Ty baker

>>102000715
>2 of my gens went into collage
feelsgoodman

File: ComfyUI_01911_.png (763 KB, 1024x1024)

763 KB PNG

omg it migu

File: ldg7.jpg (483 KB, 1999x1999)

483 KB JPG

I almost baked. Here's my collage

It's been two months, what have I missed?

>>102000736
go back to /sdg/

>>102000743
Nice collage anon

Okay, who pissed debo off this time?

File: clipboard.jpg (272 KB, 1435x638)

272 KB JPG

>>102000347
As previously mentioned >>102000305
The pic on the left is after the fist run while idling
The one on the right is while processing during the second run

>>102000771
You

double collage we are truly blessed

File: ComfyUI_00979_.png (1.01 MB, 720x1280)

1.01 MB PNG

>>102000743
Well sorry, I thought you died while baking.

>>102000820
I am not the normal baker so don't worry. I am what you call, plan C

>>102000817
kewl

File: 265700399.png (842 KB, 768x1344)

842 KB PNG

>>102000756
No, I don't think I will.

File: ComfyUI_00978_.png (1.01 MB, 720x1280)

1.01 MB PNG

>>102000831
Neither am I, I am all the way down there in the order.

>>102000858
TroonMix?

File: 00025-3953693291.png (3.08 MB, 1280x1920)

3.08 MB PNG

>>102000866
Stop posting CP

I'm not koff doe

i'm here once again to ask. does anyone know the maximum prompt length for flux?

>>102000925
>does anyone know the maximum prompt length for flux?
512 tokens

>>102000925
I think cfg plays a role.

File: ComfyUI_32683_.png (872 KB, 1280x640)

872 KB PNG

Don't cry Migu

>>102000925
I've heard it goes up to 512. Tested it myself and it seemed to include stuff I put near the 512 token mark in my prompt

>>102000689
>>102000775
>torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 314.00 MiB. GPU 0 has a total capacity of 24.00 GiB of which 0 bytes is free. Of the allocated memory 22.09 GiB is allocated by PyTorch, and 1.13 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

down from 1.59GB at least

>>102000950
nigga what, no

>>102000944
>>102000970
thanks anon, here's a (you) as payment

File: ComfyUI_00983_.png (1.12 MB, 720x1280)

1.12 MB PNG

>>102000955

>>102000998
Comfy

>>102000993
If cfg is low, I find it won't even notice the last words. Do you think it does?

>>102000998
It's very cool, but the way flux handles men and women is basically like "fitness coach"/"my thinspiration"

>>102001102
>If cfg is low, I find it won't even notice the last words. Do you think it does?
cfg doesn't raise the token limit, it's more like cfg makes the model understand prompts better, regadless on how long the prompt is

My inpainting is showing lines around my mask. Happens when I use a soft mask, and a hard mask (which makes sense, I'm masking by alpha so I assume the softening isn't doing anything)

How do I fix it and make the blending better? The comfyui examples are really outdated and the workflow is entirely different, so it's hard to go off of that

File: 00051-AYAKON_12481768768.jpg (1.94 MB, 3840x1600)

1.94 MB JPG

flux background on pony character, thought this came out well

>>102001138
makes her look like a giantess

>>102001174
Enjoy our parasites

File: 1709773358020374.png (1.33 MB, 1024x1024)

1.33 MB PNG

oh no way it's mellow mike

>>102001138
flux can do genshin? this looks amazing

File: 00003-flux.png (1.21 MB, 896x1152)

1.21 MB PNG

someone pls share a flux lora config or terminal command that i can use for a character lora in kohya_ss
i tried someone's command here on a previous thread and all i get after seven hours is generic flux faces

File: FD_00152_.png (283 KB, 384x768)

283 KB PNG

>>102001129
Do you think you will ever gen anything else? Or are you pure synthwave pcs now?

>>101999095
So Q8 is overkill since Flux model is fp16?

Yo question, is an intel graphics card good for generating images or am i fucked

>>102001280
Cool gen. Prompt?

>>102001324
I mean, it CAN generate images. It just won't be a good experience and there will be a lot of fuckery.
Currently everything is built for CUDA, so it's NVidia or bust at the moment.

