they support sd3 or flux with it?

[noble coyote]

Prompt = Pope Francis by jay ryder, in the style of surreal creatures, punctuated caricature, uhd image, humorous tableau, necronomicon illustrations, loose gestures, sparse

Prompt = I feel beginnings Sudden trembling Of silence Ordinary Over the hills The wind Sighs Deep inside itself Breathe deeply The world Waits

[noble coyote]

Prompt = john mancuso's mummy (powder) In the style of cyber punk surrealism, baroque portraiture, mirror, made of insects, kinetic art, unique framing and composition, trick of the eye paintings

Prompt = that girl in the kitchen is going crazy, in the style of humorous animal scenes, colorized, chrome-plated, caninecore

[noble coyote]

Is there a png with workflow available for this Checkpoint at all?

[bitter hearth]

yes the links are in the model discription bro

[noble coyote]

I got the w/f 😄

[rain current]

dream <> fp8 <> guff.... I don't see any differences with the dev-fp8

[noble coyote]

Stormy

[bitter hearth]

Comfy Workflow : https://openart.ai/workflows/maitruclam/comfyui-workflow-for-flux-simple/iuRdGnfzmTbOOzONIiVV

Vae : https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main/vae

Clip download clip_l.safetensors and t5xxl_fp8_e4m3fn.safetensors from :

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

work flow , vae, clip l, and your T5xxxl

just take those from the hugging face dont down load the lot

[rain current]

But is there any explanation for why the results are identical to dev-fp8? Or am I not understanding what this model is supposed to do....

[craggy crest]

what sort of machine is it being run on?

[rain current]

I don't understand the question

[bitter hearth]

sorry bro im back, yes its the dev base i trained. Try the model with font and text the main training was needed there

[sterile pendant]

Oh because that was just the go to folder for the image generation portion of the models where they didn't include text encoders or the vae.

Within comfyui, he actually renamed the load unet node to diffusion model loader or something like that. Wouldn't surprise me if he adds a new folder named it as well, but also allows comfyui to also check the legacy unet folder to prevent it from breaking old workflows and whatnot

[noble coyote]

Made it!!

10 minutes at 20 steps

8Gb VRAM, 64Gb RAM

This Flux Dream Diffusion seems to get the text right 90%+ of the time! When I use NF4, it gets 70% nearly right; and 10% exactly right!

[sterile pendant]

Yes, it's true that unets are convolutional networks. Just because they have the some hacked in transformer operations does not make them transformer based architectures. The actual unet itself is not a transformer, with regards to the ML definitions of things. Dit models are though.

[noble coyote]

Flux Dream Diffusion

[noble coyote]

Pakistan?

[young blade]

Bit warm out today

USA

[noble coyote]

Ouch!

[young blade]

In the pool right now so it's perfect 😂

88 water

[noble coyote]

Ewwwwwwww

How do you get cool?

[hallow lion]

He is getting cooked not cool im afraid in that weather

[young blade]

Lol, ac works good in the house

Was my truck temp earlier so it's always a bit warmer in it

[noble coyote]

Texas? Arizona?

[young blade]

Fluxing poolside. Comfy not fun on mobile

Oklahoma

[noble coyote]

I knew Admiral William Crowe (Bill) from Oklahoma

Joint Chairman of the Chiefs of Staff under Bill Clinton

UK Ambassador 1995

[young blade]

Think he was a senator years ago

[noble coyote]

He also appeared in an episode of Cheers (he was an actor before enlisting)

[young blade]

Born and raised here, just a big bearded redneck 😂

Who loves beer and bonfires

Cheers 😂

[ancient cape]

https://civitai.com/models/656440?modelVersionId=734447

laughs in DND

flux training is slow if you do a 2k dataset (like over 24h) - but so damn worth it

even freaking kenkus are working 😄

[mortal mesa]

Fix loras having a weak effect when applied on fp8. comfyanonymous committed 27 minutes ago

[ancient cape]

shit 🤣 now my lora looks completely different

just updated

*angry shakes fist

[sage burrow]

One not too bad one from my lora, though still missing very important aspects lol

[mortal kite]

Its just Flux dev

oh is his thumb on the wrong side? wtf

[hallow lion]

Is this the long tube in aliens where he has to go do the thing?

[torn wharf]

So flux dev doesn't know about the tesla cybertruck. it makes weird tesla model s looking trucks. cool but not what elon was showing off x premium got flux as the backend. So flux pro is running a newer model already

[mortal kite]

That tunnel was even narrower. Oh speaking of that the new Alien movie it total meh

[hallow lion]

hmmmm

[ancient cape]

what was the prompt for that?
cause damn thats cool

[mortal kite]

"a horror painting of a teen girl wearing pajamas in a dark hallway at night. The girl is running toward the camera screaming. She has a look of fear on her face. Her pajamas are torn and tattered. She is bleeding from her arm. A hideous glowing tentacle monster zombie is behind her chasing her. The monster has a huge mouth with big teeth and is smiling. by H.P. Lovecraft"

[ancient cape]

according to their website. flux.dev is a distilled version of pro (unless I misread that) (and schnell an even further distilled one)

[torn wharf]

right. pro couldn't do cybertrucks at launch either. grok2 is using flux as it's image back end though and it knows the cybertruck. they must've either made them a custom model or updated the pro version

[ancient cape]

pro couldn't do cybertrucks at launch either
oh! O:

damn

if they're actually training new versions then thats impressive. cause that thing is expensive to train

[karmic gale]

s

[torn wharf]

they also could just refine their pro base

[mortal kite]

a more "vhs" looking one

[torn wharf]

xai has something like 8billion in funding secured so they can afford it

[ancient cape]

oh yeah. twitter/X is definitely using a custom model, as it includes all of the filtered content. like it can do all celebrities that are definitely not included by default

[torn wharf]

i was doing 70's tv show and gave it a graphic text overlay, and it made the graphic way past the screen margin, just like the good ol days. it has a nice retro aesthetic that's more proper

[ancient cape]

but they probably just ran a large enough finetune over it, with all (sfw) images ever posted on X 🤣

[mortal kite]

it just can't do asciii art, which kind of sucks, I want to do more 80s computers

need to make an ascii art lora. I wonder if Pro can do ascii art

SD3 at least could do that correctly. It sucked at most other things tho 😄

[ancient cape]

🤣 tried it with my lora

[torn wharf]

i know the perfect place for a dataset. ice.org

[fleet meteor]

Any idea why flux loras on sd forge don´t appear?

