File: ldg-is-aids_2024-06-15-112510.png (2.11 MB, 1344x1728)

2.11 MB PNG

>caturday
>no cats

File: 000000_13450_.png (2.22 MB, 966x1460)

2.22 MB PNG

>>100989219
>TY Baker

>>100989352
based and vaxxedpilled
fuck pureblood retards
still waiting for these nanobots to activate

first for julien is shit

Cursed thread of discordfaggotry

>>100989361
didn't like the forcing and bs, media, work masks all that. retarded,

File: ComfyUI_temp_pmzjp_00019_.png (1.81 MB, 1280x1024)

1.81 MB PNG

I love the flycatcher strips

File: image_2024-06-15_183900583.png (778 KB, 768x1280)

778 KB PNG

just one more thing...

File: ComfyUI_temp_pmzjp_00020_.png (1.58 MB, 1280x1024)

1.58 MB PNG

Close

File: 00094-3814149481.png (1.69 MB, 1024x1024)

1.69 MB PNG

>>100989221
salted raw kohlrabi

https://cfgpp-diffusion.github.io/
another free upgrade

File: Default_Mr_Whiskers_the_M(...).jpg (308 KB, 704x1024)

308 KB JPG

>>100989282
kut kot

>>100989615
cute

>>100989602
even the cherry picked examples are bad and add other issues to the gens, another snake oil.

>trending on artstation

>>100989638
>>100989455
make her blonde.. like that scammer bish

>>100989219
Giant forehead

Is there any more new sai gossip like this? >>100986578

I know we've had the Retro Diffusion model leaked, but has anyone leaked the Aseprite extension? That seems to be it's real strength.

Is it over?

>>100990014
>I know we've had the Retro Diffusion model leaked, but has anyone leaked the Aseprite extension?
the what and what?

>>100990101
always been m8

>>100990014
no, I want that too

>>100990120
https://astropulse.gumroad.com/l/RetroDiffusion
This shit. The model was posted, but not the extension that makes it work with Aseprite.

File: ComfyUI_temp_dxcpk_00010_.png (2.07 MB, 1280x1024)

2.07 MB PNG

>>100990101
No, we are back.
This might not be on level with Dall-e3 but it can't be taken away.

>>100990148
just fyi, that hose is also a whip

File: fine.jpg (123 KB, 1024x1152)

123 KB JPG

look at the bright side

File: 000000_13453_.png (2.27 MB, 966x1460)

2.27 MB PNG

>>100989834
oooh, didn't notice, good eyes.

>>100990301
catbox pls?

File: ComfyUI_temp_pmzjp_00047_.png (1.74 MB, 1280x1024)

1.74 MB PNG

It won't make a six pointed star. But a Lora can be trained for that.

>>100990344
catbox is down atm, looking for alternate

>>100990468
thanks, post whenever np

>>100990468
is that a /sdg/ thing to not just copy the json to pastebin when catbox is down?
no wonder all good genner anons left

File: ComfyUI_00012_.jpg (300 KB, 832x1216)

300 KB JPG

Hmmm...

>>100990437
it definitely does lol (the star of david)
it made one when I was prompting Illuminati shit

>>100990477
https://pub.microbin.eu/upload/fish-koala-cat

i feel like im being gaslighted
is the whole "just add more trending on artstation and it will be good" thing a meme?

File: Refine_114727731545456_00203.png (2.02 MB, 1248x1824)

2.02 MB PNG

>>100989219
post girls lying in the grass NOW

>>100990301
I cant believe its AI holy shit

>>100990544
remove "letters" from negative and there's a doubled description in T5 encoder.

>>100990587
I use SDXL Supir to upscale as well.seperate workflow.

File: 00179-1572671910.jpg (204 KB, 1216x832)

204 KB JPG

>>100989819

File: thats-it-yes-thats-it.gif (1.39 MB, 640x640)

1.39 MB GIF

>>100990584

>>100990555
it is, about a year ago some study even tested what these common meme prompts even do, turns out "concept art" and "trending on artstation " are almost synonymous. Which is no surprise if you think about it.

File: ComfyUI_temp_pmzjp_00058_.png (1.95 MB, 1280x1024)

1.95 MB PNG

Perfect.
Btw, I am Netflix executive, if you hadn't noticed.

File: cascade_woman_lying_on_th(...).jpg (147 KB, 1024x1536)

147 KB JPG

>>100990584

>>100990636
prompt?

