Is this the blessed thread

Blessed thread of frenship

File: 1713659686552752.png (909 KB, 1024x1024)

909 KB PNG

A TV with an 8-bit Nintendo game starring Miku Hatsune. The graphics are pixel art and the game is like Super Mario Bros.

can I train a lora for flux locally on a 4070 Ti Super? Currently buying buzz on civitai but don't mind waiting overnight or etc for training

File: FD_00030_.png (2.05 MB, 1024x1536)

2.05 MB PNG

Any way to set up wildcards in comfy?

>no longer ooming with the gguf text encoder
I am blessed

>>102009841
what quant did you choose? and do you notice a big downgrade compared to the fp16?

>>102009841
can it load the smooth GmP ViT-L now?

>>102009803
>wildcards
>comfy
Kek

>>102009841
Q8 with Q8. It is mildly worse than fp16 but not enough to make me feel like I need to use it over the quants.
>>102009862
No.

>>102009888
>No.
city96 is such a codelet

>>102009888
meant for you >>102009861

File: ComfyUI_00795_.png (1.81 MB, 1152x1536)

1.81 MB PNG

>>102009737
correction: I meant 10 repeats on 20 images, 15 repeats on 30 images. I did 6-10 epochs, with 3 epochs working fine and 6-10 working really well. 1024x1024 res
I used a 3e4 learning rate and 2.5 guidance per this thread discussing settings: https://github.com/bmaltais/kohya_ss/issues/2701
I'm still really new to training loras, so just trying things out

>>102009802
https://github.com/kohya-ss/sd-scripts/tree/sd3?tab=readme-ov-file#flux1-lora-training

File: 1721226180365275.png (88 KB, 719x667)

88 KB PNG

>>102009803
i did an autistic solution where i chained a bunch of text prompt nodes filled with {random | statements | and | prompts} that i reactivate and activate depending on my needs.

What vramlet haxx to increase gen speed? I have dual GPUs but offloading clip to one of them doesn't do anything for speed

>>102009803
I use a node called Impactwildcardprocessor.
It processes wildcards on files __example__ or

{yellow|red|black|white|green|blue|purple}[/icode]

>full female anatomy
>pussies still look like what an 8 year old imagines pussies look like
WHEN?

>>102010046
>t. 8 year old

is it just one guy spamming Miku at this point wtf

>>102010060
Yes, I said that because that's how I imagined them when I was 8.
Fuck it, I'm downloading Pony Realism again.

>>102009913
How do profile shots and angles from above/below look with a lora?
Does it mess up weird angles?

>>102010069
I think you have no idea how popular Miku is, on /lmg/ she's literally the official mascot, and it doesn't help that flux seems to know her better than the rest

Requesting Miku giving birth to a black or Indian baby

>>102010085
requesting you take your meds

File: file.png (2.23 MB, 1024x1024)

2.23 MB PNG

>>102010085
I redeemed. Sorry.

>>102010098
You hate babies?

>>102010111
Interracial is disgusting

Flux perfect feet when?
I hate the feet pony does. It's not slender and delicate.

>>102010123
Nazi cuck

>>102010085
Go back to /b/, indian

>>102010167
Stop replying to me

>>102010173
Grow tf up cuck

this is not friendship

File: file.png (2.66 MB, 1024x1024)

2.66 MB PNG

>>102010173
OK sir

>>102010187
Could you be insecure somewhere else?

Sorry
>>102010218
was for
>>102010173

Miku is Asian

These miku gens are getting old like the Watson arc

>>102010251
The ones of her telling a chinese man to not take his meds and kill himself were funny desu

File: image.png (1.08 MB, 1024x1024)

1.08 MB PNG

https://files.catbox.moe/k2l750.png

>>102010268
I didnt ask

>>102010286
But I wanted to tell you

Requesting miku giving birth to a Easter Island head

from my own experiments and seeing other people's results, flux works best at resolutions approaching 2048x2048 at 30 steps. throw in a >1 cfg and I'm looking at 5 minutes for each gen on my 3060 12gb

I know for sure that img2img works better at the higher resolutions

>>102010364
I dotn really care

>>102010444
kek'ed

>>102010420
>from my own experiments and seeing other people's results, flux works best at resolutions approaching 2048x2048 at 30 steps.
I noticed some loss in prompt adherance when I went to higher resolutions than the default one (1024x1024) but I did some really limited testing so maybe I was just unlucky with the seed

File: image.png (1.2 MB, 1024x1024)

1.2 MB PNG

>>102010453

Is flux it?

File: 2024-08-21_00519_.png (881 KB, 1024x1024)

881 KB PNG

>>102010236
wrong she is a fucking vocaloid

>>102010477
1024x1024 isn't really a default, flux natively supports a large range of resolutions from .1mp to 2mp. That being said, 2048x2048 is probably too high though.

How do I prompt in ComfyUI with a randomly chosen string without needing a dictionary for the options. I want something like {red|blue|green}, but I don't remember how it's done.

