Thread 42639894

AI Art Thread #69: Nice Edition

▶WARNING
>By posting in these threads, you recognize that AI generation is a valid artistic process.

Backends:

>https://github.com/LykosAI/StabilityMatrix
This is more of a one-click manager that you can install / run the other backends though. Easiest by far. Recommended to download Forge through it and use its built-in Civitai browser to download models with.

For models not through the in-built downloader just download and drop them in Stability_Matrix\Data\Models\* (lora folder if a lora, TextualInversion if embedding, stablediffusion if checkpoint...)

>reForge
https://github.com/Panchovix/stable-diffusion-webui-reForge
I would recommend installing this through Stability Matrix.

>ComfyUI
https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#installing
Has a learning curve but is extremely customizable and usually has the latest methods / papers implemented first.

>"But I don't have a decent GPU"
NovelAI also does great pony: https://novelai.net
You could also use a service like civitai.

Models:

>Pony Diffusion V6 XL
https://civitai.com/models/257749/pony-diffusion-v6-xl?modelVersionId=290740

>NoobAI-XL Great new alternative. Is closer than novelai than anything else. Really diverse yet coherent positioning / styles, knows artist tags.
https://civitai.com/models/833294/noobai-xl-nai-xl

>ZoinksNoob, a popular new model built off ChromaXL Mix, which itself was built off of NoobAI. Very, very close to NovelAI and arguably better at many things, it does ponies very well and has much less of a "burnt-in" default style than something like Pony Diffusion V6 does.
https://huggingface.co/zatochu/ZoinksNoob

>Useful artist style LoRA:
https://civitai.com/models/317578/pdv6xl-artist-tags?modelVersionId=356175

>LoRAs by /mlp/:
https://rentry.org/ponyxl_loras

>LoRAs by /h and other useful info:
https://rentry.org/ponyxl_loras_n_stuff

>How to make your own LoRA:
https://rentry.org/59xed3
https://civitai.com/articles/91/how-to-correctly-obtain-images-for-a-dataset
https://civitai.com/articles/3522/vaIstrixs-crash-course-guide-to-lora-and-lycoris-training
https://civitai.com/articles/4100/how-to-create-a-style-lycoris-for-ponyxl-on-low-vram-locally

>Prompt tips for new users:
Grab an image you like the style of from the desired model's civitai's site and drop it onto the PNG info tab in SD, you can then click send to text to image to reuse its generation info for your own images. Change the seed so you do not get just the same image. Use tags from e621 and natural language.

>No, AI art is NOT stealing.
Here, all the usual misinfo already addressed in one easy place for your viewing displeasure:
https://www.youtube.com/watch?v=SVcsDDABEkM
https://threadreaderapp.com/thread/1564878372185989120.html
For the love of Celestia, DO NOT FEED THE TROLLS, hide their posts instead.

>Previous Thread
>>42566653

>Past Threads
https://desuarchive.org/mlp/search/text/%22ai%20art%20thread%22/type/op/

>>42639894 (OP)
Wouldn't you drink her moon juices?

>>42639907
What do they taste like?

>>42639907
Yes.
>>42640039
Blueberry milk. It has a secret ingredient.

>https://www.youtube.com/watch?v=6Adcl7nXWuU
>put image in, get vidya out
Oh boy, while the stuff shown here is janky as fuck, it still appears better than that Doom/Minecraft ai demo from last year. Now just imagine when they optimize the shit out of it so it could be usable on basic 8 VRAM GPUs.

It's spooky month.

>>42640204
>that neck
She's built like a watchtower.

>>42639894 (OP)
>AI is a valid artistic process

Lol. You are not an artist having a computer generate an image from text. You cannot even copyright AI artwork as it has no authenticity and is not an original creation. Get over yourselves AI sloppers

The bone mare is here for your bones!

Could I request a pic of poni(es) standing in middle of pumpkin field at midnight, with the spooky forest in the background?

How many tries does it take to generate a video with Wan using img2vid? I want to try it, but I kind of don't want to spend thousands of buzz on videos that end up completely wrong. Making anthro porn btw, so pony anatomy shouldn't be an issue.

>>42639894 (OP)
Any chance one of you fine proompters might be willing to gen an image like OP pic, just with Luna happily selling her bottles of liquid shadow/darkness and moon juice?
>>42640381
>"Now, you wouldn't mind parting with some of your blood would ya', sugarcube. 'Ah reckon you won't miss it much."
>NMM bone armor
I love it!

