🆕｜sd3

[hexed dirge]

Alo working on this one

[errant dust]

Is that an ideal value?

[hexed dirge]

no

16 is way better

[errant dust]

Not judging, this is noob question who plans to try

[hexed dirge]

but it worked for that and since it costs so much training I published

[errant dust]

completely grateful, and downloaded it, though haven't had time to run it yet

[vestal copper]

U

[dry wave]

it's cool, but the noise is horrible 😦

[bitter hearth]

very nice and big

the window of the car of course

gib prompt

[sacred jewel]

High quality, detailed, sharp, masterpiece, an intense explosion, showing a building collapsing amid a massive fireball and thick black smoke, with fragments of debris flying outward in a chaotic, destructive scene against a clear sky

Seed: 477049711089277
Sampler/Scheduler: Euler/Simple
AR: 16:9
Flux.1 Dev

[torn wharf]

protoball

[bitter hearth]

16:9 is nice yeah

its more interesting than square, but more reliable than 21:9

I love 21:9 but most SDXL model take a big quality hit

apart from one single checkpoint (Chinook)

[sacred jewel]

Agreed. I love 16:9 and 21:9 but the latter I use when doing WALLPAPERS for my widescreens.

"My my that's a big knife you got there hon... I swear, I did not cheat on you with your sister... now please, put the knife down... "

[sacred jewel]

I actually like the noise...

[noble coyote]

Flux.Dev

[bitter hearth]

TBH I sometimes add roughly that amount of film grain 😛

[sacred jewel]

I make heavy use of Adobe Camera Raw film grain feature... just sayin'

[sacred jewel]

these are great, Torcello

[noble coyote]

Hey thanks 🙂

[torn wharf]

me thinking "fake film grain is obvious an you should get a print of your image on real film for the true effect"

photoshop filmgrain filter is no different than an instagram filter or fake bokeh effects using calculated depth maps

[craggy crest]

and the right ratio for youtube videos

[bitter hearth]

When the girl says she's alone

[bitter hearth]

wanted to throw it at sd3

[sacred jewel]

Nice

[hexed dirge]

https://civitai.com/models/677725

I really hope SD3 will be, if it's ever possible, better than flux

but flux is amazing

really

[bitter hearth]

any suggeation on which gguf model to download ? https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main

[noble coyote]

Q8 and Q4 do well on my 8Gb VRAM GPU

[sage burrow]

Did comfyui get faster or something? My flux dev renders don't take ad long now

[bitter hearth]

ok, but i chose to grab the Q6_K for both the model and t5xx

got 12gb vram

[bitter hearth]

ERSGAN fine tuning would be ideal for this

keeping it 1x magnification

[errant dust]

The K-models are the newer and better ones. Q5_K is superb

[bitter hearth]

ahh cool, didnt know about K model being better. im already downloading them but Q6_k

[hexed dirge]

are them better than fp8 and support loras?

[errant dust]

I tried them all, The K-quantization is a higher quality one, and explained by the author in the description. In my experience Q6 doesn't have an edge over Q5.

[errant dust]

Better is always going to be a hard eval, but yes these support LoRAs just fine. In fact..... I justr ran your new loRA on an image and will share the original and the one with the LoRA

The difference was NOT small....

[hexed dirge]

I'm very curious

[errant dust]

So I ran the following prompt (workflows are in images anyhow): satirical cartoon of an artist at his easel painting a picture with chess, detailed, high quality The first images is pure Flux Dev (Q5K GGUF), the second is identical but with the Blue Future LoRA at strength 1.2. All else is identical. I was staggered at the effect.

[hexed dirge]

thanks

[errant dust]

I realized I had not specified it paint about chess, so I fixed the prompt for a new batch, but this was really interesting

I'll share the new batch when done.

[hexed dirge]

Also if you use for Sci-fi loras does it bests

[errant dust]

I will do exactly that. Though I am a hardcore chess nutter, I will do one with the guy playing some android

[errant dust]

Really this was my first use of it, and I like to test in very controlled conditions., to really get a measure of the impact on even general situations

[noble coyote]

@hexed dirge LoRAs are superior He was sorely missed during his small 'retirement'

[sacred jewel]

so good

[hexed dirge]

oh wow

[sacred jewel]

Can I have a favorite LoRA? LOLOL deee -leee-cious..

[hexed dirge]

great images and thanks

[sacred jewel]

Thanks.. .the LoRA did the heay lifting... thank [YOU]

[errant dust]

so the second batch came out also with a very pronounced style. Next will be with SciFi, but for now:

[dry wave]

up to date Comfyui should be able to handle Loras on fp8, too

[errant dust]

I will add that the AI that did the best with this prompt was Imagen 3, quite outstanding in fact. Dall-E 3... could not accept that chess was only being painted about.

[bitter hearth]

haha nice

I love chess

[real terrace]

Is this flux dev or which one?

[torn wharf]

flug dev

[real terrace]

it is a nice name

[errant dust]

What is the prompt?

[pseudo owl]

schnell at 10 steps, 768 res(performs better prompt following and text rendering at this res). Not perfect but decent.

[sacred jewel]

Flux.1 Dev

[sacred jewel]

the boy stands in front of a large black nose piece with big eyes, in the style of dima dmitriev, large canvas format, light red and red, craig mccracken, gigantic scale, haunting portraits, playful and fun imagery

[rare gust]

Sorry guys, new here, how does this work? How to generate?

[sacred jewel]

...and the Blue LoRA of course 😄

[sacred jewel]

https://discord.com/channels/1002292111942635562/1237461679286128730

[errant dust]

That is the prompt for this?

[bitter hearth]

are you guys generally using facedetailer while using flux?

[zealous verge]

lol

[pseudo owl]

no need, flux gets good faces most of the time because of the vae.