[torn wharf]

Don't think they're supported yey

[mortal kite]

figuring out the best keywords for crazy angles

[radiant ledge]

they are (with fp8 flux at least)

[torn wharf]

in forge or comfyui?

[radiant ledge]

forge

[mortal kite]

flux has a reeealy hard time with views looking upward...

[bitter hearth]

stay away, zombie, he's dangerous

[hallow lion]

lol

im just in awe that AI guns finally look more or less like guns

[bitter hearth]

Very questionable

[hallow lion]

well yeah

a gun expert will disagree

😄

[bitter hearth]

At least they look better than this

[fleet meteor]

Finally loras work on forge!

I think*

[bitter hearth]

wait until controlnet comes around

https://tenor.com/view/unlimited-power-star-wars-gif-10270127

[bitter hearth]

Not SD3 nor Flux but still want to show off my latest work. America's favorite blonde as dark magician girl

Also made the yu-gi-oh collectable card version

I'm basically gonna turn all the classic-era yu-gi-oh characters into photoreal. But I might wait until Flux finetunes + cnet come around since this took FOREVER with 1.5

[torn wharf]

didn't slim shady return with frosted tips an all and killed her chart numbers for a bit? Think ol slim is more fave.. but i guess he ded now

[bitter hearth]

apples and oranges my friend

[hallow lion]

Unlimited power will only come after perfect consistency and the abiliy to move that perfect consistency perfectly consistently (at least for 5-10 seconds) comes. This si when Hollywood will fall.

Right ow it's a joke. Pile on as many controlnet sor IP adapters as you want consistency is not there.

It has to be down to the last mole and hair

[torn wharf]

Says here that Madonna holds the most #1 hits of any recording artist, an she's blonde too.

[sterile pendant]

lmao, imgsys.org added the three flux versions (i've seen some pro ones on there, but it's not on the leaderboards yet. probably needs more votes to show) and already, they are the top by a landslide finally dethroning realvisxl 4.0

[bitter hearth]

taylor swift has the highest monthly listener count on spotify at 100M+

in fact the only one to have reached 100M

pretty safe bet she's America's favorite. Sure as hell is mine

[bitter hearth]

the skin detail is also not consistent. patience is the name of the game my friend. by end of year we're gonna have a killer finetune I'm sure

[hallow lion]

Yeah it won;t be long for sure

[sacred jewel]

she's a natural brunette

[hallow lion]

We used to say in 10 years when we talk about a new advancement in graphics. With AI its mor elike in 6 months. XD

[sacred jewel]

Anyone remember Silicon Graphics International? Those were the days 😛

[torn wharf]

so was marilyn

[torn wharf]

made the units that rareware used for DKC fkyeh i remember sgi

https://youtu.be/mMLHHxV0vWU peak sgi

[mortal kite]

Heli is a bit small...

[mortal kite]

why does it keep getting the heli scale wrong lol

[torn wharf]

i guess SGI stations were also on production of Jurassic Park and Starship Troopers, which imo still hold up. So those maybe peak

https://tenor.com/view/starship-troopers-shoot-bug-firing-gif-17413233

[torn wharf]

maybe the scale is right but she's an amazonian

[mortal kite]

death by snu-snu

[hallow lion]

Yes, the scale of things is the next big thing to tackle for the AI models. Even SORA messed this up big time in those early promo videos.

[bitter hearth]

Because it's a... Little bird

[rain current]

[flux lora experiment]

[tough oriole]

Just went thourgh 50 imgs on the imgsys rankings... it isn't even close flux demolishes.

[hallow lion]

Well dooooooooh.

[sage burrow]

I just put up my 2nd Flux lora on civitai. It is better in the showing what it's supposed to way, but still sucks lol

[hallow lion]

Well either way thanks for developeing and furtherign the tehcnology for all of us.

[sage burrow]

I can't wait until all the training data is released for flux 🙂
SD3 also 😄

[craggy crest]

you assume there won't be something before before that

[sacred jewel]

link?

[mortal kite]

Unfortunately I like to make images that are more than just portraits of hollow people looking at the camera ....

[sacred jewel]

That's awesome... you put out fantastic images!

[mortal kite]

Bottom right kind of looks like my wife's Uncle

[torn wharf]

i just wung an image through ultimate upscaler in forge. it was easy. this was just 0.55 denoise over 40 steps and megapixel squares

think i'm about done with the subreddit community. i saw a post saying flux has no creative styles at all and can't do img2img work. i told them that i dont know i think that's maybe wrong , but they insisted and then reported me for harassment. reinstating my account was a mistake. it's like 4chan there now. a place where misinformation goes to thrive.

[mortal kite]

take reddit with a grain of salt, its all 14 year olds

and bots

[torn wharf]

looking at this upscale i'm just reflecting on that from earlier, "can't do img2img" comments are baffling me

i dont think it's all 14 year olds. there are lots like me still hanging around there

[mortal kite]

pick any current social trend or hate train and reddit will be right on board. It's pathetic, they don't think for themselves

[torn wharf]

lurking in the shadows

[torn wharf]

sausage party! the new sequal series was dynomite

[torn wharf]

i can't really spot any seams on that upscale i did. remarkable.

less denoise should maintain more consistency too

[sage burrow]

It does the best img2img ever!

[limpid thunder]

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

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

Other languages
en: help
ja: help Japanese

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

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

Other languages
en: help
ja: help Japanese

[alpine summit]

Flux

[torn wharf]

is 847 isnt it?

[fleet meteor]

lol

[noble coyote]

Pope Francis says Good Morning! Flux@nf4

[noble coyote]

... trying to find the custom node which is 'breaking' ComfyUI since the new frontend ... ComfyUI works, but like swimming thru treacle!!!