>>42640838
Whatever you say, meatie.

Has anyone tried Pony V7 yet?

>>42640867
>How many tries does it take to generate a video with Wan using img2vid?
I'd say 2-3, but I'm not sure it'll let you include nudity.

>>42640381
Throw her an apple. That should distract her.

>>42641236
We need more photorealistic ponies & humanized/EQG ponies. The one thing that truly can't be done without AI help.

>>42641571
Not yet, have you? I was waiting to see if it would be released publicly.

>>42641727
>more photorealistic ponies
Yeah.
>& humanized/EQG
No.

>>42641571
*crickets*

>>42640039
Salmiakki Koskenkorva

>>42642172
Context?

>>42641236
What prompt do you use to make this?

>>42641615
According to the Civitai doc, the 2.1 version runs in-house and shouldn't be filtered.

>>42640838
>posted in the thread
Thank you for recognizing that AI generation is a valid artistic process.

>>42641236
Is the creepiness intentional? I would not want to have sex with this Applejack.

>>42642483
It's a little uncanny valley, probably even more when it's moving (see last thread's Loira). Would still bang though.

What e621 tag do I need to use to gen this kind of face?

https://twibooru.org/3184160

>>42642570
"naughty face"?
https://e621.net/wiki_pages/7855

>>42642570
ahegao

>>42642585
Oh that one might work, I knew the boorus had that tag but I didn't know e621 had it too. Thanks for the tip!

>>42642591
No, ahegao is different, usually has eyes rolling up and a smile rather than a grin/smirk.

>>42642633
To get the specific expression in your example, you could try either "naughty face" or "bedroom eyes", together with "blush, raised eyebrow, open mouth". Maybe even "looking pleasured", since that has 264k hits.

>>42642675
Yeah, I just tried with "blush", "tongue out" and "naughty face" and got pretty close. Gonna try adding "saliva string" and maybe some tags like "lustful gaze" or "one eye half-closed", see what that does.

I know this is caused by prompt bleed, but how do I mitigate, for example, petals appearing in the image when I prompt for Pipp Petals (using YiffyMix in this case) without inpainting? Do I just put "petals" in the negative prompt or does that reduce the model's accuracy to reproduce the character?

>>42642472
Oh, I see. I was thinking of WAN 2.0 over on the WAN website. I haven't used the original, so I have no idea how many tries it would take.

Another one rises!

I haven't forgot about this. I haven't had that much time and it's very tough getting everything even in the right ballpark since there's so many details that have to be right in the background. Posting this one since I thought it interesting.

>>42642172
>>42642401
PDXLv7 is released on Civitai (but only for online generation, not download): https://civitai.com/models/1901521

The comments and the usage ratio on Civitai ("Dormant", which is what the image refers to) should tell you enough.

>>42643294
Actually, update: it's currently "Buzzing" (extremely high generation usage) so I guess people are at least trying it out.

>>42643294
Download should be available next week. Hopefully that will allow people to use it in checkpoint merges with other models like Illustrious and NoobAI.

>>42641727
This one would take the very best artists months to draw.
>>42642468
This one can get a pretty decent show accurate look: https://pastebin.com/avmJkKD5 (Prompt for pic rel.)
Only possible using Invoke AI though due to incompatible Loras.
>>42642483
For that one, yes. I'm always trying to find ways to decrease the creep factor in these.

>>42643442
Thanks! Gonna play around with that a bit later.

>150 steps
Isn't that a bit excessive? Civitai's online generator doesn't even go past 50.

>>42643464
With an SDE type scheduler it gets more detailed the more steps you have, while other types of schedulers cap out at a certain number of steps. This is crucial for photorealism because things in real life have infinite detail.
I'd recommend 100 minimum for this kind of stuff.

>>42643294
>PDXLv7
>Ponies look like shit
LMAO I guess this is what happens when you change your focus to anime/ 3D/ anthro content.

