Unable to load image

Forcing Stable Diffusion to sit through millions of iterations of "this is Marsey"

We are onto the next step towards an AI generated Marseyverse!

Textual Inversion lets you train Stable Diffusion (to an extent) off of a few images. It works okay. There's already something newer and much better, but of course it's locked down by Google, so this the best thing available for a bit. I trained a few models on some Marseys to see how they'd perform.

Prompts

digital art of Marsey in the style of Marsey

https://i.rdrama.net/images/1684135087700925.webp https://i.rdrama.net/images/16841350899901626.webp https://i.rdrama.net/images/16841351050245376.webp https://i.rdrama.net/images/16841351063523526.webp

digital art of a dog in the style of Marsey

https://i.rdrama.net/images/16841351076086657.webp

digital art of an elephant in the style of Marsey

https://i.rdrama.net/images/16841351089181418.webp https://i.rdrama.net/images/16841351101713574.webp

Peter Griffin in the style of Marsey

https://i.rdrama.net/images/16841351116600697.webp

a 3d render of Marsey, octane render, cgsociety, raytracing

https://i.rdrama.net/images/16841351129956193.webp

It's great at style transfer, someone with artistic ability could use img2img + inpainting very effectively to make Marsey-style art. The embeddings are here if anyone's interested in trying it. Run this webui to use them.

And look at this picture I made using generated art. I like how Stable Diffusion gave the planet that smile :)

https://i.rdrama.net/images/16841351142239795.webp

134
Jump in the discussion.

No email address required.

its really stupid how we have this amazing technology and the first thing the people in charge think of doing is "we have to make sure our stock photos generate mullatos!"

Jump in the discussion.

No email address required.

Which is why Stable Diffusion even just releasing is very big news, the public now also has free access to a good quality network. Let's just hope we continue to have public releases of AIs in the future.

Jump in the discussion.

No email address required.

how do I do it?

Jump in the discussion.

No email address required.

check /g/

Jump in the discussion.

No email address required.

If you don't have loonix it's a bit of a pain to set up locally but doable. If you don't use it extensively just use one of the bajillion Google Colab notebooks made by community: https://pharmapsychotic.com/tools.html

(thx @chiobu for the link btw)

Jump in the discussion.

No email address required.

It's probably simpler to just dual boot. Trying ML on windows is one of Dante's circle of heck.

Jump in the discussion.

No email address required.

Basic ML is fine but with big projects you get filtered at the first wget yeah.

I only have 500gb ssd in my laptop (though it has good components for the rest) and it's already almost full... Should look into a nvme ssd soon

Jump in the discussion.

No email address required.

Neighbor windows has Linux nowadays

Jump in the discussion.

No email address required.

WSL has no Cuda/GPU support unless you want to beta test for MSFT

Jump in the discussion.

No email address required.

https://rentry.org/guitard

Just using anaconda is fine for stable diffusion in windows lol, its easy

Jump in the discussion.

No email address required.

If you have NVIDIA ot is easy as pi, there is even an exe on the subreddit.

If you have anything else you can get fricked, lol. It is possible but only for certain GPUs. Protip: no matter what they say, the RX570 is NOT supported

Jump in the discussion.

No email address required.

Use this to run SD locally https://rentry.org/GUItard

Jump in the discussion.

No email address required.

OpenAI may have a bit of secret sauce, but these things are usually based off of publicity available papers. Any script kiddie with a google cloud account can do AI these days. Thats where the really fricky shits gonna start cropping up

Jump in the discussion.

No email address required.

Any script kiddie with a google cloud account can do AI these days.

Entry level AI maybe but these enormous networks take months and hundred of thousands of dollars to train

Jump in the discussion.

No email address required.

Open source deathbot AI when?

Jump in the discussion.

No email address required.

Yeah, sane people just try to make porn of women that vaguely look like their crushes.

Jump in the discussion.

No email address required.

i wanted to make spider-man meet sonichu but you do you

Jump in the discussion.

No email address required.

>coomers

>sane people

Jump in the discussion.

No email address required.

If you think saying "I wanna see tits" to an AI is limited to extreme coomers you're sorely mistaken

Jump in the discussion.

No email address required.

:marseygasp: they do?

Jump in the discussion.

No email address required.

If you think most peoples first thought when being told "we can create a photorealistic image of anything on demand" isn't degenerate shit you might be neurodivergent

Jump in the discussion.

No email address required.

Holy frick some of those elephants are ridiculously cute :marseylove:

Jump in the discussion.

No email address required.

I showed one to my girlfriend and she wasn’t impressed. :marseyrain:

Jump in the discussion.

No email address required.

Women are haram not because they are inferior to men (this fact is unrelated), but because they do not instinctively love Marsey

Jump in the discussion.

No email address required.

Slightly in her defense I sent her this one.

![](/images/16620940908156095.webp)

Jump in the discussion.

No email address required.

Excuse me, that's not a certified Official Marsey. Please don't damage her brand.

Jump in the discussion.

No email address required.

![](/images/16620992859124708.webp)

![](/images/1662099286341858.webp)

Jump in the discussion.

No email address required.

This could be pretty good if someone cleaned it up, tbh

Jump in the discussion.

No email address required.

They're so adorable :marseyblush::marseyfingerhearts:

Jump in the discussion.

No email address required.

I did this in January and 90%+ of outputs were shit https://rdrama.net/post/35169/i-got-an-ai-to-furiously

Holy frick this AI art thing is moving fast

![](/images/164177234771.webp)

Jump in the discussion.