>>102010600
I got the combinatorial prompts node, but it outputs a string, and nothing takes a string as an input

>>102010618
Nevermind, I didn't know about right-click > convert input

>>102010600
>{red|blue|green}
that is the syntax for wildcards in comfy
>>102010618
right click the text conditioning node and change the text widget into a text input, then you can connect a string output to it

File: ComfyUI_temp_ostyt_00057_.png (1.32 MB, 1024x1024)

1.32 MB PNG

>>102010600
>>102010022

>>102010515
bigma soon

AdamW8bit or Adafactor?

>>102010600
Wildcards work just like you demonstrated. Use that in your prompt.
If you have too many of them it'll start to ignore those but try it out.

Requesting miku giving birth to a anti racist baby

>>102010646
oh no why is she looking at me like that? why is she just standing there? FEAR BONER GO AWAY
>please catbox

>>102010703
Sorry, genning 1girl (legal). Go back to your country

>>102010703
>>>/h/hdg

or better

>>>/d/ddg

>>102010729
Childbirth is not sexual

>>102010715
>boner
kill yourself

>>102010752
>he's never gotten an erection from fear
pfft not my fault you got broke dick mate. More for me i guess.

>>102010778
more what, anon?

File: file.png (2.12 MB, 1024x1024)

2.12 MB PNG

>>102010703
Here you go.

>>102010665
kohya seems to recommend adamw8 for 24gb and adafactor for 16-12gb

>>102010785
more

>>102010813
Racist

File: file.png (1.88 MB, 1024x1024)

1.88 MB PNG

This one came out pretty good

>>102010715
https://files.catbox.moe/gjhqer.png
I removed the deactivated nodes, it's a different picture but the workflow is the same.

File: file.png (1.99 MB, 1024x1024)

1.99 MB PNG

>>102010835
This thread is not for requests. But if you want tips on how to gen a particular image, by all means, ask. But we're not your particular elephant in diapers fetish providers, so fuck off with that.

>>102010869
pretty much the same kind of image, thank you very much.

>>102010813
>>102010867
>>102010895
DO NOT REDEEM THE MIKU SCHIZOPOSTER REQUESTS BLOODY BASTARD

File: 2024-08-21_00527_.png (1.29 MB, 1024x1280)

1.29 MB PNG

I gonna say it.. anime in FLUX is absolut slop, especially hands and feet. Realsim and abstract is excellent, but their anime dataset is total shit but for Miku and Sailor Moon pictures

>décolletage
joycaption thinks I went to college

>>102010907
Requesting obese miku breastfeeding a pajeet

File: file.png (2.72 MB, 1024x1024)

2.72 MB PNG

KSampler previews stopped working despite using --preview-method

>>102010974
update

File: 2024-08-21_00529_.png (1.43 MB, 1024x1280)

1.43 MB PNG

the best way I found to to make anime work was not mention anime at all but instead use loli as style keyword .. fuck you BSL

schnell or normal for 16gb vram?

File: clueless.gif (1 KB, 128x128)

1 KB GIF

Havent run Comfy in a few days, time to update everything!

>>102011009
doesnt matter they have the same vram requirements (atleast in the base version)

File: how to learn AI.jpg (580 KB, 3840x2160)

580 KB JPG

>"I think you need to join twitter, X, the whole AI industry is on X, and they're all like anime avatars"

>>102011009
Schnell is total slop. Use Q4_0. Force CLIP and VAE to CPU.

>>102011009
gguf flux1-dev-Q8_0 imo

>>102011009
this probably >>102011065

I guess I am the only idiot still running the base version on my 4090

>>102011044
>Force CLIP and VAE to CPU.
interesting, is this somewhere in comfy settings or do i need to edit an ini somewhere?

File: file.png (503 KB, 388x1140)

503 KB PNG

>>102010869
I grabbed this workflow and tested something. The bottom image uses both prompts (CLIP and T5). The middle one uses only T5. The top one uses only CLIP. Although it does change the image slightly, I don't see any apparent reason to use CLIP as well. Is there a special reason why you include it?

>>102011037
Literal who

>>102011139
>george costanza pajeet threatening miku for redeeming

>>102011107
Use this. Put it in ComfyUI/custom_nodes
https://gist.github.com/Sunderbraze/d0b0f942256965b40f54247344fea37f

File: hhhhhhhhhhhmmmmmmmmmm.png (299 KB, 1738x827)

299 KB PNG

I get the feeling this isn't normal..
>3 minutes of that was the model loading before it actually began genning
>Q4

>>102011158
thank you i'll try it out

>>102011037
i think he needs a nap

>>102011145
he's basically just making money off the back of open source. you know the plan - create a website, set it up with stable diffusion, profit, etc.
there's lots of these literal whos all over twitter.

>>102011194
we all do.

Requesting Indian scammer stealing miku breast milk from a fridge

File: image.png (2.02 MB, 1024x1024)

2.02 MB PNG

cute but too artifacted

File: 1696623651320728.png (1.11 MB, 1024x1024)

1.11 MB PNG

got a double miku, lucky

>>102011139
lmao, nice one

File: file.png (2.2 MB, 1024x1024)

2.2 MB PNG

Use classic pasta as prompt. But image is meh.

