Blessings upon thee, young anon

>>101092621
>he made it to the /ldg/ collab

>>101092621
goddamn those are some good gens

File: 1girl_noclassic.png (1.47 MB, 1024x1024)

1.47 MB PNG

A strong point of Hunyuan is that it can generate non-fuckup hands pretty well

>>101093035
Low sampling steps on either DPM2 or DPM ++2M (can't remember). I think the error in the main one might actually be the best use of AI: Just dreaming up things real people wouldn't normally imagine.

>>101092621
>General dedicated to local usage of free and open source text-to-image models
Not complaining, but why the change?

>>101092930
how much vram does it use anon

File: Sigma_02418_.jpg (2.39 MB, 2048x2048)

2.39 MB JPG

>>101093308
Change is a constant in life and sdg couldn't adapt. Sigma gave us a glimmer of hope, and HunyuanDiT handed it to everyone. This is for all local models, including SD3. The question is, can you adapt anon?

Interesting... https://civitai.com/models/528620/ttplanet-controlnet-tile-for-hunyuandit?modelVersionId=587377

comfortable thread

>>101093308
its not a change. real thread is this way
>>101091912
this thread is a dead-end off-shoot

>>101093620
I meant the last thread had
>General dedicated to the discussion and development of local text-to-image models.

>>101093620
>The question is, can you adapt anon?
But point taken

>>101093936
pedos are over there

>>101094051
Vased.

File: Sigma_02431_.png (2.98 MB, 2048x2048)

2.98 MB PNG

>>101093944
>>101093978
Upon further inspection, a fair question. Trigger happy about migrants crossing the border. Wonder why the stance all of a sudden about openness

>>101094113
kek

File: 00009.png (968 KB, 832x1152)

968 KB PNG

sdultimateupscale throwing me mat1 and mat2 shapes cannot be multiplied errors :/
oh well

>>101092679
You're famous now

Any useful Auto1111 extensions worth using?

still looking for a png comparing extension

>>101093716
Very interesting.

File: tmpj4mxjm8s.png (182 KB, 398x474)

182 KB PNG

>>101092621

File: 1706273896789.png (1.43 MB, 1024x1024)

1.43 MB PNG

>spend eons trying to recreate with SD a style that DALL-E gave me with zero effort
>come back to board and see how unclean and uncrisp it is compared to everything else
>realize you've strayed further and further from the truth
it's frustrating to have an idea dangled before you and then snatched away

>>101095118
I highly recommend doing a hires fix/2nd pass on gens
with that said, that's a good gen

>>101095118
I've seen much worse.

File: 1703551699569.jpg (982 KB, 2048x1024)

982 KB JPG

>>101092621
oh no someone put that in the collage
that was an img2img of a bing prompt
the results i've been having recreating from scratch are dismal

>>101093830
nice

>>101095163
left one is way better.

>>101095678
That's the one that bing made. Right is img2img.

File: z9.jpg (1.03 MB, 1974x1974)

1.03 MB JPG

>>101091734
I asked for wood shavings and botched the upscale.

File: z10.jpg (995 KB, 1980x1982)

995 KB JPG

>>101095728
And i think she works well with brown tones lol.

File: image - 2024-06-06T000654.850.png (1.9 MB, 1024x1024)

1.9 MB PNG

>>101095118
>>101095163
Notice that it looks like fine art, more like a painting with loose brushstrokes. Secondly, the style of the character is not anime, but more manga-ish.

Now take a look at pic related. No, that is not me trying to recreate your style. That is a pic I genned with Hunyuan a while back, while testing its capabilities. I took the prompt straight from a Dalle thread, basically it's a "manga painting". To achieve Dalle results locally, you need a model that is good at both anime and manga, and Hunyuan excels at both. So it's not over for local, Dalle is currently only ahead in prompt adherence.