>>102001324
Yes, someone wrote instructions in the ComfyUI readme.
https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#intel-gpus
The resources in the OP also have a few links, although they are a bit outdated. I seem to recall it was like around March or something where we had a few Intel Arc users asking and figuring stuff out. From what I've seen, it's more hassle free once you have everything set up than ROCm on AMD but support is a bit more limited. Which just goes to show how badly AMD is floundering with AI even though they are correcting themselves now.

>>102001324
I do t think uvr ever head anyone on /g/ say they actually bought an Intel GPU vefore

File: bComfyUI_105299_.jpg (381 KB, 768x1024)

381 KB JPG

>>102001349
p simple prompt. forgot which deity the name was but i've just been throwing random ones in then adding by Alphonse Mucha after it. been getting some good stuff.

File: ComfyUI_00993_.png (822 KB, 720x1280)

822 KB PNG

>>102001425

File: ComfyUI_00994_.png (890 KB, 720x1280)

890 KB PNG

>>102001477

>>102001168
not really? It seems to be pretty sensible perspective to me

can I train a LoRA atm on flux? I have 24gb vram.
Links/settings please :3

File: Screenshot 2024-08-21 124909.png (873 KB, 1041x747)

873 KB PNG

Using SD Ultimate Upscale I am getting these horizontal lines on my upscales. Anyone know what causes this, or more importantly how to fix it?

>>102000970
>>102000944
>>102000925
Ive gone 1000 and had it still take stuff from the whole prompt. not sure how it works

>>102000737
nice piftie

>>102001636
stop using shitty upscale methods

>>102001653
Can you make this with her facing the front? Thank you anon

>>102001671
Thanks, very helpful. Do you perhaps have a suggestion for upscaling?

>>102001031
micro bikini lora working then I take it?

File: Intel-Arc-A750-150-US-Jap(...).jpg (851 KB, 1920x2560)

851 KB JPG

>>102001540
In Japan, they were going for really cheap last year. The A750 was selling for 150 USD.

File: bComfyUI_105360_.jpg (283 KB, 768x1024)

283 KB JPG

>>102001674
im a gacha roller bro ask someone who knows how to boomer prompt with flux. i just got back into this shit today after a year break.

>>102001681
ult sd upscale can cause seams at the tiles but you can usually avoid those by raising the mask blur (at least 16) and tile padding values. those horizontal lines shouldnt be there. and ult sd upscale is perfectly fine. show me your workflow/settings. need to see it.

File: ComfyUI_00388_.png (1.45 MB, 1024x1024)

1.45 MB PNG

>>102001692
its meh and it took me like 50 million attempts to get a somewhat okay picture.

>>102001709
Yeah but also PC gaming in Japan is not really a thing

>>102001714
Lovely, I know how to prompt though and can give you tips if you want. If not, what artist did you use for this?

>>102001709
>Pull tape off the box
>It leaves residue or scrape the printing
WHY, just store them behind the counter

File: Screenshot 2024-08-21 125953.png (62 KB, 479x680)

62 KB PNG

>>102001719
Already did.
Workflow. Didn't happen on this image for some reason. I think it's to do with shadows
https://files.catbox.moe/4tgziw.png

File: bComfyUI_105361_.jpg (301 KB, 768x1024)

301 KB JPG

>>102001725
Alphonse Mucha, i'm just going through my old prompts from last year and seeing what flux does with it. i'll fuck around with boomer prompting later.

>>102001775
Lovely! Thanks anon

>>102001735
It looks like stretch wrap, it doesn't damage anythiing.

>>102001775
just feed the prompt into chatgpt and it will boomerfy it for you

File: 00069-87341868.png (2.73 MB, 1920x1080)

2.73 MB PNG

>using forge
>try to use lora
>half the time vram+shared goes to 30+ gb used, gens are insanely slow
>vram usage never comes down
>other times everything is fine
am I being retarded or are loras in flux really janky? It feels like I need to put a lora in my prompt, gen a 512x512 image and hope I get lucky. If vram settles after the first gen then I can do whatever I want and gen images at any resolution

File: Flux-20240821_024722-up-9(...).png (2.27 MB, 1344x1728)

2.27 MB PNG

>>102001721
meh. loras all over the place and I am not a patient man.
>>102001743
>https://files.catbox.moe/4tgziw.png
looks clean, good settings, mightve been an extreme case because of the color OR the lora? maybe remove lora from model path for ult sd upscale. other than that I dont see any issues, aside from your choise of upscaler but thats a diff topic altogether

>>102001814
nah shits totally fucked that was my experience

File: Flux-20240821_024220-up-2(...).png (2.17 MB, 1344x1728)

2.17 MB PNG

>>102001814
"all over the place"; "experimental".
>>102001825
there is a lot going on with that bikini lol

File: ComfyUI_01003_.png (913 KB, 720x1280)

913 KB PNG

>>102001507

>>102001825
uhhhhhhhhhhhhhhhhhh

>>102001816
Yeah could be. I wonder if it will still keep the subject without the lora in the upscale. I will try it.

