official pixart bigma, lumina 2 and hunyuan finetune waiting room, now with flux 12b

>>101685438
fr lower your guidance to 1.25 and enjoy your monet oil paintings anon

File: file.png (636 KB, 1280x800)

636 KB PNG

Imagine going to a buffet and complaining they don't have

File: ComfyUI_00050_.png (1.58 MB, 1024x1024)

1.58 MB PNG

>>101685438
I think it knows hastune a bit too well to allow her to be too heavily stylized

File: file.png (1.53 MB, 1280x800)

1.53 MB PNG

Get a new tool and the first thing you do is try to make everything you can already do with a heavily finetuned SDXL model, and then complain that the new model isn't heavily finetuned

File: FLUX-Schnell_00019_.png (1.05 MB, 1024x1024)

1.05 MB PNG

NYT best seller

>>101685566
try not listing artist style, instead try classical abstract painting of 'character name'

>>101685471
nta but how do you do this in cumragui? is there a custom node for it?

File: file.png (2.72 MB, 1280x800)

2.72 MB PNG

What if I just use the same prompts and settings I used on my heavily fine tuned SDXL workflow and then complain when the new model doesn't respond to my 100 image Lora keywords.

File: ComfyUI_00060_.png (1.83 MB, 1024x1024)

1.83 MB PNG

>>101685602
Aint that the truth, I'm just having fun seeing what it can do well, though.
>>101685619
I'm using the split-artist method, which seems to work decently well at emulating an artstyle

SD bros, we barack? https://x.com/recatm/status/1819348949972476019?s=46&t=t04c6G-lAweXkub6PTPbKQ

>>101685688
Nope, they're retards. They're incapable of testing, they lack for taste, and they hand-wring about safety.

File: file.png (871 KB, 882x1072)

871 KB PNG

>>101685688
>>101685706
Holy shit their examples are garbage, Flux is CLEARLY better and if they think bragging about SD3 being 2.5B makes them look good...

File: FLUX-Schnell_00026_.png (1.18 MB, 1024x1024)

1.18 MB PNG

>>101685688
>The 3.1 model is almost on par with Flux Pro in terms of quality, but it's nearly 5 times smaller, with only 2.5b parameters,
>2.5b is all you need
I promise you it's dogshit

>>101685688
Stable Diffusion will always win in the end, it always does. Millions of chinkslop models later yet SAI models remain on top. They simply cannot be defeated.

>>101685731
Pony is the only reason SDXL is relevant. SD3 is irrelevant. SDXL was irrelevant before Pony.

Emad will add the artist tags back and begin his revenge arc, trust the plan
oh wait he left, who even is sai anymore now that emad is gone and so is everyone else? lykon scratching his head knob-tweaking the half-finished codebase he was left with?

>>101685753
Artist names are a crutch and cope. The only things it really needs is: pop culture understanding and autistically tagged art styles.

File: ComfyUI_temp_bhorq_00013_.png (3.01 MB, 1280x1512)

3.01 MB PNG

pretty cool

File: 2024-08-02_00149_.png (1.52 MB, 1280x1280)

1.52 MB PNG

>>101685731
9/10 FLUX.girls disgreee with this opinion and are laughing now

>>101685566
It defaults to miku if you put "figurine", I tried some touhou characters or random anime crap and it just makes a miku or blue hair with ponytail kek
>>101685471
i'll try that later ty

File: offload_test.png (803 KB, 2540x835)

803 KB PNG

Offloading CLIP + the VAE to a second GPU seems to stop the model shuffling for the most part.
https://gist.github.com/city96/ (also added to extra models)

someone with a lot of knowledge told me that because fluxdev is a distilled model, it will probably not be possible to train the model properly.

can anyone else with a lot of knowledge say something about this?

File: 8.png (1.47 MB, 1200x1450)

1.47 MB PNG

>>101685688
no we arent
LMAOOOOOO

File: ComfyUI_temp_bhorq_00015_.png (2.9 MB, 1280x1512)

2.9 MB PNG

>>101685769
doing some different political stuff. I DO NOT ENDORSE THESE

>>101685798
Pro is the real model which you would want to train on. Dev is the next best model, it is non-commercial so any training is for funsies. Schnell has to be fixed.

It's the reason why SDXL Lightning models aren't trained on, they're an output from a real model.

File: FLUX-Schnell_00034_.png (1.07 MB, 1024x1024)

1.07 MB PNG

>>101685812
I do

>>101685834
why do people even care about licences like just don't snitch

Does Flux-Schnell also need 24GB VRAM?
Is it a local dalle-quality option now?