Found it - and disabled - efficiency-nodes

[bitter hearth]

I love the new front end

add node menu is less buggy for me

and can easily filter for node packs

managed to squeeze a bit more image quality out of SDXL

with SEG node

[noble coyote]

Flux@nf4

[bitter hearth]

looks like the nf4 is not bad

[noble coyote]

nf4

The lighting is realistic

[bitter hearth]

Prompt for ray tracing

[noble coyote]

Flux@nf4

[bitter hearth]

ugly

[noble coyote]

Successfully ugly? 😄

[bitter hearth]

ugly ugly

[hallow lion]

basic flux dev?

[bitter hearth]

Enough woman face prompt

We all know these models can do that well

Time to game some bit

[noble coyote]

Flux@nf4

[noble coyote]

Slowly trawling thru custom nodes - disabling three at a time - to stop ComfyUI "working like its swimming thru treacle!"

Flux@nf4

[rain current]

https://fluxpro.art/create Flux Pro Free? 😮

[alpine summit]

Flux+rpg6 lora

[sacred jewel]

Yes, exclusively Flux.1 Dev for me... I get 20secs generations using simple scheduler. Beta scheduler can vary between 20 to 60 seconds per generation. Worth it for the extra quality

[sterile pendant]

dpmpp_2s_ancestral now works with flux/sd3/aura:
https://github.com/comfyanonymous/ComfyUI/commit/2622c55aff9433d425a62e5f6c379cf22a42139e
just note that since it's a 2s sampler, it will take twice as long per step. but hell yeah, finally an ancestral sampler

[sacred jewel]

For those like me who are really dumb, where would an ancestral sampler be better than a non ancestral one?

[sterile pendant]

well it's a really complicated topic that doesn't really have any one-size-fits-all answer. the big key thing is that it opens the doors for a new set of options

[sacred jewel]

Thanks... all I know about ancestrals is that they add noise back between steps? or something like that?

[sterile pendant]

it's the in the name, the whole ancestral part. basically, yeah, they keep track of prior steps

[sacred jewel]

You thought I was [that] smart? pfft... I thought ancestral meant they were older... LOLOL

[sterile pendant]

lol...

[sacred jewel]

meaning older than modern ones 😄

[dusky thistle]

they were discovered by decoding hieroglyphics

[sterile pendant]

👆

[sterile pendant]

insert "i should call her" memes

[sterile pendant]

city96 also added in k-quants for the ggufs now. so you can use quants like q4_k_s now, which is better than both q4_0 and q4_1 and only a hair bigger than q4_0

[sullen moss]

Hmm...

[tawdry jolt]

Mia Khalifa serving food in a restaurant wearing a protective revealing leather outfit

[tight whale]

Aquascape

[bitter hearth]

these are all amazing, I guess its flux dev ?

[sacred jewel]

Thanks... yes, all Flux.1 Dev 👊

[sage burrow]

Why do you insult other people's work?

[errant dust]

Oh rly?

Will check it out

It had been asked about in the community chat at Hugging

[bitter hearth]

flux1-dev-Q2_K.gguf at 4.03 GB

so smol

[errant dust]

The question is at what price

[bitter hearth]

yeah its a deal with the devil at that point

[errant dust]

My testing had shown 5_0 to produce the best balance, and 4_1 to be the best balance (diffs are nitpicks)

[bitter hearth]

okay thanks
haven't tried flux yet I'm just taking notes about it

[errant dust]

The biggest strength of the GGUF is that it did not drop performance speed like a hammer from the sky if you exceeded the perfect 8GB limit

So you could produce images at say 1280 x 1280 and it would not become an instant turtle. NF4 is great but has that issue for 8GB cards. 1k is fast and 1280 is s snail

[errant dust]

The fine details are really were you can see the effects of greater and greater quantizaton. I have an image and settings that helped me pinpoint this

[bitter hearth]

ah yeah that's an issue because fine details are my main focus

this is why I always multiply my sigmas by 0.8

[errant dust]

I literally have no idea what that would do or how to enforce that with Flux

[bitter hearth]

I don't know what it would do to Flux either

could well do the wrong thing

[noble coyote]

New ComfyUI commit 7333216

[noble coyote]

YFG prolific and always hi-quality output!

[bitter hearth]

sample_dpmpp_2s_ancestral_RF
we can get ancestral samplers for flow models now?

[noble coyote]

My 8Gb RTX 2070 has always worked well with NF4 - but maybe I am the exception?

[noble coyote]

Dunno

[mortal kite]

that HAS to take a while to render though like what, 3 minutes

[bitter hearth]

render time depends on hardware

[mortal kite]

he literally said 8GB RTX 2070

[bitter hearth]

ah ok didn't see

[noble coyote]

90 seconds is a typical 1024x1024 render - which I am happy with 🙂

[bitter hearth]

90 seconds is fine yeah

[noble coyote]

Flux.Dev however - when I use LoRAs - can be upwards of 25 minutes

So Flux.Dev is only experimental with me

Not sure that Flux LoRAs are really necessary - prompting can almost produce the same results

[bitter hearth]

current loras are probably pretty small datasets and fast training

compared to what may come later

[sacred jewel]

Thanks for the kind words... you guys (and girls) are of course the biggest inspiration. I plagiarize often 😛

[bitter hearth]

raiding the midjourney discord for IP adapter or control net inputs is a fun thing

[noble coyote]

... oh the prompts I've "borrowed" over the years ... ! 😄

[sacred jewel]

I am NOT NEARLY as creative as the best around here... so yeah, I "borrow"

[noble coyote]

Flux@nf4

[bitter hearth]

in my experience Florence2 can steal prompts pretty effectively

[noble coyote]

Does this have a Florence w/f embedded at all?

[bitter hearth]

there's a node but I forgot its name

[errant dust]

Perhaps you missed what I said specifically. It is fine if I do 1024 x 1024. but if I try larger image sizes like 1376 x 1376, it exceeds my VRAM and shifts to snail mode.

[noble coyote]

Thanks, its good to know!

