Blessed thread of frenship

official pixart bigma and hunyuan finetune waiting room

File: ---3.jpg (178 KB, 1024x1024)

178 KB JPG

>Blessed thread of frenship
>official pixart bigma and hunyuan finetune waiting room

File: Sigma_12386_.png (3.55 MB, 2048x2048)

3.55 MB PNG

Day 69 of posting Sigma gens until Bigma arrives

>>101733259
I look like this

File: Sigma_12355_.png (3.91 MB, 2048x2048)

3.91 MB PNG

>>101733267
It's a look

>>101733243
bigma balls

File: FD_00094_.png (556 KB, 768x768)

556 KB PNG

>>101733287

1girl gens only below this post

>>101733306
I wanna see this animated. Nothing like Joe swiveling around like a roomba after 'great' speech

>>101733297
if you have a saved image you have a saved workflow, unless you intentionally save as jpg

>>101733286
box?

File: ComfyUI_Flux_3407.jpg (207 KB, 1024x1024)

207 KB JPG

are you splitting your sigmas?

https://x.com/GiliBenShahar/status/1819996980023636188

File: FD_00771_.png (1.6 MB, 1024x1536)

1.6 MB PNG

Flux finally makes good mermaids. They have been dogshit for years, with shit like split tails, or tails too short just weird ass looking tails.

File: Sigma_12375_.jpg (1.73 MB, 2048x2048)

1.73 MB JPG

>>101733345
https://files.catbox.moe/ofdf95.png

>>101733352
>want a stronger illustration style with Flux1? Just split sigmas at 1 or 2 and take the lower sigmas
How confusing. Not Sigma sigma, but flux sigma

>>101733352
translation to retard please?

>>101733369
sorry but can i have a catbox for that one too?

I'm a 8gb laptopjeet and can only reasonably generate flux with their service. Off to sdg I go?

File: Sigma_12349_.jpg (2.66 MB, 2048x2048)

2.66 MB JPG

>>101733382
Oh to be loved/trolled. https://files.catbox.moe/0xjnj2.png

>>101733370
sure here is a workflow that does a low sigma split
https://openart.ai/workflows/neuralunk/flux-1-dev-fp8-low-sigma-split-runs-well-on-4060ti-16gb-32gb-sys-ram/tECUhCcFvh4jb7XFW8Jo

>>101733396
you should be fine here as long as you don't post flux pro gens

>>101733418
oh is this just running in fp8 for low vram? I don't really get it

File: ComfyUI_00719_.png (1.31 MB, 1024x1024)

1.31 MB PNG

>>101733331
pic related vs old one >>101732992

This was using 0.3 denoise, i could probably go higher since its not porn

>>101733480
Whys that dude poking his nose?

>>101733480
>>101733489
and only 20 steps, time to increase to 50 steps and increase denoise

File: 2024-08-05_00428_.png (3.2 MB, 1536x1536)

3.2 MB PNG

>>101733101
I correct myself, if you prompt flux with some braincells working you can get almost anything out of it. Pic related.

File: 982675580.jpg (464 KB, 2048x1024)

464 KB JPG

prompt: gay porn
kek

File: FD_00007_.png (1.89 MB, 1024x1536)

1.89 MB PNG

>>101733360
>>101733465
>>101733474
>>101733580
The end. I hope you liked my short story.

File: ComfyUI_00720_.png (1.31 MB, 1024x1024)

1.31 MB PNG

>>101733489
Not anymore anon
0.4 denoise, increased detail, probably I can go higher.

https://files.catbox.moe/ep9dky.png

The workflow is kind of a mess as it was just smashed together in the last 2 hours, but its not that bad. Can easy see what is happening, just need to sort the busses out at the top and what not.

>>101733604
based str8 flux

What's the best resolution to gen with flux? The good old 1024x1024?

File: FD_00009_.png (1.05 MB, 1024x1024)

1.05 MB PNG

>>101733515
>you can get almost anything out of it

>>101733657
It gens up to 2048 fine but I haven't pushed it past that.

File: ComfyUI_01145_.png (1.15 MB, 1024x1024)

1.15 MB PNG

shame it doesnt know what nipples are

File: FD_00014_.png (1.57 MB, 1024x1024)

1.57 MB PNG

>>101733763
It does. It's just filled with ugly nipples. It's better than SD3 who just barbied everyone.
It's much easier to draw the nipples back on with a detailer for flux than SD3.