>>42643302
Early feedback seems mostly negative:
>From what I've seen so far, and from the tests I've done, it always seems to be performing worse than or at best equal to the Illustrious XL base, which isn't exactly a good sign.
>I've never been more disappointed with the release of a model in my entire life. So much wasted hype, just for subpar generations on an even worse on-site generation suite
>sigh I guess the issue here was Aura. Chroma is better and worth working with instead.
>Not even better than old SD1 models
>The model is a disaster, and if anyone gets the idea to show it in a native image comparison here, it will be deleted without comment!
>Unfortunately, this model loses out to the V6 and even more so to models based on the Illustrious.
>Getting a decent image is a struggle. Lots of bad outputs. It is unclear what style cluster to use for desired style. As I understand score_ tags do not work as expected in this version so it is unclear what quality tags to use. Maybe 7.1 will be better.
>Born as dead base model nobody asked for. Results are worse than from early sdxl anime models
>This sux, lmao.
>this would go hard if it was 2023
A few users aren't so negative:
>I see a lot of people clowning on the model, and of course some of it is warranted, but looking at the example images below, I have to admit I'm impressed.
>The hate on ts unfinished model is crazy.
>Guys please give this some time. i remember when the original sdxl was released it felt like garbage to me because i did not know what to do until loras and finetunes came out . ive never used the base models but from there things go on !
This seems useful if anyone wants to experiment with it themselves:
>The main problem is that Pony7 works with tags really worse, than Pony6. Pony7 somehow can work with long captions, but when typing everything short you almost always get bad results. It needs longer prompts unlike Illustrious.

I don't know how much stock to place in these evaluations. People seemed to have genned a lot of anime girls, and there are some futa anthros, but not a lot of actual ponies. The pony gens look pretty good, see picrel.

Kinda incredible that Astralite Heart was so autistic about the V7 tagging that they made v7 autistic itself. You need to describe exactly what you want in excruciating detail in a very rigged format to get it to produce something better than dogshit. What the hell is the point of sinking so much time and money into producing a model that isn't even an improvement over v6 at making MLP style images of ponies? That goal is suppose to be the guiding star. Congrats, your model is probably better at backgrounds (when fed a prompt with 10x more details about the environment), now try to generate a style-accurate frame of MLP. Oh what's this, your current rube-worthy "style-cluster" system means I gotta manually plow through thousands of artist-agnostic approximations of traits that an OCR tool deemed to be the "same style" so that I may have a quivering CHANCE of finding a "cluster" superficially resembling the THING THE MODEL WAS NAMED AFTER? Sounds like you've wasted the time of everybody involved. It's really telling that there isn't a SINGLE show-accurate image on the civitai page for the model.

>pic unrelated, generated by Google Gemini just now by directly asking for MLP's style in the prompt

>>42644143
>That goal is suppose to be the guiding star
I thought he sold out. Isn't that what Fictional is? There's a fuckton of money around anything AI these days, and tbdesu I bet I'd sell out too.

>>42643836
All the gens are so blurry. I also haven't seen a single really good pony image yet.

>>42644258
I figured, what's the harm in trying? So I tried to write a prompt that would produce something like the image in >>42644143. My first try was:
>Show accurate angry My Little Pony Rainbow Dash standing on a cloud at dusk. In the background is a rainbow and hazy clouds. Her wings are spread and she has one hoof raised. She wears a blue baseball cap with a logo of a lightning bolt in a cloud.
As a negative prompt,
>Photorealistic, SFM, 3d, anthro, human
Everything was left at the defaults: CFG 3.5, Euler, clip skip 2, 25 steps, 1536x1536. Seed was 1982053091. I have no idea if these are anything close to reasonable for this model. Literally this is my first attempt with Pony V7.
The result is shit, honestly. Forget about being show accurate. She has fucked up ears, twisted wings, a barrel that's too short and legs that are too long. It's a mess. Did I get unlucky? Did I prompt it wrong? Or does it even matter?

>>42644323
I used to recognize AI generation as a valid artistic process, but now I'm not sure.

>>42643689
This looks like a reverse version of Godzilla.

>>42643106
Ooh, it's looking neat. Good stuff, Anon.

>>42643836
Did you use AI to summarize this?
>Guys please give this some time. i remember when the original sdxl was released it felt like garbage to me because i did not know what to do until loras and finetunes came out . ive never used the base models but from there things go on !
"just wait for other people to fix this mess" LOL

>>42644974
AI is great at some things, but I can still cut and paste better than it can.