File: ComfyUI_01004_.png (866 KB, 720x1280)

866 KB PNG

>>102001845

File: 3071729818.png (1.13 MB, 1152x896)

1.13 MB PNG

>>102001841
>>102001850
yeah it's a little excessive, but hey

File: ComfyUI_01925_.png (1.04 MB, 1024x1024)

1.04 MB PNG

Controlnets for flux are such a fucking scam.

>>102001816
>aside from your choise of upscaler
Any recommendations? I have used this one since 1.5

>>102001880
the problem is its schizo and doesnt attach cleanly

File: ComfyUI_32694_.png (1.79 MB, 1024x1024)

1.79 MB PNG

>hmm today I will prompt "vocaloid teto"
>...

File: Flux-20240821_021817-up-1(...).png (2.14 MB, 1344x1728)

2.14 MB PNG

>>102001853
or add a 2nd lora loader at lower strength and run the model pipeline through it *pc combusts*. I split up my workflows, upscaling completely seperate.
>>102001880
well we are burning high end GPUs to gen ass&tiddies while ppl starve to death, excess is the name of the game baby.
>>102001892
ultramix balanced. to quote the maker: "This is a mixture of models based around UltraSharp and my other available models. These are usually interpolations that have separate but very helpful uses." proven to be really good. https://mega.nz/folder/3Jo2AAAa#4CGEwUM0dKu3kkaJa-qUIA
or, if your machine can handle it, try DAT stuff, like this one here https://openmodeldb.info/models/4x-FFHQDAT

>>102001709
>GTX 1630 15k yen

File: 00066-1147785344.jpg (372 KB, 1632x1920)

372 KB JPG

meh civitai rejects anything I try to upload, their filter model is too sensitive

>>102002000
2nd image, celeb, not kosher=rejected, cant even upload any celeb in a pantyhose (and a sweater). no sir.

>>102002001
so whats the trick to not get butt chins?

File: 1724202499583__000001250_0.jpg (58 KB, 768x768)

58 KB JPG

lora training is addictive

File: ComfyUI_00760_.png (1.17 MB, 1024x1024)

1.17 MB PNG

>>102002037
this is the image without my lora applied, so...loras are one way

>>102002039
so how does it actually work?
like lets say you want to train it on a specific person, lets say some celeb or whatever. do you just put a shitload of pictures of the person into the training data? or do you have to cut them out?
how does the computer know you want the person and not some other stuff thats also in the image?
sorry if stupid question

>>102002039
make sure to take some promptbreaks

File: michemashu55-Flux1_e00001(...).png (568 KB, 502x502)

568 KB PNG

>>102000715
Flux LoRA training is super Kino (once you ignore how relatively expensive it is)

>>102002054
there is a LoRa that removes Buttchins?
also sometimes when I use a Lora it still gives me Buttchins.
whats up with this Buttchin shit anyway? what is causing this?

>>102002073
>how relatively expensive it is
wdym? how much it costs?

File: ComfyUI_01936_.png (572 KB, 640x976)

572 KB PNG

Has anyone here had success the with instant x union controlnet? All I get is monstrosities

>>102002080
>there is a LoRa that removes Buttchins?
no, althought I'm sure someone could. I just mean that since my lora was trained on a specific person, it's injecting her likeness and 'overwriting' the buttchin emphasis

File: michemashu55-Flux1_e00001(...).png (569 KB, 502x502)

569 KB PNG

>>102002094
This technically isn't fit in since it's technically not local, but I use Civitai's on site trainer. It costs a minimum of 2,000 "buzz" or $2 before tax, we'll just relatively expensive compared to training a 1.5 or XL Laura which typically only cost around 500 for LoRAs with a similar amount of steps.