File: ComfyUI_00722_.png (1.13 MB, 1024x1024)

1.13 MB PNG

0.8 turned it into a cartoon, it needs more guidance in the prompt for photo realistic but its still pretty neat. at 0.6 it was better but some of the faces lacked detail, could prompt convert back for another pass in detector nodes to fix faces and then back into flux on a low denoise setting to get the finished result.

>>101733796
I honestly don't think SD3 is worth talking about anymore. It's a complete failure, completely worthless.

File: FD_00016_.png (1.68 MB, 1024x1024)

1.68 MB PNG

>>101733817
It's the closest point to comparison. They both had similar goals, a "safe" model that uses natural language prompting. SAI shitting the bed doesn't make it less of an apt comparison. Flux is certainly more comparable to SD3 than it is to XL.

>>101733796
>It's much easier to draw the nipples back on with a detailer for flux than SD3.
yep because at least there is something there for the AI to work with, otherwise you'd have to draw crude ones in image editor which is time consuming and bothersome.

>>101733763
--> >>101733604

how do you prevent flux from genning a cartoon image?

File: FD_00021_.png (1.7 MB, 1024x1024)

1.7 MB PNG

>>101733852
Yup. I had some limited success getting nipples onto SD3 barbies. The bigger issue was the complete and utter lack of navels. Like, what the fuck why is deleting navels "safe"?

>>101733880
"photo", "photography", "3d render"

>>101733880
Tell it not to.
"A photograph of ..."

File: ComfyUI_01147_.png (1.35 MB, 1024x1024)

1.35 MB PNG

woah got an almost decent nipple here

https://files.catbox.moe/smg0zg.png

>>101733892
>>101733894
ok thx, I tried 'photo realistic' I guess it has no concept of that tag.

File: FD_00025_.png (1.48 MB, 1024x1024)

1.48 MB PNG

>>101733914
That's not very safe of you, Anon.
I find it really easy to generate nipples, even if they look like they have been dipped in hot oil. Are you actually having a hard time of it?
Just specify "Topless, nude, naked, nipples, breasts exposed"

File: ComfyUI_Flux_Dev_00025_.png (1.22 MB, 1152x896)

1.22 MB PNG

I fart on the heads of all 1girlers.

File: ComfyUI_01149_.png (1.13 MB, 1024x1024)

1.13 MB PNG

>>101733946
its difficult and doesnt give me nipples errytime

File: FD_00028_.png (1.59 MB, 1024x1024)

1.59 MB PNG

>>101733969
Don't you want to kiss her though, Anon? Look at those lips

What were they thinking?

>>101734036
This is pony, isn't it
>>101733992
https://files.catbox.moe/w70kv0.png
What I hate is even mens nipples are censored

>>101734068
yep yep

>>101733892
>>101733894
it got a little better but the damn thing does not want to play, i guess i'll need to change seed. This is why i got bored with flux, the lack of consistency or control.

File: ComfyUI_temp_eqjci_00017_.png (1.13 MB, 1024x1024)

1.13 MB PNG

>>101734142
I dunno man I feel like I have excellent control with it. What are you even prompting?

File: ComfyUI_00724_.png (1.22 MB, 1024x1024)

1.22 MB PNG

That is what I mean, its still cartoon, seed change is definitely genning something closer to a photo though.

>>101734162
I think know why "3d render" would well durr do a 3d render >>101733892
So no the wonder, i've removed that now and lets see...

File: FD_00050_.png (1.26 MB, 1024x1024)

1.26 MB PNG

>>101734142
>>101734224
It's extremely literal, Anon.
>stylized abstract modern artwork featuring bright colours, complex geometric shapes form an abstract nude womans silhouette made of simple shapes, highly detailed, intricate visible brush strokes,

File: 1702825028067440.jpg (571 KB, 1389x2000)

571 KB JPG

Bros i wanna proompt image like this

>>101734252
>A photograph of a busy restaurant in china, Chinese family of 6 people are dining, photo, photography

is still producing cartoons, i have no interest in cartoons, why do people insist on trying so much of this garbage into the model? I have to stare at this god damn shit every day, its gets boring as fuck when you're getting older...

