Is this the blessed bread?

Blessed thread of frenship

>>101166584
>The midjourney pic made the collage
Uhh

>>101166626
Sorry, flew over my head.

>>101166626
which one is it?

>>101166691
Big knight girl in bottom right.

I can do a quick rebake if you mind it enough.

>>101166697
>>101166626
can you write a prompt for it? I'll try to dupe it and we'll ignore it.

>>101166739
interior, small kitchen on a space station, planet and stars and space station out the window, junk art, gadgets, post apocalyptic, star wars outpost, style of ian mcque and joe madureira, a man cooks on a stove, slice of life, light leaking through window --ar 16:9 --stylize 400 --niji 6

interior, small bedroom on a space station, planet and stars and space station out the window, junk art, gadgets, post apocalyptic, star wars outpost, style of ian mcque and joe madureira, cinematic lighting, a house cat looks out the window, star wars themed --ar 16:9 --stylize 400 --niji 6

Can we add MJ to the OP, please?

>>101166772
sure

File: UI_0010.jpg (1.21 MB, 1248x1824)

1.21 MB JPG

>>101166697
How's this?

>>101166867
the space cockpit ls also made by MJ kek

>>101166626
Was bound to accidently happen at some point with how many images are posted. I'm surprised it took this long. Shit happens.

>>101166874
Close but no cigar.

>>101166880
Half the collage is just random MJ images that were posted specifically to mog on the shitty local gens ffs

>>101166626
>freetards subconsciously recognize saas is simply better
kekarooooooo!

>>101166772
>>101166791
MJ isn't [LOCAL] diffusion.

>>101166878
JFC

>>101166884
...I'll be cooking for a while then.

>>101166900
>Half
This is an exaggeration. Anyway b8

>>101166867
>Still got an MJ pic in it
it's so over for local gen christ

File: 00013-3140994425.png (2.74 MB, 1920x1280)

2.74 MB PNG

I'm not sure I have a model that imitates that strongly of the roughdraft/painting art style of that MJ pic, but I'm sure it exists because plenty of people like that look.

Also, isn't it kind of noisy like a DALL-E3 gen?

local or bust

local is a bust*

local makes me bust a nut (because i love it so)

>>101166976
we all do

File: UI_0014.jpg (1.3 MB, 1536x1536)

1.3 MB JPG

>>101166874
>>101166915
I might need to build a workflow for this one...

File: long dick general.jpg (2.39 MB, 3202x3264)

2.39 MB JPG

>>101166584
Three times the fucking charm, one more and Imma bust a nut.

>>101166900
Point out MJs beside the knight one.

File: 00013-3140994425.png (2.74 MB, 1920x1280)

2.74 MB PNG

>>101166938
forgot to inpaint her face

>>101167046
I can vibe with this 1girl

>>101167019
There is still one more MJ - the second anime knight

>>101167052
>just as I posted rebake of the rebake
nut it is then

>>101167062
please don't fuck with me, I'm sleepy enough

so what's the deal with the open model project thing? looking at the discord, it looks like a case of too many cooks.

File: UI_0016.jpg (1.66 MB, 1024x1536)

1.66 MB JPG

>>101167089
There is no discord only Thread...

File: UI_0017.jpg (1.58 MB, 1024x1536)

1.58 MB JPG

>>101167106

>>101167106
She is the one that Slaps

File: UI_0018.jpg (1.09 MB, 1024x1536)

1.09 MB JPG

>>101167116

File: 00005-852594867.jpg (232 KB, 1176x1571)

232 KB JPG

>>101167054
hell yeah

File: UI_0020.jpg (1.09 MB, 1024x1536)

1.09 MB JPG

>>101167172

>>101167089
There are two projects: one is clearly trying to cash in on the demise of SAI, assuming they can even train a model it's going to be more censored that SAI and they have the same hemming and hawing and hand wringing. The other is just a bunch of Reddit 16 year olds larping about training a model, you won't get anything out of them outside of their list of datasets.

>>101167228
We do need good datasets for training

>>101167265
Yeah they had one that was good that I haven't seen before.

>>101167285
do you have huggingface link?

File: UI_0022.jpg (1.2 MB, 1024x1536)

1.2 MB JPG

>>101167200
I knew I recognized that artist. Wlop. for MJ to clone from.