>>102011108
Yes, there is. I usually set this workflow with several passes of SDXL to add style, and SDXL doesn't like T5 type prompts. I deleted the deactivated nodes, not to confuse you.

File: file.png (2.45 MB, 1024x1024)

2.45 MB PNG

>Love this, thanks for sharing... I have an RTX 2070 (8gb) and P40 (24gb) Not recognizing my second card cuda:1 ... cuda1 not in list. I found another update for ComfyUI and nodes... Going through updating everything now to see if it shows up... Had issues with Flux options showing up when not updated fully. But I just updated like 12 hours ago. Update: It's still not seeing both my cards... I realized I set the command line option to pick cuda 1 so it was not seeing the other card. once I removed that option they both showed up. It was weird because real world card on cuda 1 was showing as only option in multi GPU workflow as cuda0... Since my first card is only 8gb, I can't run fp16 for either option as it crashes instead of sending what does not fit to cpu memory.. So not very useful unless I can get by with using t5xxl-fp8.. But I think the text is not as good with fp8.. Another update: It seems this workflow crashes on me no matter what models I load. There is no error message, it just crashes to command prompt. Anybody else have this problem or know how to fix? Also where in the menu are the nodes listed? I can't find them. lol edit: I found them, they are under ExtraModels/Other in my setup but this is not the default location. Only there because my setup has some extension installed that changes it.
What the hell did this prompt do to Flux?

>>102010984
Did not work

File: ComfyUI_Flux_10011.jpg (126 KB, 768x1344)

126 KB JPG

>>102011314
it gave me this

>>102011342
>he pulled

File: ComfyUI_Flux_10013.jpg (73 KB, 768x1344)

73 KB JPG

>>102011314
>>102011368
and this

it seems retardedly easy to train looks with barely any effort. just throw in 20 pictures and you get perfect faces every time

>>102009913
But several anons for a whole thread were making fun of one anon that said you could train at 12GB vram, are you telling me that anon was right?......

>>102011394
https://www.finetuners.ai/post/training-lora-on-flux-best-practices-settings

Some retard was saying you can train in any resolution you want just because the model was trained with multiple resolutions. But you should use 1024 for LoRA and stay consistent.

File: 1711824191189483.png (1.01 MB, 1024x1024)

1.01 MB PNG

one more cereal box

>>102010576
If I have images that are way higher res can it auto train that at lower res, or do I have to manurally reduce it myself? Babe

>>102011414
i really need to get kohya sorted out with that sd3 branch... but the other half of me just want to wait untill the main branch gets updated

I'm afraid of getting into LoRA training because I tend to get obsessed with things, and being able to generate images of anybody is... It ticks my schizo compulsions.

>>102011414
>Good resolution: The minimum is 1024 x 1024.
>Correct ratio: For training on Flux, a 1:1 ratio is required. Crop your images accordingly and place the subject in the center
This is false and retarded and what causes loras to be less versatile than they should and apply the same to every subject in the image.
These are best practices for 1girl loras, not proper loras.

>>102011463
this man is spitting facts right there

>>102011463
Yeah I saw that and stopped reading

>>102011463
Those people do this for a living. Can you link to the LoRAs you trained so that we can judge your results?

>>102011463
*apply the same face to every subject

>>102011414
if you're referencing >>102010576 then you're retarded because they were talking about genning speeds/res, not training.

>>102011512
The article its obviously about training it to recognize specific individuals.

File: file.png (933 KB, 1536x1024)

933 KB PNG

prompt:
>This is false and retarded and what causes loras to be less versatile than they should and apply the same to every subject in the image. These are best practices for 1girl loras, not proper loras.

>>102011527
No, I was talking about a post a few days ago.

>>102011503
You are very rude, so no I will NOT. Please leave me alone!

>>102011538
what does that have to do with what I said

>>102011423
I just enable buckets and let it do the rest and it's been fine. I haven't manually cropped or scaled anything. Whether or not that's optimal I have no clue, but it does work

>>102011556
You know exactly what I mean Mr naughty boy

>>102011503
>Those people do this for a living.
they train 1girl loras for a living, yes

Any NSFW fine-tune yet?

>>102011556
I mean that the article is talking about 1girl LoRA, basically. Not meant to generate images with more than one person in them.
>>102011581
There's actually a couple on civitai

>>102011573
No, anon, I have no fucking clue. There is no logical connection between your statement and what I said. Please explain.

>>102011566
So you enable buckets, but do you have to set a default resolution to train at in the settings or something.

Yeah sounds good, although I wonder if it still buckets fine with something like 6000x4500 resolution or something

OK, I downloaded kohya. Can someone give me a list of basics to get started with this?

>>102011581
tl/dr no
there only are some nsfw loras and one guy on civitai who wants to sell you his merge with nsfw loras as a finetune

>>102011590
I'll explain to you in deep detail, but you will have to come over. I have everything prepared for you to keep this going further if you know what I mean.

