>>101978569
That's not Flux tho. It's realphotogasm. I can't use Flux properly until my new 3090 arrives :3

debo's bussy

ive done some AB and cant see a difference on outputs between this smol clip_l linked by comfy
https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main
and the extra gigabyte and change on this one
https://huggingface.co/openai/clip-vit-large-patch14/tree/main

File: 2024-08-19_00266_.png (1.19 MB, 768x1024)

1.19 MB PNG

>>101978563
thank you baker

>>101978608
the one on openai's repo is FP32 AND includes the image encoder
the one in the comfy repo is FP16 and is just the text encoder

I wish I knew what people were doing with their flux lora captions. It seems weird to feed it nlp+tags

do you guys remember the good ol days of berrymix?

>>101978646
yeah

Blessed thread of frenship

Why the upscaled image is slightly different to the image generated before?

>>101978659
shit was fun! it had so many variations, and trying to find the recipe, the lost rentrys... what a time

>>101978669
do you know how this technology works

anons whats better? NF4 v2 or GGUF-Q4
8g vram btw

In SD, one should use the same LoRA and the same prompts when upscaling an image that they were used when generating it? I think I want to lower the LoRA strength because some fuckery happens.

flux+controlnet is a disaster
is anyone having any luck with it?

File: 1707875825513742.png (1015 KB, 1024x1024)

1015 KB PNG

>go to civitai
>sort by newest and flux only loras
>always something new to try
remember when people said loras and finetunes werent possible?

File: Flux-20240818_183810-up-8(...).png (2.83 MB, 1824x1248)

2.83 MB PNG

>>101978702
you have complete freedom when upscaling an image. adjust the prompt. add things, remove things, lower the loras strength. add new loras.
>>101978675
he does not lol

File: 356722258-0c194aaf-e959-4(...).png (2.01 MB, 3835x1744)

2.01 MB PNG

what causes this level of autism?

>>101978762
coomerism

File: 2024-08-19_00269_.jpg (685 KB, 1536x2048)

685 KB JPG

>>101978702
in theory yes, lora+model for both steps, if you wanna experiment with it using lower lower lora strength for the upscale either use ComfyUI or if you use a A1111like webui use the img2img step instead of using hiresfix (its basically the same anyway

>>101978700
GGUF-Q4 is hitting the 12 GB mark too close

>>101978762
damn, thats impressive

Is there a "negative" IPAdapter extension for forge/a1111 somewhere? I know Comfy has one.

File: 92635494050.png (1.87 MB, 896x1152)

1.87 MB PNG

look at me hands guys loooook!

>>101978765
I hate both comfy and its users
>>101978779
not really, no

File: image.jpg (83 KB, 1536x1024)

83 KB JPG

>>101978762
people with this level of autism will revolutionize the film industry if they can stifle their goon instincts long enough to create long form content

>>101978762
I seen worse, but all that work for an ugly blonde bitch that is impressive

>>101978774
so I just stick to NF4 then ig

>>101978786
to me as a beginner it is.

>>101978747
yes

>>101978762
I consider this a medium/standard sized workflow. but boy what a slopfest

>workflow becomes even more perfect
>gives me pure coomer kino after kino
I will NOT go again. You cannot make me.

>>101978789
press x to doubt

>>101978801
you're a retard

>>101978805
you're a retard as well

>>101978813
rude

File: 00035-199235134.png (3.12 MB, 1280x1920)

3.12 MB PNG

>>101978813
Hurr

>>101978774
Haven't tested NF4 so can't say for sure, but it should be your best bet from what I read.

>>101978817
eat shit get cancer and die :3

>>101978820
*coomer retard

File: aseet.jpg (20 KB, 542x375)

20 KB JPG

>>101978810

>>101978826
Q4_0 is more consistent with higher quants

>>101978800
woops >>101978826

>>101978831
Sorry. It feels too personal. I'm not kidding.

>>101978826
thanks that what I thought so too considering only 8g.. the gguf stuff is too much for me sadly

File: Flux-20240815_143651-re-5(...).png (2.18 MB, 1344x1728)

2.18 MB PNG

>>101978762
love how he ruins his coomflow with that shitty upscaler lulz
>>101978831
YOU TRIED

>>101978855
You should consider upgrading your GPU at this point, and make sure it's not AMD.

>>101978876
these workflows are the equivalent of the cat-feeding machine in back to the future

>>101978876
>anonymous delivers
>newfagversion: anonymous gatekeeps
newfaggotry, not even once. I bet your "coomer kino" is hella mid tho
Bonus pro-tip: you got psyopsed by jews to be "gate keeper", because that way is you and your fellow discoord troons, aka, not a trheat

>>101978947
You fucking schizo. I don't want to share my fetish images because they're very personal, that's all there is to it.
>bet your "coomer kino" is hella mid
Possibly.