[craggy crest]

raw flux dev

[bitter hearth]

i can't seem to get loras have any effect

using the flux1-dev-Q6_K.gguf

[errant dust]

Workflow is in the pic

[bitter hearth]

got unlucky with cloud servers today couldn't get one that worked

[errant dust]

LoRA was used

[bitter hearth]

looking at it now, thanks

i see you are using t5xxl fp16

could that be a reason in my workflow cause im using t5-v1_1-xxl-encoder-Q6_K.gguf

compressing a text encoder for diffusion models is a bit of an unknown thing in terms of the effects

[bitter hearth]

so you suggest not using it?

what I would suggest is switching to cloud and using uncompressed stuff
but if that is not an option then what I would do is unload the text encoder after creating the text embeddings

a lot of people did this with SDXL, for example

[bitter hearth]

ok well im getting the black forest lab's version of the t5xx fp8 encoder

[sterile pendant]

It's really not. Semantically speaking, it just leads to an occasional token being something slightly different here and there. If you've worked with LLMs before, you'll know what I mean. Overall, they tend to be 95% the same if you're comparing a Q8 to something like a Q4km

using quants with diffusion usually just has more of a micro-seed variation effect. like some rose in the background might change shape, but all the other pixels stay mostly the same

assuming you're staying above a q4km quant

q8 is almost identical to fp16

[bitter hearth]

I think we have good published data for LLM tasks for this, where it doesn't impact benchmarks that much, but not for this particular case of using them as text encoders for diffusion models

[sterile pendant]

it has even less of an effect, actually

using fp8 versions of the LLMs are already negligible vs fp16 and q4 to q5 quants already beat fp8 quants usually

[bitter hearth]

if it hasn't been done in a recent paper I think it will be done soon, due to the sudden surge in the popularity of quants for the text encoder, but what I would like to see is FID and Clip score on MS-COCO for this

[sterile pendant]

you'll just see a bunch of margin of error trends. just like they found with LLMs when people were trying to make the same arguments about quants

a q5_k_m quant is within 0.06 percent perplexity of a full fp16 model

[bitter hearth]

oh damn, the loras are working fine now with workflow change

[dry wave]

if I remember correctly, quantizing the text encoder has, surprisingly, stronger effect on the outcomming image than quantising the diffusion transformer model

[bitter hearth]

one with vintage anime lora and the other niji lora , using the same prompt

[dry wave]

that's said, effect was still tolerable

[sterile pendant]

here's a really good graph i saw a while back that summarizes it pretty well

[dry wave]

I wouldn't use text encoder quantisation, though, if bf16 fits into the hardware

[sterile pendant]

the biggest key though is to always use the quant size that can fully fit your GPU though

[bitter hearth]

i changed the encoder to t5xxl_fp8_e4m3fn.safetensors

[sterile pendant]

if you have some 16gb gpu and 32-64gb sysmem, just use a full fp16 if you want

but just know that you'll get next to no gain in doing so vs a Q8 for a fraction of the size

[bitter hearth]

i have 12gb vram

with 32gb ram

[sterile pendant]

just use q8s then

[bitter hearth]

why not unload after creating text embedding?

[sterile pendant]

you do

[bitter hearth]

im running another test with t5-v1_1-xxl-encoder-Q6_K.gguf

[sterile pendant]

you do, but not everyone can fit full fp16 LLMs in their vram.

that's why fp8 became the norm for t5xxl

it can fit in 8gb vram

[bitter hearth]

ok this is good, i just got this lora working again with q6_k ecnoder

[sterile pendant]

but a q8 will fit in the same amount of vram and will have the quality of the full fp16 model

[bitter hearth]

im pretty sure the workflow i had before had some issue

the previous workflow wasn't optimized for flux

most stuff from older workflows won't work on flux anyway

ahh that explains it

either things need to be ported, or need a different custom node if they are ported, or are waiting for better support for negative predictions (if that ever comes)

but the nodes were taken after usual workflow, they worked but might not have interpreted flux correctly

this is very nice to see loras working

its still very early days with flux, I wouldn't be surprised if we have the sigmas super wrong for example

[sterile pendant]

flux dev and schnell do not use cfg, it's been talked about to death on these channels. you can make a faux cfg workflow, sure, but it comes with drawbacks and will likely require a second pass to cleanup any of the damage done. but just to be clear, it is not an actual negative prompt like when using a regular model that has CFG.

[bitter hearth]

I thought people had actually got to the point where they were getting an actual second negative prediction at each step, or a third prediction for perp-neg

but that it was getting off-manifold issues (CFG burn)

I haven't started using flux yet so I have not verified this yet 🤷

[sterile pendant]

it's a hack

but again, nothing is stopping people from experimenting and doing things like multipass workflows to cleanup the issues

like a second pass at 0.4 denoise would likely fix any of the damage

[bitter hearth]

ModelSamplerTonemapNoiseTest, TonemapNoiseWithRescaleCFG or DynamicThresholdingFull would be the first things to try

the new Skimmed CFG node also

[sterile pendant]

yeah those are all good choices to try

[bitter hearth]

I never make an image without tonemap

[dry wave]

to be fair: cfg is a hack in general

no diffusion model was trained with cfg

[bitter hearth]

yeah its actually a crazy thing that cfg even works

the reason I love tonemap and I think its under-rated is that tonemap lowers CFG-style burn from any source

so its a quick fix for lots of issues

and it can also fix CFG being too low (maybe a bit of a niche issue)

another thing is to switch to scheduled conditioning (only have the conditioning running for as few sigmas as possible) and switch to block-by-block conditioning (only have the conditioning injected into as few blocks as possible, or vary the conditioning by block.)
these both help a lot with SDXL, SD 1,5 etc to lower CFG burn

block-by-block perturbed attention guidance and smoothed energy guidance can also help with this if they get ported over

[low stone]

really great. what's the lora on the right? have the name of it?

[bitter hearth]

https://civitai.com/models/673620/niji-flux

[low stone]

wow those samples are fantastic

this one is nuts from their gallery

[bitter hearth]

im impressed with the number of community built loras in such short time

[low stone]

yeah, the sentiment is that flux is faster to train than even sdxl

[hallow lion]

and the loras are microscopic

I can barely find them after i download them theyre so tiny.