>>101167188
you aint him

File: UI_0023.jpg (1.18 MB, 1024x1536)

1.18 MB JPG

>>101167329
Still building workflow.

>>101167316
https://huggingface.co/datasets/UCSC-VLAA/Recap-DataComp-1B is all you need (the new LAION)

https://huggingface.co/datasets/tomg-group-umd/pixelprose - huge Gemini captioned dataset

You could probably make a foundational model with Pixelprose. And definitely could with the DataComp dataset.

>>101167378
sweet now if only this would kick in
>https://chipsandcheese.com/2024/06/25/testing-amds-giant-mi300x/
otherwise I'm stuck waiting for christmas to get my shitty 4090

File: tmpply54501.png (882 KB, 1440x808)

882 KB PNG

>>101167356
>you aint him

File: UI_0024.jpg (1.01 MB, 1024x1536)

1.01 MB JPG

>>101167367

File: UI_0028.jpg (1.23 MB, 1024x1536)

1.23 MB JPG

>>101167106
>>101167116
>>101167172
>>101167200
>>101167329
>>101167367
>>101167410
SD1.5

File: 00017-2237513411.png (2.55 MB, 1920x1280)

2.55 MB PNG

>>101167487

>>101167562
You're winning in the race so far m8

File: 00024-58630552.png (2.88 MB, 2112x1184)

2.88 MB PNG

Two for the price of one? kek

File: SDXL_0001.jpg (2.32 MB, 2048x2048)

2.32 MB JPG

>>101167562
>>101167676

File: 00025-58630553.png (3.24 MB, 2112x1184)

3.24 MB PNG

>>101167733
Going to need to change my prompt.

File: 00026-58630556.png (3.11 MB, 2112x1184)

3.11 MB PNG

whoa

File: SDXL_0003.jpg (2.36 MB, 2048x2048)

2.36 MB JPG

>>101167758
After so many gens I already forgot how the original looked like.

>>101167019
Nice

hibernation mode activate

>>101167228

Well yes there is hemming and hawing because legally they could be fucked if the law gets changed at any moment.

If its made illegal to allow prompting artist names then they have to retrain the whole god damn model. They leave loras for styles up to the community so their asses won't get sued.

Otherwise our only other option is another NAI leak but good fucking luck with that.

File: SDXL_0021.jpg (677 KB, 1024x1024)

677 KB JPG

>>101168256
Gen is taking a while now since I'm trying to do upscale of this one

>>101168417
Is this AOM? Or 1.5 it's got that sheen to it.

>>101166584
why did you take the tracker out retard?

>>101168486
cetusmix4

File: ComfyUI_00277_.png (1.06 MB, 1152x896)

1.06 MB PNG

been working on some liminal spaces what do you guys think

>>101168532
very nice and very liminal

File: ComfyUI_00320_.png (1.23 MB, 1152x896)

1.23 MB PNG

>>101168555
thanks man trying to figure out how to make it look like shit while also not looking like shit

does anyone have any upscalers they recomend for non-humanoid gens im trying 4x_nmkd_siax_200k and idk

>>101168508
Whats the artist style on that?

>>101168532
creepy (in a good sense)

>>101168486
AOM gives blurry gens, it is pretty annoying, though I bet that's a non-issue with refiner. But then you probably have to lean harder into booru tags, which don't lead to the best results in many anime models and it is best to try and make do without them, in my experience. While I didn't use AOM that much before deleting it from my system, a bunch of better models have it mixed in, so it is not a bad ingredient for mixes, apparently.

>>101168498
I dun reckon that you answered your own question there, m8

>>101168498
qrd?

>>101168921
torrent tracker was in the OP

>>101168879
>>101168921
>>101168945
>>>/sdg/

>>101168835
>booru tags, which don't lead to the best results in many anime models
Wha?

>people actually don't do 2nd passes/hires fixes
What in the god damn fuck?

>>101169058
try not using them, booru tags suck

File: 1716280215825478.jpg (6 KB, 225x225)

6 KB JPG

>>101166584
I installed SwarmUI with SD 3 and it fucking stank.

>>101169119
Good, now go use something else.

File: 1719444176104168.jpg (158 KB, 816x1325)

158 KB JPG

Is there a guide on how to setup krita+comfyui so I can feed it my base sketches and have it finish them for me? I'm comfortable with node based workflows, I've done a bunch of unreal blueprints stuff.