Now, as far as SD is concerned (at least XL, especially Pony), it's obviously a model that is not on par with either Hunyuan or Dalle, since it's not a good anime/manga base. You'd need a LoRA to achieve such results, though I'm not sure if that is even possible. I don't think you'd also have as much success replicating the style on something like SD3, because similar to Pony being an overbaked anime model, that just seems like overbaked dreamshaper to me. Your best option on such a model would be something like IPAdapter (or obviously a LoRA of just manga paintings).
Dalle style is not really a mystery, after all it's just a mixture of manga and paintings.

If you have the VRAM for it (6GB+), I'd say Hunyuan is definitely worth a shot, if not (or maybe it's too slow for you), then perhaps wait for 0.7B version.

>>101095163
left one is way better. (I'm a different anon)
Dalle 3 just rocks at SOUL, can't wait to have something that can learn how to make its outputs locally.

>Dalle is currently only ahead in prompt adherence
The eyes of your image are very poor and the whole thing lacks any kind of SOUL.
Stop pretending it's not miles behind Dalle.
I never cared about prompt adherence because I'm not looking to create some pic, I want to see what it does with a prompt, and more often than not it outputs something amazing that I could have never imagined!
Did you see my cover at the Dalle thread?
>>101075676
The Betty Boop on the left that Dalle generated is so good some anon thought it was an original picture and not AI, it's that good and no other model can come close to it (prompt in the filename.)

File: z11.jpg (993 KB, 1978x1978)

993 KB JPG

>>101095744
Last from me, got some shit to do today.

File: image_2024-06-22_082527186.png (939 KB, 832x1088)

939 KB PNG

stay safe

File: image - 2024-06-16T010935.996.png (1.46 MB, 768x1280)

1.46 MB PNG

>>101095917
t. Has never used the model.

We are roughly 90% of the way to Dalle. Hunyuan is already better at realism because Dalle simply is too censored there. Ignorance is bliss. Back to your containment thread.

>>101096302
Can Dall-E animate?

File: HunyuanDiT- Joel and Elli(...).png (1.27 MB, 1280x768)

1.27 MB PNG

>>101096314
That's the thing about it, it's cloud shit. You can't animate because the overlords have decided you can't. You can't use controlnet because they have also decided you can't. You are heavily limited across any "API" offering.

>>101092621
Got a question about training LoRA's on PonyXL using Kohya's.
There's a relatively obscure set of characters from an old 90's cartoon show that I want to make into character LoRA's for PonyXL, and nobody has done it currently. Problem is, there's almost no high quality art I can use for the datasets. Most of the fan and R34 art for them are trash or don't quite look right, so all I really have to rely on are stills from the show itself.

The res is like 720x540, and given it's 90's shit, the quality is pretty low. Upscaling makes it look worse if anything, so I'm not sure what to do. Anyone have experience trying to train LoRA's on low quality images like this and have any advice?

Keep getting an error with easydiffusion about not having write access to /dev/kfd despite being in the correct group.

>>101096434
I've done this before, but it required a lot of manual work, but this was my workflow:
1) Take 15-20 screenshots of the character from various angles.
2) Open each shot in PaintTool SAI 2. Watch a how-to video on the GUI, it's easy.
3) Trace the character, starting with the line art, then give it simple shading. Most 90's cartoons are flat colors with maybe some light shadows anyway, so it's not that hard. You don't need the backgrounds, just use a white/black background but make sure it's tagged properly as that in your captions.
And that's it. Yeah, it's a pain to have to manually draw that shit, but it's the only way. You can't upscale it without it looking like ass, and using low quality images will produce ass results no matter what you do.

File: image_2024-06-22_094926526.png (1.01 MB, 832x1152)

1.01 MB PNG

we need more emma's

Chang, what's going on with the pedo posting?

>>101096541
I don't have a stylus or anything though and wouldn't that take a long time doing them by hand?