If you use the replicate tool kit (Don't use that it injects incorrect metadata into your LoRAs) it can cost as high as $10 depending on how long you train it for.

>>102002080
some of the loras that affect faces obviously can remove it. why its there? training fuckup / we dont know how to prompt it out yet consistently
>>102002098
sick gen tho

File: file.jpg (47 KB, 416x576)

47 KB JPG

>>102002063
I'm experimenting, this is a mixed bag Lora that isn't even one specific concept. I think with the number of parameters that Flux has you can be looser with your bag of concepts. I'm trying to do a pop culture mix with some nsfw concepts.

My current format is:
[found search title] + [florence2] (with filters) + [wdv tags]

Shampoo - Carrie Fisher Photo 34038161 - Fanpop  a black and white photograph of a man and a young girl. The man is wearing a leather jacket and sunglasses, and has long hair. He is holding the girl in his arms and is smiling at the camera. The girl is also smiling and wearing a baseball cap. They are standing in front of a brick building and there is a tree in the background. The image appears to be taken outdoors. 1girl, 1boy, hat, jacket, monochrome, hetero, greyscale, sunglasses, traditional media, parody, ring, realistic, real life insert

>>102001950
Seems to be the LoRA. Removed it and it's gone. I will try it at half LoRA because it did remove some features from the subject.

>>102000715
So I finally decided to make the switch from SDXL to Flux by using forgeUI, which model on civitai or anywhere else is the NSFW king currently?

The buttchins are bad, but those nasty little snub noses get old really quick, too. Flux-chan was fine at first, but the more I see her the less attractive she is. Something like photomaker or IPAdapter would be a godsend while we wait for models that understand facial features.

>>102002154
It is very easy to make them go away, like 1000 steps on a human dataset.

>>102002120
>sick gen tho
For sure, absolute kino, but still nothing like the preprocessor input.

File: vocaloid teto.jpg (3.39 MB, 3072x3072)

3.39 MB JPG

And then I prompted "vocaloid teto" 9 more times

>>102002172
I trained a style LoRA with a bunch of characters, absolutely zero of which had snub noses and butt chins. The butt chins went away, but the noses and cheek shape did not. I'll try targeting them directly with a face LoRA.

File: ComfyUI_01016_.png (1.31 MB, 1280x720)

1.31 MB PNG

>>102002145

File: 1724203799107__000001500_0.jpg (55 KB, 768x768)

55 KB JPG

>>102002200
I'm just surprised someone hasn't just done a basic aesthetics reset and merged a new base checkpoint. Maybe it'll be me. I'm hoping someone figures out full finetunes on 24 GB VRAM.

hell sirs how do I train the loras for flux, I have my data I have my 3090 I need some hand holding

File: 00002-1390996088.png (840 KB, 1024x1024)

840 KB PNG

bill is coming for anime

>>102002239
Adding onto this with a question of my own, how much vram do you need to make a flux lora?

>>102002239
https://github.com/ostris/ai-toolkit

I cannot hold your hand any tighter than what is written in this github. send me (You) if you somehow fuck it up.

>>102002239
https://github.com/ostris/ai-toolkit
read the instructions
dead simple

>>102002270
>>102002272
do I really need this licence shit for a local thing? Did the ostris person add this in intentionally

is there way to always get prompt in your image to appear.

For some reason its always randomly appears whenever flux feels like including it. I wonder if there is a way to get it almost every time to appear

Any tips are appreciated

>>102002282
holy shit you're dumb

>>102002134
kinda thought so. those celeb gens from me here, i used the lora at a 0.5 strength for the upscale, just to keep the facial features. flux is suprisingly good at keeping the style if you dont crank up resample denoise too high. also model sampling flux, i keep those values both at 0.5 for the upscale to not deviate too much from the original, but again, seperate workflow etc

>>102002270
>>102002272
Is there any way to train locally with flux dev without a HF access token?

>>102002301
you download the model and link to the local files, you'll want to do that anyways because huggingface cache is absolute ass cancer

>>102002282
Bruh. Just make a dummy account if it rubs you the wrong way. All it's doing is saying to HF "Yo, gimme the weights"

>>102002270
>You currently need a GPU with at least 24GB of VRAM to train FLUX.1
I thought it was possible with 12 gb? And it took over 24 hours to train

>>102002343
go use kohya then and their not as simple trainer

>>102002343
>I thought it was possible with 12 gb
It's possible to train it on any Turing capable device, it doesn't mean it's practical or supported.