>>101978762
>all of this just to get some shitty result
What the point?

>>101978894
Einstein is a dog.

>>101978970
>I don't want to share my fetish images because they're very personal, that's all there is to it.
I can understand that, but you can just "switch to netural /g/ approved fetish" and call it a day

>>101978980
exactly
thank you for existing, fellow human

>>101978991
oh right, it was a dog-feeder, haven't seen it in decades
my point still stands

>>101979009
You need to rewatch BttF at least once a year. It's a rule.

>>101978970
ignore the thing, report it.

File: image.jpg (74 KB, 1536x1024)

74 KB JPG

the nsfw filter doesn't like these pink smooth booba

>>101978970
at least share your fetish with us then

>>101979019
fuck u

is anyone using flux controlnet in forge?

>>101979019
I don't have time for this, anon

>>101979033
is this even flux, you coomer imbecile?

>>101979037
>I don't have time
Yes you do. You can go a few hours without edging.

File: file.png (69 KB, 1876x366)

69 KB PNG

>>101978774
Huh?
>>101978700
Q4_K_S works just fine for me

>>101979065
I'm too busy trying to make basic things you can in a1111 on trannyui

>>101979053
it's Jīn qiú

>>101978876
What's the most decent all-round upscaler nowadays? I haven't tried anything in over a year since I've found 4xNomos8k

File: file.png (474 KB, 700x521)

474 KB PNG

>>101979053

>>101979082
kys unironically

>>101979102
how does one do that unironically?

>>101979071
NTA Speeds? Is it faster than Q4_0? Because I can fit that in 8GB, but with the GPU full, everything crawls to a halt regardless.

>>101978762
This is why I don't bother with flux. I can make 2x better looking shit with half the effort.

I think there is someone ITT that can't gen and is sour grapes about it...

File: 00040-3219050246.png (3.32 MB, 1280x1920)

3.32 MB PNG

>>101979053
No it's not flux

>>101979148
Indeed

File: 2024-08-19_00277_.jpg (574 KB, 1536x2048)

574 KB JPG

gotta say lora mixing works great on FLUX, I used just .75 of bekisnki lora here for the arcane symbols >>101979119 .. adding micro bikini lora at 1.15 makes same prompt do this

File: 1700791610484035.png (1.02 MB, 1024x1024)

1.02 MB PNG

squidward no

>>101979143
you're half-retarded

>>101979148
that's me, but that's not the problem

>>101979163
why are you shitting up this thread with this trash? isn't that the whole point of having two threads?

penis thumb

File: up.png (74 KB, 679x857)

74 KB PNG

>>101979090
its is tricky. 4xnomos8k isnt bad. siax is nice for stuff like watercolor style or drawn art in general, lots of detail, but it can turn film grain into a black pattern and brick your gens. remacri also ok but brightens the image. I usually use either siax or the 4x NMKD superscale for a slightly softer look, and ultramix balanced if I want clarity and artifact removal. the DAT ones, I dunno. not worth it usually.
tl;dr ultramix balanced

File: 2024-08-19_00282_.png (1.19 MB, 768x1024)

1.19 MB PNG

>>101979239
kek

File: file.png (79 KB, 2037x546)

79 KB PNG

>>101979122
Haven't noticed any speed changes compared to Q4_0
CFG6 with Tonemap mumbo jumbo and two loras
vs
Same but no loars
vs
just pure cfg 1 and nothing else

File: image.jpg (79 KB, 1536x1024)

79 KB JPG

can someone quickly photoshop away the cyber booger i really like this gen :(((

File: 1724094697273636_cleanup.png (872 KB, 1536x1024)

872 KB PNG

>>101979271

>>101979282
thank you wtf how did you do that and post the result in under a minute

>>101979271
>can't even do simple retouching
this is why everything about these threads sucks ass

>>101979298
he's not retarded

>>101979090
here did some testing few days ago, look
https://imgsli.com/Mjg5MTY3
>>101979239
BRO

File: file.png (173 KB, 363x344)

173 KB PNG

>>101979239
This makes me feel weird.
>>101979303
hahaha you're so transparent dude. poor thing

>>101979326
>hahaha you're so transparent dude. poor thing
what's that supposed to mean?