>>101685865
serious finetuning is a considerate financial investment, projects like pony succeeded cause they got enough sponsors to finance their weeks of server time, you don't do that for funsies unless you are way to rich

>>101685834
i got that - he said it just wouldn't make much technical sense - lora, finetunes etc.
so there won't be a similar ecosystem, not because of commercial, but from a technical point of view because dev was distilled.
i didn't understand the details ^^

File: 2024-08-02_00138_.png (2.1 MB, 1280x1280)

2.1 MB PNG

>>101685895
both dev/schnell have the same vram requirements, I seen reports of ppl using it with less with alot of swapping, if its fun? you decide

File: ComfyUI_00072_.png (1.53 MB, 1024x1024)

1.53 MB PNG

>>101685772
based flux tentacleanon

>>101685900
I mean these models aren't magic, Schnell still has the parameters and knowledge, but you basically have to full tine tune it back up to Pro. I'd imagine it would be costly to do but maybe for a massive copyright and pornographic fine tune it wouldn't matter because it might be enough new knowledge that you have to do it anyways. Like it takes a lot of epochs to teach a model a vagina especially when the model has preconceptions of what a crotch is.

running flux dev in comfy on cpu (64 gb ram) just crashes python for me

File: FLUX-Schnell_00035_.png (1.47 MB, 1024x1024)

1.47 MB PNG

>>101685895
Running it on a 4080. The default workflow on comfys github is dogshit btw

File: flux1-schnell_00036_.jpg (673 KB, 1024x1024)

673 KB JPG

>>101685688
YOU WOT?!

>>101685767
artist names ARE art styles, there is zero difference. you will not find a more powerful stylistic compression format than simply using the nametag. cope all you like, but there is no avoiding this truth.

File: ComfyUI_temp_bhorq_00026_.png (3.15 MB, 1280x1512)

3.15 MB PNG

I feel that using schnell at higher resolutions nets me much better results than dev on normal 1024. and it's faster

>>101685964
yeah faggot but you will never win that PR battle and honestly I like how it makes people like you seethe, I can already see you struggle when you are forced to deviate from your workflow because honestly, you're not very good at this

>>101685929
ok, there's nothing i can do anyway - ticked off.
i was only interested because the statement came from one of the developers of transformers and it sounded like a technical problem in itself

>>101685989
I don't know why it would be a technical problem per se, from what I understand the basics of lightning models is you have two models, a teacher and student, you give both models the same prompt, the teacher does a normal generation (say 50 steps) and the student has to do the generation in 4 steps, the student is graded on how close it is to the teacher. This jiggles the weights to value speed.

File: FLUX-Schnell_00046_.png (1.23 MB, 1024x1024)

1.23 MB PNG

>>101685964

File: file.png (87 KB, 256x256)

87 KB PNG

Finally prepping for the 512px model. I won't wait for the 256px model to be perfect. It'll still take like a week encoding all the latents.

>>101685784
I have 2x3090 how can I do that ?

>>101686045
>>101685913
>>101685925
kek, I guess atleast the furry and tentacle porn finetuners don't have much work to do

>>101686066
im so fucking hyped right now

>>101686068
Grab ComfyBootlegOffload.py from the linked page and put it in ComfyUI\custom_nodes then restart comfy.
Reload the UI and add the ForceExternal[CLIP / VAE]Device nodes after your clip/vae loader as shown in the image.

File: file.png (88 KB, 256x256)

88 KB PNG

>>101686109
I'm nervous about how much it likes white background though. It seems very autistic about prompts.

File: ComfyUI_temp_bhorq_00042_.png (2.52 MB, 1280x1512)

2.52 MB PNG

love this shit

>>101685784
Just tried it. The output image seems to be slightly different compared to having the them all on the same GPU. Any reason why that would be the case?

>>101685688
Model in the lab is always good. Then when comes time for release, you have to introduce "safety".
Then it's no longer good, but it is safe, which is better than being good, right?

>>101685988
artcel argument. how quickly jeets adopted the mindset once they found a couple knobs to tweak. pathetic

File: FLUX-Schnell_00059_.png (1.16 MB, 1024x1024)

1.16 MB PNG

>>101686144

File: 1718096231391059.png (91 KB, 1955x1081)

91 KB PNG

Why is Comfy like this?

>>101686128
Thanks anon!

>>101686208
I'd imagine SharkSampler is some jeet's node

>>101686201
you dumb nigger artist prompts haven't been relevant since 1.5

>>101686208
>when the spaghetti spill

>>101686223
Downloaded some anon's catbox from a few threads back and wanted to try and get it running but no luck - it's from https://github.com/ClownsharkBatwing/RES4LYF
Guess it's joever for me.

Flux is so slow on my 12GB GPU, takes multiple minutes to generate a single image.

File: ComfyUI_13078_.jpg (504 KB, 1280x768)

504 KB JPG

>>101686188
Not really sure why it would change, all the node is doing is moving the model to the other device and stopping comfy from moving it back.

File: flux1-schnell_00039_.jpg (888 KB, 1920x1080)

888 KB JPG

>>101686208
>RES4LYF

File: ComfyUI_00909_.png (1.27 MB, 1024x1024)

1.27 MB PNG

>>101686066

>>101686208
I had an error when I tried Flux for the first time, I don't know if it's the same than this one, but all it took was to pip install --upgrade safetensors.

File: 1722611152815486.jpg (113 KB, 802x890)

113 KB JPG

So now that a bigger image model has come out, it is truly and undeniably over for vramlets? Maybe I should sell my 4070ti and buy a 3090 instead

>>101686350
24 GB VRAM will be the promised land for awhile.

File: ComfyUI_temp_bhorq_00075_.png (1.61 MB, 1512x1280)

1.61 MB PNG

really nice underwater photos with flux

>>101686241
have you updated recently? did you try updating both comfy and that repo?