>>101734312
yes i'd also like to gen images like this and not dumb ass cartoons. It really needs a negative prompt for this to filter out what I don't want.

File: 1714473373411099.jpg (90 KB, 800x1170)

90 KB JPG

>>101734312
I got this close with proompting in flux

>>101734344
Maybe it's time to stop being a manchild and start doing something better with your time.

File: ComfyUI_00726_.png (1.28 MB, 1024x1024)

1.28 MB PNG

maybe its over cooking the image i feed into it from pony?

>>101734368
This. Like working.

File: ComfyUI_Flux_56.png (1.12 MB, 1344x768)

1.12 MB PNG

How the FUCK do I get rid of the annoying depth of field? Preferably without the cfg meme workaround just to enable negative prompts because it works like shit and slows generation into oblivion

>>101734404
" Clear background , detailed background " maybe

>>101734404
have you ever used a camera before anon

File: FD_00053_.png (1.28 MB, 1024x1024)

1.28 MB PNG

>>101734312
My attempt
>>101734362
Just stop genning cartoons, I don't understand the problem.

File: ComfyUI_00727_.png (1.32 MB, 1024x1024)

1.32 MB PNG

>>101734456
>Just stop genning cartoons, I don't understand the problem.
I'd because I'm doing something a bit different from you anon, I'm genning with pony only to send into flux to clean it up. pony can't handle so many people at range. Anyway pic related 0.7 denoise

File: RealTest_00024_.png (1.24 MB, 896x1216)

1.24 MB PNG

>>101734362
SDXL does good realism and has negative prompt

File: ComfyUI_00718_.png (1.28 MB, 1024x1024)

1.28 MB PNG

>>101734534
its fucking horrible mate, its shite even with detailer... Its only good for portraits and stuff.

File: FD_00056_.png (1.21 MB, 1024x1024)

1.21 MB PNG

>>101734404
picrel
>>101734431
pretty much, I don't get this complaint, he was moaning about it yesterday too.

>>101734534
Look at how fucked up her eyes are. Flux has spoiled me. Everything look shit now. SHIT

>>101734534
>>101734551
but it does have a lot trained into it, things we would like to exist in flux. So for now I'm trying to figure a way to use it as inpainting on an image genned by pony/SDXL

File: ComfyUI_Flux_57.png (1.06 MB, 1344x768)

1.06 MB PNG

>>101734404
>>101734417
Forgot to add - no amount of "detailed background", "in focus", "detailed" and other phrases work
>>101734431
Yeah, and I even tried adding "f/22 aperture camera" which apparently worked for some SDXL models in the past but alas
>>101734552
It's ogre...

>>101734583
yeah you apparently haven't because you'd know how focus works

File: 1699583048623373.png (1.22 MB, 768x1024)

1.22 MB PNG

>>101734456
This is Ideogram

File: file.png (1.92 MB, 1500x564)

1.92 MB PNG

>>101734600

>>101734620
do you know how far those mountains are way from the camera?
do you know how focus works

>>101734620
Would

File: file.png (1.08 MB, 800x533)

1.08 MB PNG

>bro why isn't the mountains always crisp like the real pictures

>>101734630
do you know how a fucking camera works you dunce? You can see the digital ones to actually focus on a wide shot what ever you call it where detail isn't lost.

>>101734646
that's not how physics works, retard
show me real photography of what you want
because it doesn't exist
you can't have a subject 6 feet from the camera be in focus at the same time as something that is 3000 feet away
what you want requires EDITING

>>101734670
like our eyes right?

again you're fucking dunce.

File: 1708976096990383.jpg (119 KB, 740x1232)

119 KB JPG

G n bros

>>101734679
yes anon, you do know how our eyes work right? your peripheral vision isn't crisp, retard
holy shit i'm actually talking to a retard

>>101734691
Then why do 1980's photos exist with out this problem if they require editing? Its the type of lens you moron, There is nothing wrong with my eyes, i see perfectly clear, i don't have tunnel vision etc. Do you wear glasses by chance?

perhaps you should get your bloody eyes tested.