>>102011599
Ask the guy that pops up everywhere, SEcourses or something, he would love to give you a 1 on 1

>>102011585
>I mean that the article is talking about 1girl LoRA, basically. Not meant to generate images with more than one person in them.
Then it should be titled "Training shitty 1girl LoRA on Flux" so people don't think it is best practices for all loras.
>>102011605
sex pest

>>102011593
yeah, you set the training resolution. I've been doing 1024,1024 on my 4080

>>102011628
Cool I'll do that then, keeps things simple, thanks.

>>102011619
one true grifter god

Okay, I'm gonna grab a set from elitebabes and try to make this work

>>102011638
It's at the level where you can almost respect it really lol

File: 00012-4122411394.png (3.23 MB, 1280x1920)

3.23 MB PNG

>>102011649
Godspeed

>>102011649
>elitebabes
Sounds good, I never heard of it but seem decent

>>102011686
If some day I end up producing something worth sharing, is it OK to post LoRAs here through catbox or similar? I don't know why we rely on civitai honestly.

>>102010665
My 1.5k step bake on ada was trash, but there's a good chance it had nothing to do with the optimiser, I'm still working stuff out

>>102011738
>I don't know why we rely on civitai honestly.
bandwidth

>>102011753
I meant to share between us. Fuck the masses. And fuck whoever makes a lora for fame and shitcoins.

Best practice for training:
- consistent bucket resolutions, steal the buckets from Pixart, they're the best, you crop and resize to the nearest resolution.
- these models are very good at associating captions with features, you can effectively train even on extremely watermarked images if they're properly tagged (ie: "watermark, alamy stock photo"), the most important thing is if you have a recurring element in your images like a watermark, timestamp, etc, if you make sure it's in the caption, the model will be less likely "learn" it as a feature of your subject.
-variety in poses and backgrounds with a diverse set of captions

Also with Flux, it seems less is more. The N64 lora was only trained on a couple dozen images and you can do a grabbag lora train where you train on multiple concepts at once. For example, I think you could likely do a full video game style lora with sets of images for every console.

>>102011574
>they train 1girl loras for a living, yes
nobody throwing in the big bucks for the kubrick lora

>>102011782
makes sense, 90% of the weights will already be in the base model, you're just giving it a gentle nudge

>>102011738
yeah a lot of anons have used pixeldrain too

>>102011802
I think the parameters matter a lot with Flux + the T5 encoder. The model is way smarter than you realize.

Can I train on the fp8 or do I need to download the full fp32? And I can fit it in a 3090? Sorry about all the questions.

>>102011819
I think half the loras in civit are redundant, you could achieve them with better descriptive prompts, but there's buzz to be had

>>102011782
Thank you.

File: ComfyUI_temp_vxvqs_00008_.png (923 KB, 1280x768)

923 KB PNG

is smegma really that bad. having super sec genning feels pretty good. it's just soo bitte bitte

>>102011845
People are training it like it's a shitty SD model. I'm confident you can do a full celebrity Lora with a hundred celebrities (might need to crank up the network ranks)

>>102011861
That's too good, vanilla flux?

>>102011861
absolute kino. prompt?

pixelwave flux has begun
https://civitai.com/models/141592?modelVersionId=750297

>>102009692
https://discord.com/invite/Y4aH5KubP8

Bepcord, join NOW! For your /ai/ needs

>>102011908
>Trained 33 LoRAs and merged them with the Flux base model.
Holy fuck

File: 00078-4169123338.jpg (191 KB, 1344x1600)

191 KB JPG

>>102011908
>>102011912
ugh, buy an ad

ever since redditors starting posting here because of flux, we have been bombarded by people trying to shill stuff

File: ComfyUI_temp_vxvqs_00012_.png (1.17 MB, 1280x768)

1.17 MB PNG

>>102011855
i'm retearded. the speed is more related to the step count than the model. same speed with Q8 with same step count. superfast

File: 00121-2024-08-21-cJak.jpg (2.99 MB, 2048x2688)

2.99 MB JPG

I hope nuforge can get lora block weight running again

File: image.jpg (108 KB, 1536x1024)

108 KB JPG

>we have caught you posting pro-Palestine memes on 4chan, you must come with us
>resistance is antisemitism

>>102011912
>The main goal is to wash out that AI look from the model.
Flux has an AI look?

has anyone made a lora yet that makes the character stand further than 2 meters from the camera? i hate how it puts literally everything right in your face even when you explicitly tell it not to

>>102011912
Do NOT join this server; it's a pedophile grooming server

File: 1695727575070240.png (964 KB, 1024x1024)

964 KB PNG

Looks like I'm gonna have to wait for awhile until a model that supports danbooru tags comes along. Wall socket genners are not eating good

>>102012027
Good idea, make a lora with every image of someone far away

File: 2024-08-21_00557_.png (1.98 MB, 1280x1024)