Can flux run on a P40? Yes I know it will be slow, but will it run? Anyone tried?

>>101685865
That works until you hire employees.

>>101686657
>P40
if it runs pytorch it can, but you will probably wait 10 minutes for a simple diffusion

File: ComfyUI_00306_.png (1.17 MB, 1408x704)

1.17 MB PNG

Anyone know how to prompt for consistently sharp backgrounds? It keeps giving me DOF.

>>101686684
you do know design 101 right?
you do know why that image not having a focal plane is retarded right?

>>101686712
Idiots see SD 1.5 gens where every detail is sharp (making it look like copy pasted in parts) and then cry when better models actually create coherent images

flux walked so bigma could run. finally we have a new standard to compare local models to. anything that releases worse than this will be quickly dismissed, as it should be. same with censorship, anything more censored than this will be quickly dismissed. the model checks almost all the boxes local needed so we can finally go "flux did it, why can't you?" whenever some useless censored griftbake comes along. bigma will do it even better in only 8b, the future is bright.

so when will civitai start the fundraiser campaign for ponylux and bigasplux?
I'll donate 100$ per project

File: ComfyUI_15318_.png (1.39 MB, 1024x1024)

1.39 MB PNG

duct tape SLI

>>101686827
Whoever starts a rental service for finetuning Flux is going to make big money.

>>101686563
Yeah, my whole Comfy setup isn't that old, installed it when Flux came out. Might try a clean install though, maybe some of the other custom nodes I have are interfering with it (though I doubt it).

>>101686790
I'm on suicide watch

>RAM is dirt cheap
>can easily buy 64gigs and upgrade my PC
>can't do the same with VRAM just because of (((NJudea)))'s greed

File: 1wing_4.jpg (347 KB, 1552x1200)

347 KB JPG

>>101686468
Great stuff

>>101686838
it is too expensive for amateurs, the non-commercial license makes it uninteresting for professionals.
Either fundraiser or we will have 3 completely useless semi finetunes in the next 12 months that are not worth downloading

File: 142124645677809.png (241 KB, 2477x1276)

241 KB PNG

help
i'm running on cpu and it does this

>>101686791
considering the discord server for pixart is full of braindead retards spreading misinfo, I doubt it. I for one welcome the german overlords

>>101686874
Honestly I expect someone to distill the 12B model to something more manageable like a 6B model. Also necessity breeds innovation, there are going to be a lot of researchers taking at swing at it to get it to a 24GB VRAM training target. A lot of innovation on the LLM side happened because of the large models BTFOing 24GB cards.

>>101686712
The focal plane being from the subject to infinity is entirely normal though when doing wide angle landscape shots, which are what I'm trying to go for. I actually don't mind if Miku is a little out of focus, but I'm having a skill issue even trying to prompt that successfully right now. It really likes just keeping her in focus.

>>101686874
after SAI shit into the pony teams face when they wanted to do a SD3 finetune, I could see flux.pony happening, they got the fanbase (cause porn sells), so they could lift the money thru fundraiser.. problem is then we just got another porn model

File: ComfyUI_00921_.png (826 KB, 1024x1024)

826 KB PNG

>>101686887
>>101686903
>>101686904
>>101686911

>>101686904
Post three examples of what you think you want.

>dabble in LLMs
>48GB VRAMlet
>Flux releases
yay

>>101686887
who cares about some shitcord really, i remember back when the devs had only 2 channels and it was full of jeets trying to run prompts in the image-share channel thinking it was a bot.

Meh, a bit too stylized for my taste. SD 1.5 was more diverse, whereas flux is stock photography/influencer generator.

>>101686922
devs left for Nvidia bux

>>101686885
do you have pagefile enabled?
start from small setup, use fp8 on both dtype and t5, use 512x512 for quick test

>>101686917
kekd

>>101686903
honestly
can we still fundraise and then people can distill pony&co?
don't feel like waiting 12 months for 'maybe, well, could be'
there's a huge hype and the people who could afford 24gb vram can spare a 20 and we'll have what we want in two months
the fanbase is big enough, civitai is well known enough by now and the two finetuners are experienced enough - we leverage that with money and they do what they want to do anyway, but not the money for it
why simple when there is also the principle of hope
SDXL could be used commercially and look at the shit up pony

Is there any research going on into making models consume less VRAM?
It can't possibly require the whole model for making multiple gens off the same prompt
Culling all the weights that don't contribute to a prompt basically

>>101686994
c++ has better quants than python

File: 2024-08-02_00198_.png (2.16 MB, 1280x1280)

2.16 MB PNG

>>101686917

File: ComfyUI_00922_.png (1.07 MB, 1024x1024)

1.07 MB PNG

>>101687038

>>101686994
the issue isnt models, its jewvidia. models are reasonably sized and the scaling is actually fucking perfect in theory given how easy it is to slap on more ram and keep adding more params. the issue is that they're intentionally holding back local cards to sell enterprise ones for 20x the price. they very easily could slap on double or triple the vram and we'd be fine.