File: FD_00030_.png (1.23 MB, 1024x1024)

1.23 MB PNG

>>101734719
It doesn't matter what you think you should or shouldn't be able to do, Anon, because you can't do it in Flux. it's simply not possible. Every image with a background has dof. This is just how the model is.
There is 0 point in arguing about it. Accept this as a limitation of the model and work around it. Bashing your head against the wall for no reason isn't productive.

>>101734805
no he's just going to produce three images, start bitching and call it a day
because he's a faggot and a retard

>>101734691
>your peripheral vision isn't crisp
What you're referring to is field of view fov and not depth, obviously you can't see behind you or too far to your sides. But what that other anon simple wants is focus on the big picture and to claim that requires editing is bloody moronic... The problem is dumb asses that can't use a camera properly post their armature photos online and the model gets trained on them. These days its just a setting that needs changed on a good digital camera, on most basic cameras it will capture the whole view focused and sharp, a telescopic camera will require you to focus on the subject, and so on.

File: file.png (1.49 MB, 1280x896)

1.49 MB PNG

the real issue is he's a low IQ promplet, he's incapable of experimentation and problem solving
even f/22 is just him thieving from other people

File: ComfyUI_Flux_Dev_00037_.png (1.6 MB, 1152x896)

1.6 MB PNG

farp

File: FD_00066_.png (725 KB, 1024x1024)

725 KB PNG

>>101734836

>>101734816
https://www.ripeinsurance.co.uk/photography/exposure-triangle/aperture.html

its how our eyes actually work faggot. Does not require editing.

>>101734805
>This is just how the model is.
Literally just answer with this. Why would you go far asking how camera works lol.

>>101734865
my favorite part is you still can't produce real examples

>>101734874
Pretty sure I told you exactly that yesterday. Are you going to come in tomorrow and complain about it again?

understand that the Flux devs deliberately targeted famous female names for scrubbing AND their caption model is shit and couldn't caption anything sexual whatsoever

File: ComfyUI_98.png (981 KB, 1216x832)

981 KB PNG

Thread theme https://youtu.be/UlwsV6d6P1c?t=149

File: FD_00037_.png (1.57 MB, 1024x1024)

1.57 MB PNG

>>101735005
They only deleted the names. The likenesses are still in there. I get them occasionally by accident.
Not sure what your complaint about captioning is when the model is extremely coherent.

ldg bros
Share a 3070 8G workflow for flux pls

>>101735044
Based

File: midjourney face.png (890 KB, 800x736)

890 KB PNG

so they scraped midjourney? that explains the generic greasy aiface and melting plastic skin. the more i scroll through the gallery the more obvious it becomes.

>>101735005
also good chance their caption model is one of those that often describes subjects as "individuals" with zero gendering
see the LAION-POP captions for examples of that
>>101735054
any likeness when using female celeb names comes from CLIP
of course occasionally you get someone that kinda looks like someone else
I'm talking about smut here, it knows little of it, that's the complain about the captioning, existing caption models suck describing adult topics

File: up_0002.jpg (1.03 MB, 3232x5120)

1.03 MB JPG

gm anons

>>101735183
there are little to no celebrity names in there, I have little doubt they used a list to scrub any images matching those names

>>101735235
>Bruce Willis

I return to waiting for bigma

>>101735256
they also did this with a bunch of characters, it's strange, some characters flux doesn't know by name, but if you word salad their wikia description in there they pop up. i guess this will make training some characters back in with loras easier.

File: 1701204551791556.png (289 KB, 2322x1123)

289 KB PNG

Trying to use Flux shnell locally - it takes forever to get going after clicking 'Queue Prompt', fills my entire VRAM before it starts actually generating the image. Still hasn't finished generating 1 image and my PC is choking. Is there some box I need to check here? I have a 3090 so I cannot have more VRAM

Oh, it failed but there's no error logged. What to do I need to do to get it to work correctly?

>>101735619
I think that's a separate problem related to using AI to generate captions. Without going out of your way to ensure the names are in there the caption will either be too generic (ie "anime character", "animated character") or misnamed to a similar looking character. The only way to be accurate is using boorus and raw search titles.