File: ramlets.png (2.4 MB, 1920x1080)

2.4 MB PNG

>>101169074

File: 00627-1764813909.jpg (204 KB, 1260x1680)

204 KB JPG

>>101169119
10 scoops and leave humanity behind

>>101169158
theres a new photoshop plugin but im not sure how much more difficult it is than krita. either way, if you're using scans and not digital like that pic you might as well feed them into the sampler regularly with base comfy but thats just me.
you should find guides / examples on the relevant software pages. cute pic btw

>>101169180
oh I don't touch photoshop

also if anyone wants to run that drawing through any model I would have fun seeing where it goes.

>>101168594
>upscalers
4x_Ultrasharp

>>101169158
IDK but the plugin https://github.com/Acly/krita-ai-diffusion has some documentation including how to activate the controlnets (which is something you want to do since your sketch lines need to guide the diffusion with just a prompt you can't extensively complete and hope something much of your sketch reemains)

>>101169180
>if you're using scans and not digital like that pic you might as well feed them into the sampler regularly with base comfy
Actually can you expand on this a bit?

I've got a tablet, I can draw equally with digital and traditional. Apparently the models react differently to them as inputs?

>>101169217
>Apparently the models react differently to them as inputs?
in a sense but that's not what i was referring to. typically the reason why someone chooses krita is because they want to have that sort of "AI Canvas" feel. likely you can disregard that statement.
https://github.com/AlekPet/ComfyUI_Custom_Nodes_AlekPet includes a painter node which might be of some use to you.
it sounds like you dont have much experience with imggen so you'd probably benefit from messing around with straight txt2img before jumping into what you describe. but definitely read through the link in >>101169211

>>101166874
Really bizarre anatomy. There's about 1.25 head lengths from the bottom of the chin to the middle of the breasts. She's got two more.

File: w9tchhun7er_002850_00_202(...).png (874 KB, 1024x1024)

874 KB PNG

As stupid as this sounds, can bad LoRA settings cause jpeg artifacts? Or what look to be shitty artifacting? Because I know the images I used are clean, but my samples are coming out as if I was using images with their backgrounds poorly cropped out and giving me really awful artifacting along the character lines. Picrel. As I said, I KNOW my source images are clean and free of this shit, the only thing I changed was my text and unet learning rates...

>>101169401
>unet learning rate
did you increase it too much? could be overfit

>>101169461
I actually decreased it to 0.0001 down from 0.0003, and my text encoder from 0 to 0.00005. I realized now I forgot to remove the 'train unet only' additional parameter. Could that have caused it?

>>101169401
might be a bias in the resolution bucketing of your dataset

File: SDXL_0037.jpg (1.87 MB, 2432x1664)

1.87 MB JPG

>>101166697
>>101166874
>>101167014
>>101167106
>>101167116
>>101167172
>>101167200
>>101167329
>>101167367
>>101167410
>>101167487
>>101167758
>>101167850
>>101168311
Last try.

>>101169499
Oh fuck me, you're right. I didn't use the upscaled images I had prepped. The original ones I had were 485x850 before I did a 2x upscale on them, but I used the wrong folder like a dumb fuck. Shit. That's 3 hours wasted, lol.

>>101169583
using upscaled images will degrade your output

>>101169715
Well, it couldn't be any worse than what it already shit out. And I've had success with LoRAs using upscaled images so long as it wasn't some ridiculous resize that smoothed everything out and made it start to blend together.

>>101169715
I haven't had any problems so far. Just clean images before upscaling. Artifacts are easy to remove and lineart can be smoothed

>>101169768
I think this is something you may not notice easily if you don't already have a lot of experience with digital art. upscaling line art, for example, tends to produce some common errors like radii at what should be sharp intersections of lines, truncating or distorting fine calligraphic strokes, and mangling together high-frequency strokes. this stuff all together contributes a lot of the "AI tells" like blobbed-together linear details.

if you do this, make sure you include negative tags as appropriate describing what you did to help the model filter out these degradations.

>>101169857
Lineart is so easy compared to 80s and 90s messy comic art. 2x upscale is max I do, images just become something else at some point.