[sterile pendant]

If you use the gguf versions, you can even do 2048x2048 images without ooming and the time scales pretty linearly. Like if you got 4 sec/it at 1024^2, it will be like 18 sec/it at 2048^2. I have an 2080 8gb GPU btw

[hexed dirge]

this is a lora just trained. Please try reproducing because I'm curious about the prompt

[noble coyote]

I should have added that the LoRAs I've been using can be done via prompting i.e. landscape, pop-art, art-nouveau. This looks is very sepcific and is probably better served via a LoRA

[noble coyote]

A regal figure adorned with intricate jewelry and a glowing crown sits on a throne amidst a futuristic cityscape. The scene is dominated by blue and golden hues, creating a dramatic contrast and highlighting the detailed architecture.

[sterile pendant]

oh and i forgot to mention in my last post to you that i'm using the q8 gguf, not even the smaller ones since it only makes less than a one second/it difference between q4 and q8 for me

[noble coyote]

is gguf better/faster than nf4

[sterile pendant]

yes, it's better. not sure about faster though, but it should be roughly equivalent

and it doesn't plague you with oom issues or spilling into the sysmem

even comfyanon recommends using it now and likely won't bother further updates for the nf4 format addon

from comfyanon's repo for the nf4 loader node "NOTE: This is very likely Deprecated in favor of GGUF which seems to give better results:"

[noble coyote]

This page? https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main

Q8

[sterile pendant]

yes, you need the node of his as well

https://github.com/city96/ComfyUI-GGUF

i recommend testing the q8 and q4_k_s models to see if there's a massive speed difference between them, but for me at least, there isn't

like i'll get 3.3s/it with the q4 and 4s/it with the q8. q8 is basically 99.9% as accurate as the full fp16 model

and again, that's with 8gb vram and 32gb sysmem

[errant dust]

Well, for me on the first run Q_5 K is very slow

[sterile pendant]

that's normal, it has to load the model

it doesn't do it until it hits the ksampler step

for some odd reason

[errant dust]

NO, I mean the plain rendering is crazy slow

[sterile pendant]

what card?

[hexed dirge]

With Comy and only 12GB of VRAM (64GB RAM) I get 2.6s/it with 4070 plain

[errant dust]

Laptop 4060

[sterile pendant]

have you updated the nodes today? i saw he put in a bunch of updates last night

[errant dust]

I did actually since the first time it actually crashed without the update

[sterile pendant]

i know he said something about the k quants potentially being slower "Most of them are coherent, though the speed may be slower compared to the legacy quants (for now)."

but for me, they haven't been that different. maybe try the 4ks version instead. i'd have to check the code to see if the version giving your problems has some extra expensive math in it

[errant dust]

Yeah... we are talking plain Dev speeds practically

for Q_5 K, but I restarted

it was at 35s/it

[noble coyote]

Dev is just too big for us mere mortals

[sterile pendant]

could be that annoying windows memory management issue as well, or if you don't have 32gb sysmem, it's doing a bunch of pagefile shuffling that's limited by your drive speed

and you're using the fp8 version of the T5, right?

[split bramble]

I like this one. 🙂

[errant dust]

Regardless, the normal non-K GGUFs did not have this issue at any of the sizes

[noble coyote]

Which LoRA?

[hexed dirge]

trained it today

[noble coyote]

I'm not the only one waiting for @hexed dirge to get back on his publicly-available-LoRA-horse once more! His LoRAs are a cut above!

Ethereal Grace

[errant dust]

Anyhow, a restart seems to have normalized the speed to 12s/it

Which is a bit slower than Q_5 plain, but not dreadully so. Maybe 25% slower

I will check image and then try Q_4 K

I have a plain Dev to compare to needless to say

[uncut river]

Reading this prompt is seems promising: "In an abstract realm of dense, intricate patterns, vibrant colors twist and tangle, forming an intricate maze of chaotic beauty. Each thread of color and texture weaves through the next, creating a pulsating rhythm of energy and emotion. The tangled lines blur the boundaries between order and chaos, a surreal dance of endless possibilities."

but try it in sd3, it almost gives noise

[noble coyote]

Recent commits have jarred my ComfyUI workflows - I had a nice setup 2 hours ago at 3s/it - and then i updated!!! 12s/it

[rain current]

It's the one I use (dev-Q8_0). I've done quite a few tests and comparisons, and it's the one that works best for me. Plus, it works very well with Loras.

[noble coyote]

Combing thru custom nodes and disabling - its like using a nit-comb on your kiddie's hair!!! 🙂

[errant dust]

I thought the quantized models were incompatible with LoRAs

[pseudo owl]

they are compatible I believe but its just that most things didnt support it

[noble coyote]

... just d/loading Q8-GGUF - its on a very slow (or very busy) - link

[noble coyote]

Poetry

[rain current]

gguf q8 - 24steps

[pseudo owl]

A pretty nice thing I found is just flux schnell seems to perform better at 512x512 res instead of 1024. Text on images is flawless and prompt following is slightly better as well. Same seed, same steps, 512 is left and 1024 is right
Prompt: "A monkey holding a sign that says "Well it seems like flux is better at 512 resolution then 1024?" and on the top right it says "Here is more text to prove that.""

[noble coyote]

Does GGUF go into Checkpoint or Unet?

[sterile pendant]

unet or the new diffusion_models folder

guess comfyanon heard my words about the folder needing to be renamed yesterday

[noble coyote]

... don't have that folder ... yet

[errant dust]

Ok, so the plain rendering of Q5 K took 6 mins. By means of comparison, NF4 or Q4_0 took roughly 2.5 mins. But plain Q5 was no speed demon so this is not unexpected. First analysis shows Q5 K to be 99.5% identical to pure Dev. You need a microscope to see the diff

[sterile pendant]

update comfy and run it to make the folder

[errant dust]

and pure Dev took... well, forever

I placed in Unet and the node finds it

[sterile pendant]

"Enable non blocking transfers in lowvram mode." oooh that might solve some issues https://github.com/comfyanonymous/ComfyUI/commit/73332160c8c9843876680fb04f037793c73d55b6

[noble coyote]

I updated to the new commit about 90 minutes ago - it broke my w/f - so I did a hard reset to the previous commit

git reset --hard 2622c55

So no diffuser_models folder ...

[sterile pendant]

well how did it break your workflow?