>>42644143
the fact you can wrangle it to have anything similar to friendship is magic style at all, is better than base illustrious. if it can do poses or if loras are significant enough it might be good. pretty useless right now without a local download though, trying to 'figure it out' through online gens sounds awful.
all that said, ponyv6 wasn't good either, but made a great base for loras. i assume there is some legalese reason v7 is the way it is so hopefully there is a 'friendship is magic' addon lora

>>42645530
It should be strictly better than v6 out of the box. Is it?

Grok video is pretty fun
http://haystack.yerf.org/GCOB5S.mp4
http://haystack.yerf.org/3w4Fnm.mp4
http://haystack.yerf.org/dthfIy.mp4
http://haystack.yerf.org/DQp45h.mp4
http://haystack.yerf.org/3iREax.mp4
http://haystack.yerf.org/BCQeGP.mp4

>>42645548
At ponies? Doesn't seem like it.

>>42643302
Update: after the initial high usage yesterday because of the release, it is once again "Dormant" (not currently being used for generation).

The fuck is a "style cluster"?

>>42646454
I think it's a group of artist styles that Astralite thinks look close enough to count as the same thing. However, with a system like this, it means a model will never fully accurately reproduce a certain artist's style (likely on purpose).

Honestly, I don't even need the artist tags (I use LoRAs for this, tags are unreliable when using low CFG scale) but the decision to remove them just to please traditional artist (who are going to hate you regardless) is truly braindead.

>>42646229
>Wasted 100k+ dollars on this
How? Did he donate to them or something?

>>42646528
Also, isn't using numbers for these "style clusters" incredibly susceptible to prompt bleed?

>>42643294
Discord trannies are working overtime to boost this pooprompt model.

>>42647432
Even disregarding the subject of the images, the right image just looks bad. I mean, look at that hand.

>>42644143
>You need to describe exactly what you want in excruciating detail in a very rigged format to get it to produce something better than dogshit.
>when fed a prompt with 10x more details
>>42644323
It's because he trained the trashmodel on gpt describing the image, kek.

>>42647625
I may be in the minority in saying this, but I much preferred the tag format of the previous model (and all other models I have used). It just seems way more intuitive - I just list everything I want to see in the image (and what I don't in the negative prompt) and usually get something good with a couple of tries. It may just be the autist in me, but using this kind of flowery language to describe what should appear in the image just seems like a ton more work.

>>42646528
>that Astralite thinks look close enough
I thought he used some kind of clustering algorithm for this. It was always a stupid idea, because you can't make useful clusters without knowing what you're clustering, and you can't do that automatically without a model, and the whole point of making clusters was to help train the model. It's a chicken-and-egg situation.
>>42646820
It depends on how it's done. If you give each cluster name its own token, then you'll be fine, because the token is functionally a one-hot encoding of which cluster you're in. But I suspect he didn't do that, and in all likelihood, he did cause prompt bleed.
>>42647798
I want both tags and natural language. There are things you can't express with tags, like, "She wears a blue baseball cap with a logo of a lightning bolt in a cloud." using tags. But if I want a picture of Rainbow Dash, then I want to just write "Rainbow Dash" and have it know what I mean. When Chroma first came out, I tried prompting some MLP characters, and in order to get them right, I had to say things like, "Rainbow Dash, a blue pegasus pony with magenta eyes and a rainbow mane and tail" and even then it sometimes generated a pony that didn't quite look right.

>>42646528
That does sound awfully lot like a regression.

b

>>42648319
Artist recognition isn't the only thing that's regressed. For my second ever generation with Pony V7, I tried:
>Pony Twilight Sparkle sits in a plush easy chair in a well-lit library. She is reading a book. Her horn is glowing, and she is concentrating on the book. The back wall of the library is several adjacent tall bookshelves all packed with colorful books. The scene is in a show accurate My Little Pony style.
Same negative prompt and parameters as before.
In a way, it's better: Most of the picture is background, and nothing in the background jumps out as wrong until you look closely. Except maybe the chair because it looks like it's been upholstered with horse hair. But Twilight looks absolutely terrible. Her horn is like a crushed paper towel tube, and the glow is weirdly irregular. Her head changes height at the horn. Her eyes are slightly derped, her nose is a misshapen blob, and she doesn't really have a mouth. Her limbs don't have hooves. And the style is nothing like show accurate.