>"AI tells" like blobbed-together linear details
Blobs mixed in fine lines yeah. Can't really escape them with 1.5 and sdxl

File: grid.jpg (1.26 MB, 3264x960)

1.26 MB JPG

its the details, where it lacks

>>101170029
try that again on sd3
supposedly the 16ch vae improves fine detail a lot

>>101170043
yeah. hopefully anon figures out how to get a 16ch vae going for pixart. the devs seemed to allude that there's more to it than just that but idk

>>101167106
Cool

>>101166584
>Nightly Edition
yup

Remember when Emad was based?
https://www.nytimes.com/2022/10/21/technology/generative-ai.html
>Emad Mostaque, the founder and chief executive of Stability AI, has pushed back on the idea of content restrictions. He argues that radical freedom is necessary to achieve his vision of a democratized A.I. that is untethered from corporate influence.
>He reiterated that view in an interview with me this week, contrasting his view with what he described as the heavy-handed, paternalistic approach to A.I. taken by tech giants.
>“We trust people, and we trust the community,” he said, “as opposed to having a centralized, unelected entity controlling the most powerful technology in the world.”

File: ComfyUI_temp_rdaoy_00007_.png (1.76 MB, 896x1152)

1.76 MB PNG

k value = 134

File: our_future.jpg (1.07 MB, 1536x2256)

1.07 MB JPG

>>101170957

>>101170957
And then he shot himself at the foot by cucking his models for no reason (no law forced him to do that)
>Do Not Obey In Advance
>Much of the power of authoritarianism is freely given. In times like these, individuals think ahead about what a more repressive government will want and then offer themselves without being asked. A citizen who adapts in this way is teaching power what it can do. – Timothy Snyder

>>101170957
runway sd1.5 leak was oct 20
guess his beliefs didnt stand the test since he still tried to have them taken down

>>101171103
He was just lying at the time to get enough hype for the VC money, we as users are just tool for him, once he got the money he could show us his true colors without much consequences

File: ComfyUI_temp_rdaoy_00013_.png (1.86 MB, 896x1152)

1.86 MB PNG

>>101171080
>no law forced him to do that
i understand the trepidation, but its litigation not law that pushes these guys away from it. here in our armchairs its easier.
that being said i never think about him.

>>101171184
Are you using a artist?
They have a very distinct style, i like them.

>>101171207
ipadapter with ~20 input images no artists in the prompt

File: grid.jpg (796 KB, 1792x1152)

796 KB JPG

ty anon

>>101171253
pretty metal

have a nice day anons, im off to work.

>>101171367
Cunt

>>101171367
take care

official pixart bigma waiting room

>>101169139
Elaborate then; enlighten me, oh wise one.

File: ComfyUI_04448_.png (1.32 MB, 1024x1024)

1.32 MB PNG

Okay, I was out of the game for almost 6 months. What do I have to catch up with?

>>101172008
nothing, SD3 was released and it was a fiasco

>>101172008
nothing much really, sd3 came out and it's shit, pr disaster, awful license, blah blah blah. we have alternative models like pixart and hunyuan which gained a huge surge in popularity after stability shot themselves in the foot so local image gen won't be dying anytime soon. the pixart guys also said they work at nvidia now and are training a bigger model.

I'd like to add I've been focusing on AI music made with Udio, that was a new renaissance for me, but I want to go back to my roots. Too bad the service is going to get nuked soon thanks to you know who.
>>101172037
Yeah, I heard it's illegal to train. WTF?
>>101172044
Thanks, I'll check them out.

>>101170067
You probably need to retrain the whole thing. A VAE that significantly different probably means a whole different latent space. So the UNet is converging through a completely different space.

>>101172065
It's not illegal to train, you just need to pay a monthly fee. Also something about the outputs of SD3 and its finetunes potentially being SAI's property (not explicit but a potential legal consequence of the license), civitai banned SD3 because of that.

>>101169001
what does that have to do with /sdg/ mongo?

>>101172008
Not much, Pony is still the queen of models.

https://huggingface.co/aipicasso/emi-2-5
What is this? A new base model?

>>101172469
XL finetune

>>101172491
I see, it wasn't on civitai so I thought that was a new thing or something

>>101172469
>このモデルの特徴として、Danbooruなどにある無断転載画像を学習していないことがあげられます。
>Stable Diffusionと同様のデータセットからDanbooruの無断転載画像を取り除いて手動で集めた約3000枚の画像
>Stable Diffusionと同様のデータセットからDanbooruの無断転載画像を取り除いて自動で集めた約50万枚の画像

haha....