I have to look hard

I was expecting 1GB+ monstrosities. But no they're minuscule.

[low stone]

I'm kind of wondering if those sample images are using additional loras. Not a single image on the gallery or sample images includes the prompt, and I'm not getting out of my stuff what they are. Mine don't look anywhere near as good.

[bitter hearth]

raw sd3

fried sd3

[bitter hearth]

yeah those look amazing, im gonna try to render some with that lora and see if i can get anything close to those texture

[bitter hearth]

here is one with niji lora.. i upscaled it too,

[low stone]

he's getting that, I'm getting this with his lora... something is up...

that's his prompt, he uploaded his workflow, which truly doesn't have any other loras in there.

[bitter hearth]

the lora settings is not applying in that workflow it seems

i just had that problem few mins ago

i had to change the workflow that is optimized for flux

now all the loras are looking nice with that workflow

with clear distinctions in style using same prompts

[low stone]

so i'm basically doing the same thing but the lora isn't doing anything now.

[bitter hearth]

oh i dunno about that sft

im using gguf model

[low stone]

that's the original one that flux published.

that's straight from their huggingface.

[bitter hearth]

yeah but not sure if lora worked with it

[low stone]

every other lora does.

[bitter hearth]

odd

[hallow lion]

Perkele.

[bitter hearth]

this image you showed above is not displayng the niji style right

[low stone]

totally agree.

[bitter hearth]

what in the french did you say

[bitter hearth]

btw if you used gguf that's better optimized for quantization and smaller in file size

[hallow lion]

Haista Vittu.

[low stone]

yeah but i want to run the full version, not quantized.

[low stone]

see, when i put on the cute 3d lora, works immediately

[hallow lion]

Nice. Theres fingers galore but AI enthusiasts are used to that by now.

[bitter hearth]

heh

flux is better

shouldnt have those typical anomalies

[low stone]

I wonder if it needs the word anime... or animated. I added anime and the lora came to life.

maybe it doesn't work on photographic images. have to have some kind of illustrative style first.

[hallow lion]

On the flip side we are slowly enetring the time when AI images are becomign almost 100% perfect.

this lady looks good enough

[bitter hearth]

i got this cute 3d cartoon lora

[hallow lion]

Can;t find much wrogn here at first glance... Maybe nothing.

[low stone]

I am well versed on the cute 3d lora. 🙂

[hallow lion]

Her right hand maybe a bit too twisted.

[bitter hearth]

sure but i mean i didnt have to prompt for the style between that one or niji

[hallow lion]

Yeah its backwards.

[low stone]

yeah, i have no explaination. it's the first lora that didn't just auto work for me.

[bitter hearth]

whats your prompt ?

[low stone]

there's no question it's working now though. here's an example without the niji lora, and with.

same prompt and seed for both

niji_flux, anime, Super muscular dark blue rabbit-man with glowing eyes, wielding a futuristic black and tan submachine gun, stands confidently amidst chaos. He wears a bright red, sleeveless jacket with fluffy white fur trim and black, fingerless gloves. An enormous explosion engulfs the background, with fiery plumes and thick smoke rising high between towering city buildings. Dollar bills swirl around in the fiery, neon-lit air as crowds of people flee in panic, their shadows stretching on the wet, reflective streets. Overhead, the night sky is obscured by smoke as cables and lampposts add to the urban destruction. The city lights flicker against the backdrop of the explosion, highlighting the rabbit-man's immense strength and determination.

[bitter hearth]

the right side image looks better now

[low stone]

totally

it's definitely working now

[bitter hearth]

try adding these quality prompts to see if they accentuate the image tone detailed, cinematic, 4k, textured

4k might not work correctly

[bitter hearth]

cinematic makes some visible difference

[low stone]

yeah that looks great

between niji and cute3d on the same day, it's like christmas

[bitter hearth]

nice

i noticed some of the samplers that worked great in schnell nf4 doesn't seem to work right with gguf models

[low stone]

schnell nf4... that must be really fast. i've used the regular nf4 v2 of dev which works well. didn't know there was a schnell.

[bitter hearth]

https://huggingface.co/duuuuuuuden/flux1-schnell-nf4-v2/tree/main

[low stone]

very cool, thanks

[bitter hearth]

no problem

[low stone]

I have to see what this lora does with cute3d mixed in

[uncut river]

i've made my fractal lora into a FLUX version, check it out if you like (3d) fractals:

https://civitai.com/models/615164/yhnrfractalflux

[low stone]

cute3d lora / cute3dlora + niji

[uncut river]

since it's a weird style, it probably works nicely with other flux loras

[errant dust]

Glad it is working for you now

[bitter hearth]

thanks for sharing that workflow, i edited it a bit 🙂

[errant dust]

And powerful. Andreac's Blue Future LoRA is super cool and has a deep impact on images, despit being, as you said, really small in file size

[low stone]

wtf there's more?!

oh god...

ya know... what's funny is that they posted screenshots of that one JUST before they took it all down when they were doing sdxl loras. I was 1 day too late. now they go and post it for flux.

all these months later.

[pseudo owl]

There are so many flux loras.

[hallow lion]

Things are moving fast.

[errant dust]

https://civitai.com/models/658958/blue-future

[pseudo owl]

Half illustration lora which is pretty nice imo.
https://huggingface.co/davisbro/half_illustration

[hallow lion]

Future VR will run on advanced Flux models. More real than real is our motto.

Flux 9000 because reality is overrated and boring.

[errant dust]

Heh. a 42-y.o quote. I wonder if that was in the original short story

[hallow lion]

looool

that was mnore human than human

no no this is more real than real

its more all encompassing

[errant dust]

You know that although he denied it at first, Clarke later admitted that the name HAL was in fact to insinuate one step ahead of IBM. Each letter being one before

[hallow lion]

yeah

it sure was!

I can't let you run this Lora Dave.