File: 00006-914023985.png (968 KB, 1024x1024)

968 KB PNG

I lived long enough to fulfill my true otp.

Now this is art.

I'm going to add onto the pile of flux lora questions. Has anyone attempted a style lora and if so, what are you doing with your dataset / captioning? Feel like I'm doing something wrong either with my dataset or parameter settings

File: ComfyUI_01026_.png (722 KB, 1280x720)

722 KB PNG

>>102002401
NTR Will have a whole new meaning
>Nigga Tyron Rape

>>102002420
yes. first try does literally nothing, second try does figuratively nothing. i have nothing useful to contribute.

>>101998125
different 12gb lora anon here and kohya recommends updating pytorch and just using sdpa instead, so I do that. considering flux training is an overnight affair it's not really worth the slight speed boost of xformers vs sdpa anyway imo

>>102002420
The LoRA training mafia will kill me for telling you, this, but if you're just training for a style, you don't need to caption the dataset at all.

>>102002450
we kill you for it because if you don't caption it looks like shit. try one with and one without captioning, the difference is so drastic that the no caption option shouldn't be considered an option

>>102002469
>the difference is so drastic that the no caption option shouldn't be considered an option

I've yet to be convinced otherwise. More often than not, your shitty captions just taint the data further.

File: 00073-zavychromaxl_v80-20(...).jpg (731 KB, 2304x1664)

731 KB JPG

>>102002458
>>102002464
thats some good shit.

File: ComfyUI_00989_.png (1.16 MB, 720x1280)

1.16 MB PNG

So, any help with this? >>102000789
Or suggestions?

>>102002469
what if you train a lora only to be applied thru model not clip

>>102002420
>>102002450
>>102002469
There some interesting info on Flux captioning here
https://civitai.com/articles/6792/flux-captioning-differences-training-diary
For his style lora he found WD14 captioning the best

File: ComfyUI_01032_.png (774 KB, 1280x720)

774 KB PNG

>>102002523

File: ComfyUI_01033_.png (722 KB, 1280x720)

722 KB PNG

>>102002534

>>102002420
I use local joycaption with a chatgpt edited script to queue captioning of every image in a folder then spit out individual .txt files. then I review them and fix any outstanding errors. just finishing off the last few epochs of a mix of that + booru tags over a 500 image dataset

next test will be trying the --enable_wildcard thing and doing both joycaption and booru tags

If you had to pick one for flux I'd just do joycaption. if you want to be really autistic, my other plan is having chatgpt edit joycaption so that it automatically reference the datasets booru tags to help create the boomer prompt. would probably help tagging accuracy, but requires you pretag your dataset with booru tags

>>102002518
>im at the office and everyone here is laughing at this thread

>>102000715
what's a good laptop with cuda? I want big power laptop

File: ComfyUI_01034_.png (758 KB, 1280x720)

758 KB PNG

>>102002555
>Checked and keked

>>102002389
you know what I meant nigger, I'm talking about training within a couple days on 12gb, not a couple millennia on an actual turing machine

>>102002571
For ai no laptop will have enough vram to be worth genning on. Just use --listen on a PC and gen from your laptop that way. Otherwise forget it.

File: 00006-11879257.png (1012 KB, 1024x1024)

1012 KB PNG

>>102002423
the future is looking bright, bros. The new age of memes is arising.

>>102002590
i dont have a pc and dont like them bc i cant carry it to school :/

File: ComfyUI_01036_.png (675 KB, 1280x720)

675 KB PNG

>>102002591
>saved

>>102002583
nta but if you keep your dataset under 100 images you can train at 1024*1024 to 30 epoch in around 12 hrs with Adam or lion and cosine. 4-8 at 512*512
- t. 12gb vramlet

>>102002450
>>102002469
>caption
>don't caption
I guess I'll give it an attempt but that does sound like a bad way of doing things
>>102002524
>WD14 caption
Also haven't given this a fair try
>>102002552
That's currently what I'm doing (minus chatgpt) and that doesnt seem to be working

File: ComfyUI_01019_.png (969 KB, 1280x720)

969 KB PNG

>>102002605
>school
Right, it's summer

>>102002605
then buy an online service or rent, they don't make laptop GPUs with the amount of VRAM you want to comfortably gen flux with. you might be capable of genning with one that costs upwards of $3-5k slowly, but at that cost you could have two 4090 24gbs in a PC.