>>100990679
>Nigger Jew
Can this exist? kek

File: ComfyUI_HunYuan_00047_.png (1.43 MB, 1024x1024)

1.43 MB PNG

Bros, I think I've taken the Hunyuan pill

>>100990699
https://en.wikipedia.org/wiki/Beta_Israel

>chinese shills wake up

>>100990730
>says the SAI chill

>>100990730
It's 2am in China, retard

File: Screenshot from 2024-06-1(...).png (69 KB, 862x602)

69 KB PNG

>>100990704
>no controlnets yet
:c

>>100990738
>>100990743
oh shut the fuck up no one is buying that zhang

>>100990764
https://www.timeanddate.com/time/map/

>>100990770
i know what a timezone is but you want to act like persistent shilling cares about the time of the day? are all chinks this retarded?

File: ComfyUI_HunYuan_00051_.png (1.45 MB, 1024x1024)

1.45 MB PNG

>>100990730
Go back to the discord Lykon

>>100990787
debo you are killing /sdg/ with this :(

File: file.jpg (20 KB, 374x365)

20 KB JPG

>>100990704
>bros i think i've taken the hunyuan pill

File: Screenshot from 2024-06-1(...).png (167 KB, 917x834)

167 KB PNG

>>100990756
oh, never mind

where my 1.5 chads

File: 00189-2015856396.jpg (163 KB, 832x1216)

163 KB JPG

>>100990691
https://civitai.com/models/478568/black-souls-style-and-pony-xl?modelVersionId=532361

are we back or is it over?

File: tmpca5r50b1.png (657 KB, 768x768)

657 KB PNG

Getting a "realistic" noncartoony creature is hard in Pony, despite having source_furry, cartoon and anthro in the negatives

>>100990938
never left

File: ComfyUI_00110_ (1).jpg (484 KB, 1728x3840)

484 KB JPG

hello

File: resolve_out.webm (2.29 MB, 1024x1024)

2.29 MB WEBM

>>100990879
I still use 1.5 for AnimateDiff.

>>100990948
There's realpony and everclear, although the results can be uncanny.

>>100990956
that's my gf

>>100991001
kek

File: ComfyUI_temp_ialkp_00009_.jpg (477 KB, 1728x3840)

477 KB JPG

>>100991015
n-no?

File: tmpxfm6e6ac.png (662 KB, 768x768)

662 KB PNG

>>100990966
Might keep that in mind, though it looks like at least one of those is meant for human(oid) subjects

>>100991001
We're back! omg..

File: comfy_res_o_00014_.png (873 KB, 768x1152)

873 KB PNG

Does the license actually prevent people from saving SD3?

>>100991148
yes, ponyXL creator said he won't make an SD3 funetune, so no v7
he's making a second SDXL version called ponyv6.9

File: ComfyUI_SD3_0066.jpg (2.35 MB, 1664x2432)

2.35 MB JPG

the backside of astronaut helmets will never be fixed

>>100991148
>>100991169
but that's partly because it requires a paid license for commercial use, and ponyXL guy offers cloud generation services to recoup training costs
so technically someone could do it, but the monetary incentive is gone

>>100991148
only if you agree to it

>>100991148
no it's fud

File: ComfyUI_00016_.jpg (444 KB, 1216x832)

444 KB JPG

>>100990879
Home

File: ComfySD3_00009_-1718477882214.png (772 KB, 2050x2050)

772 KB PNG

Morning anons

>>100991229
sexo

>>100991229
Is that JM sneaking back in?

File: comfy_res_o_00073_.png (916 KB, 768x1152)

916 KB PNG

>>100991169
>>100991199
>>100991252
>>100991287
Why are SAI so content with shooting themselves in the foot

>>100991413
stop exposing me

>>100991414
it's their right I guess, if this was important then it wouldn't be the disaster it turned out to be

File: ded.webm (152 KB, 1024x1024)

152 KB WEBM

>>100991414
I think they were bleeding money hard (and talent too I think, Comfyanon bailed on them), so they needed a more aggressive monetization strategy. Maybe more of a "grab what you can off this sinking ship" strategy?

File: angel_0020f.jpg (1.45 MB, 1792x2304)

1.45 MB JPG

>>100991414
They're not making any money like most AI companies and the bills came due. You can blame their CEO for squandering the (ample) money they had on bad project direction and not securing additional capital by being obviously clueless when he spoke to potential investors. Now they have no choice but raise capital somehow or they go belly up.

>his gens are only good when hes using a good lora

>>100991567
>nooo, you can't use the best tools available!

>he needs to rape the models weights to compensate for his bad prompts

File: 00000-4152924225.png (1.27 MB, 1024x1024)

1.27 MB PNG

>super saiyan bayonetta
I definitely see the Toryiama influence, but I don't really see any Bayonetta in here at all [spoiler]besides the titties[/spoiler]

>>100991694
>>100991721
neat

File: 00552-2042263744-9.5-20-2(...).jpg (1021 KB, 992x1792)

1021 KB JPG

>>100991806
thank you

Can I get a quick rundown on sd3

File: 00001-849212417.png (1.39 MB, 1024x1024)

1.39 MB PNG

>super saiyan pikachu
what the fuck

>>100991803
>A1111 naming system
That's not SD3, is it?

