Playing around with AI Image Conversion on Sierra titles

[cosmicr]

I've been having a blast converting Sierra images to modern art pieces using AI.

See below some examples:














I have found that it works best for fantasy style scenes and landscapes. Luckily, a lot of sierra games feature these!

[ZvikaZ]

That looks amazing!

Can you further explain how did you do that?

What are your next plans?

[Kawa]

You are aware, of course, that Maniac Mansion was not a Sierra game? 😸

[MusicallyInspired]

I'd love to know how this was done!

[cosmicr]

I threw Maniac Mansion in there for fun, as I am also a big fan of LucasFilm(Arts) games.  Looking for any other suggestions.

This is done using a Machine Learning platform called Stable Diffusion with Image-to-Image. You basically provide an image and a description, and the model will produce a result. You can then fine tune the result by feeding it back into the input, update you description as necessary, and repeat as many times you like. There are a lot of parameters that can be changed to influence your result. I also upscaled the images using AI and for the Passionate Patti Picture I ran it through GFPGAN which makes faces look more realistic. You need a decent GPU for quick results.

Here's a bonus image from Betrayed Alliance:


If anyone has any requests let me know, I love this stuff. It's not limited to this style either. It can do painted, pencil sketch, 3d rendered, other styles I haven't thought of.

[Collector]

You could try the pics from KQ4 for a possible VGA conversion.

[doomlazer]

I like that skeleton Sludge pic and the skull cave turned out well.

How close can you get to the KQ4 box art if you start with Rosella, the Unicorn and Goons on the path to the castle?

[Doan Sephim]

Wow! Didn't expect to see a rendering from my game on here! Really cool stuff and thanks for sharing it. I really dig the KQ 4 one at the top, too

[Cloudee1]

Those are absolutely exceptional.

[bokkers]

Great renditions! Keep em coming folks. Would love to see some SQ3 scenes. Cheers!

[cosmicr]

Hehe, it certainly has come a long way since DALL-E and Stable Diffusion 1.5 (Which is what I used for those images in my original post)

As soon as the new OpenAI image generator came out the first thing I tried was some sierra screenshots...



 

 

 

The only thing I don't like about the OpenAI generator is that it's not open source, and it seems to put a brown hue over every 2nd image...

[russdanner]

awesome

[lskovlun]

The thing I've always wondered about this scene is if it's all automated, why are the lights on all the time?

[Andre999]

Great stuff. I think such AI image conversions definitely can be used to create VGA-like versions of Sierras parser EGA games. Either by using SCUMMVMX (little work) or creating genuine SCI01 VGA parser games. Some challenges, though, like the generated images moving the camera, modifying objects, use different colors in different images, etc. But still possible, I would think.     

[cosmicr]

There area other models like "Depth Anything" by Meta that could be used to re-create the priority screens. If you were REALLY keen on upscaling/recreating a game using AI, it could be done, albeit not without any manual work.

Upscaling sprites and animations (views) is much more difficult though.