1.98 MB PNG

>>102011890
>>102011895
no Nicolas Roerich lora with abit of Saruman lora
>https://civitai.com/models/667290/nicholas-roerich
>https://civitai.com/models/670669/saruman-from-lords-of-the-ring-flux1-d-lora?modelVersionId=750780

but the style is Roerich, thats a relatively unknown painter from the early last century, I was wtf an anon made 500 mb lora for flux of him? and it slaps

File: 00210-652653370.png (1.95 MB, 1360x768)

1.95 MB PNG

https://civitai.com/models/659761
holy fuck that one looks good

>>102012050
>500 mb lora
That's not a good thing anon

>>102012061
what a bell end

>>102011926
Told you it sucks kek

>>102012027
That's against the best practices, anon! https://www.finetuners.ai/post/training-lora-on-flux-best-practices-settings
If they don't take 80% of the image your lora is gonna be shiiiiiiit!

>>102011912
that is hell on earth

Is there even a single lora that actually needs to be 500 MB?

File: Capture.png (9 KB, 977x78)

9 KB PNG

>>102012073
you have such a rich vocabulary anon

File: 2024-08-21_00559_.png (2.45 MB, 1280x1024)

2.45 MB PNG

>>102011895
ow and prompt is just:
>a painting in style of Nicholas Roerich of Saruman, cinematic lighting, highly detailed, sharp focus, Global Illumination

>>102012073
yea theoretically, but this one just works.. it spits out Roerich exactly his style .. I

>>102012092
it's a conspiracy I tells ya, the hard drive making people are in control

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

1.23 MB PNG

>>102011912

>>102012061
time to make porn with it

>>102012070
Why not?

File: media_GVg7sRvWgAAEiMt.png (108 KB, 2560x2560)

108 KB PNG

https://xcancel.com/ideogram_ai/status/1826277550798278804
Good news for the guy who makes auraflow, he has better images to scrap now kek

File: ComfyUI_05118_.png (1.49 MB, 1024x1024)

1.49 MB PNG

>>102012061
finally my Migu is unslopped, can't believe a 18mb lora can change a model style this hard

>>102012043
sadly i dont have the vram for it
i have to hope someone else shares my frustrations enough to make it work
>>102012081
probably resembles best practices of training a model too
i wonder if a lora would be enough to fix it when most of the training data is probably person in center front of environment and not person inside environment with description of their specific position in the frame

>>102012238
Doesn't ideogram use stock photos as inputs for controlnet?
I used it a few times and often I'd get four images with the exact same pose and layout just like I'd get when using canny.

>>102012263
I have 24GB and I want to learn to train LoRA's, but it's very difficult to find a step by step for kohya with flux that isn't youtube engagement bait trash.

File: ComfyUI_05119_.png (1.39 MB, 1024x1024)

1.39 MB PNG

>>102012061

File: 1699068527503039.png (864 KB, 1024x1024)

864 KB PNG

>>102012238
am i understanding correctly that this has no open source model like flux dev?

>>102012309
Also kohya install fucked up my drivers and now nothing works

File: bogged matrix.jpg (17 KB, 200x232)

17 KB JPG

>>102012354
>he redeemed requirements.txt?

>>102012363
Worse. I downloaded and installed a .deb file from the nvidia site

>>102012346
yeah Ideogram is simply an API like MJ or dalle

File: ComfyUI_05120_.png (1.52 MB, 1024x1024)

1.52 MB PNG

>>102012061

File: bogdanoff meme1.jpg (20 KB, 400x400)

20 KB JPG

>>102012386
>He redeemed the deb?

sorry for your loss linuxfag
>i learned that the hard way before i re-re-re-switched back to win10ltsc

File: file.png (545 KB, 1024x1024)

545 KB PNG

>>102012394

File: ComfyUI_temp_oeuqn_00015_.png (1.32 MB, 1216x832)

1.32 MB PNG

Flux love for Bokeh is exaggerated.

File: supermetroid.png (1.28 MB, 1536x792)

1.28 MB PNG

What I love about Stable Diffusion 1.5 models is that they would infer a lot from a few text and would be very creative about it. Here's the prompt:
>little girl in Super Metroid.
Flux just fails terribly with the most boring thing, see, I could describe everything in the picture and get a much better version, but Liberte just hallucinates everything and delivers, and a different seed is going to give me a completely different generation.
I got used to this, I didn't have an image in my mind that I want to see on the screen, I have a prompt and want to see what a model does with it, so being able to tell the placement of the enemies and the color of her dress and describe how they are and having Flux draw them that way is useless to me because I didn't even know what would the model give me with such a prompt, but Liberte delivered and Flux seems like step backwards.
Can this even be fixed with finetunes? A "be creative" Lora?

File: ComfyUI_temp_mihah_00065_.png (3.15 MB, 1360x1600)

3.15 MB PNG

Is there an alternative vae for flux on comfy? some of my gens collapse when i use that vae, when I use the integrated one with forge (fp8) image don't have this trouble

>>102012483
that's because flux doesn't know enough styles so it outputs only what it knows

Just updated my comfy and downloaded the T5-q5_k_m. The speedup is phenomenal. Using just 4GB of RAM and 12GB of VRAM, this is better than SDXL in terms of memory, at least. If speed with lora can be improved, it will beat SDXL in all grounds.