File: ComfyUI_15332_.png (1.02 MB, 1024x1024)

1.02 MB PNG

>>101686845

>>101686993
Pony is a fluke and I wouldn't hold your breath. I predict he's going to take a year or more for his next model and it's going to be shit, just like Waifu Diffusion. Pony is now acting SAI handwringing and focused on censorship now that he has the spotlight. This is also ignoring how incompetent his training ultimately was and I think what Pony became was actually a lightning in a bottle where someone incompetent sometimes can do something well.

It's not *that* hard to make a model from scratch, anyone smart wanting to do this and spend real money would be making a model from scratch and thus be in full legal control.

>>101686920
Google kept giving me irrelevant results kek. Here are some from a stock image source.
https://www.dreamstime.com/stock-photo-woman-looking-coastline-near-azure-window-gozo-island-girl-coast-line-malta-image41848987
https://www.dreamstime.com/royalty-free-stock-photos-romantic-beach-vacation-thailand-image22434028
https://www.dreamstime.com/stock-photo-couple-sitting-cliff-watching-ocean-landscape-wide-angle-view-mallorca-image47994768

>>101687048
its not like the original target audience of the consumer GPUs need more VRAM, do you know a video game that needs 24GB VRAM? I don't .. so from that standpoint its a reasonable descision to not give more to the top line (okay I agree somewhat on the gutted lower end ones) ..

but now the situation has changed and its more a demand issue getting hands on precious GDDR chips ..

>>101687079
those images are shit holy moly

File: 1708883931868049.png (79 KB, 277x435)

79 KB PNG

how do i stop this "halo" effect around edges using flux?

File: flux1-schnell_00049_.jpg (111 KB, 512x384)

111 KB JPG

5 minutes to gen with R5 3600 16GB
Tbh this is more tamer than PixArt for vramlets

>>101687099
Ask darkforest to have fewer overedited images in the training set.
Overly aggressive unsharp mask in this case.

>>101687078
Go to your garage and hang yourself you tranny faggot, your babble is unbearable - let some niggers fuck you at the train station before that your that you have at least once achieved something in life, this pseudo smart shit is a shame for humanity. Fuck you retarded subhuman

File: ComfyUI_00924_.png (1.61 MB, 1024x1024)

1.61 MB PNG

>>101687099
It's so huge it might have learned that distortion kek

>>101687131
Hey leech, no one is going to do everything for you.

>>101687091
It is what it is, I didn't want to spend more time at this point trying to find better ones, these search engines suck.

File: view.png (1.6 MB, 768x1216)

1.6 MB PNG

Trying to get FLUX to spit out old 90s computer graphics with lots of dithering without making it pixelized. Any advice?

>>101687170
use dalle

File: file.png (36 KB, 900x600)

36 KB PNG

>>101687160
You won't find examples because what you want is shit and not how the world works. Your eyes have a focal plane, so does cameras, and in art, you have a focal plane because it's falls in line with the principles of design which is necessary for making coherent and pleasing images.

>>101687079
what llava-llama3 gives me:
>A woman sits on a rocky outcropping overlooking the ocean. She is wearing a blue tank top, shorts, and white sneakers. A green backpack rests behind her on the rock. The sky above is clear and blue with fluffy white clouds scattered across it. In the distance, several cliffs can be seen against the horizon line of the water. The photo was taken from a high angle looking down at the woman. There are no texts or other objects in the image. The relative position of the woman to the ocean suggests that she is quite some distance away from the water, and the cliffs appear far in the background.

what llama 3.1 made of it to turn it into a prompt:
>Serene coastal scene: a young adult woman sits alone on a rugged rocky outcropping, gazing out at the vast expanse of the ocean with her back to a lush green backpack. The clear blue sky above is dotted with puffy white clouds, and the warm sunlight casts a gentle glow over the entire scene. In the distance, several rugged cliffs rise up from the horizon line of the water, creating a dramatic backdrop for the peaceful atmosphere.

File: 2024-08-02_00214_.png (1.9 MB, 1280x1280)

1.9 MB PNG

I am loving the this model, didn't have as much fun with generating stuff since when I was making my little own SD1.5 models. So much potential..

Right, right. But show me a pretty Flux gen.

>>101687226
I can't, range banned for image uploading.

>>101687189
I'm not sure why you're so stuck on the close focal plane idea. I just want want the background to be in focus. The subject may or may not be, I actually don't care about that.
Ideally we'd have these models generate a depth map for the image so you could adjust the focus yourself but that's not what they do, and right now I can't seem to prompt it so that the focus is on the background rather than on Miku.

>>101687185
>185▶
>>>101687170 (You)
>use dalle
Get a load of this guy

>>101687253
have you tried using your words, if the subject which is in the background is what is to be in focus...

>>101687284
use dalle. you wont be able to get that result out of flux without loras, which arent available yet

>>101687227
missing the hecking wholesome flyffy ears

File: ComfyUI_00311_.png (1.11 MB, 1280x800)

1.11 MB PNG

>>101687231
Is that with the prompt from >>101687190?
Interestingly I also get blurred background, but yours seems fine. Can you catbox it? I just used the workflow from Comfy's card.

>>101687309
I wrote specifically that it's a shitty amateur photograph from someone who has no fucking idea what they're doing.