[errant dust]

Funny how every single scifi computer somehow ends up playing chess in the movies, whether 2001, Wargames, or even The Thing

[hallow lion]

lol

Chess is like the smart game so a computer playing it is ecovative

Hey.

Upgrade your Dog now so you cna play chess with him.

[errant dust]

at least it was then the reference

[hallow lion]

Today ti will be more soemthing to do with chatbots and aI

My dog pooped all over the floor and kept poopingin the living room. I thought having a doog will be more fun. But then I upgraded him and now he beats me at chess. All in all a good investment.

[low stone]

release the roombas!

[errant dust]

Dog? Oh I can see someone doesn't know the classic Russian joke. So I will share it.

A man meets up with a friend and sees him concentrating on a game of chess.... with his dog. The friend is taken aback. "It can play chess? That's incredible!"
The friend replies, "It's not so impressive. I beat him two out of three."

[hallow lion]

loooool

[low stone]

did you try that schnell nf4 with the loras? doesn't seem to work for me.

[errant dust]

As a rule Schnell won't anyhow, but regardless, NF4's don't. Not even the Devs

It is the nature of Schnell: the guidance is locked in, and LoRAs require being able to manually change it

[bitter hearth]

lora wont work with any of the nf4 models

[errant dust]

But in any case, NF4 has not been able to work with any of the LoRAs. The explanation was that it was because it is quantized, but GGUF is also quantized and works. Go figure.

[low stone]

https://tenor.com/view/disappointment-gif-21343121

[low stone]

niji lora / niji combined with flat color lora / niji combined with cute3d lora

niji does bring emotion to the faces.

adds some dynamic elements over simple fca or cute3d

[hallow lion]

He is ready to go back to pen and paper.

[bitter hearth]

nice, im gonna try that idea and load up multiple lora too and see what i come up with

[low stone]

that rabbit one, with flat color anime and niji

yeah this is the anime i kept trying to get before.

yeah, killer combo

[bitter hearth]

think this will work?

[low stone]

I've never done it that way. I've put the lora into each other. flux->lora1->lora2

Don't know what that node collector does.

[bitter hearth]

it collects the nodes and outputs them all together

[sage burrow]

I'm catching up now, which one? 🙂 https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main

[bitter hearth]

try any of the recent Q_K model

[low stone]

i do it like this.

[bitter hearth]

ahh i see now, both our method should work i think

[sage burrow]

Which one?

[bitter hearth]

this is what i got with a combo of wlop and mj

[bitter hearth]

i downloaded Q6_K

[sage burrow]

Thank you 🙂

So does GGUF need yet another workflow?

[bitter hearth]

yeah just 2 changes in the node from the usual unet workflow

[low stone]

needs this node to load it

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

[sage burrow]

Thank you 🙂

[craggy crest]

your push mower blades won't work with the riding mower, even though they are both mowers

[bitter hearth]

you can try this workflow

actually try this 2nd file

[sage burrow]

Thank you 🙂

[errant dust]

Super cool diagram, thanks. I do have a question. What is the difference, if any, between the bf16 T5 encoder and the fp16 encoder. Or is it the same? And what about the 32?

[sterile pendant]

Any of them will work, but keep in mind that right now the operations are bottlenecked by some slow python math operations until he can push an update to llama.cpp. like with my 2080 8gb GPU, the difference between Q8 and 4km is 4.1sec/it and 3.6sec/it

So I just stick with Q8

And it won't oom you either

[craggy crest]

fp16 is shorthand for floating point to 16 places. where did you see something called bf16

[errant dust]

https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf

[sterile pendant]

And for the t5 ggufs, I just use q8 as well

[craggy crest]

that's an entirely different kettle of fish

[errant dust]

https://huggingface.co/Kijai/t5-large-encoder-only-bf16

[sterile pendant]

Using both Q8 flux and Q8 t5 will entirely fit in 32gb sysmem and won't hit the page file

[errant dust]

(which does not work BTW)

Ok, but is there any qualitative benefit with the 32-bit model (which is 19GB)

[sterile pendant]

Think it settles around 16gb after all the initial loading

[errant dust]

Is that the driving diff between Dev and Pro for example?

[bitter hearth]

hmm these loras have some downsides, some of these loras are messing up fingers

[errant dust]

I've been using the standard fp16 T5 encoder until now, as I found it made a diff. WHich is why I wonder about the 32-bit I see

[low stone]

I'm finding that there's a higher chance of messed up fingers on the niji one. probably wasn't the most discerning with the source images.

[bitter hearth]

yeah ... the loras have rooms for improvements to maintain the quality of the flux base models

[low stone]

that said, out of 10 images generated, 3 had messed up fingers.

[bitter hearth]

i wonder how loras would impact text

[low stone]

seems to work.

[low stone]

so the question is, one has the gguf.... how do you load that gguf as a clip?

the gguf node is for image models.

nevemrind, another node

[errant dust]

The answer is... it worked without a hitch. the f16 model of GGUF gave an identical result to the standard fp16 one

I tested, I am now DLing the 32-bit model to see if 1) it works on my rig at all, and 2) if there is any change

[low stone]

I'll spare you the download

fp16 / fp32 of t5

[hallow lion]

My favorite is the boring people lora. 😄

What about ya'lls? XD

[low stone]

@errant dust

f16 / fp32 of t5

there's no question the fp32 is beter.

the shoes are now furry toes, the lines on the airplane now make sense...

the wire supports on the wings make more sense on the fp32.

[errant dust]

It is a small example but yes, it is not quite identical. I am already DLing, but whether or not my rig can handle it (and the time to use it is remotely reasonable) remains to be seen

I have found the diff between fp8 and fp16 to be quite pronounced

[low stone]

yeah there's far less of a diff between fp16/fp32 and fp8, so across a ton of generations, probably not worth it.

[errant dust]

In your first set, the diff was barely noticeable

[low stone]

yeah in the first one, the only thing worth mentioning was the text on the "for" was better on the fp32.

more distinct letters

the one with teh cat was a giant monster combo of several long prompts.

so if you're not doing massive prompts, probably won't be of benefit.