>>101096756
You don't need one. SAI 2 has a curve and pressure tool so you can do it with your mouse, moving the points around to line up curves perfectly, then adjusting the pressure afterwards to give sharp/smooth end points. Like I said, it's easy. And no, it doesn't take that long to trace and color a simple character once you get the hang of it. Maybe half an hour or less per image once you know what you're doing. If you're making it for Pony XL, make sure the images are a minimum of 1024 pixels. And like I said, you only need about 15-20 images per character. Up to you, but I guarantee the method works for obscure 90's cartoon characters that have no good HQ images to source.
Tracing in SAI2: https://youtu.be/zYQkJSpTpbs?si=3hBRjkNtAmSYeKaH&t=181

https://civitai.com/articles/5840

>>101096820
>those answers from the lawyer
wellllllllllllllllllllll fuck
stay based civitai

File: ia_00021.jpg (487 KB, 1256x2712)

487 KB JPG

you're all cursed too....
this was supposed to be a refuge
instead, it is a

>>101096807
Nice, I'll give it a shot. Thanks bro.

>>101096820
>Below is a screenshot of the email we received back from our lawyers
>this means that for the time being SD3 will remain banned.
lmaoo, get fucked SAI, and thank you the gigachads at civitai

File: KING.jpg (8 KB, 210x240)

8 KB JPG

>>101096820
>Civitai is actively committed to helping creators monetize their work, it would be severely irresponsible of us to promote this model with the license as is.

File: image - 2024-06-22T043120.335.png (1.43 MB, 1280x768)

1.43 MB PNG

>>101096820
This is just a huge win for /ldg/. Burn SAI, burn. And let alt models flourish.

File: file.png (7 KB, 1860x52)

7 KB PNG

>>101093317
Here on RTX 3070 Ti, just enough

>>101095117
but why?

>>101096820
huh kek

>>101097256
me bargaining my soul with the furry devil for pixel generation powers

File: KEK.jpg (81 KB, 3494x496)

81 KB JPG

>>101096820
>Even civitai employees are making fun of SD3
BASED

interesting time we live in

>>101096820
>corpo flying the faggot flag has huge influence over the local scene
>>this is a good thing!

>>101095036
wtf she's so hot as a teenaged tomboy

I don't like the sound of that post >>101097329

>>101097351
I'm talking Kanbaru energy.

oh boy. are we going to get a whole nother round of SAI trolling? someone wake up memeanon

>>101097323
No one is perfect I guess, but I still appreciate that civitai is willing to kill SD3 once and for all, fuck SAI and their cucked models

SAI fucked themselves on this one. They could have simply not included such a gay license or better yet, they could have listened to the community and walked it back. But we all know they don't give two shits about the community.

>>101097359
>Kanbaru energy
>Monogatari
oh no, that's a high power level to begin with
>>101097365
I've been summoned, just give me a while for the sleepy-b-gone juices to flow in, so the creative ones can flow out. Besides, it's been relatively calm as of late. If shit hits the fan, it's best to just ignore off-topic posts and focus on something worthwhile instead.

>>101096807
why not use i2i and prompt for shading etc and use low cfg?

>>101096820
It sounds a bit performative on their part, to put pressure on Stability.

>>101095502
What the fuck are those eyes

>>101097501
it's BS, civit are censoring jews and most fall for this out of hate for sai

>101097501
>101097562
samefag btw

>>101097423
this, they got what they fucking deserved, I won't cry on their grave

>>101097470
You want it to look as consistent as you can for the initial training images, and the vast majority of 90's toons are flat colors. Plus, it doesn't matter because once you train the character LoRA, you change the art style via prompting and other LoRA combos anyway, you aren't stuck with the style. You only want it to learn the design itself.

>>101097584
Schizo

>101097677
get new material

should I keep using forge or go back to a1111?
I just moved to a new hard drive and now i'm getting some kind of dubious ownership error when trying to launch forge

>>101096820
>caring about lolcences
Kek what cuckolds

>>101097673
Fair enough. SAI 2 looks pretty interesting, super lightweight.