>>101979326
>This makes me feel weird.
because you're a homosexual

File: image.jpg (84 KB, 1536x1024)

84 KB JPG

im on linux anon i cant even draw a circle in gimp
i also wouldnt be so into genning if i knew anything about digital art or photoshoop

>>101979298
content aware fill in photoshop

File: ComfyUI_00491_.png (1.34 MB, 1024x1024)

1.34 MB PNG

>>101979326

File: 00034-2928813351.png (3.43 MB, 2304x1536)

3.43 MB PNG

>>101979303
>>101979314
Not helpful, idiots

>>101979298
IOPaint
https://www.iopaint.com/

>>101979264
What are the differences between KS and 0?

>>101979353
Nah, I checked.
>>101979368
Wholesome.

>>101979361
this is EXACTLY what I mean
and it's a really big problem

>>101979369
>retarded coomer
checks out

File: file.png (731 KB, 2403x1539)

731 KB PNG

>>101979375
Less ppl change = closer to the original model

>>101979242
that's not your usual resolution, nice image though

>>101979254
ultramix and ultrasharp ruin texture detail. it's a fucking mess and anyone who has moved past SD1.5 should not be using it. stop falling for sharp edges while losing actual details.

>>101979369
based Coomer

>>101979361
educate yourself

>>101979413
wat

>>101979429
they're not going to

>>101979375
If you can't fit the model + the image being generated in your VRAM, you're not going to notice a difference in speed. They're all going to be slow as fuck, so you might as well go with Q8_0. That said, anything below Q4 in my experience produces lots of artifacts and is unusable (perhaps they need more steps).
The different quants only matter if choosing a smaller one over a larger one will allow you to fit everything in VRAM. Then you will notice a massive speed improvement.
This applies to SD gens too. If you make the image too large, at some point performance drops significantly because it's genning partially in RAM.

>>101979441
Since the smallest usable (Q4) quant is basically 8GB, there's no way to gen without offloading to RAM with an 8GB GPU.

>>101979410
but the more ppl change the more they stay the same so i have no idea what you're talking about

>>101979361
>linux users can't even using a basic graphics program

HAHAHAHHAHAHA

I THOUGHT YOU WERE THE SMART ONES????

WINCHADS WIN AGAIN

File: image.jpg (84 KB, 1536x1024)

84 KB JPG

>>101979369
>Not helpful, idiots
it was the same person, look at the post times
>IOPaint
>https://www.iopaint.com/
neat

>>101979429
i need to learn video editing skills and after effects more than image editing skills. if you have a good youtube series for that then link it

slightly offtopic but it's absolutely shameful how gimp hasn't improved their ui in like 20 fucking years. it felt and looked like shit when it came out and it still feels and looks like shit now.

>>101979502
>i need to learn video editing skills and after effects more than image editing skills. if you have a good youtube series for that then link it
LOL

File: FLUX-20240819_212537-fix_(...).png (1.29 MB, 1024x1024)

1.29 MB PNG

>>101979414
completely different beasts, those two. but hey.
and yes ultrasharp-o should be avoided.

>be GIMP
>be not have non-destructive editing
>be laughing stock

>q2_k
GGMLQuantizationType.F32 471
GGMLQuantizationType.Q2_K 304
GGMLQuantizationType.F16 5

>q8_0
GGMLQuantizationType.F16 476
GGMLQuantizationType.Q8_0 304

What do these numbers mean? Why is Q2 using f32 and f16? And would it be possible to tell comfy to use f8 for the q8 quant?

File: 997V4PCbP-CBq2N-Eb_6F_def(...).png (540 KB, 1024x768)

540 KB PNG

https://civitai.com/models/662230/jessica-cloudlet
/weg/bros we fuckin WON

>>101979514
>gimp
imagine being too stupid to pirate photoshop

>>101979525
the Q quants have a mix of precisions
>>101979528
horrible goblin

>>101979528
Post it

File: 1716884838769835.png (923 KB, 1024x1024)

923 KB PNG

here we go, better society miku:

File: 1687708147643745.png (114 KB, 1536x864)

114 KB PNG

>>101979528
>>101979553

File: 1724096015611214.png (50 KB, 949x363)

50 KB PNG

>a porn actress lora
>mfw

File: file.png (175 KB, 421x319)

175 KB PNG

>>101979581
get this ugly mutant out of my sight

>>101979441
This is not entirely true. I'm testing a bunch of them, and the smaller the quant, the faster the gen. All on 8GB. Q8_0 was 40% slower than Q4_K_S, while not looking significantly better (it did look different, so precision is lost).

>>101979624
Are you VRAM limited? I noticed no difference in speed between a smaller quant and the base model on 24gb vram

>>101979617
https://www.youtube.com/watch?v=1BDhOxX6r2A

>>101979369
IOpaint is really neat, remove backgrounds in gens, remove artifacts, and also inpaint. Saves a lot of time rather than just using photoshop when it works.