>>102012483
>finetune vs barebones base model

here we go again. every model release lmao

>>102012483
eh you're comparing a basket full of apples to a basket with nothing in it and getting frustrated that the basket is empty
you're right though 1.5 was a beast with bullshittery, i really respect it for that.

>what's liberte?

>>102012501
more like it knows to many styles and only outputs what you tell it, while sd15 models style knowledge was extremely limited but got creative with what it got which was mostly the popular themes everyone wanted anyway

File: oops i blotched.png (894 KB, 1024x1024)

894 KB PNG

>>102012394
into the trash it goes

>>102012418
not even a waste, i think they're bullshitting eeeeverything. the evals are based on human ratings, and checking out the website i see absolutely nothing that suggests people should be chimping out over what ideogram is generating. i would guess it's some kind of (illegal due to licensing?) iteration of the flux dev base model, maybe also implementing google's t5

>>102012569
lol no retard

>>102012547
I KNOW HER!!

>>102012582
>i would guess it's some kind of (illegal due to licensing?) iteration of the flux dev base model, maybe also implementing google's t5
it's so cute when newbies make stuff up

>>102012609
present your theory and let's see how cute we can be

File: fs_0050.jpg (373 KB, 1920x1080)

373 KB JPG

>>102012596
with flux loras possible at home, about to waste several hours so I can gen her in the first pass instead of spending 30 seconds detailing with a 1.5 or XL lora

>>102012520
we also got the useless linux v windows shit here >>102012415
/g/ is going to /g/

>>102012670
cry about it it's just banter. If you can't handle the heat get out of the incredibly autistic kitchen.

>>102012415
I use both. I just had to restart to use the new drivers.
>i learned that
You didn't learn much if you've had to run away from an operating system with your tail between your legs because it gave you a hard time.

>>102012680
having trouble with context buddy? You'll get it eventually. We believe in you!

>>102012654
sovl

>>102012701
>>102012707
why do linuxfags get hurt this badly over the tiniest potential infraction? i didn't even mean anything by it.
Funny how the bogposting hurts no one until linux is involved.

>>102012596
Poo in the loo sir

>>102012620
Hooo-lee shit, catbox? Wanna motorboat that tummy.

File: 1711185957057412.png (135 KB, 906x694)

135 KB PNG

I'm really scratching my head how to train flux with Kohya. Am i supposed to take this command and change all the paths to the models? But then i'm missing the .toml file which has all the lora training data. Where do i get that? And do i have to create a new copy of my kohya folder and hack it so that it doesnt automatically replace the torch 2.4 version with an older one?

>>102012654
>Flux LORAs possible at home
Interesting, how so? How bad are the compute requirements?

>>102012714
its you. I didn't give my OS or support for either. Context is still eluding you.

File: ComfyUI_05127_.png (1.71 MB, 1024x1024)

1.71 MB PNG

>>102012061
holy sovl

File: FLUXenhanced.png (1.14 MB, 1024x1024)

1.14 MB PNG

>>102012550
>eh you're comparing a basket full of apples to a basket with nothing in it
Is it, though? I put what Liberte gave me in here:
https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha
To get a caption and put it on Flux, and it delivered... so it can do this fine, the question is why can't it infer this from my prompt like Liberte did? There is something wrong with the way it processes concepts. It had to be this long:
>This is a vibrant, digitally rendered illustration in the style of a classic video game cover. At the center is a young girl with short, light brown hair and a determined expression. She wears a bright pink, sleeveless leotard and matching boots, which are adorned with small, metallic details. The girl's skin tone is fair, and her eyes are large and dark, giving her a focused and slightly worried look.
>Surrounding her are four humanoid robots or mechas in various poses. To the left, a large, green robot with a metallic, armored body and a menacing expression stands with its arms outstretched. To the right, another green robot with a more compact and agile build crouches, ready to spring forward. In front of the girl, a smaller, green robot with a more insect-like appearance and bright red eyes stands, its arms extended. Behind her, a large, pink, octopus-like robot with glowing yellow eyes and a menacing grin looms, its tentacles wrapping around the girl's body protectively.
>The background is a deep purple with a bright, yellow, circular light source at the top, creating a dramatic contrast and enhancing the futuristic, sci-fi atmosphere. The ground is a sandy, rocky texture, adding to the alien
I can't believe my main use of SD1.5 models is producing pictures to get a description to use in Flux and get something usable.
This can't be right.
>what's liberte?
It should be the best fineture of the SD1.5 architecture, it never got popular because, like Flux, it can't do NSFW, and unlike other finetunes, it doesn't do fetishes either.

>>102012719
I can't long history

>>102012718
How did you know I just took a shit?

>>102012753
Am i talking to Conker here? Are you having a bad fur day? Calm down. Context sensitive or not chiiilll.