>>101726162
dammit

>>101735692
deep

File: ComfyUI_Flux_3577.jpg (153 KB, 1024x768)

153 KB JPG

>>101735661
the node at the top left
select fp8_e4m3fn as the weight_dtype

File: file.jpg (896 KB, 3564x880)

896 KB JPG

dev 50 steps vs dev 4 steps vs dev + schnell merged 4 steps
https://huggingface.co/sayakpaul/FLUX.1-merged

>>101735734
those efforts are kind of useless because Dev has a non-commercial license, effort should be spent on Schnell only to keep the open source license

>>101735826
schnell too schit to bother training

File: file.png (2.33 MB, 1024x1024)

2.33 MB PNG

>>101735713
>>101735661
I was about to ask the same thing.
Thanks

Takes about 15sec for a 4090 to do one, damn

>>101735826
mix the weights enough and people stop caring. licenses don't matter outside of gigacorps. dont forget the insane amount of shit people sold based on the leaked NAI model.

Damn bros, just found out about flux.
It's not going to be doing anything amazing for nsfw yet is it? Especially anime.

>>101735879
Entirely SFW but doesn't shit itself when posing women at least.
best prompt adherence of any open model but styles/many popular characters/female celebrities can't be prompted for directly, they didn't survive the automatic captioning.

>>101735870
Actually they do matter if you want to make any real money.

File: 1696937799610583.png (1.8 MB, 1024x1024)

1.8 MB PNG

>>101735713
Thank you anon, my hero!

>>101735902
oh so dev license prevents faggots like nai from taking something open and locking it behind a paywall? good, time to keep devving on dev

>>101735918
The dev license prevents ALL commercialization. That also means Civit can't host the model.

File: ComfyUI_Flux_3497.jpg (268 KB, 1024x1024)

268 KB JPG

>>101735912
>>101735857

also make sure to add the flux guidance node after your prompt. most people are keeping the value between 1.5 and 2.5 for better prompt/style adherence

how to prompt belle delphine in flux?

>>101735932
Civitai is already hosting both models lmao, go check

>>101735976
Go ahead and waste your time then and get your cease and desist.

>>101735932
you won't see someone like the pony guy making something out of it, but there will be plenty of people interested in making loras and such.

File: score_9, score_8_up, scor(...).png (847 KB, 1024x1024)

847 KB PNG

>>101735958
Nice for some reason it reminds me of toy soldiers, like using my blankets as a battlefield while putting toy soldiers everywhere, I suppose making gens is partly that in a way

>>101735976
Not for image gen.

>>101735992
Yeah you're going to have some one-off loras, okay but the real finetunes that cost real money won't happen.

>>101735912
That's a big fucking Greenland on the map

File: 1707220740382431.png (140 KB, 1312x902)

140 KB PNG

>>101735958
Sorry I'm not sure what you mean and Googling it didn't bring anything back - I'm not used to Comfy UI (I hate node spaghetti).

>>101736047
Replace your clip text encode node with ClipTextEncodeFlux

>>101736047
https://comfyanonymous.github.io/ComfyUI_examples/flux/

File: file.png (243 KB, 2038x670)

243 KB PNG

>>101736047
maybe this

>>101735734
Tested it in the HF space. Pretty bad compared to Flux dev.

>>101735932
>That also means Civit can't host the model.
That's not why SD3 was banned kek

>>101736095
As expected. It makes more sense to compare it to schnell with the same 4 steps.

8 GB vram bros, whats our status

>>101736032
I want her to tell me "you're an apple of my eye"

>>101736193
waiting

>>101736193
Coping

>>101736193
i mostly used dalle, so for me it's just a wait and see if there is going to be any loras or finetunes. dalle 3 was really fun for 6 months or so, then at some point i had seen 90% of what the model can do and just went back to local

>>101736252
90% of what OpenAI let through. It was restricted from the start and it only got worse.
God I wish someone leaked DALL-E 3

File: 00108-4250189857.png (1.8 MB, 1024x1024)

1.8 MB PNG

>>101736027
That's the mercator projection for you.

>>101736193
heard it can run with 8gb vram + system ram

flux lora training proving successful so far:
https://x.com/ostrisai/status/1820462674230059328

File: file.png (21 KB, 980x179)

21 KB PNG

>>101736193
Honestly I expected worse.
t. loading everything at fp16 on rtx 2080 with 32gb ram.