File: 00032-2165958027_cleanup.png (3.13 MB, 1280x1920)

3.13 MB PNG

>>101979664
Yes, it's my favorite tool

>>101979651
Yes, this is a matter of VRAM limits. I assumed it didn't matter how large the limit was as long as the model didn't fit, but that's not correct.
>>101979651
>the base model on 24gb vram
I'm waiting for a 3090 with 24GB. Are you saying I can expect to run fp16 at full speed? What timings do you get?

>>101979664
photoshop has the same functions integrated now, numbnuts

>>101979679
Hopefully they eventually update it to use Flux, it will be smarter and better (though slower, so give option to switch).

>>101979688
IIRC, the base model is 22 GB. Wouldn't you want to run a quant so that you have more spare VRAM for the image, upscaling and other stuff?

>>101979651
>I noticed no difference in speed between a smaller quant and the base model on 24gb vram
That makes sense since it works by doing a dequant pass of a tensor and then that result goes through the matrix multiplications and so on as normal. A slightly reduced amount to read still has the same 16 bits/value to write followed by the same heavy math.

>>101979690
I'm not a paypig and I use GenP so I can't use AI tools in PS. You wouldn't want to send your workflow to (((Adobe))) anyways.

File: FD_00090_.png (1.09 MB, 1024x1024)

1.09 MB PNG

>>101979112
Does it just know Tifa or is that a LoRA or just good proompting?

File: 00037-1773544073_cleanup.png (3.16 MB, 1280x1920)

3.16 MB PNG

>>101979728
IOPaint is a standalone tool, it can be used for any image, AI or not

>>101979764
I mean the underlying model they use is SD.

File: image.jpg (73 KB, 1536x1024)

73 KB JPG

enough 1girl time for cool neon cars

>>101979624
OK, this is interesting. Q4_0 looks virtually identical to Q8_0, while being 20% faster (on my 8GB) than Q4_K_S. So I see no reason to use the latter at all.

>>101979748
I mean the pirated version, you don't even need firefly

File: ComfyUI_03662_.png (1.11 MB, 1024x1024)

1.11 MB PNG

>>101979688
To give you a benchmark, I gen an image in 15 seconds on Comfy's basic Flux Dev example: https://comfyanonymous.github.io/ComfyUI_examples/flux/ with weight_dtype on the model set to fp8_e4m3fn. 1.52it/s.

>>101979729
Yeah I have dropped down to the Q6_K quant as it gives me a lot more headroom to work with for LoRA's, controlnets, upscale models etc while having very minor variations from fp16.

>>101979783
d...debo?

What does this node actually do?
I have it in my workflow because it was there by default but I don't know what it means

>>101979264
They added support for LoRA's?

>>101979825
It prevents you from posting images.

>>101979825
ANON, THE IMAGE, THE INTERNET FAIRIES DROPPED IT

>>101979783
Before you move onto cars could you catbox that neon girl?

File: Screenshot 2024-08-20 074941.png (8 KB, 291x151)

8 KB PNG

>>101979825
retard forgot pic

>>101979763
new lora. go on civitai and filter by flux and sort by newest, theres new stuff to try all the time.

>>101979842
Shamelessly copied from another anon (who apparently shamelessly copied it from reddit)

>"base shift is a small, consistent adjustment that stabilizes the image generation process, while max shift is the maximum allowable change to the latent vectors, preventing extreme deviations in the output. Together, they balance stability and flexibility in the image generation.
Using a dog as an example:

>Increasing Base Shift: If you increase the base shift, the generated image may become more consistent and closer to the intended form (a clear image of a dog) with less variation or noise. The dog might appear more stable, with well-defined features, but it could also lose some subtle details or become slightly repetitive in texture.

>Decreasing Base Shift: Reducing the base shift could introduce more variability, allowing for finer details or more nuanced textures to emerge. However, it might also make the image slightly less stable, potentially introducing minor artifacts or inconsistencies.

>Increasing Max Shift: By increasing the max shift, the model has more freedom to explore the latent space, potentially leading to more creative or exaggerated interpretations of the dog. The dog could end up with more exaggerated features or a more stylized appearance, but it might also risk deviating too much from a realistic representation.

>Decreasing Max Shift: Lowering the max shift would constrain the model, leading to a more controlled and realistic depiction of the dog. The image would likely remain close to a typical dog appearance with fewer unexpected variations, but it might lack some creative elements or subtle uniqueness."

>tldr "artistic freedom slider"

>>101979809
K models work in comfy? I tried a K model once and I got an error, Q4_0 and Q8_0 are fine, maybe it was forge, I prefer comfy cause it's faster and I dont get lag with kmodel memory management stuff.