File: ComfyUI_01037_.png (594 KB, 1280x720)

594 KB PNG

>>102002633
From blacked to greended

File: parrish.png (865 KB, 768x1024)

865 KB PNG

>>102002420
I made a maxfield parrish lora with fal, 35 images captioned with florence. seems to work all right.

>>102002627
>That's currently what I'm doing (minus chatgpt) and that doesnt seem to be working
what training settings are you using, dataset size vs amount of steps?

>>102002653
the last pill is the greenpill. natural progression for the anime girl.

>>102002696
>from tiny blurry mobile thumbnail I think this is a realistic Viera
>cool!
>expand image
>nightmare fuel
I feel both cheated and cursed

>>102002660
Someone linked https://github.com/bmaltais/kohya_ss/issues/2701#issuecomment-2297761417 so I started there
50 images, 5 repeats, 15 epochs
adamW8bit
cosine with restarts
LR 0.0004

File: MarkuryFLUX_00059_.png (1.42 MB, 832x1216)

1.42 MB PNG

Rumor says that DALLE 4 is set to be announced in several weeks.

File: ComfyUI_01041_.png (840 KB, 1280x720)

840 KB PNG

>>102002699
The rabbit hole goes deep
>>102002712
This whole AI is stuff is nothing but nightmare fuel to begin within

>>102002744
when's the porntune

>>102002718
with 5 repeats you definitely want a higher LR than that otherwise you probably won't see decent results until epoch 23-30. I'm not at my PC right now but if you're still around in a hour or two I'll drop the training settings I like and you can try them if you want

>>102002744
DE-4 will be disappointing, I guarantee it. Especially with the election coming, they'll be terrified of Trump memes.

>>102002804
Well it might be announced, doesn't mean it will be available before the election. Just like voice, sora, and the full GPT-4o model.

>>102002771
Thanks for the tip, I'll hover around and fiddle with some things

File: ComfyUI_01044_.png (1.01 MB, 1280x720)

1.01 MB PNG

>>102002833
>No to fighting!
>Yes to LOVE!

>>102002605
My laptop is the biggest piece of shit ever. It has less RAM than my phone. But it still gens on a 3090 because I can remote into my real PC from anywhere. In fact, I'm genning on it right now.

>>102002863
oh no T.T

>>102002844
good luck anon. that thread you linked had someone saying 8bit training felt "undercooked" at that low LR as well, so definitely try increasing LR

File: ComfyUI_01049_.png (925 KB, 1280x720)

925 KB PNG

>>102002870
Nooooo! They could have made love instead!
Why did it have to come to this!

File: ComfyUI_09036_.png (1.3 MB, 1400x800)

1.3 MB PNG

Flux base model, no lora can do a lot of video games. Pretty cool

File: file.png (455 KB, 2337x1057)

455 KB PNG

daily reminder no one here can add furniture to an empty room like this site: https://www.virtualstagingai.app/

this trash didn't work btw lol

File: 00062-crystalClearXLPrime(...).jpg (1000 KB, 2304x1664)

1000 KB JPG

>>102002934
MUSCLE CARS BRO

>>102002984
Not again... please leave

>>102002984
daily reminder that you need to get lost. "Discussion of free and open source text-to-image models". do you see "adding furniture to empty rooms" in there, somewhere?

>>102003003
>>102002991
Don't acknowledge him. Don't feed him. Don't reply to him. Report and ignore.

>>102002991
I just got home lol

>>102003003
honestly can't tell if troll or actually retarded

>>102003010
lmao

File: bill gets back with miku.png (903 KB, 1024x896)

903 KB PNG

>>102002909
>>102002856
in the end love conquers all.

File: 157e0028.webm (2.39 MB, 1360x752)

2.39 MB WEBM

>>102002893
>>102002934
>>102002985
Also, Luma 1.5 dropped today. Did a test between these two images and it actually turned out coherent.

>>102002984
I don't like the obsession but I wish people posted more screenshots of their workflow, it's actually very cool to look at and it makes me wanna fiddle with my own workflows a bit more

>>102003028
I'm objectively making these threads better by pointing out how many retards that don't know shit live in here 24/7, that's a FACT

File: ComfyUI_01050_.png (930 KB, 1280x720)

930 KB PNG

>>102003017
Not with this green bitch, I meant love with that other lovely handsome green lad, why did he have to die for her!