[errant dust]

Honestly the diff with for is really not a factor. In photography we call that pixel peeping, The airplane image was more interesting

[low stone]

fp16 / fp32

the strut on the left side is correct in the fp32, wrongly angled in the fp16.

the scarf is a single thing in the fp32, it's split in 2 in the fp16.

if i was doing it professionally, i'd use the fp32, otherwise not.

[coral sable]

ETA on new SD model? SD3 2B remake if I'm not mistaken

[sage burrow]

2 weeks

[errant dust]

The airplane propeller in fp32 is completely wrong

[real terrace]

any big difference?

[low stone]

wrong..... or better!

🙂

[errant dust]

I'm pretty sure it is just wrong. 3 blades and an obvious fourth is missing

[low stone]

because it's spinning so fast.

can't see it.

[errant dust]

right, just the other three

Houston, we have a problem. Only one of the four airplane blades is moving

[sage burrow]

Did it actually work for you?
Are you just trolling people by posting models that don't actually work?

[desert garnet]

2 weeks

[sterile pendant]

Like I said earlier, you'll mostly only see minor differences in same seed outputs. But it's pointless to nitpick on a per seed basis. Put against a large enough sample size with random seeds and you'll see a 50:50 blind a/b preference rate in images, between even Q5km and fp16, +/- some small margin of error.

The things you should compare between quants and fp8/16 are mostly things like did they follow the prompt correctly and are there serious collapses in quality and complexity

And the answer is no, so long as you stay above Q4_k_m

see how much difference there is just by adding a , , to the prompt with the same seed? this is essentially an analog to what you're seeing between different size t5 models and quants. just tiny little microvariations in the embeddings that get created, but they usually converge on roughly the same data

(quick BS example using sdxl)

[sacred jewel]

Yup, but with the Blue LoRA

[craggy crest]

no. pro is the full model. dev has had a distilling process run on it

[sacred jewel]

Who was first? Star Trek 4D chess or 2001: ? hmmm

[sage burrow]

Since the above posted jsons for GGUF posted here were useless 💩 , here's one that actually works. In case anyone needs one. It's not any faster than Flux Dev though.

[fleet meteor]

interesting style 👀

[errant dust]

Star Trek was not a computer, it was Spock vs Kirk

[sacred jewel]

Thanks, it is with the wonderful Blue Future LoRA

[sacred jewel]

https://tenor.com/view/the-goon-dodgeball-movie-ben-stiller-white-goodman-gif-14923367

[bitter hearth]

Boat

[sacred jewel]

"Honey, I am hoooommmeeee"

Ballz

[bitter hearth]

good ball

[hallow lion]

You can never go wrong with ballz.

[hallow lion]

🤣

Emad (yes it's Emad I dont care) will redeem himself with a model that can do 101% consistency. And then finally AI will become a thing and not just a gimmick. Hollywood execs will then have to change udnerpants.

Right now their underpants are safe.

[craggy crest]

he didnt' create any of the models. he's not a programmer or a researcher - he's the vision and driving force

[craggy crest]

you really need to start reading twitter

[hallow lion]

Make a model that forces the hollywood executive udnerpants into the washing machine.

Meanwhile it's flux flex time.

[craggy crest]

hollywood is already way ahead of what the general public can get their hands on. and they likely have contracts with openAI for sora

[hallow lion]

Aki Kaurismaki said Hollywood is dead.

[craggy crest]

well aki is wrong

[hallow lion]

I say it will fully die when AI gets off its ass and makes consitent models.

then it will die 100%.

[craggy crest]

it won't. but what WILL happen is that the creatives in hollywood won't be stuck taking pennies from greedy studios

[craggy crest]

just go use Scenario if you need consistant. they have that licked

[hallow lion]

Until we don;t have 100% consistent models characters, backgroudn and assets AI imageyr is just a one off gimmick for youtube thumbnails and maybe stock footage.

no one take sit seriously

Wanna kick ass and change the industry of entertainment forever: deliver consistency.

[craggy crest]

you REALLY gotta stop living in fantasy - i hate to tell you this but movies are already being made. with highly consistant characters. you've got you rhead in the sand, and assuming a lot of incorrect stuff

[hallow lion]

One off images are pretty much there by now. Video is catching up. Just add consistency and then Hollywood is dead.

[craggy crest]

here, go watch this https://x.com/david_vipernz/status/1826163511011209392

David J Morris (@david_vipernz) on X

If you want a new method for getting AMAZING CONTINUITY with ZERO IMAGE PROMPTS an no extends with #LumaDreamMachine v1.5 Watch (and listen) First half is the result - 2nd half How I acheived it. Only the audio was done outside @LumaLabsAI #GenAI #Emoji #MicroMovie #PromptShare

Twitter · Aug 21, 2024, 7:44 AM

and hollywood is not dead, will not BE dying, and is already ahead of this

[hallow lion]

We need consistency down to the last mole on the face type of consistency.

I need to arrange a scene like a movie director.

Put the furniture where I want to. And light it exactly the way I want to.

And so on.

[craggy crest]

dude, no we don't. you think that stunt dubles are 'consistant'? you think people are actually that 'consistant' when you film the from various angels? they're not. the mind extroplates

[hallow lion]

When AI can do that then we have made it and the world will change.

[craggy crest]

the AI CAN DO THAT, you just can't get it to do that yourself

you have to start reading twitter and see what's actualy available. not just reading this discord

[hallow lion]

In comfyui I can generate each asset and character separately and using SAM cut them out and put them in a different background and then generate the background also and then use IC Light to relight it. So we are almost there.

But it;s not quite there.

And video is not there yet, at leats not local video.

I absolutely don;t care for anything non local. It;s pure trash if you can;t run it locally.

U use up your credits in 5 seconds and the results are never what you want and it's also heavily censored most likely.

[mortal mesa]

gonna need one of those hollywood budgets

[hallow lion]

So Kling just sit down.

Go home.

We need local.

[hallow lion]

Absolutely not.