>>42650603
For comparison, I asked ChatGPT to generate an image using basically the same prompt:
>Generate an image of pony Twilight Sparkle sits in a plush easy chair in a well-lit library. She is reading a book. Her horn is glowing, and she is concentrating on the book. The back wall of the library is several adjacent tall bookshelves all packed with colorful books. The scene is in a show accurate My Little Pony style.
I wouldn't call it perfect, since her right ear and hoof are missing, but it's a lot better.

>>42647200
>>42647560
What a hot mare.

>>42650603
What a fucking disgusting chair

>>42639894 (OP)
https://vocaroo.com/1klQ5mM7hTtH

>>42646528
Removing the artist tags absolutely crippled the pony model. AI art looking the same is the biggest complaint, and modern Pony literally enforces it.

And dumbest of all, like you said, the haters are going to hate the AI anyway. Removing the artist tags destroyed the model for literally no reason.

>>42650978
Everypony stand up to /mlp/'s national anthem

>>42650603
That Twilight looks like the very first models that vaguely knew what a pony looks like.

>>42650992
Another dumb decision was the "opt-out system" he employed. No doubt tons of pony artists filled out that form to have their work excluded from training. And there is literally no need. Unless an artist can prove you trained the model on their images (they can't) they have no legal basis to complain anyway. In fact, the legal basis they're using is on very shaky ground to begin with: an AI getting trained on images is the same way a regular human is "trained" to copy a style (which by itself cannot be copyrighted anyway), namely by "observing" and "learning" (the result of that process being the weights of the models). There is literally ZERO data of the original artists' work in the model. Like the AI misinfo link in the OP says: "It doesn't have the textbook, it studied for the exam!" But most importantly of all, these people you are trying to appease hate your guts anyway, and they are probably overjoyed to see him ship a model that's worse than the previous one (meaning they actually got what they really wanted - which is simply to see AI fail).

>42650992
>42651681
>NOOOOO I can't just plug in my plagiarism anymore?
cry about it.

>>42652593
Thank you for recognizing that AI generation is a valid artistic process.

>>42652606
>troon logic
thank you for admitting you're a tranny

>>42652928
I see /jdb/ has been cooking.

>>42653449
>source: my ass

b

Fictional is a fucking slap in the face to his OG fans. The pics of ponies fucking suck hard, like intentionally bad. Please don’t look and give that prick a download. Just take my word.

>>42650603
I wonder if Astralite removed all screenshots from the show from his dataset.

>Rarity desperate to stay relevant by uploading a clip of her AI generated likeness.
>Not realizing she uploaded the wrong segment
>Trying to convince herself no-one will notice the change.
https://files.catbox.moe/p7nokp.mp4

>>42656587
wow, so now we can redo all post s3 shit without twilicorn

>>42656587
Very nice. Looks quite authentic.

>>42645991
>https://haystack.yerf.org/3iREax.mp4
WTF!?
Are you saying me you just prompt something like "Cute children Applejack inviting me to llok at the stars" and it gives you a video with background music and a voice close to her!?

>>42650609
Why doesn't it works for me?

>>42658060
I just put in "applejack invites you to watch the night with her"
Found it more interesting that i managed to replicate Applejacks voice

>>42658140
Damn, technology really goes fast...
We had pic rel a bit more than a year ago...

I was wondering, can you feed the last image of your video to it and generated the follow up?
I know it won't be perfect, but I'm curious...

>>42658174
Sure, but it's a lot of redoing since control is limited
https://haystack.yerf.org/XIM9DD.mp4
Aside from that it's pretty good and rather uncensored
https://haystack.yerf.org/OE7uaA.mp4
https://haystack.yerf.org/QH9Cxf.mp4
https://haystack.yerf.org/VfA1Ty.mp4
https://haystack.yerf.org/a8Tz2L.mp4
https://haystack.yerf.org/2xaIBV.mp4
https://haystack.yerf.org/BlC8NI.mp4
https://haystack.yerf.org/uhZjmS.mp4

>>42658271
Not flawless but it works, nice.
Also, I'm not into child or questionable age.
But you Cadance one was really interesting.
Do you have a free acount?
I don't even know how grook works, I don't have an X account.
I will try to run Wan in local when I have some time tho.