File: file.png (7 KB, 402x97)

7 KB PNG

>>101979874
did you update the node?

>>101979852
Celeb sloppers have already taken over the Flux 1.D feed
Onsite lora training was a mistake!

>>101979882
is that for unet loader (gguf)?

>>101979861
so it's cfg but better?

Is there a node for both cfg and flux guidance? How the fuck you add both nodes separately?

>>101979891
through sheer quantity there will be good stuff even if some is slop, the swift one was quick and that works fine.

File: file.png (34 KB, 765x272)

34 KB PNG

>>101979897
Yes

>>101979861
this reads like GPT slop

>>101979912
whats the difference between q6_k and say, q4/q8?

I can't figure out why Flux fp16 slows down to a halt after using it many times on Windows (loading, unloading by turn off server. Using PC for tasks such as browsing, watching videos and some photoshop editing). After doing all this I close all tabs and at some point when I gen with the model the first step becomes very slow to the point I think it's frozen and I have to cancel which also hangs. Then I have to restart my PC to fix this. This is on a 3090 and I know I have enough vram space, does anyone know what could be causing this?

>>101979785
Moreover Q4_K_S looks significantly different to Q8_0.
>>101979809
>I gen an image in 15 seconds
It takes me 15 minutes to gen that same image... lol
>I have dropped down to the Q6_K quant
Alright, I'll use that one myself when I get the card. I can't fucking wait.

>>101979785
SUMMARY
On a shitty 8GB card.

Q8_0      100%
Q4_K_S   146%*
Q4_0      164%
Q3_K_S    156%*
Q2_K      151%

* Same seed produced slightly different results (also different between them).

Q2_K produces garbled images, so I wouldn't use it. Q4_0 seems to be the fastest. Perhaps because it doesn't use mixed precisions and my AMD card likes that? I have no idea. But based on these numbers, I should I use that one.

File: 1696707062335437.png (1.09 MB, 1024x1024)

1.09 MB PNG

>>101979557
heh, this time with a disney lora

When all the porn focused diffusion generals switch to flux, then I know it's ready. No one yet uses flux for any type of porn or soft core stuff.

Is it possible to do a proper finetune of flux on only 12gb if you're willing to wait ridiculously long?

>>101979957
Porn diffusion was a mistake.

>>101979938
Wait more, it needs to load first time

File: image.jpg (133 KB, 1536x1024)

133 KB JPG

welcome to slop city

>>101979841
>Before you move onto cars could you catbox that neon girl?
i cannot but i can give you the prompt. its base flux anyways

>This is a photograph of a young Russian teen girl with a fair complexion and long, straight blonde hair, styled in a simple, natural manner. She has a slender build and is wearing a luxurious, white fur coat with a skimpy pink bikini underneath that adds a touch of glamour and sophistication to her appearance. Her facial expression is neutral, with a slight hint of seriousness or contemplation. She has large, expressive eyes with a soft, natural makeup look that highlights her features without being overly dramatic. The background is a dimly lit, modern setting with a neon pink and purple glow, suggesting a nightclub or a trendy, hipster venue. The neon lights create a vibrant, almost surreal atmosphere, casting a pinkish hue over the entire scene. To the left, there are indistinct figures of people, blurred due to the low light and the distance, adding to the sense of a bustling, lively environment. The teen girl stands out prominently against the background, her fur coat contrasting with the vivid neon lights. The overall style of the photograph is modern and artistic, capturing the essence of contemporary urban culture with a touch of glamour and mystery.

File: ComfyUI_03624_.png (1.35 MB, 896x1152)

1.35 MB PNG

>>101979930
https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main

FP16 - Base Flux model, which should have the most accurate/best gens.

Q8 > highest similarity to FP16 - highest amount of VRAM to run

Q6_K .. Q3_K_S

Q2_K > lowest similarity to FP16 - lowest amount of VRAM to run

So basically you're trading decreased VRAM usage for 'worse' gens.

>>101979938
i think its a bug with loading models or something, after loading a new model the next image genned is super slow but if you interrupt and restart it's fine. best to change model, wait until the process is done in the console, then click

>>101980002
Thanks I'll take it

File: 2024-08-19_00284_.png (1.19 MB, 768x1024)

1.19 MB PNG

>>101979957
erotica without nudity is definitely possible, but I guess that's not what the coomer generals look for

File: 1694754154696941.png (1.05 MB, 1024x1024)

1.05 MB PNG

simpsonize lora:

When training a LoRA on multiple subjects should I include images of the subjects by themselves in the training data?