>>102002984
kek, i knew right away but i didn't want spoil it so i knew you'd comeback, you got trolled or you can't even setup a workflow like that lol

What does Comfy dequant GGUF to? Does it do it to FP16?

>Heh yeah, I train on 12gb of vram
>Sure, It's rank 1
>Sure I set the resolution to 256
>But it trains
>Heh

>>102002893
>>102002934
badass af

>>102003043
miku got tired of shrek. she knew that shrek is dreck. this is why he died. shrek is truly love, he offered his life for her love.

>>102003050
bitch plz, none here has the IQ to troll anyone
the retard that posted that actually thought that was useful just because of the florence segmentation
which is often the case with comfy users who clearly have zero experience using graphic software and can't even do basic photoshop tasks
anyways, as a I said yesterday about a 100 times, the inpainting was the challenge, not the segmentation

File: ComfyUI_01058_.png (865 KB, 1280x720)

865 KB PNG

>Please consider the following

>>102003101
>Shrek is dreck
Shut the fuck up your stupid motherfucker. Do you have any idea what you're saying? SHREK IS NOT FUCKING DRECK.

>>102003081
I train at 1024*1024 with dim 8-16 and it only takes up roughly 9gb of VRAM and is done training by the time I wake up the next day. could probably leave it running a bit into the morning and do 32dim if I needed to. weak bait

File: ComfyUI_01057_.png (840 KB, 1280x720)

840 KB PNG

>>102003101
Shrek is Wrek, because he will wreck the shit out of that bitch

>>102003146
>no gen
I'm sure your loras exist and look great lmao

>>102003169
turing machine anon is bitter

>>102003169
I'm not at my PC but I'll post it in 30 or so once I am, sure. this trolling doesn't even make sense

>>102003195
>phone poster
I'm sure you train all the time

File: ComfyUI_01065_.png (875 KB, 1280x720)

875 KB PNG

>>102003131
No more love, that bitch has to die

>>102003195
>I'll post it in 30 or so
That 30 was an ETA on his gen after he hit generate.

>>102002632
college

File: Flux-20240821_042509-up-9(...).png (2.51 MB, 1344x1728)

2.51 MB PNG

>>102003196
is your lora on civ? hand it over son

File: ComfyUI_01067_.png (909 KB, 1280x720)

909 KB PNG

>>102003231
Eww it has a dick

File: ComfyUI_temp_ghpfc_00001_(...).jpg (2.39 MB, 3768x2512)

2.39 MB JPG

Flux upscale of a MJ starter image, base with no lora

Any guesses for how much it would cost to create a Flux fine tune on Runpod?

File: ComfyUI_temp_ghpfc_00002_(...).jpg (2.54 MB, 3768x2512)

2.54 MB JPG

>>102003280
now again with the Eldritch oil paint lora for Flux
I really really like this lora, basically solves the slopped flux art style for me

>>102003200
you never leave the house?

So I feel like 32GB RAM is no longer enough. I keep maxing it out and causing my gens to go shit slow.

File: ComfyUI_01069_.png (836 KB, 1280x720)

836 KB PNG

>Ivan listen
>if gun is hold like this and fire
>gun will fire both bullet and gun
>enemy confused won't know how to dodge

>>102003289
depends on the size, easily over 10k if a poony sized dataset

>>102003306
I definitely don't use 4chan when I leave my house

>>102003330
must not do it very often then. lots of lulls in day to day life where you got nothing better to do than look at your phone. kinda weird you seem to think trolling via the PC is somehow a superior expenditure of your time desu

>>102003343
I can multitask on my computer. You're out in public being fat, ugly and on 4chan.