>>42658373
I still have Twitter Blue but it should work with the free accounts too I think
faggot
https://haystack.yerf.org/EDCSxF.mp4

>>42656363
Very likely. Not to mention probably 90% of the original pony booru dataset because of his "opt-out program". GG Astra, really.

>>42658441
Well, thanks for the pointers, I will ponder about it.
But I think I will try the local route before, it's more customizeable.
Thank you.

>>42658077
I have no idea why, but some people/countries(?) get filtered harder than others. I tried to ask ChatGPT to write me some greentext like found on 4chan a few times and even when I told it to leave out slurs or rule-breaking content it often got filtered. Meanwhile, some other dude did and uploaded AI-generated greentext that literally contained "-fag" and didn't.

It's just one more reason AI should never be left into the hands of corporations. They can't even be unbiased with their filtering.

>>42658568
>I'm not a pedo. Btw, I'll gen locally because, uhh, because it's more customizable, yeah!

>>42658271
>https://haystack.yerf.org/BlC8NI.mp4
What the fuck? The fact the Sunset video came through already surprised me, but this beats all.

Hope you're using a VPN. Based, though, even if I'm not really into loli that much.

>>42658930
Usually the less human something looks the higher the chance
https://haystack.yerf.org/gSoKVQ.mp4

>>42658077
I think it's because I went through ChatGPT. It might be calling Dall-E on the back end, but ChatGPT doesn't seem to care that I'm asking to reproduce a copyrighted style.

As an experiment, I tried generating another image through ChatGPT and comparing it to Pony V7. For Pony V7 I used the prompt:
>A show-accurate My Little Pony Pinkie Pie stands in front of a kitchen counter. She wears an apron and a chef's hat. She holds a whisk and is quickly stirring a mixing bowl full of batter. Globs of batter are flying out of the bowl and landing on her face and body and on the counter. Next to the bowl is a cake pan. In the background is a kitchen with a stove, oven, range hood, and cabinets. A saucepan with frosting is on top of the stove. The oven is preheated to 350 degrees. The cabinets are open and their shelves are stocked with spices.
For ChatGPT, I used the same prompt except I started it:
>Generate an image of a show-accurate My Little Pony Pinkie Pie
ChatGPT didn't complain at all. It generated picrel. There are a bunch of errors that make it not that impressive. The whisk is two-headed. The geometry of the kitchen makes no sense. The color of the stove isn't consistent, and if you zoom in on the stove, you see that the shape of the burner is messed up by the whisk. I could also argue about some of the artistic choices made here: The globs are more "suspended in midair" than "flying", and Pinkie looks like she's paused in the middle of mixing.

>>42659328
Here's Pony V7's attempt. Pinkie is balanced on one hoof on the counter. The stove burners are a mess. The saucepan isn't shaped like a saucepan and the frosting isn't on the stovetop. In the back left corner around the range hood, the cabinet is all messed up. The chef's hat doesn't fit on her head. The cake pan isn't shaped like a cake pan, and the batter is in a cauldron in a mixing bowl. The area behind and around the cauldron is messed up. I do like how the batter really looks like it's flying out of the mixing bowl, except for the giant glob of semen spurting from Pinkie's head. And, as always, the model has no idea what "show accurate" means.

>>42659365
It's actually surprisingly good for PonyV7. The examples I've seen on Civitai are pretty much all terrible, even with VERY simple prompts. That doesn't make it good though, especially if this wordy flowery nonsense is the only way to get semi-decent results from the model.

>>42659401
Actually scratch flowery, your prompt isn't flowery, it's mostly descriptive.

>>42656876
Right now due to all the restrictions in place (and failed results due to the filter flagging it as third party), it's kind of a proof of concept. It can be achieved, but by how and whom remains an issue.

As it stands, I see it as a milestone of comparison for open models to try and match as the tech gets better and smaller. Even if an upcoming open version is worse, if it's small enough to fine-tune on consumer hardware then we have the pony data to refine it further towards perfection; mares.

>>42657687
Very, and getting Rarity was by random chance. The prompt only nudged towards the show without flat out requesting it plainly. Got Fluttershy, Pinkie and Rainbow in another, but not nearly as authentic.

>>42639894 (OP)
>Yup, that's me. You're probably wondering how I ended up in this situation.

>>42660261
Realism with just enough SFM in the mix to feel off. The spear is remarkably coherent, though. Good job on that!