>>102012758
isn't that just how flux prompts?
that's the type of prompt i got from this anon >>102010869
and im even using it with XL for the hell of it kek

File: 71OMly2pizL._AC_UF894,100(...).jpg (128 KB, 835x1000)

128 KB JPG

how would you recreate this movie poster?

>>102012773
Aww. Well, what kinda model/LORAs did you use?

>>102012758
>There is something wrong with the way it processes concepts
Nothing wrong, it was trained on long prompts so it learned long prompts.
In the future, with better VLMs, they could have levels of verbosity with small prompts mentioning just the important bits but that introduces a different kind of bias.

>>102012784
>Am i talking to Conker here?
I thought I was the only guy on earth who knew that amazing game, based anon

>>102012804
Pony autism mix

>>102012749
https://github.com/ostris/ai-toolkit

I tried with kohya too and failed. ai-toolkit seems much straightforward for people who don't want to learn how anything works and just want it to get it to work. Just read the readme here.

>>102012814
Everyone on the planet knows that game sir

>>102012822
i downloaded aitoolkit yesterday but it said it required 24gb vram, and i only have 16. it also didnt like my version of flux-dev for some reason

>>102012846
>24 gb vram
Damn, I'm fucked.

>>102012749
Correct, it doesn't explain how to use the damn thing
you don't need the toml, you can add the tag

--train_data_dir "path/to/dir"

which should point to the parent folder containing your training set, and the folder with the actual images in it should be named

[repeat_count]_name category

, so for example

4_Swimsuit Clothing

>>102012846
Ah. I have 24GB. Sorry. I think you can offload to RAM with the low_vram setting in the config file tho.
>it didn't like my version of flux-dev
The script downloads it from huggingface with your token. It's all explained in the readme.

>>102012050
>Nicolas Roerich
it's amazing how modern some artists from the 20s and 30s were. same thing with music in the 60s and 70s. some sound like they're from 2015

File: 359631364-e7e7d182-b3c9-4(...).jpg (447 KB, 1024x1024)

447 KB JPG

This guy is the biggest fucking faggot in this entire space. Inserts himself into every possible discussion. Sells himself as some kind of "AI expert" yet constantly begs for help and asks extremely basic questions like "what is the difference between mixed precision and full bf16 training". People spoonfeed him, and then through sheer trial and error he eventually finds some lora settings that kinda work okay. He then takes all this info he (probably incorrectly) learned and writes blog posts he paywalls. Then shills his shit everywhere he can. Won't even share the config he used for the loras, you gotta subscribe to his patreon for that, so you can learn to become a finetuning expert like him.

I hate this AI influencer / hustle culture so goddamn much.

>>102012869
oh yeah i didnt want to download the file again, guess i have to.

>>102012860
thanks. i'll try this

>>102012609
i knew you were a coward, too afraid to battle in the marketplace of ideas...

File: ComfyUI_temp_oeuqn_00026_.png (1.38 MB, 1216x832)

1.38 MB PNG

>>102012802
ctrl+c and ctrl+v
or try describing with a lot of detail it in flux, it won't get 100% of it right but it will surprise you.

File: ComfyUI_temp_vxvqs_00029_.png (480 KB, 768x1280)

480 KB PNG

>>102012802
the kinoest i got

>>102012904
>i didnt want to download the file again
You need to download the full big boy repo from BFL

>>102012784
>isn't that just how flux prompts?
Wasn't prompt handling supposed to do better? How can it be worse than 2 years old models at it?
We should have gotten "being able to place things where you tell it and with the descriptions that you tell it" as an EXTRA, not INSTEAD of the creativity we used to have for short prompts.

>>102012903
>I hate this AI influencer / hustle culture so goddamn much.
amen anon, fucking amen

>>102012913
I can't beat shitty ideas when you handle them freely like that.

Please stop posting outputs from the lora based off images of your own face to LDG
Thanks

File: ComfyUI_temp_vxvqs_00028_.png (556 KB, 768x1280)

556 KB PNG

>>102012919
yep

>>102012903
He's a genius.

>>102012814
Why are people so excited by a character from Diddy Kong Racing? It's like getting excited about Geno.

File: xl ghost leg.png (472 KB, 744x836)

472 KB PNG

what the FUCK are you doing SDXL, giving me a right spook here

>>102012929
Just fire up nemo or mini-magnum and ask it to augment your prompt with prose or to add details to it.
Mmh. I wonder if there's a node for comfy that can pass prompts through an LLM for processing according to another custom prompt. That would be cool

>>102012903
>get piano lessons
>complete lesson 1
>start a business as a piano instructor, you'll always be one lesson ahead of your student
easy money

>>102012872
The music from 2015 sucked so that sounds like an insult.

>>102012964
>Why are people so excited by a character from Diddy Kong Racing?
kek, so you're telling me I should be so excited about Banjo and Kazooie because they were also a character from Diddy Kong Racing? ;'(

>>102012964
Geno? You mean the character based on magic girl from Yu-Gi-Oh?