>>101097786
I still use Forge. Perhaps the next A1111 update will be the one I've been waiting for

>>101097823
anon, the pony fags spent tens of thousands of dollars making ponyXL, I think it's fair he deserves to get that money back, we can't advance in this field if people can't monetize their work

>>101097728
>get new material
yeah, SAI should get a new licence I agree with you on that

>>101097947
AH is a cuck too for caring about lolcencing, just do it who's gonna stop you?

File: Money.jpg (407 KB, 1962x1584)

407 KB JPG

>>101097999
>who's gonna stop you

>>101098012
No licensing around AI is legally enforceable

>>101098032
that's not how it works, if you decide to use their models, it's like you signed a contract, so they have the power at the end

File: Licence.png (191 KB, 800x529)

191 KB PNG

>>101098032
The lawyers say otherwise.

>>101098042
not how it works, and that shit definitely isn't enforceable outside of the jewS of A

>>101098063
>falling for the civit publicity stunt

>>101098099
so that's your counterargument?

File: torch.jpg (89 KB, 1321x508)

89 KB JPG

what is this torch error I wasn't getting this before

>>101096820
if SD3 was an amazing model, civitai wouldn't ban this for sure, they know they're taking zero risk as SD3M is too cucked to be saved in the first place

>>101098104
SCascade has the exact same license and civit isn’t banning it.

>>101098142
that's a fair point, so what's your theory behind all of this?

>>101097954
The fact that they haven't made any statements about it AT ALL is insane but as >>101097309 puts it these are interesting times
>>101098142
>SCascade has the exact same license
>exact same
Quote it for anon?

>>101098151
Already said it, publicly stunt and exhibiting their power to ban whatever they want to the applause of the easily fooled.
Don't cry when they start banning other models that produce harmful content, or come from a company they have beef with

>>101098173
but then why didn't they make this publicity stunt on SD Cascade aswell? It was the first model with such licence

>SCascade and SD3 have the same license
anon has said this before but i am highly suspicious that theyve even read it themselves

>>101098181
Because Cascade didn't have any hate towards it and banning it would've been seen as overbearing like it is?

I can confirm that SCascade and SD3 have significant differences between their licenses.

>>101098191
>Because Cascade didn't have any hate towards it
So basically civitai went on the side of the users instead of the company? That's kinda based

>>101098185
>>101098194
https://huggingface.co/stabilityai/stable-cascade/blob/main/LICENSE
https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/LICENSE
you're free to post the line that's different and causing the issue

>>101098235
Examine the document yourself. Do you see how it is different?

>>101098202
More like fishing for brownie points since SD3 Medium is shit. They'd turn this policy around in a heartbeat if 8B released

>>101098242
>Do you see how it is different?
no, because it isn't?

go back

>>101098289
yep, I agree with that, but we know we're not gonna get the 8b model, so civitai is 100% winning the brownie points with that move

>>101098242
what's different? give us an example anon

>>101098320
It's uhhhh, there's less line breaks!!

>>101098289
The implication is that Cascade is a good enough model to warrant a disregard for it's license and that is simply not the case kek

>>101098352
>and that is simply not the case kek
Except it is because it shares the license and I can still upload cascade models to civitai?

>>101098202
>>Because Cascade didn't have any hate towards it
no one cares at all about Cascade its barely relevant

>>101098376
>no one cares at all about Cascade
and the reason no one care is because of the licence

>>101098376
all the model trainers are moving to cascade since sd3 was a failure

>>101098376
that's a shame because that model has a better anatomy (and nipples) than SD3
https://civitai.com/images/9756975

>>101098142
Stable Cascade is not available for use with the commercial Stability License though.

File: 1718653780188478.png (349 KB, 1024x1024)

349 KB PNG

Why does anon rarely post cascade images or even discuss the model

>>101098427
because we know it won't lead to anything as no one is willing to finetune that shit do the non commercial licence

>>101098427
Cause SAI shilled SD3 at its release to kill their last good open base model

>>101098427
The only non-retard reason is that the images look too smooth due to the extreme compression.
It's a step back even from the SD1.5 VAE.