[noble coyote]

Went from 3s/it to 12s/it

I've been having node clashes - been disabling over 100 of them - seems like IPAdapter-ComfyUI and efficiency-nodes-comfy are the culprits

The minute I disabled these two - the speed went back to normal

[sterile pendant]

yeah that's probably the main problem then. well for now, you can test by saving your current workflow as a backup. then update comfy and make a blank basic workflow to test

[noble coyote]

But the recent commit knocked the speed back down

[sterile pendant]

a lot of addons use monkeypatches, so they might f with the rest of the systems under the hood, if they aren't up to date

[noble coyote]

OK, Q8-GUFF had d/loaded ... pip install --upgrade gguf

[uncut river]

at least on thing sd3 is a little better at than flux, grainy photo stuff

[sterile pendant]

you installed the city96 gguf addon with the manager, right?

[rain current]

This lens flare is very realistic

[noble coyote]

No - via git clone

[sterile pendant]

ahh, you're installing the requirements.txt then

[noble coyote]

Yes

All requirements satisfied

[sterile pendant]

and when you load the model, the node will be something like gguf, just search for that. everything else would be the same as the nf4 workflow, just replace that one node

you'll still use the same dualclip loader for the clipL and t5

[errant dust]

and VAE

[noble coyote]

I'm using ae.sft

[sterile pendant]

is the node

[errant dust]

How did you revert to a previous Comfy?

[noble coyote]

git reset --hard 2622c55

[errant dust]

That reverts?

[noble coyote]

Indeed it does

[errant dust]

anyplace specific to do that?

[dry wave]

🤩
nice to hear that you make Loras again!

[noble coyote]

In my cmd terminal

When in the ComfyUI folder

[hexed dirge]

thanks

[errant dust]

ah that was what I was asking

the install folder or the sub-ComfyUI folder?

[noble coyote]

X:\ComfyUI

The main mother folder

[errant dust]

ok, I asked because my main folder has three folders: Update, the Python and ComfyUI

[noble coyote]

OK, you have portable then

[errant dust]

yes

[noble coyote]

I use system ComfyUI

[errant dust]

I don't use the portable stuff, but that was the starting point

My 'reply' was:

C:\Users\silve\Applications\ComfyUI_SD3-Flux\ComfyUI>git reset --hard 2622c55
HEAD is now at 2622c55 Automatically use RF variant of dpmpp_2s_ancestral if RF model.

Sound right?

[noble coyote]

Yes its exact

[errant dust]

it does say this was released today you know

Loading: ComfyUI-Manager (V2.48.6)

ComfyUI Revision: 2563 [2622c55a] | Released on '2024-08-18'

[noble coyote]

What do I do for a VAE now I am using the GGUF UnetLoader?

[errant dust]

Use the standard Flux VAE

[hexed dirge]

is there an fp16 dev that works with forge?

[noble coyote]

Sourced from where?

[errant dust]

Flux

[noble coyote]

The GGUF UnetLoader has no VAE out ...

[errant dust]

NO, the original Flux VAE

via the VAE node

[noble coyote]

The one hiving off the checkpoint loader?

[errant dust]

or just the plain VAE node

[torn wharf]

flux vae should be a separate file and you load it with Load VAE

[noble coyote]

Got it!

"We're up and running Q8-GGUF!" 😄

[errant dust]

https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main/vae

[torn wharf]

the filename by default is AE.safetensors

https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/main/ae.safetensors?download=true

[sage burrow]

Does anyone have a simple flux dev (not ns4) wprkflow which uses a lora?

[noble coyote]

I am getting 7.5 s/it - still its probably node conflicts - even after a back-commit

NF4 was 3s/it

[errant dust]

What was the commit for yesterday?

[noble coyote]

2622c55

[errant dust]

the reset was still dated today

That one is dated 8-18

which is today

[noble coyote]

It was updated about 2 hours ago - sorry, 2622c55 was updated about 10 hours ago ...

[errant dust]

WHich is why I asked about yesterday

[noble coyote]

And a new 7333216 commit is in place

Don't know about yesterday's commits - mebbe there's a list somewhere?

Q8-GGUF

5 minutes

Which is about 3.5 minutes slower than nf4

[turbid grotto]

I heard that with gguf it can be possible to run multi-gpu, is that true?

[torn wharf]

I really think that flux upscaling kills the need for 4xultrasharp style upscalers

those are fast but the quality aspect lags

if you only have 8gb of vram, try using the q4 gguf. it should all fit into memory then

the speed differences between nf4 and q8 are probably due to your system's current state. maybe you have more browser tabs open this time.

[noble coyote]

I'm getting advisory warnings ... \ComfyUI\custom_nodes\ComfyUI-GGUF\dequant.py:8: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor)

???

[torn wharf]

tha'ts more for the developer to see

[noble coyote]

OK

[torn wharf]

something pytorch changed and wants devs to know about

[uncut river]

hm, not quite

[noble coyote]

Q8-GGUF

[hallow lion]

That would be neat as then you could upscale and refine with sdxl or sd15 faster. 😮

[bitter hearth]

Cute

It's watercolor right

[uncut river]

no, it's complicated. prompt was something like:

{day|night} time, {flat colors|realistic lighting}, {front|back} view, (Perched on a sunlit hill, the intricate Spanish coastal town unfolds like a painting. Tourists wander narrow, winding streets, their footsteps echoing against brightly colored facades. Balconies adorned with vibrant flowers overlook the shimmering sea, while the scent of salt and citrus fills the air. Laughter drifts from bustling cafes, where locals and visitors savor the town’s timeless charm:0.32), (signature text "NOEDEL" in the bottom right corner:1.23)

so maybe a photorealistic prompt description together with "flat colors"

[sacred jewel]

Love it. Is this with a LoRA?

[hexed dirge]

yes

[sacred jewel]

Private or publicly available?

[hexed dirge]

private

[frail shoal]

Had to install forge to use flux dev nf4. Worth it. Very fast, i'm even upscaling.