>>100991897
why would it be anon unless they were using sdnext

File: 00002-2298414908.png (1.52 MB, 1024x1024)

1.52 MB PNG

>>100991897
i have no idea how to update, so i'm using a really old version.

>>100991916
Are you using SD 1.4?

File: super saiyan mewtwo.png (1.58 MB, 1024x1024)

1.58 MB PNG

>>100991933
Not sure, honestly. How do I check?

>>100991706
are you talking shit about me again?

>>100991838
It's ass because they're trying to suck up to investors who are afraid of boobies with the resulting neutered model being shit and making SAI look like a bunch of retards

>>100992042
What checkpoint/lora? That looks sick

>>100992111
Is it the investors or payment processors?

>>100992146
it's a very old SD1.5 merge I have on my pc
I don't think it's anywhere on the web
sorry about that

>>100990141
>>100990014
Where is the link for Retro Diffusion model?

so checking in is there a training fine tuning scriopt for 3 yet?

Have you guys heard about the swedish game studio that offers $15,000 to whoever gets to fix img2img and inpainting in Layer Diffusion? I just saw a few people talking about it on a forum but I can't post the link here because it's a lolicon forum.

>>100992488
yes

>>100992513
where

>>100992007
It should be on top of the webUI.

>>100992553
Why do you care? You aren't going to do anything with it. If you can't even perform a basic google search you sure as fuck aren't going to train something.

I spy tft doing some more morgana action

>>100992569
Cry

>>100992648
:'(

>>100992160
I suspect they proactively botched the model out of an extreme abundance of caution and not because anybody in particular forced them to

File: succ_0027f.jpg (863 KB, 1792x2304)

863 KB JPG

What is the difference in SD3 between clip_g and clip_l? I know the other one is supposed to be for boomer prompting. Comfy's example puts the same thing in both g and l and it is background info

>>100992973
AFAIK it's the same setup with SDXL and people performed some tests and it yielded it is better to mirror the prompts in clip l & g.

>>100993020
heyyyy there's the ottermaid

File: tmpiwpxl4jj.png (641 KB, 768x768)

641 KB PNG

Lineart quality is crap, but I like the pose and the creature on the left here

>>100991317
>morning
you live in ozzy?

File: horses.jpg (1.32 MB, 2688x1536)

1.32 MB JPG

>>100993012
I see, that makes sense, thanks

File: tmp1_sgd8c3.png (689 KB, 768x768)

689 KB PNG

There's not many pictures with a mermaid with an otter tail

File: stormy_coast.jpg (1.13 MB, 2688x1536)

1.13 MB JPG

>>100993291
An ottermaid, nice

>>100992973
it's the third one for boomer prompting, I think?

>>100993606
It's the same prompt for all of them because that's how it was trained

>>100991706
>He's unhappy that no one's downloading his garbage finetune

Just got the news on forge after spending much of the last two weeks getting back into SD. Is the only option for fancier samplers and decent performance comfyUI now? My eyes glaze over seeing the squiggly lines and boxes.

>>100993805
sdnext is an option
a1111 is getting the performance tweaks from forge so there's that too

File: meadowlands.jpg (954 KB, 2688x1536)

954 KB JPG

>>100993821
>a1111 is getting the performance tweaks from forge so there's that too
Really? That is good news

>>100987013
In the last 6 months we got PonyXL with is a significant improvement to the SD model.
We also have multiple DiT offerings, proper competitors for local diffusion.

File: rise_milf.png (1.17 MB, 864x1152)

1.17 MB PNG

bimbo rise kujikawa milf bros (literally just me), we're back...

File: Untitled.jpg (66 KB, 767x349)

66 KB JPG

I forget what causes this kind of artifacting, does this seem obvious to anyone?
I thought it was the cfg or step count being too high but I adjusted them and it's still happening.

>>100993805
Just copy someone elses workflow. You don't have to tweak it if you don't want to.

File: duck.jpg (1.43 MB, 2688x1536)

1.43 MB JPG

Ok in these landscapes I've seen basically no artist reactivity from SD3. Haven't tried any super old ones like 1800s and earlier but I think they curated most all the living artists at least from the dataset. Or maybe they just all opted out like Greg R did by now.