>>42660261
If one holds her spear by the chin, it frees up the hooves for pugilism.

>>42660261
>cheesy-legs shadow
sussy

Bump with an old one.

>>42658582
>It's just one more reason AI should never be left into the hands of corporations.
Nothing should be left in the hands of corporations. Corporations used to be alright when they had an end goal, at which point they disbanded. Now they grow like cancerous tumors and drag society down with them. No higher purpose, no soul.

>>42664072
You may not like it, but AI would never achieve the level we have now without the crazy corporate capital investment.
Fortunately the technology will still be there when the bubble bursts.

>>42664355
Yes it would. There was no need for Big Tech to get involved. All the important AI development came from startups, and PDXLv6 is the prime evidence that a huge corporation is not needed for development, just a group of dedicated autists.

There we go, finally done with this. Once again couldn't get as deep and bright colors as I would've liked but oh well. Difficulties of doing something other than "1girl, simple background"

>>42658077
you have to tell it to "use workarounds" for words and to avoid copyrighted names

>>42664733
No, it wouldn't exist without Big Tech. Without Big Tech's GPU purchases, the hardware would never have reached the sophistication it has how. They manage the data centers where those GPUs are housed, and those data centers operate at a scale that individuals can't match. All the software stacks on which the models run (Torch, Tensorflow, Keras, etc.) are funded and often directly developed by Big Tech. What we do in this thread wouldn't exist without them. So when the bubble bursts, AI development is essentially going to halt.

>>42664733
>PDXLv6 is the prime evidence that a huge corporation is not needed for development
That's like saying an indie game on Unreal Engine is proof we don't need big corporations making games lol.

Well, Civitai is dead. Any alternatives that don't require payment?

>>42665152
Yeah toss the craptop and run it local.

>>42665152
The best part of the announcement is how they say with a straight face, "Civitai.com stays freedom-first."

>>42665308
"We remain freedom-first, some people just have to pay for that freedom, that's all."

>>42658174

If you think this is good let me tell you about Sora 2. I am surprised no one is talking about Sora 2 with the animation and Rainbow Dash’s voice fitting perfectly but 4 Chan does not allow audio. This was before the content violations though. And he being in a church this is kinda fitting for Rainbow Dash I don’t know why.

>>42658599
Online is fast, but restricted, limited in generation and or length, and can shut down/be unuseable any day.
I prefer local because I want to queue up some prompt for the night and wake up in the morning enjoying the generations before wageslaving.
Not everything is sexual or deviant, you coombrain.

>>42665152
I was a bit miffed that I burned through over 10,000 buzz last month on gens, but now I'm happy I did. Still have more than 12k blue buzz left that is pretty much worthless now though.

>>42659365
Complete fucking slop. Even the batter flying out is wrong and just looks like it's being ejected out of the center whisk instead of the force of the whisk being used. Pinkie is also levitating half-on and half off the counter with broken anatomy while her hooves are sized for a midget.
All flash no substance 2/10 gen

>>42665100
We really don't though. Have you played any AAA games recently?

>>42665515
Sorry, I was just joshin' ya. I don't care at all what anyone gens, it's just a fucking picture.

>>42665515
What the hell do you prompt that takes all night to gen?

>>42665016
AI development has already come to a standstill though.

>>42664748
Looking pretty good though. Thanks, Anon.

>>42647432
>train on fucking auraflow of all things
>turns out a bad foundation means you make a bad finetune
Shocking, I know.

I genned an /SS/ oc with the name I got, but I kinda feel bad for the people doodling in the threads so I never posted it. They're having fun drawing and I don't want to ruin it with something much higher quality at a fraction of the effort that would likely also start a spergout.

>>42662481
The lost Picasso...

>>42668658
Self censorship is still censorship

>>42669374
News flash bucky; if you didn't order it, you're not entitled to it. Creators have the right not to release what they make.

b

>>42670560
>if you didn't order it, you're not entitled to it
Anon likely meant something along the lines of, "You are your own obstacle" or "Don't worry about it."

>>42665100
>PDXLv6 is the prime evidence that a huge corporation is not needed for development
>That's like saying an indie game on Unreal Engine is proof we don't need big corporations making games lol.
Blender is a community-driven project under the GNU General Public License (GPL). Blender Foundation is a nonprofit which is funded primarily by donations. People use the free program everyday to make pictures and animations.
Your argument falls flat.