[dull star]

https://civitai.com/models/659150/zdzislaw-beksinskie-style-flux-lora?modelVersionId=737548

Excellent

[dry wave]

I remember you don't like civitai for good reasons - but any chance you will publish your loras? Maybe on huggingface instead?

the artstyle is awesome. But also: this extreme fine details would probably not be possible with SDXL

[hexed dirge]

maybe, yes. Now I'm testing them but can think about huggingface

the problem is that with civitai I got token for training

with HF no

[dry wave]

ah, I see. i train my stuff on my graphic card. Fortunately, with all these optimizations that's possible even with Flux

[hexed dirge]

well I have 12GB 4070

not suitable for training flux loras

[dry wave]

hm :/ yeah, probably difficult

[fleet meteor]

Could anyone got the flux dev q6 working on forge? The q8 .gguf work fine but the q6 continue giving me errors

[shy stirrup]

@heady patrol a simple wedding-ring with a broken metalik heart and a red gem as blood

#artisan-2 a simple wedding-ring with a broken metalik heart and a red gem as blood****

#😊｜co-creators a simple wedding-ring with a broken metalik heart and a red gem as blood

[uncut river]

"gabriel the angel of death comes down for your poor soul"

Death agreed long ago specifically not to interfere with the living, but from time to time he just can't resist to fool around a little bit...

[errant dust]

This is amazing, from a LoRA I assume. Tell me, would you be willing to post on some site? Also, would you be interested in doing Frank Frazetta? I have a good number of his stuff digitalized. Though not labeled, which is fixable.

I mean his fully illustrated comics

not the 40 odd covers

[hexed dirge]

At the moment I'm testing that Lora. After I should think about the best place to post. Also since i use buzz it costs 2K buzzs for training and anyone can.

[errant dust]

I am unfamiliar with the service. Can you link?

[hexed dirge]

https://civitai.com/models/train

[errant dust]

I have a dinky 8GB laptop 4060 so.... training locally is not an option

[hexed dirge]

also for me

but online it costs.....

[sterile pendant]

"It fixed some perf issues but caused other issues that need to be debugged." might be relevant to the issues you were having and that you had to rollback for:
https://github.com/comfyanonymous/ComfyUI/commit/6730f3e1a362d5f3ed44f8541517b03356e7bf0e

[bitter hearth]

vast.ai is consistently the cheapest option
for anything training/inference related

[errant dust]

True, but I wouldn't know where to begin to train on it

[hexed dirge]

with that link.. Upload images and select flux.....

[bitter hearth]

if you can just get to the level of linux knowledge to open up a working python environment on a fresh machine
then you can use diffusers

[errant dust]

Yeah.... You could have simply said, "If you learn Chinese and Japanese and Aramaic, and pray to Buddha, it is easy"

Meaning Neon, not Andreac

[bitter hearth]

I agree its a pain

[hexed dirge]

no. It's just easy.

[bitter hearth]

fighting servers

[hexed dirge]

ah ok

[bitter hearth]

also, not sure which country you are in
cloud is better value the closer to Sweden you are

not sure why but Swedish internet is insanely good and that goes for their datacenters too

[errant dust]

@bitter hearth I generated hundreds of GB to train chess NNs for well over a year of nearly 24/7 macines on Vast.AI, so I am familliarwith the challenges of using it. It is a monster pain, and you'd need to have all the tools and whatnot well set and oiled

[errant dust]

Brazil, and nothing here is easy or cheap in terms of that sort of thing

SOme is though. Plain internet is obtainable at decent rates

but the rest is ugh

Anyhow, I did resolve the speed issues with the new GGUF models on Comfy, by roling back to yesterday and setting --lowvram

[bitter hearth]

ah I see
South America situation must be very different yeah

[errant dust]

Don't even think of lumping all the countries here as one. I don't mean that out of some nationalistic statement, just that the economic and technical realities for each are as disparate as Europe to remote Africa

[bitter hearth]

ah okay yeah I see

[errant dust]

Anyhow, gguf Q4_K t (4.7s/it) is still slower than NF4 (3.5s/it), but is also better quality. Q5_K (6s/it) produced a near match of pure Dev, but the sample size of one needs at least a couple more to show this is a reality or a lucky coincidence. Will test Q6_K first

[bitter hearth]

sounds good

[dry wave]

I would love to have a frazetta lora. If you are willing to share the data I could train one.

my current approach of training frazetta is to train on sd 1.5 created images with frazetta in the prompt - but the sd 1.5 inages only roughly approximate his style

[errant dust]

More than happy to, but likely need to breakdown the full page illustrations into single images.

[hexed dirge]

yes

[torn wharf]

sd ultimate upscale. i used the forge extension but its probably so good on comfyui too. i'm tellin you. this was barely any effort. just .. ugnn.. i just threw settings at it and this is easily the best results i've gotten from the extension

time to make a plan an get serious with this. flux is a seriously high quality tool that opens up so much potential

[bitter hearth]

that's really high res yeah

[torn wharf]

just told it do 4x , 0.55 denoise, megapixel patches. i think i had seam fix on too , but it made the two versions, one without seam fix and one with and i can't tell which is supposed to be fixed

might've been 0.51 denoise

[torn wharf]

vintage photograph from a waterpark in my hood growing up. before and after flux ultimate upscale at 0.48 denoise

lol it added a ton of boulders to the walk way, but that's actually realistic for this place

[bitter hearth]

crazy how well that works

[torn wharf]

https://imgsli.com/Mjg4OTE2 it's so baffling that people are running around reddit claiming flux can't do img2img

use the link. it shoudln't have embedded. it's a comparison site

[dull star]

apparently it can do inpainting too?

don't know to what extent or how well

[bitter hearth]

with comfy any model can do inpainting

I don't understand why reddit makes things up

the idea of a diffusion or ret flow model that can't do img2img makes no sense

if it can't make an image from an image then how is it going to make an image from latent noise? 🙃

[torn wharf]

yeah forge should be set up to do inpainting now. img2img wasn't working earlier this week but it sure is today

[bitter hearth]

there's so many different inpainting methods now

gets confusing

[hallow lion]

Hard to pick the best approach.

[torn wharf]

the guy argued with me and said that he had talked to one of the flux creators on reddit and they told him that they didn't condition the model for img2img and it woudl always produce hazy outputs no matter what. i was like "uhhhh.. this sounds like \my dad works for microsoft\ " then he reported me for harassment

the whole sub reddit makes no sense

[bitter hearth]

wow yeah that's bad

[torn wharf]

afaik the podcast that just came out is the ONLY public posts BFL have done outside of their website

i guess thye made 3 twitter posts too

[bitter hearth]

what's the podcast?