>>100994234
Cfg too high. Compensate with more steps or lower cfg

>>100994330
Dead or alive, doesn't matter they aren't recognized. Styles and artistic movements are also out or at least the model is much less reactive. If you want a specific style other than anime and perhaps cubism, you have to find a different way to do it. I am using the output of the SD3M model as basis for img2img with an SDXL based model.

>>100994395
Thanks Anon. I had it mixed up and lowered the step count instead of raising it. After bumping it up a smidge it seems fine. (was doing 30 when I needed 32).

File: 1699721682636982.png (445 KB, 512x640)

445 KB PNG

>>100987865
The more control you have over style, the better the model is if you actually care about style. Inherent artist styles, loras, ipadapter, vibe transfer, reference only, stylize, embeddings, model merges, you name it. All of these things compound on each other giving you many orders of magnitude more control over styles, meaning you can be more creative and make cooler stuff. But external styling is extremely rigid and annoying to work with compared to inherent styles and without inherent styles you have many orders of magnitude less control unless you're an autist that creates/downloads hundreds of loras and somehow catalogues them or makes wildcards with them. And literally no one does this, I use more loras than most so I know this nightmare and I hate it.

Were you around at the beginning when people were using handfuls of artist and celebrity names in their prompts? Whenever you added or removed 1 of the thousands of terms(you could even use random names) your image or the subject's face/look would usually totally change. Giving you ridiculous amount of variety in how your pics looked. It was easy to make new styles since we had all the artists catalogued pretty well and if you knew about art or just googled some type of artist you could approximate on that style quickly. I really liked how you could blend anime with classical or cook up new CG styles. But now we basically just have model mixes and loras and as a result AI art is a lot more bland and samey than it was when you could effortlessly craft. When you use a lora or even 10 loras in comparison it feels more rigid because you're just overlaying some style over the entire model, instead of exploring a region of the base model's latent space that looks a certain way. It's hard to explain but modern models without too many artist calls are insanely lobotomized when it comes to potential artstyles and character appearances.

>>100994445
Above is related to your post too.

>>100994445
I've had more luck describing the techniques and styles used rather than solely relying on artist names.

Right now I'm currently trying to make a wildcard which captures most of these techniques and visual effects

File: 1694617394236911.png (441 KB, 448x768)

441 KB PNG

>>100994669
I ran out of space but I wanted to also say that this is why anyone removing names or artists from a model is an absolutely soulless retard that doesn't give a single fuck about art. It doesn't stop people from making loras which is what they pretend to want to avoid, it just makes the models infinitely less visually interesting and mutes the potential creativity people can express. Forcing people to rely on models only exasperates people using very specific art styles, because they have no other choice since no one wants to weigh 5+ loras against each other, it's annoying. They'll just use 1 or 2 leading to stale gens. People using NAI are a great example, they have inherent styles and the model is capable of some amazing images but most of them barely use 1 which makes the images generic.

>>100994824
>it just makes the models infinitely less visually interesting and mutes the potential creativity people can express
I disagree, it was a crutch but your point is noted.

>>100992351
https://huggingface.co/LostMedia/RetroDiffusion

File: ComfyUI_SD3_0076.jpg (1.62 MB, 1664x2432)

1.62 MB JPG

>photo of a dog and a cat both standing on a red box, with a blue ball in the middle with a parrot standing on top of the ball. The box has the text "SD3"

I got tired of bad hands and did one of these

File: 1713628149376103.png (447 KB, 512x640)

447 KB PNG

>>100994853
Why do you disagree? It's just like SD3's anatomy issue, well it's actually a much larger issue. But with SD's censorship, when you remove NSFW your anatomy is going to be worse. But when you remove a style, you remove a part of the digital brain that your generations can swim around in. I say it's worse because it's not just one style because you can compound those styles onto each other, creating a near infinite amount of completely unique styles that no one has ever seen before. When you use loras you're using a style people have seen before. And if you use loras together you'll still probably be able to see aspects of existing art styles. But with inherent styles, since the signal on them is lower, that isn't as much of an issue.

This affects even the biggest model Dall-E. The outputs are usually generic looking because you don't have a lot of control on the styles.

File: ComfyUI_17897_.png (1.27 MB, 1024x1024)

1.27 MB PNG

ZootShill here. I think my favorite thing about the model at this point is how almost any gen could conceivably be from it, e.g.

File: ComfyUI_17892_.png (1.32 MB, 1024x1024)

1.32 MB PNG

>>100995016
But also! Same model, both direct gens, no lora, no nothing

>>100994970
What I don't get is why can't people fix this with finetunes.
SAI gives us new technology and a botched base model, but people could continue training the model to add everything that was removed or not in there in the first place, just pick all those artists works and let the model learn their styles.
Like Pony but with the intention of making the model produce all these styles and concepts.
Dalle 3 is incredible because of all the knowledge of characters and concepts it has, I wonder how long will it take for open source to approach it, it's like it almost wouldn't need loras.