File: Sigma_02730_.jpg (2.42 MB, 2560x1536)

2.42 MB JPG

Good morning anon

File: UI_0033.jpg (2 MB, 2432x1664)

2 MB JPG

>>101172588
morning

File: TheFutureWeWerePromised.png (2.61 MB, 1344x896)

2.61 MB PNG

>>101172931
I see that you're trying to replicate the Midjourney picture kek

File: UI_0035.jpg (1.21 MB, 1824x1248)

1.21 MB JPG

>>101172941
yep, I can say that I'm doing better than MJ did

File: oc8rnjwd7m7b1.png (2.23 MB, 1536x768)

2.23 MB PNG

>>101172979
can you do better than this tho?

>>101172987
Give me the prompt

>>101173003
I found this picture here
https://reddit.com/r/midjourney/comments/14gbdd9/anime_characters_in_the_style_of_bladerunner/

https://shuweis.github.io/ResMaster/
>Diffusion models excel at producing high-quality images; however, scaling to higher resolutions, such as 4K, often results in over-smoothed content, structural distortions, and repetitive patterns. To this end, we introduce ResMaster, a novel, training-free method that empowers resolution-limited diffusion models to generate high-quality images beyond resolution restrictions.
That's really interesting, I always wanted to make 4k pictures but all I got was duplicated humans kek

File: UI_0038.jpg (2.25 MB, 2432x1664)

2.25 MB JPG

>>101173014
><Character from Anime> in the rain side profile in the city, in the style of synthwave, magenta and cyan, extremely detailed art, rebecca sugar, shiny/glossy, hyper-detailed illustrations


also noticed that 1536x768 res. (don't need to upscale it)

Here is the first image from same model >>101172979

File: UI_0039.jpg (629 KB, 1216x832)

629 KB JPG

>>101173059
SIGMA
>(sigmax ** (1 / 7) + y * (sigmin ** (1 / 7) - sigmax ** (1 / 7))) ** 7
output

File: eu7mb4viyuzc1.jpg (2.49 MB, 5824x3264)

2.49 MB JPG

>>101173059
not bad at all, how about that one
https://reddit.com/r/midjourney/comments/1cpqmb6/cozy_space_station_apartments/

File: UI_0040.jpg (727 KB, 1216x832)

727 KB JPG

>>101173059

File: UI_0041.jpg (676 KB, 1216x832)

676 KB JPG

>>101173085
isn't that
>>101166753

File: UI_0042.jpg (1.47 MB, 1824x1248)

1.47 MB JPG

>>101173125
SIGMA
>>(sigmax ** (1 / 7) + y * (sigmin ** (1 / 7) - sigmax ** (1 / 7))) ** 7
output

File: UI_0045.jpg (1.37 MB, 1824x1248)

1.37 MB JPG

>>101173132

File: UI_0046.jpg (1.52 MB, 1824x1248)

1.52 MB JPG

>>101173183

>>101173202
>>101173183
there should be a city outside of the cockpit

Is this threat appropriate to make a request?

>>101173257
what request do you want to make anon?

>>101173257
All depends on the request, I might be able to help out.

File: UI_0051.jpg (1.16 MB, 2016x1152)

1.16 MB JPG

>>101173230
Not really.

>>101173410
if you want to reproduce that >>101173085 then yes really

16h training wasted because wrong settings. Live, love & laugh.

>>101173418
Nope, even in OPs post he wasn't able to reproduce similar image twice

>>101173279
>>101173295
Thanks!
I've been really into the 'barber beats' aesthetic recently. It's a musical style of mostly plundered beats and the cover art that usually goes with it is really cool. It has some vaporwave elements. For some examples you can scroll through rate your music: https://rateyourmusic.com/genre/barber-beats/ or just look on bandcamp or youtube or whatever.
A while ago someone was generating some really cool similar things.
I'm not really asking for anything specific, no specific color scheme or elements, just more general barber beats type album cover art.

>>101173459
because in OPs that's also a Midjourney picture

>>101173461

>Macroblank

GOAT

why do pony and other SDXL models seem to fuck up hands even more than 1.5?

File: UI_0054.jpg (604 KB, 1344x768)

604 KB JPG

>>101173473
You have a valid point, and I'm local doing things better than it can.

>>101173496
delusional

>>101173492
Probably because you are using a bad model or bad lora.