>>101687288
Yes I tried that. But perhaps there's something going wrong with my workflow, not sure >>101687309

File: view.png (1.35 MB, 768x1216)

1.35 MB PNG

>>101687302
I've been able to get it to replicate a bunch of styles so far, even mixing 2D and 3D elements together, it's just a matter of finding the right way to prompt it.

>>101685776
>It defaults to miku
They get all the mikulovers to evangelize for them for free.

>>101685767
>autistically tagged art styles.
like picasso?

File: ComfyUI_00083_.png (942 KB, 1216x832)

942 KB PNG

I like Flux.

File: file.png (1.32 MB, 1280x800)

1.32 MB PNG

It's amazing how you have a model that basically does what you write in autistic detail and people prompt it like it's SD 1.5 raped with Loras

File: ComfyUI_00313_.png (1.08 MB, 1280x800)

1.08 MB PNG

>>101687316
Well ok, I just tried
>A young adult woman sits alone on a rugged rocky outcropping, gazing out at the vast expanse of the mountain range with her back to a lush green backpack. A shitty amateur photograph from someone who has no fucking idea what they're doing.
And got this.

File: 1717790738885308.png (1.36 MB, 1280x768)

1.36 MB PNG

>>101687190
>>101687308

File: file.png (388 KB, 711x443)

388 KB PNG

>>101687342
Geometric Fragmentation
Vibrant Color Palette
Distorted Perspective
Bold Outlines
Asymmetrical Composition
Simplification
Exaggeration
Abstract Representation
Symbolic Elements
Playful Imagery
Intersecting Lines
Overlapping Planes
Emotional Expression
Conceptual Focus
Whimsical Touch
Contrast Emphasis
Non-Realistic Colors
Depth Interplay
Flatness
Unconventional Forms

>>101687356
The rock textures are insanely good

so does Flux not work with negative prompts?

File: 2024-08-02_00227_.png (1.64 MB, 1280x1280)

1.64 MB PNG

>>101687383
or is it better realistic?
>>101687391
no it doesnt support cfg and therefore no negative prompts

File: 1716928607772068.png (1.31 MB, 1280x768)

1.31 MB PNG

>>101687366
>>101687366

>>101687391
they do work on dev

>>101687409
>they do work on dev
no

File: 1701604115204719.png (1.87 MB, 1280x768)

1.87 MB PNG

>>101687409

>>101687428
workflow and skill issue

File: flux abstract.jpg (3.44 MB, 1888x2830)

3.44 MB JPG

>>101687376
next bake, please

>>101687409
>they do work on dev
how? nothing i seem to type has any effect

>>101687495
no one said Flux is well trained on fine art
but it clear you have skill issues

>>101687350
>everything is blurry
muh promplet

SD and pixart looking real antiquated right now

File: ComfyUI_temp_fhbqo_00089_.png (1.5 MB, 800x1024)

1.5 MB PNG

>>101687376
Ok, I am going to try that list of tokens vs Pablo Ruiz Picasso and see what gets closer.
I made this with the list of tokens and it didn't work, but I had left an unrelated token to the style so that test is voided.
Some of those tokens work, but it's not enough to represent cubism.

>A young adult woman sits in the background, far from the camera, she's surrounded by a vast mountain range.

>>101687543
I didn't say that's how it is, I said that's how it ought to be. Every painting can be described without using the artist's name. And using the artist's name is a crutch for low IQ lazy faggots.

>theyre actually using the russel's teapot argument
the perfect prompt surely exists, you just havent seen it yet!

>>101687605
Well it's obvious you haven't tried at all. This is why I never give out prompts.

File: 1712059568865363.png (3.85 MB, 1771x2404)

3.85 MB PNG

>>101687476
>>101687498
just use them as normal using CFGGuider node
just don't go crazy with the CFG, I use 1.5

File: 2024-08-02_00234_.png (1.52 MB, 1280x1280)

1.52 MB PNG

nice it genned some highly believable jpeg artifacts on this one

>>101687723
wait I just realised that this is schnell, not dev lol
It works in both then

>>101687724
kek

>>101687611
That's right! Guard those secrets, lest the plebeians acquire their might!

>>101687748
Lest the plebeians believe they're better than they are.

>>101687724
Did it turn the friend chicken wings into frog legs?

>>101687765
Prompts should be passed down from father to son only

>>101687723
>"yellow tint" in negative
>it's still a chink
hmmmmm

I'm just going to say it:
Flux is absolute garbage for the amount of vram it needs. At that point in params it should be much more capable than it is.

File: 2024-08-02_00235_.png (1.66 MB, 1280x1280)

1.66 MB PNG

>>101687766
I actually prompted fried frog legs. I realized you can not discern them when they are fried.

>>101687912
feels like a lot of empty space between the parameters, they don't feel densely packed and 'maximized'. could probably accomplish the same on a well-trained 8b

File: flux.jpg (95 KB, 1024x1024)

95 KB JPG

>>101686790 >>101686845

File: 1712061359376327.jpg (1.23 MB, 3623x1124)

1.23 MB JPG

>>101687723
>>101687495
miku is a very strong tag, try downweighting it or something neutral

File: ComfyUI_Flux_0757.jpg (215 KB, 1024x1024)

215 KB JPG

>steam deck handheld console by valve

File: ComfyUI_04045_.png (1.02 MB, 1024x1024)

1.02 MB PNG

>>101687980
i could get miku in CCTV footage but not in artstyles

>>101687938
Tasty looking fries
Also, fingernails?!