>>100995016
Googling ZootShill gives me no results to I don't know what you're talking about, create an entry on urbandictionary or something.

>>100995016
what model ?

How do you guys handle converting between 1.5 and Pony/SDXL resolutions? I set up a workflow to convert Pony 864*1152 to 1.5 1024*1536 by upscaling 1.33 and cropping, but the crop really bothers me.

>>100995034
The amount of training required is prohibitively expensive to do for free.

File: ComfyUI_SD3_0083.jpg (1.5 MB, 1248x1824)

1.5 MB JPG

not the best freckles but I've seen worse. definitely not the best ripped shirt tho

>>100995064
its zootanon (zootvision model)
I tested it before and it has (well had it might have changed in the recent update) a heavy bias towards wolves/dogs when trying to generate animals.
Outside of this it's an interesting 1.5 model

>>100995078
Upscale by an exact resolution instead.
I go from 512x768 to 768x1024 to 1024x1536.

>>100995092
what do you get when prompting for "skin imperfection" rather than "freckles" ?

>>100995118
Aren't you supposed to only use one of the standard SDXL resolutions? Or is that not a big deal anymore?

>>100994970
I don't disagree and I am not trying to practice any SAI apologia here just working within the constraints we've been given. Surely you can look at it that things have been taken away, and yeah they have, but it also doesn't mean they didn't provide the tools in order to replicate similar style, aesthetics, composition, visual effects, and overall vibe of an image.

>>100995116
Thanks, will try it out. For versatility LiberteRedmond was a model I loved, too bad it basically was censored on the NSFW department, so I'm hoping someone some day makes a NSFW version of LiberteRedmond (I tired but it's either good NSFW realism and the animation goes to heck, or good animation NSFW and the realism goes to heck, and having 2 extra models that aren't versatile defeats the point.)

lmao it wouldn't
here anyways:
https://civitai.com/models/490451?modelVersionId=573612

File: ComfyUI_SD3_0086.jpg (1.52 MB, 1248x1824)

1.52 MB JPG

>>100995122
definitely more subtle

>>100995076
https://civitai.com/models/490451?modelVersionId=573612

>>100995086
I mean, Stability AI could do it because they were funded and people donated money to them like crazy and they got investors.
Why nobody else could do it? Heck, why don't the guys at huggingface or Civitai try to become the next SAI like this?
Stability AI did this and that and swam on money, yet nobody else followed suit.

>>100995116
This is the kind of feedback I repeatedly have said I want people to leave in comments on the page lol. I didn't like intentionally train it on wolves and dogs more than anything else. Are you saying it replaced something that wasn't one in your prompt with one, though? Anyways maybe for V6 I'll put some more focus into highres animal photography, it's not something i've actually focused on yet, been moreso addressing like landscapes, people, and highres human-on-human NSFW, and so on.

>>100995179
poor girl lost her thumb in the mixer

>>100995208
I was using wildcards for the animal generation. So it could have been a heavy association with "creature" to dogs/wolves.

File: 1701312038608589.png (288 KB, 720x480)

288 KB PNG

>>100995034
Because it's a big data problem. It's like turning SD into an anime model, if you want to do that well you basically have to use all of Danbooru and even then it's not perfect. And then people who do this, like Pony, fuck with the styles which to me is just pissing money down the drain. He's training on the images but with fucked up tags that get the model confused and retarded and unresponsive, it's so stupid.
SAI supposedly removed at least 80 million images from the dataset, you can't even guess how many less styles there are in the model now, styles that people could have stumbled upon and merged with others. Like how many classical painting styles are there? To make a model that could express all of them would be like making a model larger than Pony Diffusion and instead of SAI having that in their model by default which they could have done, they removed the nodes because they're completely retarded.

>it's like it almost wouldn't need loras.
Right. That's what a base model should be. Something that emboldens creativity and gives you tools to create any number of original works. The base model SD is going for now is practically useless for this, it's main use case appears to be this>>100994914 for marketing or some shit. Dall-E too, it knows a lot but it always pulls you towards a relatively few amount of styles because it's writing your prompts based on what you ask, it's not listening to your prompt precisely.

I have no idea when open source will be capable of making a really cool model that knows everything but I'm not really hopeful after this release. Lately I've noticed people are excited for those Chinese models but I don't see how they're much different than what we have already. Base models and 999 loras just clog up hard drive space, it's gay.

>>100995144
But I want new styles, not copy pasted vibes that's what I'm trying to get at. I want it to be easier and easier for more people to make unique gens. We're moving away from that.