>>101098481
I really don't get those retards, instead of focusing on one single model and make it great, they're making a shit ton of useless projects (LLMs, audio shit, several different imagegen models in the same time)

>>101096820
>click link
>immediately see /lgbt/
>close page
who else?

File: 1719053170699849.png (1.75 MB, 966x1460)

1.75 MB PNG

>>101098627
vu will be living in a world ruled by the trannies
vu will be happy

>>101098848
nice

Nice

(you) are awesome
now make a square circle

(You) are nice.

File: 1692661875667287.png (2.51 MB, 1024x1536)

2.51 MB PNG

1girl

>>101099367
>(1girl_trisomy21:1.2)

File: 1713222481883792.png (2.44 MB, 1024x1536)

2.44 MB PNG

imagine the sex

does auto1111111 support sd3 yet?

>>101099652
Not officially, we're waiting for the sd3 branch and the dev branch to merge in a new main release, since dev branch apparently comes with optimizations the likes of Forge.

File: 1715963840982377.png (1.56 MB, 1024x1536)

1.56 MB PNG

1girl bump

I love 1girl

>>101099685
that can't come soon enough

>>101099068
>>101099121
>>101099190
TY anon

>>101092621
What do you use to make colleages like this?

>>101100130
>What do you use to make colleages like this?
https://www.befunky.com/create/collage/

File: 1691484283120251.png (1.51 MB, 1024x1536)

1.51 MB PNG

i love dolphin shorts and leotards so much bros

>>101100249
>dolphin shorts
didn't know there's a name for that, also funny they're both named after an animal

>>101100263
>didn't know there's a name for that

yeah, i learned it from some thread few days ago

File: tmpbt3pp7hp.png (671 KB, 770x1000)

671 KB PNG

>>101100276
good couple of threads ago I learned about dutch angle and skindentation, my life hasn't been the same ever since

File: UI_0002.jpg (1.04 MB, 1536x1536)

1.04 MB JPG

>>101100249
This one for you bud

There should definitely be a site that I can upload a bunch of files to do lora in browser. Who is making this?

>>101098132
Did you sandbox your shit and disable networking?

>>101100663
i think you can do that on civitai, it costs 'buzz' whatever that is

File: 1712436243193.png (1.38 MB, 1200x1080)

1.38 MB PNG

>>101100249
>>101100263
>dolphin shorts
oh yeah i've been using those
they can get a little... unexpected sometimes

>>101092621
How does NovelAI's "vibe transfer" work? It's pretty good at recreating characters from just a few images, especially their outfits, something Lora seems not too great at

>>101100894
I think it might be a more advanced version of something we've seen in IPadapter.

>>101101413
isn't the latest ip-adapter pretty good at that already? plus 2v something

>>101100702
Does Civitai have any restrictions to making LoRAs out of real people? like not celebs but like people that look normal/ordinary?

Wait for Kohaku to deliver Hunyuan training in sd-scripts: https://github.com/kohya-ss/sd-scripts/pull/1378
Meanwhile I'm still wrangling with my script on ipex...

File: tmppabee8t2.png (1.63 MB, 1280x1280)

1.63 MB PNG

>>101100249
don't mind if I do

File: 1701142859150102.jpg (811 KB, 3024x1728)

811 KB JPG

Last train to Redwall

How do you guys manage to get widescreen resolutions like 1344 x 768? TensoRT only lets me export for up to 1024x1024,768x768, or 512x512. Furthermore, models seem to only be trained with those.

Imagine just trying it out
Imagine not using the functionality if it doesn't allow you use a certain latent resolution

>>101102594
I'm still on the Forge branch of auto, with my 8vram laptop GPU, and it.. just works. I type in the resolution, press generate, ???, profit. Unless you mean TensoRT in particular, but then again, the models do normally support vertical and horizontal ratios. I just deleted a note with a list of all the SDXL supported resolutions, and 1344x768 was one of them.