>>101979957
>it's not good unless i can coom
We really live in a society.

>>101980011
I see, so the idea is to just cancel and try again.

File: Flux-20240813_154026-re-7(...).png (2.65 MB, 2304x960)

2.65 MB PNG

>>101979861
damn right I did, and its important info.
>>101979980
good face at that angle

>>101979957
https://files.catbox.moe/w254n0.jpg

>>101980065
I think I'll be using SD/SDXL for the foreseable future for coom.

bros how can i tell if my lora is improving?
trying to improve accuracy of a character's outfit, 2500 steps now and there's not a huge difference other than the general aesthetic looking less "sharp"

>>101979930
_k means slightly improved versions of q quants. For instance q4_k_m is better than q5_0, and so on while requiring less vram. q8 is almost lossless while q6 and on is still high quality. Below q5 is when you start to see very noticeable regression, though this applies to LLMs.

>>101980115
I'd say you start seeing noticeable artifacts on Q3 and below.

>>101980090
I have to use regional prompting to get good public nudity images with SDXL (SD1.5 is right out)
Flux has much less concept leaking and more consistent environments, I can just prompt for it.

https://civitai.com/models/661864/side-view-blowjobs-for-flux-photographic-and-anime
How the fuck did he manage to make it look sloppier than the shittiest sd1.5 models?

>>101980146
I personally would never touch a q4_0 quant but it's all personal preference, after all GPTQ is 4bit.

File: ComfyUI_03666_.png (1.65 MB, 1024x1024)

1.65 MB PNG

just noticed today that i must have accidentally clicked and switched my clip from fp16

>MFW MY LAST 700 GENS HAVE BEEN MADE WITH THE FP8 CLIP

>>101980148
>public nudity
I can see how that would be limited.

>>101980188
Throw it all in the garbage.

>>101980188
So...placebo?

>>101980170
Flux is ass

About to attempt making my first LoRA for Flux - if I have a very large dataset (10,000+ pics), what's the most efficient way to train on 24GB of VRAM so that I can fit in multiple epochs? Or is that a pipe dream?

>>101980188
That's pretty easy to notice kek

>>101980170
trained 100% on AI and
>all photorealistic images were tagged with 'photo, realistic'
the man is retarded

File: file.png (893 KB, 865x900)

893 KB PNG

>>101980223
you can enjoy my long list of failgens here kek

so much of adjusting the guidance, the base shift, the prompt.

fucking kill me, wonder how many fingers and toe issues i could've avoided with a better CLIP

>>101980234
i mean it's just one node in a massive workflow that i never have any reason to touch

>>101980249
Do just like 99% of civitai model makers? Except they don't disclose and everyone rates those models pretty highly, and it's worst possible AI slop/aesthetic you can imagine.

>>101980262
>massive workflow
Sounds like bloat

>>101980286
post catbox and we'll compare

>>101980262
Nice. I love this shit.

>>101980233
>if I have a very large dataset (10,000+ pics),
exactly how many concepts are you trying to train into it? If it is any less than 1000 you have way too many images for a lora.

>>101980188
BRO

>>101980173
brap

>LOOK AT MY MASSIVE WORKFLOW BRO
lmao

File: 1701564535648499.png (1.43 MB, 1024x1024)

1.43 MB PNG

inkpunk lora miku, neat aesthetic

My workflow is girthier than yours.

What temperatures do your GPUs reach? Mine's (junction) sits around 93ºC/199ºF. How hot can I expect a 3090 to get in this scenario?

>>101980188
what if you use fp8 model weight and fp16 text encoder?

>>101980033
i think that might help if you also train the text encoder... perhaps it even helps if you stop the text encoder training earlier sort-of when it has been trained and then bake the unet for a bit longer.

>>101980352
that's what most people are running

>>101980365
ive had good text even with fp8 encoder, how much better is fp16? I assume better for more complex text based prompts? (news overlays etc)

>>101978563
>weird old man fakes the support of taylor fucking swift
>AI gets regulated and we never get to use it again

File: image.jpg (131 KB, 1536x1024)

131 KB JPG

i need a cyberpunk store name that's not as cringe as AQUA FUSION
rude and funny names allowed

>>101980343
Around 75ºC never seen it reach 80ºC
4070 ti Super

File: Flux-20240813_155811-re-5(...).png (2.77 MB, 2304x960)

2.77 MB PNG

>>101980331
not bad-o
>>101980342
my workflow is modular-er

>>101980412
My case's cooling might be shit. I assume a 3090 gets hotter by default than a shitty AMD right?

>>101980390
You can try to come to my home and regulate.