No email address required.

Honestly wouldn't be surprised if twitter artists start closing shop by the end of this year. Digital art isn't real art regardless so I don't really feel bad.

Jump in the discussion.

No email address required.

Just wait a few months until it can do audio... or much longer until it can do animation... or real-time interactive scene rendering, and eventually replaces traditional 3D rasterization renderers for PC and VR. Shit's going to get fricking weird.

Mark my words, there's going to be /tv/ shitposts in 3-4 years where someone - just with a few hours on a consumer GPU - replaces every actor in Blade Runner with nude Emma Watson, including their voices.

Jump in the discussion.

No email address required.

  • The ones that are just Marsey are good and I'd definitely recognize them as marseys although they're kind of scary Dark Marsey.

  • The elephants looked pretty grotesque but that's gotta be one of the hardest animals since it's got a trunk.

  • The Peter Griffin marseys are bizarre because it clearly understands that he's an overweight white guy wearing a white shirt, but there's stuff that comes out of nowhere. Like the red tie that keeps appearing. It looks about 70% minions, 20% Dilbert, 10% Family Guy.

Jump in the discussion.

No email address required.

Dishelved middle aged Marsey

![](/images/1662095324605636.webp)

Jump in the discussion.

No email address required.

:#marseymariachi:

Snapshots:

Textual Inversion:

something newer and much better:

https://i.rdrama.net/images/1684137475889827.webp:

https://i.rdrama.net/images/16841374772519884.webp:

https://i.rdrama.net/images/1684137478574746.webp:

https://i.rdrama.net/images/1684137479834751.webp:

https://i.rdrama.net/images/16841374811705165.webp:

https://i.rdrama.net/images/16841374824964042.webp:

https://i.rdrama.net/images/16841374838994088.webp:

https://i.rdrama.net/images/16841374852300017.webp:

https://i.rdrama.net/images/1684137486535178.webp:

here:

this webui:

https://i.rdrama.net/images/1684137486923866.webp:

Jump in the discussion.

No email address required.

!blackjack100

Jump in the discussion.

No email address required.

holy shit artists ARE fricked, those elephants are adorable

Jump in the discussion.

No email address required.

This one definitely says cute twink.

![](/images/16621042886144254.webp)

Jump in the discussion.

No email address required.

The planet, holy shit

Jump in the discussion.

No email address required.

Dope :marseyclapping:

Jump in the discussion.

No email address required.

This is some of the best marsey generation content yet. There are a lot of gems there that artists could probably clean up. Also marsey planet is too cute

Jump in the discussion.

No email address required.

Artists don't need no stinking help, especially from some random person on the internet. If you don't like it, then don't look at it. And Mars is NOT cute.

Jump in the discussion.

No email address required.

Mars is probably top three cute planets.

Jump in the discussion.

No email address required.

Why are you so obsessed with cuteness? Is your life that devoid of meaning that you have to find it in the planets? Get a grip.

Jump in the discussion.

No email address required.

I am going to get a grip on your peepee :horny:

Jump in the discussion.

No email address required.

If you think you can handle me, come and get it.

Jump in the discussion.

No email address required.

Expression on this one reminded me of blini cat and now I really want someone to make a :marseyblini: for smug posting

![](/images/1662114532353337.webp)

Jump in the discussion.

No email address required.

LPT: just install it on linux. Annoying af on windows


:!marseybarrel: :marseybarreldrunk:

Jump in the discussion.

No email address required.

not sure u did it right tbh v ugly

Jump in the discussion.

No email address required.

Can someone clip this one. It expresses something.

![](/images/16621229103900776.webp)

Jump in the discussion.

No email address required.

:#marsoyhype:

Jump in the discussion.

No email address required.

Looks like it's coming to fruition:

https://rdrama.net/post/99653/jobs-up-for-grabs-a-timeline

The AI Grim Reaper claims another casualty.

Jump in the discussion.

No email address required.

The dog ones actually capture the Spirit of Marsey much better than the cat ones. Those are hideous.

Jump in the discussion.

No email address required.

Globohomo graphicdesigncels on suicide watch, AI-generated cartoons unironically have more soul. Day of the graphics card soon.

Jump in the discussion.

No email address required.

YAY PITYII THINI J

Jump in the discussion.

No email address required.

What vram does your gpu have? Did you train on windows/linux? How many Marseys did you give it to train (I think I read 5 are recommended)?

Jump in the discussion.

No email address required.

I rented an A6000 from vast.ai and used Linux. I think 20gb is the minimum.

One of the models was trained using 5 images (upscaled with ESRGAN) and another with 28, the second model performed much better. Just added both sets of training data to the Mega link

edit: The steps were another spot I differed, the best results came after ~25000 rather than 3000-5000

Jump in the discussion.

No email address required.

Hmm I always forget that renting is an option. I heard you can kinda get it running with as low as 10gb but that 20 is recommended, yeah.

I don't think the upscaling was necessary? SD is using a unet after all.

Jump in the discussion.

No email address required.

The upscaling gave SD a cleaner style to pull in. Most of the Marseys are a bit blurry https://imgsli.com/MTIzNTAz

Jump in the discussion.

No email address required.

You are right, I somehow even forgot that Marseys aren't 512x512, which is the recommended input, and I somehow assumed you would upscale it even higher.

The unet cnn in sd does convolution and would basically inverse any upscaling you did very fast. It extracts most of the information from even more convolved layers.

Jump in the discussion.

No email address required.

Link copied to clipboard
Action successful!
Error, please refresh the page and try again.