>>100995231
hmm, interesting. what exactly was the non-wild part of the wildcard, then?

>>100995092
>>100995179
What's your style prompt here? Looks like a proper photo (minus the deformities)

File: ComfyUI_SD3_0084.jpg (1.65 MB, 1248x1824)

1.65 MB JPG

>>100995263
>selfie
>probably taken from a phone

File: ComfyUI_17899_.png (1.49 MB, 1024x1024)

1.49 MB PNG

>>100995253
also yeah I think I will put some focus into highres photos of various animals, I'm getting what I ask for when I prompt for whatever but the rate of limb or other errors is certainly higher than things I HAVE trained for, of course.

E.G. attached is a ZootVision Epsilon elephant prompting for like the most realism the model can do, not bad but the ears are weird and arguably its missing a tusk, and the image overall is less realistic than it'd otherwise be if i'd trained on that subject. So something for me to look at for sure.

File: grid-0149.jpg (956 KB, 2496x3456)

956 KB JPG

>>100995253
here's a grid from some of my testing
here's a pastebin of the metadata since catbox is down https://pastebin.com/grquxH4g

SD3 LoRAs are starting to come out.
Seems like it must be pretty easy to train then?
Doesn't seem to have support in OneTrainer or Kohya though so, how to train?

File: sunflowerhedgehog.jpg (1.06 MB, 1992x1120)

1.06 MB JPG

>>100995208
>I'll put some more focus into highres animal photography
Could you make it good at animals in other styles generally? We still don't have a model that outputs something like Dalle 3's Sunflower Hedgehogs (see picture) and it's good at photorealism. Maybe it's impossible and Dalle 3 just decides to switch to its cartoon model or its photo model depending on the prompt, but I still dream of a model that can react to animals the way it does (children book illustration style as default, photorealistic animal only if you specify it.)

File: Screenshot 2024-06-16 120943.png (3.97 MB, 2121x1149)

3.97 MB PNG

>>100995311
forgot pic

>>100995234
>But I want new styles, not copy pasted vibes that's what I'm trying to get at. I want it to be easier and easier for more people to make unique gens. We're moving away from that.
Again we're at an agreement here too. Though we already (model agnostic) have a lot of tools at our disposal to create new styles. How many are actually using prompt interpolation? How many even know about it? It's an extremely powerful tool (one of many) that I don't think people are aware of or have explored.

In the end, like you're eluding to, you have to play to the lowest common denominator and make things easier rather than more difficult for the masses to be creative or express their creativity.

>>100995359
I could take him

What would you be doing with the last year of your life if SD never existed

>>100995302
thick dickskin elephant

>>100995234
I agree with everything you say, except I'm optimistic: over and over I've seen heroes come out of nowhere and give us what we want. Like the hero that leaked novelai and changed the world.
Because the technology already exists to give us what we want, heck, Stable Diffusion 1.5 already got plenty of styles and variety and character knowledge, it just doesn't know what to do with them properly, and when you train it to do some things well it forgets how to do others at all.
All we need is someone figuring it out and being able to produce a model that fulfills our dreams without so many loras, maybe our unknown hero is already cooking...

>>100990733
Is this her?

>>100995372
Good question, but there was a 6 month gap where I basically produced nothing before SD released, even, before craiyon released, it was like I was burned out, in the last year I've done more things than in the previous decade put together, so SD filled the gap and motivated me to... generate stuff.

>>100995372
Wasting a shit load of time hand drawing reference images for 3D modeling instead of just generating 100s in a couple of a minutes and picking the best ones to sculpt.

>>100995453
Without SD and AI in general I'd probably just be obsessing over viral evolution. I haven't recovered from burnout but the world might not have much time anyway, so I'll 1girlslop until oblivion.

>>100995498
one man's slop is another man's treasure

File: 1708105938348075.gif (3.88 MB, 896x624)

3.88 MB GIF

>>100995365
The bar is just way too high with those tools. I have over a terabyte of models and loras(and I just purged a lot too lol) so I've explored lots but others won't do that. And even with that terabyte I know that a good base model with inherent styles is better than 5tb of loras.
>How many are actually using prompt interpolation? How many even know about it? It's an extremely powerful tool (one of many) that I don't think people are aware of or have explored.
Well when you have the option people will use it, when SD released we all had nothing else so we were swimming through latent space and it was cool and fun.
It's not just about the masses fiddling though. It's something that really benefits the enthusiasts too. And those enthusiasts are the people who inspire the masses. Model merging is a good example, gigabrains started that and there are now like tens or hundreds of thousands of base models made by the plebs that followed in their footsteps. That's one way of creating your interesting unique style. My problem is that it's obviously much more annoying than playing with inherent styles.