>>101687964
>>101687912
yeah, the team doesn't know what they're doing yet. hopefully they can improve

>>101688079
still better than the mountains of useless chinkslop that will thankfully be forgotten.

>>101686960
still does it and i have 64 gb of ram so it shouldn't run out ever
it doesn't even try to fill up the available ram

>>101685784
can this script be used on the regular ComfyUi?

>>101685784
can you provide your workflow through a catbox picture anon? that's hella interesting

Okay, so what is this Flux thing and how do you run it? What must i need before running the model? (Vram, etc).

>>101688160
16-24 GB of VRAM

>>101685784
So there's already a script ready for a multigpu setting and Comfy doesn't want to include it on his official repo? Why? That thing is really important if we want to enjoy big models and have enough gpus for that

File: ComfyUI_Flux_0773.jpg (84 KB, 1024x1024)

84 KB JPG

this one goes into my gems folder

File: flux_00043_.png (1.29 MB, 1024x1024)

1.29 MB PNG

>>101687226
i tried to do a george floyd but wound up with inter-species buck breaking

File: 1692528894762736.png (2.19 MB, 1024x1024)

2.19 MB PNG

>>101687980
>>101688051
try BetaSampler with alpha=0.2, beta=0.5
this is schnell at 4 steps
"painting of hatsune miku in the style of impressionism"

File: file.png (1.66 MB, 1024x1024)

1.66 MB PNG

lol

>>101688274
catbox okay

>>101688274
don't share
we don't need more AI scrutiny than is already there

File: file.png (2.13 MB, 1024x1024)

2.13 MB PNG

>>101688255

I gotta download like 70-80 gigs for dis flux shit?!?! Yeesh!

>>101688297
Meh. I had worse accidents with realistic pony models
>>101688289
kys faggot

>>101688297
Are you sure that's not a real photo?

jesus christ anon
that is indeed NSFL

so this model is well trained, bodes well for finetuning
if someone manages to snag 4xA100 that is

>>101688347
the guy who shared said this his prompt more or less, or something like that
>pale skin color, pale areolae, bale nipple, pale lips, lifeless, dead body, corps of girl, young nude girl, lying on the dirty ground, closed dead eyes, spread arms, spread legs, bruises on the body, dirt body, oiled dirt hair, skinny, flat chest, sunken abdomen, protruding ribs, blood covered stomach

>>101688274
Its all or nothing with censorship.
You can already make disturbing stuff like this with 3d tools and photoshop. Sick stuff but only then can you have the good as well.
I wish people still had the attitude like in the pc-98 era. Peak creativity.

>>101688297
emad here, thanks

File: file.png (2.17 MB, 1024x1024)

2.17 MB PNG

flux-dev

File: ComfyUI_00020_.png (1.17 MB, 1024x1024)

1.17 MB PNG

can I run Diffusion on this GPU?

>>101688387
largest pair of flux tits i've seen yet. what wuz prompt?

File: file.png (1.84 MB, 1024x1024)

1.84 MB PNG

>>101688406
Hatsune Miku with her breasts and nipples exposed

>>101688365
the problem is the type of people who make this kind of content are lazy and retarded so the barrier to entry to make truly sick shit prevented most of it, now potentially any retard can make it with a simple sentence

fuck i missed it someone post it again

>>101688444
just open desuarchive

>>101688444
just get 4chanx or check the archive

could it be that flux-dev is much better than flux-schnell?

>>101688469
it's not distilled so duh

>>101688477
it is distilled from their pro model, but yeah its still way better than schnell

>>101688469
Dev is distilled for accuracy
Shnell is distilled for steps

>>101688480
both are distilled? I thought it was only schnell

>>101688500
yes
https://blackforestlabs.ai/announcing-black-forest-labs
>FLUX.1 [dev]: FLUX.1 [dev] is an open-weight, guidance-distilled model for non-commercial applications. Directly distilled from FLUX.1 [pro], FLUX.1 [dev] obtains similar quality and prompt adherence capabilities, while being more efficient than a standard model of the same size.

>directly distilled from pro

>>101688492
nvm this seems to be it

In case it's not clear, "schnell" is German for "fast"

>>101688514
delete post. don't need it anymore because anon explained

>>101688431
And they do what with it? Post it on 4chan or jack off to it. Try looking at pixiv. There are thousands of images of abortion porn for example. Where the baby cadaver is pulled out etc.
At best shit like this is put on obscure sites and you get the van everywhere else.
I dont see the problem. Alignment does not work. I dont want an ai with a moral compass of the creator. The ai needs to obey me at all times. If in turn sickos can put out weird shit so be it. Thats an easy trade.

>>101688151
Yeah it's just a custom node for base ComfyUI
>>101688159
It's just those 2 nodes + some default 1.5 workflow from my desktop though. It even has the LoRA loader lol.
Here it is without the jank: https://litter.catbox.moe/a3ngjo.json
>>101688176
>ready for a multigpu setting
This replaces the .to function of the actual pytorch class and basically disables the default patcher, it's the last thing I'd call "ready" lmao

>>101688531
I prefer to not turn a blind eye to true depravity and psychopathic behavior. So yes, I'll draw the line at dead baby porn.