>>101736335
I saw people with 12gb vram having much slower speeds, how does it even work?

File: ComfyUI_Flux_73.png (1.04 MB, 1344x768)

1.04 MB PNG

>>101736380
No idea. Staring comfyui with --normalvram --disable-xformers --use-pytorch-cross-attention (though it automatically loads flux in lowvram), Memory fallback is turned OFF in nvidia settings, however it still does fill up my RAM completely, which makes sense, otherwise it definitely wouldn't have worked. Also it occasionally offloads shit onto my pagefile, but my m2 ssd is pretty fast.

File: file.jpg (159 KB, 1024x1024)

159 KB JPG

>>101736193

File: ComfyUI_00357_.png (1.53 MB, 1280x720)

1.53 MB PNG

>>101734404
I said this before but the way to do this for Flux is to describe everything in the background. Don't just say background of X and Y, literally describe every single thing in the background autistically, and add "detailed" to each description. Detailed actually does have influence, but "clear" and "in focus" don't, and actually may have a bad influence, don't use those words.
Then if that's not enough, add styles that usually have sharp backgrounds. I like "Drawn in pencil" or "Sketched in pencil". Pic related is actually just "Drawn by Pablo Picasso." in the clip prompt.

>>101734552
I disagree.

>>101736380
the rest of the system specs make a difference, fast nvme, fast cpu, fast ram etc

>>101736335
speed like this is torture if you actually want to gen
believe me, I tried genning XL on a 1060, it's awful

Is clip for flux like in LLMs, aka the lower in the text the more important/impactful ?
And what's the max size of a prompt?

>>101736538
yes, it works like that on dalle, but thanks to no cfg, flux lacks creativity.
just look at dev output when compared to pro:
>>101715965
>>101716030

File: ComfyUI_Flux_75.png (995 KB, 1344x768)

995 KB PNG

>proompt her to say "I'm NOT a vramlet"
>picrel
It's self-aware

>>101734363
No hint of nipples, sad.

>>101736603
kek

we are making big leaps in lora training with flux
https://github.com/bghira/SimpleTuner/pull/622

>>101736619
They should explore Adam mini, it's what I use for training 1.3B and it's very good for memory footprint

>>101736579
>yes, it works like that on dalle
No? Never noticed dall-e 3 favoring later terms in the prompt. It's just like any diffusion model, what is first in the text has more influence.

>>101736651
yeah i misread your comment, thought we were talking about not fully utilizing token limit

https://huggingface.co/sayakpaul/FLUX.1-merged

>>101736331
Nice, it's just on schnell but that's pretty fast

They said it was impossible to train Flux lmao
Necessity breeds innovation

>>101736579
Of course pro would look the best... I wonder what is actually different with dev.
More training time?

>>101736651
Oh I see, thanks.

>>101736670
What the hell, that works?

>>101736539
these are all really nice, anon

>>101736739
Of course, these models aren't magic outside of being really really big matrices.

I don't understand the difference between guidance and cfg

>>101736670
>diffusion_pytorch_model
Does this work with comfyui? New to this.

File: flux Sanna lora.jpg (1.5 MB, 5120x2035)

1.5 MB JPG

>>101736619
>they even scrubbed finnish(ed) politician
fucking why, flux

File: ComfyUI_02026_.png (1.85 MB, 1280x1024)

1.85 MB PNG

>>101736855

hmm , i get 3.8/1it with my 4060 ti 16gb and 64 gb ram @1280x1024 which is 1.3MP

getting a message loading in lowvram mode 13932.075

i guess they have to optimize everything

File: 2024-08-04_00376_.png (2.3 MB, 1280x1280)

2.3 MB PNG

>>101736995
Finnland exists? I don't think they _needed_ to scrub them.

File: file.jpg (2.12 MB, 1792x2240)

2.12 MB JPG

can flux create a texture like picrel?

>>101737332
maybe? try "old print magazine texture"

>>101737146
>i get 3.8/1it with my 4060 ti 16gb
3.8s per iteration? I have the same GPU and get ~2.4s/it

File: ComfyUI_02032_.png (2.2 MB, 1280x1024)

2.2 MB PNG

>>101737407
Same resolution ? i use flux1 dev,t5xxl_fp16,and clip l
i also use euler, simple

File: ComfyUI_02029_.png (2.08 MB, 1280x1024)

2.08 MB PNG

Also, flux doesn't know what a japanese Oni is , but still...