[torn wharf]

https://a16z.com/podcast/the-researcher-to-founder-journey-and-the-power-of-open-models/ it shoudl be on the various pod networks too. play, apple, spotify, i dunno how half it works

its pretty good and worth a listen

[bitter hearth]

ty

[torn wharf]

lower denoise and a more descriptive prompt is helping to upscale this image a lot.

5 min it'll be done. brb

[sacred jewel]

Would you be willing to share best settings for Flux Training on CivitAI? I want to give it a shot... never done it before. Unless the defaults are fine...

Thank you 👊

[errant dust]

Apparently Flux Pro is very pro-aged people since apparently this image is of a "YOUNG man":

All I can say is that if you are 25 and look like that, plastic surgery is an option

[hallow lion]

He has seen too many AI generated hands.

[torn wharf]

at lower denois it left a lot of the low quality artifacts, kept the crowds smilar, but also changed my favorite part the arbutus tree

trying forge's multidiffusion integrated now. its working good too!

[hallow lion]

Those sliding tubes are a proper Moebius strip.

You'd end up in the 4th dimension sliding thru those

[torn wharf]

thats a real park that closed when i was a kid. this is from before my time.

[hallow lion]

😮

[torn wharf]

though i ddi get to experience all fun before it died. those are legit

here's another angle from teh banana pool . these kinda waterparks were pretty common around canada

[torn wharf]

yeah screw it. i'm tossing the ultimate upscale script. multidiffusion is all you need. i used 0.41 denoising on this one and it barely touched the low quality artifacts of the original, but i can already tell it did a LOT better of a denoise and in a faster amount of time on account of less vae calls

[bitter hearth]

Average Smallest American car

[sacred jewel]

prompt was simply:

Cadillac Escalade Interior

[torn wharf]

whats nice about flux is cars dont have some kind of concept interior where the passenger has some gunner seat with a steering wheel too. they're proper now

[bitter hearth]

But that sounds way cooler

[torn wharf]

i was kinda thinking of that cool intro scene in the new ghost busters with the sweet ecto 1 gunner seat mod when i typed it

it is pre cool

[hallow lion]

that gunner seat was neat

[bitter hearth]

idk what I'm doing

[hallow lion]

https://huggingface.co/InstantX/FLUX.1-dev-Controlnet-Union-alpha

hmmmmm openpose!?

[sacred jewel]

INSANE how it just generates a high quality image from such a short prompt... just like that.

[bitter hearth]

Meanwhile me on sd3 doing random crap

Feel the holiness of the cookie

[hallow lion]

You sure know how to get the most out of SD3.

[bitter hearth]

Best model

[errant dust]

Not to mention its fine sense of irony. "average SMALLEST American car"

[hallow lion]

Hey latest AI models come with a humor module.

[errant dust]

Yes, I saw when I asked for a young man

[hallow lion]

loool

I must be messing up coz my flux upscales are a blurry mess

[torn wharf]

just get a GMC denali. they're teh same thing with less brand prestige tax

[sacred jewel]

And much different interior. I prefer the Escalade dash... the new Electric one has an insane dashboard

[torn wharf]

you know you got an escalade .

actually the new chinese electric SUVs have been catching my eye and making me say wtf

the yangwang caught my eye first because, well, the name. the yangwang u8. wild looking

(real photo for sweet interior reference)

[bitter hearth]

Real photo of the ballpocalypse that happened in a closed off country

[torn wharf]

yeah thats it. boxxie

[sacred jewel]

Can someone enlighten me on the proper way to use Flux LoRAs? I see some workflows call for ONLY the model pipeline and others have the full Model and CLIP in the pipeline...

ROTFL... this was with the Arnold LoRA LOOL

[timber dagger]

a guy with a hero mask

[sacred jewel]

this what Flux.1 Dev knows about this car:

prompt

yangwang u8

[torn wharf]

Snake Pliskin in : Escape from Ballsylvania : Balls out

[timber dagger]

/generate a guy running

[bitter hearth]

https://tenor.com/view/gorba-za-gif-25393309

[hallow lion]

It's Sawnick the Hedge Dog.

[acoustic kite]

i love how oddly specific we can get with prompts now with flux 😂

[torn wharf]

anyone remember april editions of game mags?

[errant dust]

One thing that has become quite visible is that the text parser of Flux Pro can produce strikingly different output. At first I wrote this off as just RNG doing its merry dance, but no longer.

[bitter hearth]

sonic after he browsed 4chan

[errant dust]

More specifically, Pro's text parser produces many times far more interesting, if not always as precise output

I am attributing this to the parser since I have not seen anything in Dev and Pro that suggests vastly different quality image output per se

[remote holly]

Sanics !

[bitter hearth]

@remote holly qanics!

[bitter hearth]

What's that on its mouth

[errant dust]

"The treasure is buried in the ....." <croaks>

[acoustic kite]

he didn't wanna run alone

[torn wharf]

he probably did but she just invited herself. typical. amirite?

[errant dust]

So to illustrate what I meant about Pro and Dev, here is an image by Pro (note that 2 out of 4 times it fails to produce any text at all), and the images are all in these lines:

To contrast, here is Dev (tried multiple images, but all are similar):

The prompt is: A captivating image featuring a young man seated at a desk, diligently working with a pen in his hand, his silhouette barely visible against the backdrop of a mesmerizing fantasy scene. The double exposure technique reveals a lush green forest, a valiant knight, and a female mage casting a spell, all seamlessly blended together. The words "Imagination and creativity are the mind's wings" are elegantly written in large bold letters on the side.

size was 3:2 or on Dev 1280 x 800

[torn wharf]

pro sacrifiiced prompt adhearance there. probably an aesthetics guidance thing. the dev gens had the double exposure technique down pat

and really, if we're talking about text, just generate the image and then use any editor from 1999 or newer to render text on it.