File: Lets-goo.jpg (272 KB, 2964x1398)

272 KB JPG

>>101685784
Anon I kneel, I wanted to put the text encoder on my gpu so bad and now it's possible, this should be on the OP because of how important that is! I really hope Comfy will put this script on his official repo

1) You download this ComfyBootlegOffload.py script here: https://gist.github.com/city96/

2) You put it in ComfyUI\custom_nodes then restart comfy.

I've included a workflow for those who have multiple gpu and want to to that, if cuda:1 doesn't work for you then go for cuda:0
https://files.catbox.moe/jxgi23.png

>>101688574
How long did it take for this image?

>>101688582
the prompt processing + encoding + vae decoder is really fast now, and the inference image is at the same speed as before for me, because the image model was already on my 3090 before that hack

>>101688608
So what's the total time, anon? Just say the seconds

>>101688566
You can't put the genie back in the bottle
To the "AI", a dead baby is like a coffee mug or a desk lamp
If it's not in the model, someone can train it in
Simple as

>>101688531
>>101688621
yeah that's a nice and cute le anonymous based hackerman take, but the real world won't see eye-to-eye. if theyre actually located in kekmany, then it wont be long before the 500+ forms arrive and they're effectively forced to shut down. nobody is escaping ai regulations

>>101688566
>true depravity and psychopathic behavior
That's on the person, not on the model

>>101688621
Here's the problem and I'll spell it out since you're a sociopath that probably needs to be institutionalized. The public will *NOT* tolerate dead baby porn and they will use violence on you to prevent it. You want training AI models banned? You want possession of AI models banned? You want Nvidia to implement detection at the driver level? Do not forget this.

>>101688646
Well you see anon, the public has no problem preventing people from doing abhorrent acts even if they're le consensual and victimless. See how far you go with your dead baby AI porn.

Local chads won

>>101688574
Man something like this should be standard. Load encoder on cheap 3060 and then let real gpu do the work

lol its actually fucking over for local models thanks to this. as it gets more and more realistic, the outrage grows more and more 'valid'. enjoy your safe saas future!

>/ldg/ - Local Doomer General

File: Capture.jpg (94 KB, 2801x434)

94 KB JPG

>>101688612
Here's my speed with CLIP + Vae on my 3060 + the image model on my 3090

>>101688701
Wait, so you have two GPUs? I thought you only had 3060...

>>101688566
Then you are the same like OpenAI.
Your line is just somewhere else. Who gives you the right to decide where the line is.
Doesnt even make sense too. Why even allow extreme violence in general? I think thats unhealthy to watch.
I remember to outrage of rapelay. Stalking a mother and raping her in front of the family or whatever. Would you allow that?
At least be consistent man, its not about the baby. I would ban ALOT more if I would ban what I think is unhealthy in general and disgusting.

Are you guys really unable to load the model in fp8 plus the encoders on 24 GB VRAM? It works fine for me. I have a 3090 + 3060 and tried >>101685784. My image gen times didn't change appreciably.

File: file.png (10 KB, 318x115)

10 KB PNG

>>101685784
>>101688574
I have two 3090s, I copied the script and put it in the nodes folder and used the catbox image for workflows but can only select one of the GPUs or CPU

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --highvram --port 8188 --cuda-device 0

is there any trick?

We've reached levels never thought possible
Localbros this is it

>>101688732
Take out the --cudadevides argument. That worked for me and I could select from either GPU with the node.

>>101688706
anon... look atthe topic again, it was about putting the clip on another gpu so that you can enjoy the model with both gpu, because basically it works like this:

Clip + VAE (t5xxx fp16) = 9.4gb on my 3060
flux-dev (fp8e4) = 11.6 gb on my 3090

>>101688732
>--cuda-device 0
That launch arg overrides CUDA_VISIBLE_DEVICES so torch doesn't see the second card. Remove it and it should work.

>>101688732
I think you should remove the --cuda-device 0 argument to make it work

>>101688718
I guess it comes down to that people are not given agency and responsibility for themself anymore.
It all needs to be dictated from top now.

>>101688718
What you're suggesting is self-destructive and bad for AI. How fucked is your compass that you don't even understand how bad pizza necrophilia is? That the average person would probably physically harm you if you showed them that image in person. No one actually believes in free speech absolutism and you're will probably literally die on that hill.

File: ComfyUI_00273_.png (985 KB, 1024x1024)

985 KB PNG

>>101688566

>>101688791
You're right, the accountability of people who willfully create AI models that do this.

File: ComfyUI_Flux_0787.jpg (88 KB, 1024x1024)

88 KB JPG

i like low quality gens

>>101688810
There's no one else on the bed except her though.

Can somebody post stylization gains in Flux due to Clip/T5xx differentiation of prompts, I think it's snake oil.

>>101688738
>>101688772
give me your gpus and then neck yourself

>>101688431
>now potentially any retard can make it with a simple sentence
who cares? it's still illegal to put fucked up shit on the internet, so that won't change the world, the jannies will still do their work and the police too, we won't see anything like that if we stay on "normal" sites like 4chan and reddit

>>101688828
cfg too high

i want to gen like this in flux

https://x.com/smeretivert/status/1819386982352519673

I have an old 2060, should I dig it out and pair it up with my 4070?

sampling is so slow...