File: flux_00314_.png (1.56 MB, 968x1120)

1.56 MB PNG

>you seem distracted. do you like rockets that much, cadet?

>>102012975
That's a typical boomer smartass comment that doesn't track in reality. In the real world, you can't bullshit your way through an industry this active for long.

>>102012941
i mean if you know more than me i'm genuinely curious as to what you actually think ideogram's model came from, even if the answer is boring

>>102012919
How did you prompt that teenager? With me it's either adult or child, nothing inbetween.

>>102013003
unless you're brown or a woman, but that's another topic

>>102012987
not an insult, just meant how forward they were. seems like everything past 1970 was just copying and true insight and creativity just died completely.

>>102012964
>Diddy Kong Racing
I'm sure everyone on 4chan love this game simply because of the funny indian accent they made on one of the main characters
>plz saar select your vehicule saar do the needful

>>102013011
nta anon, but maybe specify age, don't say young

File: 00061-3950541503.png (1.32 MB, 1216x832)

1.32 MB PNG

>>102012903
It gives me a little chuckle when I see him everywhere I go, like man him again.... but I don't get the emotional impact you had just now. Does this mean I have become numb?

>>102013017
the 80's were good anon, that decade was also creative, it started to become shit in the 90's with the rappers, samples and shit

>>102012802
just feed it to a vlm and work from there

>>102013011
nta It's a complicated thing. You need a lot of trial and error to get a feel of what works with any particular context, and they all look germanic and shit

File: ComfyUI_temp_oeuqn_00031_.png (1.46 MB, 1216x832)

1.46 MB PNG

>>102013011
Lucky. I get a lot of variance in age, no matter what I prompt. Specially next to a man, it was more consistent before with just the girl and I didn't do anything special.

>>102013038
90s was still good, there was still hope about the future getting better in the air. But after 2001 things went downhill in general

>>102012229
wasted RAM and space aside, in my experience bigger loras take too many details from the training data, like for an artist style lora it's more likely to also influence subjects and other shit

model is about to finish downloader. am i looking forward to a glorious future of fetish porn lora enjoyment, or a night of frustration and long hours unable to use my computer? only time will tell

>>102012990
They had their own game and their fans claim it was superior to Super Mario 64.
It got me wondering what game would be praised if Mario 64 was Banjo Kazooie and vice versa, would people remember Banjo Kazooie as a classic or do people care mainly about the character that is the protagonist?

Bakery delivery, it's bread...
>>102013088
>>102013088
>>102013088

>>102013087
When tagging for a lora, is it preferable to describe everything in the picture even if it's unrelated to what you're training for, or just focus on the thing your lora is for?

>>102013098
>They had their own game and their fans claim it was superior to Super Mario 64.
Conker also had his own game anon

>It got me wondering what game would be praised if Mario 64 was Banjo Kazooie and vice versa, would people remember Banjo Kazooie as a classic or do people care mainly about the character that is the protagonist?
I played both back in the days I still prefer Banjo and Kazooie, this game is so unique, mixing platformer and puzzle game was a genius move, no one even come close to that

>>102013072
Could you still share the prompt?

>>102012061
>289 steps

what the fuck

>>102013000
rockets, thrust, liquid propellant, expansion, heat, nozzles, suction, injection, blast off

>>102013117
>Conker also had his own game anon
Ahh, that would explain the excitement about finding another person knowing about such a niche game, I'm glad the Diddy Kong Racing characters got their own games, imagine if that happened to Mario Kart 64 and Donkey Kong got his own game!

File: ComfyUI_00024_.png (411 KB, 512x512)

411 KB PNG

>>102013038
films from the 80/90s are my favorites and i hate films from the 70s. i think filmmaking and graphic design reached their peak in that age, like possession and in the mouth of madness. but creativity wise it was lesser than previous decades. thats bound to happen when a craft reaches such high level of polish. after that it really all went to shit. i read an article once that analyzed film plots and themes and found that films where the most diverse in 1950s, and after the decades they would understand which ones sold better and stopped producing the weird ones

>>102013108
Describe everything I think

File: 1724265631296__000000000_6.jpg (93 KB, 1024x1024)

93 KB JPG

>>102012483
I don't think your theory is correct. It's just that the model won't render things it doesn't know about.
Look at this one (unquantized model)
>hipster man with a beard, building a chair, in a wood shop
It adds details just fine.

File: ComfyUI_temp_vxvqs_00049_.png (776 KB, 832x1216)

776 KB PNG

>>102013064
how do i do that?

>>102013338
Go to a page like this one:
https://aichatonline.org/gpts-2OToA97Vhr-Describe-Image
And upload your image, you'll get a prompt, see what it generates with it, modify it to get what you want.

>>102013188
>imagine if that happened to Mario Kart 64 and Donkey Kong got his own game!
but Donkey Kong got his own game, it was Donkey Kong 64 ;_;

>>102013003
that guy has 12k subs on patreon

File: ComfyUI_temp_vxvqs_00055_.png (1.05 MB, 832x1216)

1.05 MB PNG

>>102013424