[sacred jewel]

https://tenor.com/view/shrug-frank-sheldon-christopher-walken-the-outlaws-fine-gif-25275641

[sacred jewel]

Blobstyle LoRA

[craggy crest]

flux with no loras

[torn wharf]

trying to make the beastie boys but flux is like "Austin Butler as THE BEASTIE BOYS"

i'd watch it too

this is closer but even if i really lean into hard descriptions of each of them, zeroing in on one album in particular, can't do it . doesn't know beastie boys very well. thinks they're just generic boys but they're not. they're not generic boys at all!

this ones like timothe chalmamet stars as mike d

could he pull it off? from muad'dib to wonka to michael diamond?? i have hope

[rain current]

testing 2mp`(ggufq8_0)

[mortal kite]

kind of looks like Q-Bert a bit

[sage burrow]

It has officially been confirmed, Flux can infact create male anatomy! Needs a lora to do so though.

[mortal kite]

Prompt: "a comic drawing of a software engineer sitting at a desk with an exasperated angry look on his face. He is pounding his fist on the desk. There is a high end workstation PC on the desk. The monitor is displaying a pair of female breasts sticking out. He is shouting 'I was looking for Breadth first algorithms!'. A female employee is standing behind him looking at the PC monitor with a shocked look of horror on her face and she is fainting. She is shouting 'HR!!!' "

[sage burrow]

You can get two different speech bubbles with accurate text?! 🙂 🙂 🙂

Look what I found today!
It would probably be handy to perfect some SD3 images
https://huggingface.co/spaces/SkalskiP/FLUX.1-inpaint

[mortal kite]

Apparently! I tried it and it pretty much worked

I will add the prompt to the post

and I had to get a run gens to get a really good one of course, but only about 5 or 6 times

[sage burrow]

I was also curious if you had prompted the black censorship line or not, now I know 😄

Now I'm going to try ithat layout on some other themes!

[bitter hearth]

Cx

[torn wharf]

my favorite of the upscales. keeps the most quality. has just enough jpeg to be believable.

[bitter hearth]

Big tubes

[torn wharf]

#🆕｜sd3 message the orignal pic is here. it was a real park

[rain current]

https://civitai.com/models/660906/crying-cat-flux1-dev?modelVersionId=739589

[bitter hearth]

[sage burrow]

lol

[torn wharf]

everyone doing flaming gpus. missing out on the stuffed pillow gpus. throwback to my first day with flux when i was tryin to break down the walls of expectations

[torn wharf]

if you think about it, the brain is just a lot of balls with tentacles

[sage burrow]

I've been using flux1dev_v10.safetensors this entire time. Is that the best one for low vram? I've seen so many come out that I've completely lost track now

nf4 doesn't help for testing my Dev loras, do not trying to use it currently

[sacred jewel]

Finally getting around to training a Flux LoRA

[torn wharf]

brain balls

[torn wharf]

is that the spy dataset??

when flux gets ip adapter dataset building is going to take off

[sacred jewel]

Yup... I just uploaded Epoch 5... totally used default settings from CivitAI. Not sure what GREAT settings should be

https://civitai.com/models/660770?modelVersionId=739419

and with LoRA at 1.00 strength

[torn wharf]

has anyone else tried stuff like prompt editing on the newer models? it never seems to work like it did before with sd15 and sdxl. like [cat|dog] style prompts

heres a [horse:dog:0.6]

[torn wharf]

it seems to blow out the details when you shift the attention. here's a [horse\🐱0.5]

doing the alternating steps approach like [horse|cat] just completely blows out the subject

[mortal kite]

Nope I added that myself

[torn wharf]

if cfg is set to 0 it does nadda

i its probably really incompatible considering the t5 and parallel text network

[mortal kite]

Another horror genre

[sage burrow]

I tried it a few weeks back and it didn't seem to make a difference. I should try again. What could be interesting is using it when using Claude or GPT as a prompt enhancer, it's likely that by now they would know what that means... Goign to try that I think

[torn wharf]

my best results were low guidance and 2cfg. i think its a case of the code and the architecture not lining up

[mortal kite]

Juuuust a little adjustment:

[torn wharf]

POSTULATION. This is a ball.

[mortal kite]

No spine, no problem:

[sterile pendant]

Friendly reminder that chiropractic "treatments" are pseudoscience

[craggy crest]

prompt: Narrative realism

prompt: Narrative

[young blade]

why is knit so good in flux

[craggy crest]

add the word 'rabid' to that

[bitter hearth]

whoah the knit is great

this is so cool

must have chucked a lot of it in the training data I guess

[bitter hearth]

I like the effects of this prompt a lot

[torn wharf]

not what i prompted at all but its pretty. trying different extensions on forge

[bitter hearth]

looks nice

[torn wharf]

oic . i got one button prompt turned on still

[craggy crest]

me too 🙂

flux does hands

[noble coyote]

Beksinski good

[gray tundra]

||Cacca||

[noble coyote]

I'm sure it can be done better - but here is one (barely) after about 20 tries SD3 (w/f in png)

[bitter hearth]

with the new comfy ui execution inversion

what you can do is set it to retry until a vision model says its correct

[craggy crest]

do what?

[torn wharf]

Conditional loops

[bitter hearth]

run the generation with different seeds
sorry I forgot to actually say what should happen in the loop 😄

[noble coyote]

Using Dice_AI's Platinum SD3 DD Checkpoint

Has his leg gone thru the bath? Or is the bath growing around his leg?! 😄

[sullen moss]

SD3 is great for a magic mushroom trip. LOL

[noble coyote]

AuraFlow0.3

[silk solstice]

This is just super gorgeous

[placid belfry]

i am sharing this here because i cant share it in general channel

[bitter hearth]

dev or schnell?
and is it fp16, nf4 or a gguf quant?

ignore the second question if you don't know

[sage burrow]

Edit to add: it's mostly glif, flux, html, then save web page as image than flux superpowers

[dull star]

lmao

is this some glif where it generates an amazon product page for you

[sage burrow]

It is, then my remix of it lol

[bitter hearth]

we rly need them to publish the flux paper
to find out what sorcery produced such amazing text

[dull star]

I want to know what they captioned the dataset with

[sage burrow]

It's via gliff, with Claude sonnet help etc etc

[bitter hearth]

ah okay yeah

[sage burrow]

Just hit the remix button and it's all in there https://glif.app/@LadyLalita/glifs/cm00yuxa500026doktv9xznfu