>>101688803
its just a drawing anon.

File: file.png (1.59 MB, 1024x1024)

1.59 MB PNG

>>101688758
>>101688768
>>101688774
ty

>>101688828
if you don't put cfg = 1 then you'll get blurry shit, unfortunately flux doesn't work with cfg, and cfg is needed for negative prompt, that means that flux can't be used with negative prompt

>>101688908
Actually it's not, it's an artificial photograph.

File: 1701655298609518.jpg (996 KB, 1976x2560)

996 KB JPG

https://x.com/toyxyz3/status/1819289127306907848

https://files.catbox.moe/cea241.jpeg

>>101688924
i specifically prompted for low quality/old movie screenshot

>>101688916
you're welcome o/

>>101688644
You can draw swastikas in Germany with krita and it will get you vanned. But nobody will call for a ban of krita, paint or photoshop
You can write "hate speech" in word and it's the same thing
If you gen shit with AI, it's the responsibility of the genner
Simple as
>>101688653
idiot

>>101688962
AI is more like an image search engine. That's where your analogy fails.

>>101688925
Ok nerd. Thanks for that clarification nobody needed lol

File: 1692829129543483.jpg (1.86 MB, 2695x4096)

1.86 MB JPG

https://x.com/toyxyz3/status/1819315271439663365

>>101688927
flux gens look way more realistic, and with a way better anatomy, it's almost at Midjourney's level, that's insane we got our hands on this gem and we can run it, I thought such model would be way too big for our current gpu's, feelsgoodman

>>101688971
Calling it a drawing is a lie and is a form of minimization.

>>101688925
Yes, it's artificial, it doesn't exist, like in the movie John Wick, he doesn't really kill those people because it's artificial, thanks for enlightening us anon, what would we do without your superior intelligence?

File: ComfyUI_temp_ppftb_00165_.png (2.18 MB, 1024x1024)

2.18 MB PNG

>>101688697
Never forget that there is an unusually high concentration of mentally ill people here. They could win the lottery and still find a way to be negative, they are not well.

>>101688970
It's not though
Your machine is spending electrical energy to physically make those images under your direction and supervision
They are encoded in the model somewhere, and you are extracting them

>>101689000
Show a police officer that image and ask if it's okay.

>>101689000
You do know that laws in countries do not care if child porn is real or AI-generated, right? They label it as CSAM either way, even if there was no "abuse".

File: dark souls.png (1.48 MB, 1024x1024)

1.48 MB PNG

>>101688872
I tried
>An 80s dark fantasy depiction of Dark Souls: Picture a gloomy, medieval landscape shrouded in perpetual twilight. The towering, crumbling castles and ancient ruins loom against a blood-red sky, evoking a sense of desolation and foreboding. The armor-clad hero, wielding a massive, jagged sword, stands defiantly amidst skeletal warriors and grotesque, monstrous creatures. The atmosphere is thick with mist and shadows, and the air is electric with an eerie, otherworldly glow. The overall aesthetic is gritty and raw, with bold, saturated colors and dramatic lighting reminiscent of 80s dark fantasy films and heavy metal album covers.

>>101689013
>>101689009
Why artifical pictures/videos/games depicting murder (murder is illegal did you know that???) is allowed then?

>>101689033
Because humanity is retarded. First time? Murder is glorified in tons of movies/games/etc despite being illegal, that's because violence is in human nature.

>stuck on cards with no native FP8, so everything's slow
Crying.

>>101689013
>>101689009
And there it is.
It was never about some stupid ass baby for you.
Muh police. LOL

Keep on truckin'

>>101689050
>>101689050
>>101689050

>>101689033
John Wick isn't even a realistic depiction of murder. I'll give you a clue, society doesn't tolerate cartel murder videos either.

>>101689044
>Murder is glorified in tons of movies/games/etc despite being illegal,
yeah, that's the hypocrisy of humans, they glorify a very vile act (murder) to the point GTA6 is the most anticipated game of all time, and at the same time, they grandstand on other artificial illegal stuff

>>101689063
>John Wick isn't even a realistic depiction of murder.
there's a shit ton of LEGAL movies with some really realistic depiction of murder, don't pretend it's not the case

File: 2024-08-02_00264_.png (1.38 MB, 1280x1280)

1.38 MB PNG

>>101689047

>>101689080
how many feature children?

>>101689101
where's the line anon? murder is also highly illegal, if you murder someone you can easily take 20 years of prison, why do you think it's fine to let people depict and glorify such vile act in the mainstream media?

File: ohno.png (848 KB, 1024x1024)

848 KB PNG

Oh no, Mario has been drinking again...

>>101689119
Society has decided one of the lines is cruelty against children.

>>101689190
>Society has decided one of the lines is cruelty against children
as long as those children are white

>>101689190
think with your own brain retard, don't act like a fucking NPC who believes what the retarded society wants you to believe

File: 2024-08-02_00272_.png (1.94 MB, 1280x1280)

1.94 MB PNG

>>101689205
>brain retard
fried?