>>101980411
2DN slop is so boring in a flux thread

how do you load flux1-dev-Q4_0.gguf? it doesn't show in the load diffusion node

>>101980454
https://github.com/city96/ComfyUI-GGUF

here you go lad

>>101980450
this is not 2DN

>>101980497
sorry, all the slop just blends together

>>101980429
What kind of 3090 do you have? I had a FE 3070 ti previously and thing was like a radiator. Still I don't think you should be hitting 93, that seems a little high

>>101980497
do NOT download this node, it creates mustard gas

Does anyone have an example of an image-to-image workflow for flux?

File: file.png (403 KB, 1292x824)

403 KB PNG

>>101980534

>>101980318
1 concept I guess? Literally just want the model to get a certain type of outfit right, even down to the smallest details like patches/emblems which it currently messes up into distorted garbage, hence my desire to train a LoRA.
Pics are basically just batch exports from some chat channels I'm in that are dedicated to sharing pics of that sort of thing, batch captioned with JoyCaption.
Is 10 images really enough for that? Seems like an awfully small dataset for getting ALL the details right, but then again I don't know much about training. I suppose I could curate the dataset a little, depending on how many pics I'd actually need.

>>101980518
According to what I've read, it's normal for the AMD to hit 93 during heavy loads (it stays at 65 on idle).
>what kind of 3090 do you have
I should have it by the end of the week. It's an Asus TUF GeForce RTX 3090 Gaming OC 24GB GDDR6X.

>>101980557
>10000 images of one outfit
nigga what

>>101980575
More like 10k images of the general period, a lot of which include it.

>>101980555
Thanks but I should have been more specific, I meant image -> CLIP -> image.

>>101980568
Ohhh, I thought you HAD 3090 that was hitting 93. No idea about AyyMD cards, but yeah unless you live in a desert you should see better temps

how THE FUCK are doing this?
https://www.virtualstagingai.app/

the results are way too good

>>101980511
>modify this prompt and set it in an abandoned bunker instead of the beach
>girl instantly becomes 5 years younger without touching her description
What the fuck Flux.

>>101980636
>/ldg/ - Local Diffusion General
take your imagefx slop out of here

>>101978762
useful workflow if you don't have lora for the girl

File: ifx125.png (1.22 MB, 1024x1024)

1.22 MB PNG

@101980642
have another whilst you rant on everyone nig*bo

>>101980636
dude weed lmao

File: ComfyUI_00717_.png (1.6 MB, 1216x832)

1.6 MB PNG

>>101980646
trying to generate real girls is fucking weird imo

>>101980343
the fans will barely spin up if you nvidia-smi -pl 240. there's no reason to run at full wattage.

>>101980607
HELLO? IS THIS ON?

File: file.png (802 KB, 480x768)

802 KB PNG

>>101980625
I think it's the smiling. I changed her from smiling to simply "content" and she aged again.

File: ComfyUI_temp_mkpud_00032_.png (2.3 MB, 1360x1600)

2.3 MB PNG

>>101980051
thx

>>101980607
>>101980712

buy a fucking ad if you want (You)s so badly you attention whore

>>101980727
I thought for a moment that was a Poppy LoRA and I feared for muh dick

>>101980727
I can find a million REAL pictures of Anya Taylor Joy just like this
Celeb sloppers OUT!

>>101980729
I'M ASKING HOW TO DO IT BECAUSE I DON'T WANT TO PAY FOR THE FUCKING SERVICE YOU SUBHUMAN FUCKING RETARD

>>101980607
i'd say a model finetuned on furniture and segmentation/a series of masks. certainly cool but i don't see how it's not misrepresentation
>Fully MLS compliant: No misrepresentation
the only possible use case for agencies would be to misrepresent that the apartment is a furnished rental when its not
it's going to be immediately obvious and take some explaining during viewings

https://civitai.com/models/647237?modelVersionId=743342
Is it normal to sell other people's models on civitai? It's not even a finetune of any sort, he just reuploaded all the quants from huggingface

>>101980750
send them an email and ask you antisocial sperg

>normalnudes lora put a dick on my girl
It begins.

>>101980727
inpainted face?

what is the best upscaler? any good workflows?

>>101980751
what?
I'm asking how are they doing it
I srsly doubt those fucking retards invented anything revolutionary, they must be using the same fucking tech everyone has access to

>>101980752
pay for the bloody model basterd bitch saaaar DO NOT redeem the buzz

>>101980762
literally get cancer and die

>>101980789

>what?
>I'm asking how are they doing it

>i'd say a model finetuned on furniture and segmentation/a series of masks

can you not read? or do you just expect to be spoonfed down to the very last drop?