>>101173492
None of them train to do hand or feet,

>>101173520
explain how?

>>101173575
you're the one that claimed that your picture is better than this >>101173085
it means that you have the burden of proof anon
https://en.wikipedia.org/wiki/Burden_of_proof_(philosophy)

After all this can we please finally add MJ to the OP? We've proved it deserves its place here - fuck, it's even made it into the collage several times now

>>101173496
>>101173085
im a local guy but anyone claiming that whack stormtrooper gen is better than the gen with an entire city and floating ships out the window is out their damn mind

>>101173699
Sure!

Where can I download it so I can run it on my machine?

>>101173719
Discord mostly, and I know it's not Chink shit but hopefully that isn't the barrier it used to be for this thread now that we're accepting SDXL gens

>>101173699
MJ isn't a local model.

File: UI_0057.jpg (768 KB, 1280x768)

768 KB JPG

>>101173603
>>101173716

>>101173733
Sure, and SDXL isn't Chinese, but that's clearly not the rule anymore. Half the thread is MJ gens, there's MJ gens in the collage and we have an entire discussion showing off how gens like this:
>>101173739
Could clearly be improved

>>101173731
No I don't want to download or get an account, link me to the download.

>>101173731
>>101173752
all local models are welcome to /ldg/
the whole notion of "SAI models are not welcome" is nonsense
its not about chinese models its about getting away from the drama and avatarfagging

>>101173753
>I don't want to download
>link me to the download
Anyone else getting sick of retards trying to join in normal conversations? Thought we made this thread to avoid this. Please, /sdg/ is but two clicks away.

>>101173085
Damn, so we're that far from the closed source models huh? That's just sad man...

>>101173752
Read the title of the thread. LOCAL DIFFUSION GENERAL. If it's local and it's a diffusion model, it's allowed. But nice try SDG troll. What, is your general too dead right now?

>>101173752
what are you talking about? read the thread name, sd is a local model, unlike mj

>"local" diffusion general
>full of MJ gens
>try to bring this up
>no we dont post MJ gens here
Fucking hell

go to school

>>101173765
Give me the model.

>>101173784
When will someone make a fucking MJ thread already? Dalle3 has one :(

>>101173796
I can barely even understand the message you're trying to get across. I'm going to refer everyone back to these two clauses and hopefully that will make it clear your mind is about as functional as your workflow:
>I don't want to download
>link me to the download

>>101173784
Oh no a troll like you came in and started posting MJ gens something that is impossible to prevent on an anonymous forum? Maybe Discord is more your speed.

>>101173798
I've been posting MJ gens here generally, they're typically getting reasonable attention and I manage to get into the collage fairly frequently, though I do wish I had more time to spend on improving my WF

>>101173825
but this isn't the right place to do so, there should be a thread only dedicated on Midjourney imo

>I've been posting MJ gens here generally
stop doing that

>>101173820
If you're looking to join the discord, we can help with that! We're actually having a movie night soon and hoping to get some openRA games going with all the folks there!

File: cc.png (620 KB, 900x934)

620 KB PNG

Are there any new tools for background removal?? Rembg is so terrible, it's easier to do it manually in photoshop

What the fuck is MJ? Who the fuck put Michael Jackson in the freezer?

>>101173838
>>101173839
Eh? This place clearly is fine, nobody is sperging out except you two and whoever bakes the threads clearly approves enough to keep them in the collage. Most of the posts/gens itt are discussing MJ already.

>>101173809
You're right. Enjoy.

>>101173859
>hehehe I'm such a clever troll
go back to sdg

>>101173864
Why? MJ is not an SD model

saar please, read the thread name

>>101173868
It's not a local model.

>>101173739
Delusional

>>101173446
Been there. Just have to double check every setting

File: file.png (167 KB, 256x256)

167 KB PNG

cat

>>101174143
It's like inkblot test

>>101174143
yes this is cat

>>101174143
Chinese cat

is there an ipadapter model for pony?

>>101174483
>pony
>>>/g/sdg

>>101174143
this is a 1girl

>>101174483
sdxl controlnets have been pretty rough

>>101174543
Because you need a supercomputer to train anything on SDXL so even casual researchers can't do anything with it.