>>102003245
its a girl, i checked. but you are right. nice rifle! excellent trigger discipline, too. but how does she shot without the lower part of her finger

>>102003351
>excellent trigger discipline
her finger is literally on the trigger

File: file.png (2.31 MB, 1024x1536)

2.31 MB PNG

kiinda late to the party but forge can run flux, I got this migu in 14mins with my 1070ti. Although without using the hiresfix its almost 3min for a 512x768
using fux dev q5_1 gguf, randomly downloaded one i havent tried the rest

>>102003280
looks pgood, mind if I ask how you upscaled?
>>102003362
look again, lower index finger segment missing

>>102003362
and yet she hasn't shot anyone yet, very disciplined

File: ComfyUI_00776_.png (988 KB, 1024x1024)

988 KB PNG

>>102003231
>is your lora on civ?
naw, I only just started training loras, so just messin' around and trying different settings

File: ComfyUI_01070_.png (868 KB, 1280x720)

868 KB PNG

>>102003351
>>102003362
Trigger safety is for pussies, here we shoot at everything that moves, even at yourself sometimes

File: liter flux dev.jpg (234 KB, 1547x1024)

234 KB JPG

Holy shit bros... I did it.... I trained a Flux LoRa to produce near carbon copies yet response to plain English prompting in flux. Trained twice cause originally I trained batch 4x on PonyXL. Obviously I can't do more than 2x batch on Flux Dev due to VRAM requirement being gigantic. Figured I need to run 4x to achieve the same thing. I notice that Flux did not overcook. If anything, it is undercooking. LoRa strength 1 does absolutely almost nothing. Now strength 1 has an effect after running the config twice. Now just need to figure out the right config to remove the need to run 20 hrs or more.

Config
https://github.com/bmaltais/kohya_ss/issues/2701#issuecomment-2294611159

Additional notes:
>RTX 4090. %50 power limit
>Ran 2x with the same training config.
8 hrs + 12 hrs ( Around 200~ images, added more images and continued)
>Used danbooru style tagging
>Around $4 USD in power use.

tl:dr, Fuck it, just train multiple times bro.

>>102003405
Would have been cheaper and faster to buy buzz and train it on Civit.

>>102003387
ESRGAN upscale of your starter image to desired resolution
then feed the upscaled image into an "Ultimate SD Upscale (No Upscale)" node connected to Flux model, 0.25-0.35 denoise, euler 6 steps, tile size approx. 1 megapixel

>>102003383
anon, nice miqu but its time to get a better card

File: Flux-20240821_032201-up-3(...).png (2.31 MB, 1344x1728)

2.31 MB PNG

>>102003401
thats the spirit.
>>102003383
I admire your dedication. seriously. anything above 30 seconds / gen and I'm out.
>>102003405
very nice.

File: ComfyUI_09088_.png (1.55 MB, 1400x800)

1.55 MB PNG

>>102003383
>1070ti
poor soul

>>102003450
Why are women like this?

File: ComfyUI_01075_.png (1.33 MB, 1280x720)

1.33 MB PNG

>>102003456
Like what?

File: Flux-20240820_004119-up-4(...).png (2.45 MB, 2112x1056)

2.45 MB PNG

>>102003425
thank you, appreciated. only 6 steps? interesting. flux does tremendous resampling and inpainting. 0.6 denoise, voila, there is your new hand.

File: liter flux dev 2.jpg (368 KB, 1842x1163)

368 KB JPG

>>102003423

Sir, this is a local diffusion general. I am going to eat good cause I got my own configs and machine.

>update comfy to try the gguf t5
>loras now load as weird static noise, even ones that worked yesterday night
I'm going to kill myself

>>102003489
low steps because low denoise
that's the rule for img2img, (both for Flux and for stable diffusion)

when doing img2img you calculate steps based on the denoise percentage, so if your model's normal total steps would be 20, then for 0.25 denoise you reduce it to 5 steps (20 * 0.25)

woe, he pulled!

>>102003532
who pulled what?

File: ComfyUI_01080_.png (1.03 MB, 1280x720)

1.03 MB PNG

>What did you say anon?
>My gun looks unrealistic?
>What a pleb
>https://x.com/ClownWorld_/status/1825502502370058307

File: ComfyUI_09111_.png (1.49 MB, 1400x800)

1.49 MB PNG

Wonder if anything new is going to come out of flux like controlnet came out of SD.

File: liter flux dev 3.jpg (296 KB, 2404x1199)

296 KB JPG

>>102003405

Flux obeys English prompts very very well.

Delivery just arrived and it's some fresh bread...
>>102003576
>>102003576
>>102003576

>>102000715
Me in the bottom left

>>102001640
when it's over the limit it starts to compress, do you think it followed everything from the 1000 tokens? I'm not sure of that