The same way music production is now a bedroom affair.

film production will be too.

give it less than 3 years.

and I mean blockbuster michael bay stuff

not 3 character moronic drama bullshit.

That takes place in a bedroom and 2 toilets due to budgetary constraints.

[craggy crest]

you're not getting local

[hallow lion]

There are madmen out there preparing their workflows right now

[craggy crest]

you're not getting local - you can't RUN the thing local

[craggy crest]

by dec 2025 - one prompt, full movie, streamed live

[hallow lion]

AI enthusiasts always say "wow this is amazing but it's not quite there yet. But almost." With Flux was the first time I felt that as far as one off image generation this is probably enough. What's left is consistency and video.

[craggy crest]

and you aren't running that local

[hallow lion]

And were done.

All this stuff will have matured.

[craggy crest]

you don't have the hardware, you can't afford the hardware, and you can't afford the elctricity to run the hardware, for that

[bitter hearth]

So little balls

So much text

[hallow lion]

Ballllz: the movie!

[errant dust]

I'm almost certain that there is no greater source of education than Twitter

[craggy crest]

twitter is where the actual data scientsts are, and where they post papers and information. so ...

that's what they use for communication

[hallow lion]

The Hollywood executive underpants are safe for now but they are correct to be vaguely unnerved.

[craggy crest]

why are you so obsessed with their undergarments

[errant dust]

Hollywood uses AI and tech at every level, from writing scripts to post production.

They aren't Amish

[desert garnet]

yea i can tell the new borderlands movie was writen by a chatbot

madame web too

i could list all them here in case you want

[errant dust]

Might be, but I don't think you need to blame a bot for that crap. People are quite capable of writing garbage without any help

[desert garnet]

maybe there was bots writing movies since the 80s it all makes sense now

[errant dust]

https://www.hollywoodreporter.com/movies/movie-news/hollywood-ai-artificial-intelligence-cannes-1235900202/

[craggy crest]

naw, just hollywood execs telling the writers what they had to write

[desert garnet]

question would be with borderlands movie did they tell them to make a good movie or a bad movie

[hallow lion]

looool

All Hollywood movies written by AI. Allegedly

Cue the "AIwould have done a better job" comments. Which would be the correct comments! 🙂

[craggy crest]

all a hollywood exec cares about is that the movie 'makes money'

[desert garnet]

cant wait for the next big summer movie, Huge Mama starting Dwayne the rock Johnson wearing a fat suit and rescuing her kidnapped dog while making fart jokes,also Kevin Hart will be his partner in the movie

[mortal mesa]

part 10

[desert garnet]

😮

[noble coyote]

AuraFlow0.3

Flux.Dev and Blue Future LoRA

[noble coyote]

Flux.Dev and Blue Future LoRA

[alpine summit]

Sd3

[bitter hearth]

that blue and orange Flux lora looks really nice TBH

I find default flux style to be too high contrast and saturation for my taste

but the blue and orange lora is enough to tone it down a bit

would be cool to get a soft pastel-colour lora for flux that looks a bit like SDXL with very low CFG

gonna wait a month to see if someone else trains something that is a bit like this LOL otherwise I might do it

[sacred jewel]

Nice

[sage burrow]

SD3 would be amazing for Gumby images! 😉

[bitter hearth]

Sd3 has best balls

[sage burrow]

My lora creation is going very well 😄

[bitter hearth]

No

Your lora?

It can't be good

[sage burrow]

I'm working on #4 now 😉

[low stone]

Sweep the leg!

[sage burrow]

Btw, Flux is good at bows and arrows 🙂

[noble coyote]

Flux.Dev and rbcharc0al and Blue Future LoRAs

[low stone]

You would be correct. 🙂

@noble coyote this reminds me of the fairly large spider like this that sets up outside my office window the last few nights as soon as I turn on the light. It knows bugs are coming for the light and he's hungry.

[sage burrow]

That's seriously awesome 🙂

[sacred jewel]

They seem happy

[bitter hearth]

when i tried it, it was very bad lmao

[hallow lion]

Prepare your butts for the "Matrix Reorganized" - summer of 2025. By chat GPT Jr.

Directed by Grok 4.

[grand lodge]

Where do you produce these images?

Can you share any link?

[bitter hearth]

Actually it's fairly decent at bows

Breaks the string most of the time but works lmao

[low stone]

This was the prompt: fca_style, niji_flux, anime, detailed, cinematic, textured, extreme camera angle, gopro hero shot Sure, let me paint the scene for you:

In a vibrant, dynamic illustration, you see an orange cat with a confident, relaxed posture standing at the archery range. This isn't just any cat – it’s cool, badass, and ready to win. The cat is wearing stylish, dark sunglasses that reflect the targets ahead.

Its fur is a bright, fiery orange, almost glowing, giving it an energetic and daring vibe. The cat is drawing back a bow, aiming with precision. You can see the tension in its muscles, the sharp focus in its body language.

The bow itself is sleek, high-tech, with glowing accents hinting at some advanced technology. The arrow flies through the air with a trail of swirling energy behind it, adding a magical, powerful element to the scene.

Above the cat in large, diagonal text, you see "Becky93" written in bold, dynamic font. The letters are surrounded by the same swirling energy that trails behind the arrow, making the name look powerful and important.

The background shows a lively archery competition, with targets lined up and competitors in the distance. There’s a sense of motion and action, capturing the intensity of the moment.

The overall image is fierce, colorful, and full of energy - encapsulating the cool, badass spirit of the orange cat named Becky93.

[bitter hearth]

When I tried it, it was with more than one person

So it didn't know what the hell to do x3

[bitter hearth]

Too much text

[low stone]

hah flux can take it.

I'm so glad we're not still talking about 77 token windows.

[bitter hearth]

I didn't leave school to come write essays to AIs

Wait

Flux is ahead of the game

Extra arm holding the next arrow

[low stone]

just need the ai to write essays to the ai.

[bitter hearth]

that's an upgrade yeah

Flux knows what makes a good archer- 3 arms

[low stone]

is that a quantized version?