>>101174143
kitty :3

IPadapter for pixart when

>>101173132
nice

File: file.png (91 KB, 907x355)

91 KB PNG

>train LORA overnight using kohya
>get back to pc
>everything worked as intended
>try to train another lora
>now it cant access cuda drivers
>restart pc
>still cant access cuda drivers
what the fuck happened?

>>101173839
Probably a larp since it really only happened in the last thread and he maybe posted 4 total

>

File: ComfyUI_PixArt_00116_.png (2.22 MB, 1024x1024)

2.22 MB PNG

>>101176026
best post in the thread

File: PA_0018.jpg (2.79 MB, 2432x1664)

2.79 MB JPG

>>101176151
He said enough

>>101176026
unpopular opinion and many will hate to hear it, but you're 100% right.

>>101176178
kek

>>101176026
brap

File: Hunyuan1719509557-cfg7.5-(...).png (1.74 MB, 1280x768)

1.74 MB PNG

>>101173059

What are you using to gen all this thread?

File: ComfyUI_PixArt_00119_.png (1.99 MB, 1024x1024)

1.99 MB PNG

>>101177181
saar, i am using the pixart.

File: Hunyuan1719512061-cfg6-st(...).png (1.5 MB, 1024x1024)

1.5 MB PNG

Fuck it, this is the closest I got to the subject. The eyes! Why can't it learn the eyes!?
I don't know wtf in Kohaku script https://github.com/kohya-ss/sd-scripts/pull/1378 that he got his training to converge so quickly. Maybe I should download it to study in depth

File: 01751-1914074388.jpg (223 KB, 896x1200)

223 KB JPG

turns out that sigma gens can be pretty damn good for lora datasets since they are lacking the usual sd jank

anything local that can match nai in terms of inpainting?

>>101178296
>match
>nai
>match
Anon, nai is old and busted. Pony merges blow it out of the water.

>>101177939
Sigma has its own kind of jank. Cool idea but inbreeding is never the correct route.

>>101178331
when inpainting? i might as well be using 1.5, even with fooocus.

>>101178420
Yes? Why the fuck would you use 1.5 tho, not to mention faaaagus.

>>101178472
1.5 inpainting models dont completely lose the style of the rest of the image, while fooocus does an alright job, im not a fan of its interface, and im not sure if its inpainting model is compatible with pony.

File: Mass_Effect_Character_Cre(...).jpg (57 KB, 740x416)

57 KB JPG

Is there any way to influence facial features, like nose lengths, chin length/depth etc like a lora or something? Creating characters is really hard since you don't have many ways to finetune facial appearence bar some very general concepts like skin color etc.

>>101178546
What? They work perfecrly fine. The only difference is that there aren't that many controlnet models for sdxl, but they still work just as well.

File: 00006-2887668082.jpg (397 KB, 2096x1568)

397 KB JPG

>>101178366
Yeah. Loras are probably the best use case

>>101178595
you can add weight modifiers to regular shit like that. Just do something like (large chin:1.2) where the 1.2 is the strength modifier.

>>101178812
>Just do something like (large chin:1.2) where the 1.2 is the strength modifier.
maybe it depends on the model but it feels like these prompts make little to no difference, I usually get sameface

>>101169001
Fucking retard, moving to the torrent tracker is exactly what /ldg/ needs. Imagine trusting the frauds at Civit

>>101178993
Just make one yourself

File: nai inpainting example.jpg (515 KB, 1620x1024)

515 KB JPG

>>101178596
i wish i could get results like this locally.

File: 00012-3208744832.png (2.18 MB, 1744x1304)

2.18 MB PNG

>>101178825
Very nice. If you could prompt for people with muscular physique with similar style I'll borrow them

>>101179212
Looks like nothing special. I use inpainting all the time, even on the images that were not genned by the model. People underestimate how powerful it is, especially with controlnet.

>>101179163
I did

>>101179464
good! have fun

what is the point of this faggot thread and /sdg/ when both are dead

>

>>101179601
if you are not happy, then talk to the mods

>>101179423
is there a better controlnet model for inpainting than inpaint_v26.fooocus?

File: file.png (26 KB, 561x327)

26 KB PNG

thoughts on training on 1216x811 images instead of 1024x1024? i've always trained on 1:1 ratio pictures and i'm not sure how this is gonna work.