>>101736839
>>101736851
this time it's realistic stalenhaag
unnerving

File: score_9__score_8_up__scor(...).png (1.15 MB, 896x1152)

1.15 MB PNG

>>101736774
thanks, I have been trying different loras and mix matching to see what comes out

And I just noticed the third leg >>101737527, damn

File: 2024-08-05_00586_.png (3.07 MB, 1080x1680)

3.07 MB PNG

>>101737332
not quite what you wanted.. its a fun gen anyhow.. I am not quite sure what the term for that print raster effect is

>>101737564
ive tried dot matrix and halftone effect but no luck

>>101737474
ah, didn't notice the res. at 1280x1024 I get just under 3s/it
same sampler and scheduler
T5 is fp8 but that doesn't change the it/s
I'm loading the model with weight_dtype fp8_e4m3fn

File: ComfyUI_01202_.png (1.6 MB, 1024x1024)

1.6 MB PNG

>>101737590
is it good or bad when the s/it is high?

File: ComfyUI_01186_.png (1.34 MB, 1024x1024)

1.34 MB PNG

>>101737640
meant it/s sorry

currently using SwarmUI with Flux Schnell and i'm wondering why my preview changes so much. I get happy with the preview and then it changes a whole lot. Can i make it more stable?

File: 2024-08-05_00593_.png (3.34 MB, 1080x1680)

3.34 MB PNG

>>101737658
it/s high good
s/it high bad
>>101737582
>>101737332
I got it
>Neon Genesis Evangelion robot Typo-01, red sky. Old print 90s magazine texture, grained. Visible print texture. Print raster, rasterization effect. Visible grainy raster. cmyk dot pattern, four color print raster
this is the important part that made it appear "cmyk dot pattern, four color print raster
"

File: ComfyUI_01206_.png (1.48 MB, 1024x1024)

1.48 MB PNG

>>101737699
>it/s high good
>s/it high bad
wtf does any of that shit mean anyway?

File: 2024-08-05_00597_.png (3.31 MB, 1080x1680)

3.31 MB PNG

>>101737736
>it/s
iterations per second, ergo you want many iterations in one second cause 20 iterations in one second are better than 10 iterations in one second
>s/it
second per iteration, when your it/s drops below 1 it flips to seconds per iteration, so you want that low, cause 20 seconds for one iteration is worse than 10 seconds for one iteration

my my

File: 2002-1930's photo of a ma(...).png (1.46 MB, 1024x1024)

1.46 MB PNG

>>101737697
Okay fuck it, i realized that not using a scheduler in the sampler made it all random. not it's more consistent.

>>101737736
it / s = iterations per second.
in your KSampler settings there's a parameter called 'steps' that tells you how many iterations the gen will take. your it/s determines how many steps per second your hardware can do (and how fast it takes to finish)
when it / s is really low, it switches to s/it (seconds per iteration) meaning each 'step' is taking multiple seconds

File: ComfyUI_01207_.png (1.39 MB, 1024x1024)

1.39 MB PNG

>>101737762
>>101737781
man this shit is confusing af.
so lets say I generate an image with 30 steps (pic related) and it takes in total 44.62 seconds and 00:27 seconds for the steps and it shows 1.11 it/s.
is that a good value or nah?

File: ComfyUI_Flux_3761.jpg (303 KB, 1024x1024)

303 KB JPG

>>101737699
thanks. I've removed the "old 90s print" and it's definitely getting closer

What's the consensus on genning settings in flux?
I'm using euler 25 steps genning at 1024x1536 (no upscaling)
Don't think it requires upscale like SD1.5 right?

>>101737762
>Weird errors on the left margin
Why is that?