>>100995432
I hope so but I'm in doom mode right now.

>>100995476
That's fucking cool anon. AI should be used to help with other arts/projects.

>>100995372
Video games and movies have been boring so maybe I would have become extremely radicalized by politics.

>>100995308
what version were these done on, for reference?

>>100995550
This was done in gamma

>>100995316
i can for sure do more dalle training. "by dalle3" and "by midjourney" are actually things it always recognized, although there's not nearly as much highres data there ATM

>RAanon cooking as usual
always a treat to see

File: Screenshot 2024-06-16 123456.png (3.59 MB, 1909x944)

3.59 MB PNG

The T5 part of the prompt is the only bit that matters.
Left has the same verbose prompt in l, g and t5,
Right has the same t5 prompt as left, with l and g only being key words.
Why is it split into 3 separate prompt sections when literally only one of them matters?

>>100995597
My experience has been the opposite. The clip l&g are take priority over whatever is passed to t5.

For a crude test, put in "alien" in the t5 and "woman" in clip l & g and see how it goes.

File: RA_2_00213_.jpg (778 KB, 1920x2808)

778 KB JPG

>>100995579
Thanks, honestly this is just old stuff I never posted.

File: SD3_16624_00018_.png (1.68 MB, 1024x1024)

1.68 MB PNG

>>100995632
Hmm, you might be right, regardless why the fuck do we have 3 when there's next to no value? I am just going to link a primitive to all 3 and use that because it's so pointless.

>>100995566
to go into a bit more detail, the underlying basis (like the checkpoint that I first released first as "Alpha", is a Very Good™ "full anime model" that I won't directly name cleanly merged 50 / 50 with a Very Good™ "full photorealism" model that I won't directly name, injected at 0.6 strength with CluelessC's `hll6.3-a10-eps` Lycoris and also quite a number of unreleased Lora's I trained myself, at various strengths.

From there, I've continued to iteratively train and inject on top of each released verson, which is proving to be very cost-effective and practically effective I think. In total there's *at least* 20,000 images worth of custom training by me in there, as of the most recent version.

>>100995666
>I am just going to link a primitive to all 3 and use that because it's so pointless.
This is probably the best route until we can figure out how to wrangle these to be honest

>>100994824
>>100994669
prompt for those elf cuties

Also can you gen the following:

sfw (masterpiece), best quality, ((by Franz Xaver Winterhalter)), ((Albert Lynch)), [[Serge Marshennikov]], (((absurdres))), (big breasts:1.4) woman, (navel focus:1.2), (silver hair:1.3), (wide hips:1.3), (facing viewer:1.3), detailed face and eyes (obese:1.1), vineyard, 
Negative prompt: 
nsfw (mutated hands and fingers))), ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed face))), ((ugly)), ((bad anatomy)), (((bad proportions))), (((extra limbs))), extra face, ((double head)), ((extra head)), (((extra feet))), monster, (text), (logo), [blurry], (penis), borders,  Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 4032651720, Size: 640x1024, Model hash: 2241e516, Denoising strength: 0.7, Eta: 0.67, First pass size: 0x0  berry mixed with wd1.2 at 0.3 

>>100995714
that's purdy

>>100995700
I think they came from some ancient model mix like nai+wd and I'd have trouble finding them again, do you really want them? And no thanks all my old models are in the closet right now.

>>100995700
Damn that's some old school prompting style. Why extra brackets instead of x:1.2 ?

File: SD3_16624_00021_.png (1.48 MB, 1024x1024)

1.48 MB PNG

>>100995700
this prompt in SD3

>>100995700
SD3
https://files.catbox.moe/s6bnyq.png

>>100995786
would

File: 1703601343118161.png (475 KB, 448x768)

475 KB PNG

>>100995747
>Why extra brackets instead of x:1.2 ?
They came from a time before that precision was added, we could only stack brackets. Those old prompts are broken because of this unless you enable the "old emphasis" something or other option.

>>100995785
>>100995786
delete this

>>100995786
Wow it actually put the nipples in the right place.
This was the first gen I got but I cherry picked the "better" one
https://files.catbox.moe/7nmebv.png

File: ComfyUI_17901_.png (883 KB, 640x1024)

883 KB PNG

>>100995700
NTA by ZootVision gave this for your (somewhat lopsided DESU) prompt lmao

ima SPECIFICALLY train it a bit on Franz Xaver Winterhalter for V6 just for keks I think

>>100995741
yeah it was WD 1.2 with SD 1.5 but with the old merging method

>>100995747
Extra brackets was the old way of doing stuff like x:1.2 back in the day. I just grabbed that prompt from my word doc of favourite gens that another anon posted.