>>101102594
tensorrt 1344x768 works with XL

>>101102828
me and the boys on our way to prank someone's doorbell

File: gagimageData.png (1.82 MB, 1024x1024)

1.82 MB PNG

looks like a lot of onetrainer sigma lora trainings still fail for me for some reason, either they collapse to black or do this

>>101103108
visualization of my last braincells trying to survive

>>101103216
Could be something like that. This isn't the best training data but it worked for full sigma checkpoint training as well as SDXL/SD LoRa, it failing with suspiciously many attempted settings/optimizers and so on is ... suspicious.

>>101103340
>>101103137
>>101102852
Is this based on the Vampire Hunter D movies? Style looks familiar.

>>101103345
yeh https://myanimelist.net/character/1020/Leila

>>101103108
try training in fp32 yet?

>>101103108
Official trainer best trainer

File: file.png (73 KB, 979x512)

73 KB PNG

Let see if this time is better...

>>101100223
>>101100744
>>101102296
>>101102442
>>101102852
>>101103716
I like

>>101103767
The official Hunyuan res appears to be 768x1280 for portrait, 1280x768 for landscape, per a statement on their github you may get poor results using hunyuan outside these resolutions, and I think that may also apply to training.

File: sxt4dzluz58d1.png (248 KB, 2352x982)

248 KB PNG

>The Juggernaut guy is considering finetuning pixart
We're so back!

>>101103868
Don't worry, different from SD, Hunyuan also use image size as embeds

File: 1714914308889111.jpg (1.27 MB, 3024x1728)

1.27 MB JPG

>>101103787
ty anon

File: Sigma_02473_.png (3.41 MB, 1536x2560)

3.41 MB PNG

>>101103767
You get a hype

>>101103929
You get a hype

EVERYONE GETS A HYPE!

>>101103929
I think he's making a mistake, like the pixart devs aren't finished pretraining their models, he should wait a bit more before going in the finetuning part

im so fucking hyped right now bwos

>>101104187
finna brap, finna brap @ u're

>>101104246
too dam hot to eat

Florence 2 is really good at captioning. Could see a workflow of automatically making (lora) dataset captions by just generating them with it desu.

>>101104367
>Florence 2
can you run it locally?

File: file.png (32 KB, 748x253)

32 KB PNG

>>101104404
yeah
https://huggingface.co/microsoft/Florence-2-large

>>101104367
is it better than CogVLM? Can it caption NFSW?

>>101104367
Didn't Meta also release their own captioner? I wonder how it compares to that.

>>101104188
If anything it would serve as a proof of concept. But you're correct that it seems the pixart devs are not done.

>>101104340
sus cloud

>suddenly struck with Lord of Vermillion
ded.jpg

>>101105332
lol

Slow Sunday wait it's Saturday

>>101103108
you could work with those kind of outputs desu
something something abstract something something

>>101105423
I'm pretty sure /sdg/ is on its 3rd thread today

>>101104466
It's clinical with NSFW descriptions but doesn't ignore them either. You will get things like a nude woman with her legs spread for an sensual image. Its general captions are good and what makes it good is you can do them at 0.7s, full long form captions and it's decent enough quality, enough for the purposes of fine tuning. But the main advantage is pure speed at the expense of control and a tiny bit of accuracy. It essentially only has three options, short, shorter and novel for caption output. It also has no pop culture knowledge.

>>101105545
yeh and 0 posts worth reading like usual I bet

>>101105648
haha fuck /sdg/

>>101105648
yep, I wonder why I'm wasting my time there, I got my news on those threads but when nothing happen it's just a frustrating experience

>>101105648
>>101105678
>>101105684
Samefag

>>101105332
funny guy

>>101105648
>looking to read posts on an images thread
You okay there?