>>101179958
If you have buckets enabled, it doesn't matter. That was the whole point of them, so you could train on other image sizes without getting fucked

>>101180220
>it doesn't matter
are the results the same or does it just make it not crash or whatever? all my models have been trained on 1024x1024 images. if i dont have to crop everything it would save me a huge amount of time.

>>101180241
The results are the same. You don't need to be cropping, it just helps narrow down the focus of your training in attempts to limit it just to that subject.

>>101180220
The point of bucketing is to avoid unnecessary cropping.
>>101179958
You'll be limited to generating in that resolution

>>101179958
Are you way better off multi cropping your dataset at sane ratios.

>>101180250
surely smaller images are faster to train right? less vram usage? i always remove the the backgrounds around the subjects so having a bunch of empty space around them seems wasteful.

>>101180264
i usually generate portrait images on models trained on 1024x1024 and they have worked well so far.

File: Sigma-output.jpg (2.03 MB, 2088x3132)

2.03 MB JPG

>>101179353
>prompt for people with muscular physique
Okay

>borrow them
Just take them man

>>101180286
>i usually generate portrait images on models trained on 1024x1024 and they have worked well so far.
Doubt

File: Sigma-output.jpg (1.91 MB, 2088x3132)

1.91 MB JPG

>>101179353
>>101180408
I have men-y

File: 7stvln.jpg (98 KB, 1074x1057)

98 KB JPG

>>101180452
>>101180408

File: Sigma-output.jpg (1.39 MB, 2088x3132)

1.39 MB JPG

>>101180468
Just gym bros, but here, for you

>>101180495
that's better, thank you anon :3

File: Sigma-output.jpg (2.82 MB, 1748x3576)

2.82 MB JPG

Testing wideness

File: Sigma-output.jpg (2.26 MB, 1812x3576)

2.26 MB JPG

6 panes of bump

>>101180699
very wide, very wide

File: Sigma-output.jpg (1.28 MB, 1876x3192)

1.28 MB JPG

>>101180774
It gets wider

File: Sigma-output.jpg (2.51 MB, 2068x3192)

2.51 MB JPG

>>101180774
>>101180785

File: Sigma-output.jpg (1.03 MB, 2068x3192)

1.03 MB JPG

Dat thumbnail tho

>>101180892
this is cool

File: 0.jpg (649 KB, 1024x1024)

649 KB JPG

>>101180983
thanks

>>101181196
>>101181240
You finally got a graphics card to run them yourself?

File: image - 2024-06-27T182041.461.png (1.45 MB, 920x1296)

1.45 MB PNG

>>101181286
yeah

>>101181685
Let's see how good your workflow is. Your Christmas prompt back at you.

Dancing as a stripper laura kinney, nightclub, showgirl, gift wrapped body, ribbon around body, ribbon bondage, santa hat, "candy cane in her mouth", athletic, green eyes, long black hair, abs, detailed face, strip club, pole dancer, Christmas decorations

File: image - 2024-06-27T190907.796.png (1.79 MB, 992x1400)

1.79 MB PNG

>>101181750
first one without nipple peeking out

File: image - 2024-06-27T191031.655.png (1.79 MB, 992x1400)

1.79 MB PNG

>>101181799
lol

File: image - 2024-06-27T191707.981.png (1.38 MB, 1384x776)

1.38 MB PNG

ribbon bondage prompt usually taped up her mouth

File: image - 2024-06-27T192340.376.png (1.76 MB, 992x1400)

1.76 MB PNG

last one

>>101182089
pretty sweet

File: file.png (43 KB, 708x507)

43 KB PNG

>>101177480
Aha, maybe this is the reason, not using the variance terms, hmm... But that's not a good thing at all

File: UI_0004.jpg (1.96 MB, 1536x2304)

1.96 MB JPG

>>101181799
>>101181815
>>101181860
>>101181913
>>101181964
>>101182015
Very NICE!!!
Glad you could join us.
Don't know if you recall this one.

>>101181964
>>101182015
What model is used for these two?

>>101182203
I think I have it saved on my tablet

>>101182222
nice quads

Zonkey pony realism

>>101182239
Thanks, I recognized it but just wanted to confirm it.

File: image - 2024-06-27T194127.250.png (2.01 MB, 992x1400)

2.01 MB PNG

>>101182239
that last one might have been to spicy

picrel is now damnponyxl

Any news on generating realistic images of naked women that don't look like they are made of rubber? Is pixart good yet?