>>101980396
Grugg's radio shack

>>101980812
fucking imbecile, kys
you're not saying anything

File: 1695452853025919.png (1.31 MB, 1024x1024)

1.31 MB PNG

pixar migu

>>101980827
anon if you know enough to know i'm (((not saying anything))), then you know enough to go *do the thing* instead of whining like a retard asking to be spoonfed

>>101980795
Glorious

>>101980844
>THAR USHING MOSKS AN ISHTABOL DEEFFUSHON
you're not saying anything you literal fucking retard lmao

>>101980789
yes, probably finetunes, segmentation, masking, the same tech everyone has access to. not sure what more you want. do you want me to build the an equivalent service for you? not going to happen unless you pay me a lot more than the $25 it would cost you to use their service for a month to generate the images you want to sell or rent your property. if you can't afford $25 you certainly can't afford a custom solution

File: loraTesting.jpg (3.8 MB, 3891x2500)

3.8 MB JPG

Testing out using captions in a LoRA (pony based model) along with tags

>>101980855
who rattled you cage this morning?

>>101979441
I don't get it. Q2 supposedly fit 6 GB vram but I don't see speed difference, even with fp8 I got same speed.

>>101980867
imagine being this stupid

>>101980872
your whore mother

>>101980855

why don't you spell out for the class exactly what you'd like to know anon? i think they made it pretty clear.

>>101980884
no explain why you mad all day huh?

>>101980884
i can't, why don't you tell me what it's like

>>101979090
this is getting dated, but posting it anyways.
https://phhofm.github.io/upscale/multimodels.html

Pick two or three. There is no silver bullet upscaler.

>>101979780
it does allow you to change models. I haven't played with it enough to see what is possible.

>>101980886
what part of "how are doing this" you don't get, you subhuman imbecile?

>>101980892
because your whore mother

>>101980895
imagine being this stupid unironically

>>101980874
I have 8GB and although it can technically fit the model in it, it can't fit the image it's generating. 6GB is even smaller. You get the same speed because your RAM is doing a lot or most of the work (not your VRAM).

>>101980914
look d*bo we are not moving to your samefag hugbox, sorry not sorry

File: ComfyUI_03695_.png (1.07 MB, 1280x768)

1.07 MB PNG

>>101980914
yes, what exactly do you want the "how are they doing this" answer to entail, anon? they've given you an outline - what *precisely* do you want from this interaction?

>>101980929
i don't give a shit about discord tranny drama you subhuman piece of shit

>>101980914
what part of "finetunes, segmentation and masking" do you not understand? short of breaking into their systems or conducting extensive research that's my best guess at how it's done. how would you like it to be done if not like that?

File: 1701881887012605.png (1.36 MB, 1024x1024)

1.36 MB PNG

>>101980836

File: ComfyUI_Flux_9691.jpg (153 KB, 1024x1024)

153 KB JPG

i should go out more.

>>101980934
i want a PRECISE explanation, you fucking retard
they're OBVIOUSLY using some kind of furniture lora with fucking mask, you fucking imbecile
you're not saying ANYTHING

@101980942
>d*bo

>>101980944
see >>101980960
you faggots are so fucking useless, no wonder you spend all day here posting anime trash
literally fuck you lmao

File: ComfyUI_00830_.png (1.43 MB, 1216x832)

1.43 MB PNG

>>101980960
>>101980934
spoon(((fed))) me

File: image.jpg (131 KB, 1536x1280)

131 KB JPG

>>101980815
>Grugg's
ok

>>101980977
you have no clue how to do it and can't even begin to explain it lmao

>>101980994
i'd say a model finetuned on furniture and segmentation/a series of masks

>>101981012
>i'd say a model finetuned on furniture and segmentation/a series of masks
see, you have nothing lmao
just say "I don't know"

they pay jeets to GIMP the furniture in, that's it, that's all, that simple

Fresh bread that can be bumped...
>>101981026
>>101981026
>>101981026

>>101981021
how do you think it works anon?

>>101981033
if I knew I wouldn't be here talking to retard losers who clearly know way less than me

>>101980960
ask me to elaborate then instead of freaking out
i'd say they are segmenting the areas of the room, floors, walls, windows, there will be a component for room design that estimates the size of the room and chooses an appropriate layout, then masks are generated with the size and location of the item of furniture, a finetuned model is used to generate the item of furniture
in short, they're using finetuned models, segmentation and masking
just pay the $25, you're not going to be able to make it yourself

>>101981066
>just pay the $25, you're not going to be able to make it yourself
see you don't know, it's fine
just stfu

>>101981095
>just stfu
we all wish you would d*bo