>>101103108
Onetrainer does undocumented stuff behind the scenes that you have no control over. I don't recommend using this trainer.

>>101105985
1girls growing ripe this season, I forsee a good gen harvest

What measurements are you going to compare next?

File: file.png (87 KB, 1538x674)

87 KB PNG

>>101104367
holy shit bro, this shit is so good!
Thanks for reminding

File: 1697442224644619.png (997 KB, 1216x832)

997 KB PNG

>>101100744
Same

>>101092621
local LLM anon here. Is there anything like Dream Machine locally yet?

>>101106054
>What measurements are you going to compare next?
Maybe influence of steps on inpaint/img2img behaviour.
>>101106184
>Is there anything like Dream Machine locally yet?
https://github.com/hpcaitech/Open-Sora

>>101106293
>playing with soap
that would've NEVER even occured to me, I'd sooner prompt for someone to eat it

>>101106302
Kek, I'm sharing the artist, not the prompt

>>101106217
>>Is there anything like Dream Machine locally yet?
>https://github.com/hpcaitech/Open-Sora
I'm guessing it's shit because I never see anyone post funny videos from it

>>101103501
To begin with, that OOMs on settings that should be normal.

>>101103664 >>101105901
Looks like it but the official trainer also ain't it given how it doesn't have the same optimizers, training image manipulation stuff and so on (not gonna program it all myself).

>>101106452
this is cool, can you do a water park?

>>101106452
gay man lives here

>>101106414
model?

File: file.png (103 KB, 979x512)

103 KB PNG

Let see how well this florence caption works, heh

>>101106446
>To begin with, that OOMs on settings that should be normal.
well, is there an gradient checkpointing option?

>>101106573
>well, is there an gradient checkpointing option?
Already activated. I could easily train pretty larger batches while doing other stuff on the other trainers, looks like here I can do batch 1 at best when terminating everything else. Guess I'll have to take it if it works but that's a huge difference comparatively.

>>101106626
never mind even that OOM'd, I guess OT will just not really work.

File: image - 2024-06-22T191645.983.png (1.32 MB, 768x1280)

1.32 MB PNG

>>101106517
HunyuanDiT

>>101106384
a good video model that makes funny garbage is good?

>>101106573
>Let see how well this florence caption works
seems pretty great

File: image - 2024-06-22T192227.466.png (1.69 MB, 1024x1024)

1.69 MB PNG

>>101106480

>>101106757
awesome

https://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiT
^^^ this thing supposed to work?

>>101107009
I think it was supposed to work but HF can be overloaded or w/e so just run it locally.

you guys can have debo back

>>101107284
>>101107246
hello kitty kawaii pink latex

>>101106830
nipple department says this is an offence

>>101106757
woah, not bad at all

>>101106785 >>101106877
nice style

File: image_2024-06-23_021657357.png (1.24 MB, 896x1280)

1.24 MB PNG

>>101107368
this is a christian channel , no pleasure knobs allowed

>>101107243
Lol

>>101106785
Nice tits, too bad that's a man.

>>101107402
>nice style
ty, it's just 90s animes

File: image_2024-06-23_023854292.png (1.34 MB, 896x1280)

1.34 MB PNG

>>101107352

>>101107432
;_;

>>101107659
Oof

i want to train loras goddamit, whos bingus to i have to suck for a gpu? preferably a time traveler with a 50 series.

>>101107739
its stretchy dont worry, will fit if shes preggo too

File: file.png (95 KB, 979x512)

95 KB PNG

God damn Intel XPU, running Florence on it will give garbage results after some time. Now I'm forced to run Florence in CPU in order to get stable results.