I never get that.

[bitter hearth]

not my image

Flux pro
whatever that means

Not local

oh dear this was flux pro?

[low stone]

wow, flux pro made 3 arms... well ok then.

[bitter hearth]

This is the full thing

[low stone]

oh ok, well they cheap out on the background characters.

[bitter hearth]

elf on the right stole the bow of the elf on the left

And arm

[bitter hearth]

"kids drawing" lmao

That kid is good

[bitter hearth]

[bitter hearth]

LOL the giant ice bird is funny

[bitter hearth]

I want a big silly bird

I have 2 smol ones

[coral sable]

source?

[sage burrow]

So what is this crap? "[Crystools ERROR] Could not get GPU temperature. Turning off this feature. The operating system has blocked the request." Windows forgetting who's computer this is again?

[noble coyote]

Uninstall nVidia driver, and try an earlier version

https://www.youtube.com/watch?v=QhvKvtdtYFY

[sage burrow]

It still seems to work even after that message. Do you think it would be better as is, or with an earlier version?

[noble coyote]

Leave as is until it definitely stops working

Flux.Dev and rbcharc0al and Blue Future LoRAs

[sage burrow]

Anyone know how to train Flux checkpoints? I want to do some experimentation. I don't even need to create a giant one, since it will just be merged with the main Flux Dev one later.

[bitter hearth]

will you train one for flux with 15 images

[sage burrow]

If I have to do it locally, then that would be the plan lol. If with a rented GPU, perhaps an entire 100 images ROFL

[bitter hearth]

He is ADMIRING ART

[sage burrow]

Which lora is this? I need that style for my new pixiv account...
KIDDING 😄

[noble coyote]

Flux.Dev

[bitter hearth]

not balls this is blasphemy

[dusky thistle]

here's a ball lol

another

[errant dust]

I have uncovered something quite strange and must be shared for those who use GGUG models. I tested this multiple times to be sure. The Q6_K model does not play nice with LoRAs. Q5_k is fine and enjoys all the benefits, but not Q6_K. I cannot explain why, but it is not in doubt.

I have not tested all the GGUF models, so this comment applies only to them for now

Here is Q5_K and Q6_K with no LoRA (workflow included). Both are twins for all practical purposes.

I then ran the exact same Q5_K and Q_K but with the Art Nouveau LoRA (I tested other LoRAs BTW):

Q6_K has some modification, but it has nothing to do with the Art Nouveau LoRA.

in both cases the strength was 1.2. I do get a reaction from Q6_K if I increase the strength to 1.4 though:

[bitter hearth]

you have some nasty talking tone. and that workflow works fine as i've personally tested it. i'll take a note not to share shit with you

[desert garnet]

emad balls

[noble coyote]

When we get 32K monitors, t5xxl-fp32 will come into its own 🙂

[bitter hearth]

@errant dust i found something surprising, the workflow i was using which is Flux1 unet gguf lora upscale workflow took a bit longer to render image than this simplified workflow flux1. gguf upscale simplified workflow and it's not really complex flux specific workflow .. render time with upscale is significantly faster.

here is another example of the simplified workflow at 117 seconds
beautiful landscape, dreamy cottage with chimney, wooden fence, water stream, brick roads, lush trees, evening sky, setting sun. anime style.

tried the ART lora from https://huggingface.co/XLabs-AI/flux-lora-collection in this image

[bitter hearth]

i love the power lora loader from rgthree 🙂

[bitter hearth]

you don't feel the need for separate weights for clip strength?

[bitter hearth]

i can set the lora weight, but what do you mean clip strength?

there's two strengths for loras in civit
the lora strength and the clip strength
the lora strength changes how much the diffusion model weights are changed by the lora
and the clip strength changes how much the text encoder weights are changed by the lora

hmm

not sure about clip strength that it is required unless its specific to flux

for some loras it probably doesn't do anything

but the reason im using this power lora node is for ease of use with selecting loras and setting their weights

its nicer than chaining 11 load lora nodes yeah

with disney lora
a miniature village of cute cottages and colorful mushrooms with the evening sun casting an orange glow over the hills.

disney lora sounds cool

[bitter hearth]

yep dont have to reselect

I love old school disney like snow white

or bambi

funnily enough in SDXL, using R2D2 loras is about the same as just using base models

cos every base model has lots of R2D2 in the training data

one of the R2D2 loras don't even make more R2D2s it makes more C3POs 🤔

this happens a lot because people train their lora on photos of both R2D2 and C3PO together

the model incorrectly learns that R2D2 means both R2D2 and C3PO, so the lora is worse than the base model LOL

speaking or R2D2 ...

used a combination of lora on that one...

[errant dust]

Thing is, I don't use any of the internal upscalers as I use Topaz which is both the best and fastest, albeit external

[bitter hearth]

cool, i hear florence2 is pretty good too

i havent used any of those yet but saw some videos

about the workflow and apart from upscaling, the ksampler with simplified nodes processes data faster

[errant dust]

I'll be honest, I don't see how any nodes would speed up my workflow rendering, which amounts to: Load Clip, process Text, Load model (and LoRA if applicable) render image.

Anything else should take a tenth of a second at most

[bitter hearth]

i dunno how that's faster but it is, also the workflow from https://comfyanonymous.github.io/ComfyUI_examples/flux/ has complex node workarounds, which is kinda confusing too

[errant dust]

I'm familiar with the base workflow

The more complicated and advanced workflows you may find on Civit, are for people who like to fiddle around with their rendering to no end or want other features. Mine is pretty bare bones

[bitter hearth]

i like the idea for coming up with more creative ways of doing things, but between the two workflows one is much easier to follow

and that comfyui example workflow adds the additional nodes for the flux model exclusively but you can do w/o them

[coral sable]

Emad ball told you that?