>>100995799
Yeah I know where it's from, I was here back then, just wondering why you never upgraded it to the new style.

>>100995813
"but", not "by"

>>100995807
I got lucky with the nips. the others are more in line with your results

>>100995820
>just wondering why you never upgraded it to the new style.
I remember that period. It was aids and didn't always work. Some of our prompts were weird, well most of them were.

>sd3
>prompt in different language
>it generates people which predominately use said language
what do they mean by this?

what is the best current way to do accurate pixel art? Not just a lora but making sure it has same size pixels and shit

>>100995700
Pixart
https://files.catbox.moe/qtj00b.png
https://files.catbox.moe/up2c1s.png

Heres one for the team:

(masterpiece) (best quality) (highest detail), (watercolor:1.1), (jean-marc nattier:0.8)(sugimori_ken:1.3), ([Scarlett Johansson|Emma Watson] [ginger] (milf:1.3) (large breasts:1). dynamic pose)
Negative prompt: 
(((loli))), out of frame. (bad anatomy). (((bad proportions))). (bad hands). (mutation). (deformed). blurry. (fleshpile). (flesh merge). (tiling). (tile). (glitchy). text. error. (missing fingers). missing legs. missing breasts. (mangled fingers). (mangled legs). extra hands. extra legs. (extra digit). (fewer digits). fat. ugly. lowres. asian. worst quality. low quality. normal quality. jpeg artifacts. signature. watermark. username. artist name. loose clothes.
Steps: 23, Sampler: Euler, CFG scale: 6.5, Size: 768x1280, Model hash: 38c1ebe3, Denoising strength: 0.61, Clip skip: 2, ENSD: 31337, First pass size: 0x0

I left out the end part of the prompt which is just clothes, model is anything v3
What will you make her wear with yer model ? o.O

>>100995916
very village has a slampig

>>100995925
every*
god damn it

>>100995016
I don't like its default face desu

>>100995938
yeah the gen looks better with WD 1.2 mixed with berry mix.
Works too with SD 1.5 with WD 1.2

I miss the good old days D:

>>100995916
I will die on the hill that Pixart is overrated as fuck for being something that literally cannot run without more than 20GB worth of text encoders. Like it's SIGNIFICANTLY heavier resource-wise to run locally in practice than SD3, the number of parameters of the image model by itself is totally irrelevant

>>100995954
Only an issue if you try running FP32. FP16 works fine and uses less than 6GB VRAM.

>>100995118
>I go from 512x768 to 768x1024 to 1024x1536.
What's the process? Why not straight upscale by 2x?

>>100995954
Tencent's Hunyuan is Bokehed to hell melty mediocrity for the most part also, and I'm not even one of those people who irationally hates even Tencent's provably legit open source contributions, it's just honestly not that good

>>100995989
Because then I would get deformities. I will try find my old 1.5 workflow for you, it's been a while since I've run a 1.5 model.

finally someone else who realizes it
>>100751068
models have regressed. what we have no is more clear but less aesthetically pleasing. the worshipping of ugliness. AI is dead.

>>100995980
VRAM isn't the concern, system ram is. You are fucking retarded (or perhaps just aggressively American) if you think that the majority of people running SD locally have 32GB or more of it.

>>100996018
Not American and 32GB is the minimum for a gaming PC. Are you actually running 16GB RAM? You know RAM is fucking cheap as shit?

>>100995989
>>100996000
https://litter.catbox.moe/wfy5bn.png
Enjoy my spaghetti

>>100996018
>32GB ram $80
>GPU $1K+
nigga what

>>100996018
wtf are you smoking

baaaaaaaaaaaaaker

>>100996062
OK I will bake nobody fucking do anything

>>100996036
"minimum for a gaming pc" yeah ok there bud, in what what universe is this true unless you have a laughably unrealistic view of reality. Lke anyone who would actually claim this is a complete fucking moron, end of story.

I forgot the title because I am an actual retard never let me bake again

>>100996092
>>100996092
>>100996092
>>100996092

>>100996079
>I am exceedingly poor from a third world country
OK

>>100996100
doesn't matter <3

>>100996079
oh yeah also, next you'll be telling me that VRAM *amount* is the ONLY thing that matters. It "makes no sense" that a GTX 1060 is several times slower no matter what than a GTX 1660 Ti in Comfye, despite their equivalent VRAM amounts, apparently, this is really something I've seen people claim, seemingly not understanding the differences in architecture and other specs between those two cards.

>>100996075
I was about to bake with RAanon because he was the only person to make something good in the entire thread instead of SD3 shit but oh well

>>100996135
just got here, who is RAanon?

>>100996044
Holy spaghetti