>>101095831
Looks amazing, what was the prompt for this? All anime gens I did with Hunyuan came out garbage, or straight up like western art, so I assumed it wasn't trained on anime in the end

File: Hunyuan1719096857-cfg7-st(...).png (1.42 MB, 1024x1024)

1.42 MB PNG

>>101108273
>All anime gens I did with Hunyuan came out garbage, or straight up like western art, so I assumed it wasn't trained on anime in the end
What, not that guy but here is my gen, absolute not garbage. Hunyuan can do anime straight out of the box. Remember, this is a chink model, bro!

prompt = "Artstyle oil painting. Sakuma Mayu from Idolm@ster Cinderella Girls. Wearing a dress with long sleeves and ruffles at the hem. Sitting on a bed. In front of a wooden bookshelf filled with books. danbooru tags: brown_hair, blue_eyes, hairband, ribbon, breasts, short_hair, bangs, earrings, bow"

>collage anon asleep

>>101108416
Looks like it.

>>101108459
>>101108459
>>101108459

File: image - 2024-06-22T200425.089.png (1.13 MB, 1024x1024)

1.13 MB PNG

>>101108273
Hunyuan base is capable of many different styles in anime, E.G. you can prompt by a particular artist such as Makoto Shinkai or Hayao Miyazaki
Manga artists
Kentaro Miura
Takehiko Inoue
Tsutomu Nihei
etc...

Even I don't know the full list of artists. I have tested a very limited amount and posted about it here (the above). If you want to see some interesting prompts, https://imgur.com/a/hunyuandit-0vrZEn0

The keywords I tend to use are
anime,
anime screenshot (sometimes)
and depending on what I'm going for, I might add "aesthetic", "cute", if it's a single subject prompt adding white or plain background to negative might help if you don't want that. There are many variations of keywords and styles you can use, I recommend to look around for guides how to prompt Niji because the models can be prompted very similarly (appending anime to the start of Hunyuan's prompt). You'll find that they're not prompted the same, as you have to be more creative or precise with Hunyuan but you can probably get similar in aesthetic results and have more control over Hunyuan. Aside from that look for popular anime/manga to get an idea of what to prompt and get more control over styling (aside from training LoRA).

The particular prompt I used for that was an experiment,
>From below, heavy impressionistic brushstrokes, detailed manga painting by Yoshitaka Amano: 15 yrs Siberian woman (raccoon dog ears and tail, long straight black hair, fringe, brown eyes, fur-lined siberian attire) sitting before a campfire, night. Her expression is soft and relaxed.

You may or may not get consistent results with such a prompt. Also I haven't tested anything below 40 steps and I tend to go as high as 70 sometimes. Sampler is ddpm (from the demo as I use TensorRT for faster inference).

>>101093308
I was in your shoes a couple weeks ago. Turns out the other thread is just 5 discord trannies using that channel as their public discord channel. Note that this thread has less discord faggotry

File: image - 2024-06-22T173857.766.png (1.21 MB, 768x1280)

1.21 MB PNG

>>101108570
Keep in mind one of the limitations of Hunyuan is you can only do 77 tokens in your prompt, so ensure your prompt is an concise as possible. (no more than 30-40 words in English). Chinese translations for anime prompts also tend to perform very well.

File: 1 (2).png (369 KB, 512x512)

369 KB PNG

I used to use nemusona's waifu generator from a while ago, but that's been down so I want to try and run stuff locally. I know it used to use anything v4.5, but beyond that im pretty lost. I tried fooocus, but those dont come out the same, even with anythingv4.5 as the refiner. I was thinking of trying to learn comfyui, got any tips on getting started (specially in relation to making the style of picrel)

>>101109127
>as the refiner
What's it using for the base gen? Also anythingv4.5?

>>101108570
Keep in mind it doesn't know any artist perfectly and the quality of your gens may be lowered depending on the subject matter, so a finetune/lora is probably best to actual nail a particular style.

>327/138
:(

File: 7 (6).png (358 KB, 512x512)

358 KB PNG

>>101109303
on foocus, I could use anythingv4.5 as the base as anything isn't sdxl, I would get an error when I tried to set it.