I’m afraid there’s a problem… Let’s make another try. Tell me a scene, what should it contain? I’ll try to sketch the items and feed them to the AI.
I want to play that game! 
2 Likes
I’m hoping to make it about 20 rooms (maybe about the size of Melee Island if you discount the forest maze).
I just added the inside of the restaurant, so there are 5 rooms now.
It’ll basically be Boybrush trying to figure out how to break Chuckie’s spell.
4 Likes
if you want me to pixelize with AI, just send me the originals
2 Likes
Try to draw a staircase like the one in the entrance hall or the library in Maniac Mansion. (Please draw it from your memory or like you would draw it.)
The staircase seems to be difficult to create for the AI. So I would like to know if it is able to draw a “real” staircase.
Very, very, very impressive!
The only thing I don’t feel “comfortable” with are the transitions between the scenes. For example at 00:07 in the opening scene Boybrush is walking into the background. In the next scene at 00:08 he is standing on the right of a huge entrance. The way in the first scene leads to the right border of the screen. So I would expect that Boybrush appears on the_left_ side of the entrance. Beside that, you can’t see the huge building in the distance in the first scene. Maybe due to these two things my brain has problems to connect those two scenes to a “real” landscape.
How do you have generated these backgrounds? Would it be possible to tell the AI to insert the building at the horizon in the first scene? Or would it be possible to create a new scene that connects both places?
1 Like
Man you’re asking me to draw the most complicated thing on earth, a spiral staircase, out of memory 
Also, I have no memory of any object in my brain.
I used a reference, and I made a drawing as quick as I could, to replicate a realistic workflow.
I’ll feed it to the AI and the problems will be apparent. I expect either the colors will be wrong, or the shape will be wrong.
1 Like
img2img
img2img , init strength around 0.2 (which means it respects my colors too much, and respects my lines moderately):
img2img , init strength 0.09 (it improves my colors, using the monkey 2 style, but it disrespect my lines probably too much, though they are closer to monkey 2 style)
third attempt: this is actually pretty good, it respected my lines enough and still improved the colors:
control net , edge to image
Now let’s use control net (which means it ignores my colors completely, except those given in the text description) , edge to image, weight 0.88 (which means it respects my lines a lot)
control net , edge to image, weigth 0.59 (doesn’t respect the lines a lot)
control net edge to image, weigth 0.34 (it starts lookiing like monkey2)
control net edge to image weight 0.23:
weight 0.16
this one is pretty good:
controlnet , depth to image mode
controlnet , depth to image, weigth 0.83:
weight 0.59
weigth 0.19:
I’m unsure if the test was passed. We should describe a location completely. The more objects there are, the less likely it is that all of them will be acceptable…
3 Likes
Yes, you should do this first. It’s very impressive where you can get just with the AI. And the style fits.
I would assume the same. But it would be interesting to see, if the AI is able to put things in the background and/or is able to create a joining scene.
Is it possible to refer to existing pictures in some of the AIs? For example can you tell one of the AIs to “create a forest with a hotel at the horizon that looks like the hotel on picture XYZ”?
1 Like
I haven’t come across something like that so far… there is a thing where you can take an image and have AI “expand” the image in multiple directions to fill in surroundings, I don’t know if it can be prompted with what to fill in though.
I think there is AI which can remove specific things from an image, maybe it’s possible to replace one element with another, that would be useful.
Also once I’m done I’ll post all the AGS files for this game, so if anyone else wants to make their own MI3a and wants a sort of template to get them going, they can tinker with what I’ve done or use it as a starting point.
I used AGS for what I have so far (PowerQuest was a bit too complex for me as a beginner), and there is this online book that walks you things in a really nice way, so I’m just following it and adding a bit to the game each day - GAME DESIGN WITH AGS - Game Design with AGS
1 Like
The results are better than expected – and very fascinating! Thanks! The staircases are “usable” in most cases. 
Although this is a little bit creepy:
There are several good ones! The AI came even up with some ideas I never had thought of, for example this “tree”(?) house:
Can you now feed the AI with the Maniac Mansion staircase? I’m curious what you’ll get with one of the the original in-game scenes and how they differ from your sketch.
This might not work due to the very colorful books. But maybe you can use the sketch:
Or this alternative scene:
I tried this source image:
as it is black and white; i need to use controlnet (which invents the colors)
edge to image:
wei 0.97
weight 0.8
weigth 0.56:
weight 0.3:
depth to image:
w 0.97
w 0.91
wei 0.75
weight 0.33
conclusions?
this was an interior, but I think exteriors are much more difficult to get right…
img2img, varying the strength:
control net, edge to image, varying weights:
THis is amazing:
control net, depth to image:
control net, pose to image:
2 Likes
King Graham is going to have a tough time on those
2 Likes
Thanks! The results are partially weird, but also interesting. My favorite is this:
It’s interesting to see that the AIs are having trouble to construct a usable staircase, but they are aware of David standing in front of the staircase.
And it’s also interesting so see that your sketches are leading to “better” results than Gary’s sketch.
I would think it’s the other way around: Most of the interior stuff has an impressionistic and/or expressionistic look. Landscapes seem to be easier to generate. Maybe the AIs were trained with expressionistic paintings and photos of landscapes…
They should add that to Stair Quest II…
1 Like
Nope…
for this reason, enforcing a consistent style across pictures is hard. You need to train with pictures that share a very similar lighting and palette. (no day and night together, for example)
1 Like
I think it’s time to train with monkey2 part2, which is all daylight, but I need to take screenshots and I don’t have savegames … I can’t replay all of part 1 
1 Like
Just to clarify (for my dumb brain): The picture is the output of the AI?
I seem to recognize the king’s quest style?
1 Like