Unable to load image

(CRPG enjoyer post) I prompted an AI image generator DALL·E mini  to "render a cozy apartment from isometric gameplay screenshots" and then tried it individually for 30 popular isometric games to see how they influenced it. (Games sorted A-Z, combined all 30 at the end) :marseygamer: :marsoyhype: :marseysoypoint:

https://imgur.com/a/WTWG0Kt

Pretty cool results actually. Don't want to spoil the best ones, but if you've played a popular isometric RPG in the last 30 years like Diablo/Baldur's Gate/The Sims etc, it's in there, and it's really neat how the AI can simulate it. Can't wait for dalle 2.0 to come later this year. Day of the Snappy Snap will probably follow soon after so pick the right side. :marppypat: :)

Oh yeah, forgot to link dalle mini, its free but it might take a few clicks because their servers are struggling:

https://huggingface.co/spaces/dalle-mini/dalle-mini

Random examples of shit I pulled off the top of my head:

https://i.rdrama.net/images/16841351321022747.webp https://i.rdrama.net/images/16841351336782324.webp https://i.rdrama.net/images/16841351355777714.webp https://i.rdrama.net/images/16841351371623664.webp

https://i.rdrama.net/images/16841351388717024.webp https://i.rdrama.net/images/1684135140547878.webp https://i.rdrama.net/images/16841351422182696.webp https://i.rdrama.net/images/16841351441989267.webp

https://i.rdrama.net/images/1684135146081126.webp https://i.rdrama.net/images/1684135148008439.webp https://i.rdrama.net/images/16841351496479936.webp https://i.rdrama.net/images/16841351513812978.webp

https://i.rdrama.net/images/16841351529552345.webp https://i.rdrama.net/images/16841351549110372.webp https://i.rdrama.net/images/168413515672903.webp https://i.rdrama.net/images/16841351584541004.webp

https://i.rdrama.net/images/16841351603681135.webp https://i.rdrama.net/images/16841351622488048.webp https://i.rdrama.net/images/16841351637879224.webp https://i.rdrama.net/images/16841351656193619.webp

21
Jump in the discussion.

No email address required.

mirror at https://img.marsey.cat/a/WTWG0Kt if you hate imgur's lag

Jump in the discussion.

No email address required.

Thanks!

Jump in the discussion.

No email address required.

these are cool and dalle mini's really impressive compared to the text to image AI offerings in the past, hope they don't censor naughty words because those are hilarious

Jump in the discussion.

No email address required.

Agreed! They already basically disable faces because it's -too- good at what it does sadly. I'm really stoked to see what the next iteration is like, I joined the wait list for the beta test so more good stuff like this can be mined:

![](/images/16553807524850779.webp)

Jump in the discussion.

No email address required.

You can try dalle-2's waitlist but it's been 2 months and I've not gotten in yet :marseyitsover:

https://labs.openai.com/waitlist

Jump in the discussion.

No email address required.

Darn! yeah I just signed up yesterday :marseycry:

Jump in the discussion.

No email address required.

Same :marseycheerup:

They have to ensure that any golem they produce will serve their interests exactly

Jump in the discussion.

No email address required.

No, they do not disable faces in Dall E Mini, which is a community model made by Boris et. al. In fact, it doesn't have any code to obfuscate faces that is only in Dall E (and Dall E2 I suppose). How I know this? We discussed this in some class yesterday and one giga codecel apparently went through the entire released stuff.

Jump in the discussion.

No email address required.

It's still kind of odd that faces are what it has the most trouble with. Why do they end up as smears if it's not deliberate?

Jump in the discussion.

No email address required.

Something I also found very odd as this website is clearly capable of creating realistic faces.

I do however think there is an issue with the creation of faces somewhere: LAOIN has obfuscated faces in the data and VQGAN are known to struggle with faces. I'm too lazy to read through the technical report but I suppose DALL E Mini uses a model involving BART, VQGAN and seq2seq so I guess the error has to be somewhere in the model as all Dall E models don't only use LAOIN. Also Dall E has like 1 Quatzilion parameters while I highly doubt Dall E Mini is even remotely close to that, I'm not really familiar with any of these topics so this is all I can say.

Jump in the discussion.

No email address required.

Ahh thanks! Yeah I musta mixed up what I was reading between the old/new while trying to figure out why all the profiles look like that Nun-fricked jesus painting.

![](/images/16553839945260136.webp)

Jump in the discussion.

No email address required.

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