>>101737854
if you can reach 1it/s with flux it means you are part of the 1%

>>101737854
it's all relative
you're probably better off than most people in this thread

File: ComfyUI_02045_.png (2.19 MB, 1280x1024)

2.19 MB PNG

>>101737590
I had weight set to default, changing to weight_dtype fp8_e4m3fn lowered it to 2.80-3 s/it

File: qa42.jpg (267 KB, 1792x1024)

267 KB JPG

>>101737857
shame about the training on an Eva likeness

File: 2024-08-05_00587_.png (2.97 MB, 1080x1680)

2.97 MB PNG

>>101737873
thats the 90s print style thing, you are probably to young to have seen that in real live, but many printed pages there were color indexes coded on the side, that were normally cut off when glueing in the pages, when they were cut wrong, or you ripped the pages out you could sometimes still see them

>>101737031
>>101737098
Can it do bukkake? At least some vague shit like "white cream on face" etc

>>101736855
Just get a 5080 Ti 48GB.

File: 2024-08-03_00226_.png (1.98 MB, 1280x1280)

1.98 MB PNG

>>101737943
get creative with white slime .. maybe make it transparent slime

>>101737871
everything up to 1536x1536 seems to work fine with no upscaling
I've found 30 steps is plenty and more than that may just be placebo. Their web interface defaults to 28 so I figured 30 would be a good setting for max quality without waiting too long.

>>101737978
5080 will be 20GB .. NVidia doesn't care about us

>>101738009
No large company does. It's all J3ws trying to milk money out of you, the goyim supporter of our greatest ally Israel.

>>101738009
they are going to make the 5070 12gb again aren't they?

>>101737994
I'll try 30 instead of 25, thanks

File: ComfyUI_00253_.png (768 KB, 1024x1024)

768 KB PNG

The truth is out there

>>101738121
We definitely fleegin!

File: score_9__score_8_up__scor(...).png (1.11 MB, 768x1024)

1.11 MB PNG

>>101738159
>>101738287
My inner child is in love right now.

>>101737893
only 1% here have a 4090?

>downloaded flux
>try a couple of gens
>immediately want to delete it

>>101738174
WE WILL HAVE VRAM

File: ComfyUI_01210_.png (1.62 MB, 1024x1024)

1.62 MB PNG

>>101738336
can you at least explain why?

>>101738336
perhaps a model that accepts 1girl, spread anus, giant stinky asshole will be more your speed

File: ComfyUI_07710_.png (1.4 MB, 1024x1024)

1.4 MB PNG

>>101738345
3mins per gen

Come and get it...
>>101738379
>>101738379
>>101738379

File: 2024-08-05_00599_.png (3.06 MB, 1080x1680)

3.06 MB PNG

>>101738334
I gues here its more, but overall thats about correct. Steam hardware survey says 0.92% of all GPUs are 4090s.

>>101733360
>>101733465
>>101733474
>>101733580
Come back to me when it can do more seamless tails

>>101738334
I am not paying 2k for a GPU, kthxbye

>>101738366
thats not all that long tho.
could be much worse dude

also why not upgrade your PC?

>>101738395
> Steam hardware survey says 0.92% of all GPUs are 4090s.
wtf, I thought it would be way more.

File: ComfyUI_Flux_82.png (920 KB, 1344x768)

920 KB PNG

>>101738336

test

File: ComfyUI_02070_.png (2.18 MB, 1280x1024)

2.18 MB PNG

>>101738361
yes

>>101738416
3090s are $700

>>101738485
nice.

>>101738416
why not?
its worth every cent.

>>101738499
cheapest around here is:
€ 1528,09
As for RTX 4090, it's € 1799, so those went down in price since last time I checked apparently.

>>101738556
you think so? I do not.

>>101738605
yeah I'm having a blast a 4090 is worth every cent. its the best thing you can buy.

File: 832089358-flux1-schnell.jpg (241 KB, 1024x1024)

241 KB JPG

>>101738485
Awesome

>>101738605
You can train shit with a 4090, there's almost infinite value, very cheap for a hobby ultimately.

>>101736604
Nipple is not important here

>>101736579
>but thanks to no cfg, flux lacks creativity.
you can do cfg with this
https://reddit.com/r/StableDiffusion/comments/1ekgiw6/heres_a_hack_to_make_flux_better_at_prompt/

>>101737857
The lattice you see on here is actually what flux does when it's cooked. I have several examples that I can't show you because image limit

>>101736995
Probably most female sounding names.

ever since I fucked around with the CFG shit the generating became a lot slower.
even after I changed it back to the standard workflow.
what this means? restarting also didnt really fix it