Who's your favorite AI artist?

I'm leaning either Goblintide or Celuniamoon. I thought Goblin had it for sure until Celunia dropped this fucking coco ass that had me beyond diamonds.

https://www.deviantart.com/celuniamoon/art/OPEN-Coco-pommel-Goth-FERAL-1246021707

I wish she'd do that style more. It seems like she prefers some chibi one.

>>42673067
Are you for real? Celunia makes the simplest, most generic babby's first PD6 anime style lora slop.

>>42672345
The point was that PD is just a finetune of an existing commercial model, which required orders of magnitude less effort than creating a model from scratch.

>Blender Foundation is a nonprofit which is funded primarily by donations.
From corporations mostly.
Also Blender is not development of a new technology, it's just reimplementation of old shit.

>>42673211
So are new AI models. Again, we never needed Big Tech. The most important AI model, SDXL, was developed by a startup. Sure, they springboarded current AI development, but anything that came since then from Big Tech is nothing but cucked shit we don't need like Gemini or Bing, which won't even allow you to gen a copyrighted character. Again, we DO NOT NEED Google or Microsoft. Stop buying their lies.

>>42673067
>Who's your favorite AI artist?
Me. Nobody else makes exactly the fetish shit that I like (it's hardly even a fetish, but nowadays it's sadly considered one). Or at least that I could make before Civitai decided to shit itself.

>>42673959
You are autistically ignorant of the actual work and resources required. SD was not developed on a bunch of RTX cards in someone's garage.

>>42673998
Are you denying that Stability AI is a startup?

>>42674002
No. They are a startup created after Big Tech already invested in AI computing infrastructure, without which they would not exist.

>>42674020
>after Big Tech already invested
Sure, if you wanna go that route, they'll always be invaluable in your mind. Here's a hint: it's not just Big Tech that invested in this technology.

>>42674038
the guy probably believes nobody was using computers before microsoft windows came along. not even realizing its actually these monopolist companies that kept technology back for decades

>you NEED us goy!!

>not into mlp
>never used AI before
>try pony realism
>because it's called pony realism, I prompt "a pony opening her mouth with her tongue out"
>horrific uncanny gen of a photorealistic pony showing me her throat appears
>didn't even ask but she's drooling too
>proceed to goon for 5 hours to chernobyl ponies covered in liters of cum
what a night
that's all bye

>>42674054
Typical annitranny

>>42674049
>macroshaft winblows suxxxxxx!
lol it's not 2004 anymore, grandpa

>>42673972
>Nobody else makes exactly the fetish shit that I like
Do you like hairy mares?

>>42673067
Anyone who does more than just generic shiny anime style clop.
But I guess that's where the updoots and coombux are so that's what gets made.

>>42674136
Yeah, exactly that. It's ridiculous that this is considered a "fetish" nowadays while pseudo-cunny is considered "decent and respectable clean porn for well-adjusted individuals such as myself" but that's the way it is.

>>42674253
Me too, AI is a game changer.
I was thinking about putting up my mostly anthro gens somewhere, but I'm ashamed of not cleaning up the signatures and artifacts.

>>42673067
>Who's your favorite AI artist?
Myself, of course.

>>42673175
Celunia makes stuff that doesn't look like ai at all. You are nuts. 90% of the shit I see on tantabus is AI garbage style. Celunia comes along and drops these adorable mares that remind me of old ncmares style. Its clearly completely different.

She even made fun of it in her recent poll since she posted generic ai slop style loona that everyone favd to death.

>>42650978
lmao this slaps

>>42676075
>She

>>42666972
I'm okay m8, no hard feelings.

>>42666998
Oh I don't know, I'm still trying to install wan2GP...
But knowing that I have a potato PC and a carboard GPU, it will probably take the night to generate a 3s video...

>>42675034
That's one thing Celunia fails at is details. I love the styles she does, but people like you clearly have some kind of better way of generating smaller detail.

But fuck if I don't gel perfectly well with her specific brand of pantied pony fetish.

>>42640381
Whoa what style prompts or lora for this hatching?

>>42669774
>Daybreaker
Smug.
>Nightmare Moon
Grumpy.
That's fitting on every level.

bump