[desert garnet]

yes the round ones

[coral sable]

Interesting

I'd like more sources

[bitter hearth]

i should avoid upscaling every image i render

[bitter hearth]

simplest workflow to upscale an image you wanna keep

[craggy crest]

@hallow lion you were talking about consistancy https://new.reddit.com/r/StableDiffusion/comments/1f0b45f/flux_is_a_gamechanger_for_character_wardrobe/

[sage burrow]

Now I'm wondering how Flux upscaling my SD1.5 models would look!

[bitter hearth]

how do i get to do image2image with flux in comfy?

ive dragged in an image but it says unable to find workflow

[hollow swift]

lora doest seem to work with gguf 4sk, any solution for this?

rtx 3050 laptop

[sage burrow]

MJ is offering 25 free generations this weekend, in case anyone wants to... compare? Create ref images, or?

[sacred jewel]

remember img2img is not model related.

[sage burrow]

I tried GGUF, but it wasn't that much faster than regular Dev for me , so unfortunatley I have no idea. Hopefully somoene will show up who uses it.

[hollow swift]

do you use the default fp16 dev model

[sage burrow]

Not if I want to see the images in a temely manner lol. I use the 8 version 😄

[obsidian junco]

I was gonna ask if anyone could explain the difference to me between upscaling with your model/ui that created the image vs using something like Waifu2x that can upscale any picture

I'm guessing the on-model upscale actually kind of re-runs it and increases detail in a way W2X can't

[sacred jewel]

While I am not sure technically, I always assumed that upscaling models are trained to just do that, upscaling.

Where a Diffusion model will always diffuse/inference...denoise...

[bitter hearth]

i reworked my upscaler workflow, using ultimate SD upscaler which basically enlarges the images 4x times w/o any quality loss cause i dont wanna upscale when im doing txt2img, that way i can only upscale the ones i want to keep

[bitter hearth]

saves me great deal of time to make images

[obsidian junco]

idk comfy you have to explain this in peasant terms to me

[bitter hearth]

which webui do you use?

[obsidian junco]

all I know is, W2X can upscale in seconds, but I tried upscaling in Fooocus once and it was taking so long I just stopped it

all I got is Fooocus

[bitter hearth]

ok so fooocus is probably like auto 1111

there is a very simple approach to upscaling individual images in auto 1111

but i've stopped using all those webui, only use comyui lately

[bitter hearth]

Some fever dream stuff

[errant dust]

As a follow-up, I tested on plain Q8_0 and no amount of strength from the LoRA produced any effect at all. I was quite surprised. So for now the conclusion is Q5_K works the best with LoRAs at even strength 1.2, which is fairly normal, and Q6_K will work only with a higher LoRA strength value.

[errant dust]

This first is Q8 with the LoRA strength at 1.4. It is literally as if there were none at all. By comparison, next is Q5_K with the same setting and LoRA at 1.2 strength.

The LoRA being Art Nouveau, the second clearly got the LoRA magic

[errant dust]

My how many arms you have grandma

[short thicket]

I'm running a Q4 The photo above is .5 strength. This is at 0.

[errant dust]

.5 strength of what?

[short thicket]

a lora I'm working on. Currently on epoch 3.

[errant dust]

Q4_K or one of the earlier ones?

[short thicket]

FusionDS_v0_Q4

[errant dust]

Never heard of it. Link?

I am referring to the GGUF builds

[short thicket]

https://civitai.com/models/630820?modelVersionId=736086

This is gguf

[sage burrow]

MJ 6.1
I miss flux already! 😄

[errant dust]

Gotcha, so one of the early ones

[short thicket]

Might work with the FusionDS Q8. Might be worth a shot.

[errant dust]

https://huggingface.co/city96/FLUX.1-dev-gguf

Is the original repository

[errant dust]

I did not know anyone had tried to make a GGUF of the Dev+Schnell models. Interesting. I am partial to the normal Dev versions to be honest

[sacred jewel]

Cascade?

[dusky thistle]

yep

using one of my loras for these

[sacred jewel]

fantastic

[mystic thorn]

a dog

[short thicket]

I'm currently waiting for epoch 4 of a style lora I'm working on. Here's some progress from epoch 3 from strength 0 to 1.5. I'm pretty happy with the progress made so far.

[low stone]

is it robot focused? what's the "style" ?

[short thicket]

Kind of like an eldritchpunk

Some samples from the training dataset.

[low stone]

ok i see. yeah searching for that on the internet didn't show anything like what you have, I think people just have accounts with that name for random stuff

[low stone]

hey what's the prompt for that? woudl like to try that for a command I'm making

[short thicket]

It adds some nice detail to photographs at lower strengths though. This is 0 vs .75.

[hallow lion]

Today is Giger day on the SD3, I mean flux?, channel.

[sacred jewel]

LOLOL

[low stone]

definitely. been playing with lots of loras the last few days. flux has gone from very good to great real fast

[short thicket]

I know! I feel like it's caught up to SDXL in terms of style. I need to find out how to train a concept though. I tried to do one for a gorilla press wrestling move but it failed. It gave me great wrestlers, but none showed the wrestling move.

Kept getting things like this even though I trained a token to it.

granted the dataset was only 12 images.

[low stone]

I've read about people complain about trying to train porn stuff. that if the model doesn't have a certain pose or interaction, that they've had issues getting it to latch on. That said, civit has had no trouble showing me seemingly successful loras. 🙂

[short thicket]

Yeah. I can't complain. That was only my first try. I will get it. 🙂

Epoch 4 is looking promising for the style test. I'm halfway there.

[sage burrow]

What's SD3?

[sacred jewel]

Very nice!!

[hallow lion]

In spite of the legs being all over the place this is cool.

[sacred jewel]

nightly ballz

[hallow lion]

You can never go wrong with ballz.

[sacred jewel]

https://tenor.com/view/amen-anchorman-will-ferrell-gif-10865461

[sacred jewel]

Looks like ipndm and ipndm_v samplers work well with flux too.

[hallow lion]

This tech will change the world.

It's like a wild horse.

Or wild... small horse...

for now!

[craggy crest]

did you watch the video i pinged ou about earlier?

[hallow lion]

But ponies grow up.