Gemini in Modeling

[Chris333]

My AI-foo is not so good.




But much better than last time:

[Chris333]

No idea why, but this one turned out sooo much better   

[peteski]

I still think that the graininess (the halftone dots of the low-res photo) might have something to do with the inability to produce good results.  I could be wrong  . . .

Out of curiosity, try using this version.
[ Guests cannot view attachments ]

[bbussey]

Also it can determine dimensions and features more easily from the 3/4 higher view angle. If the AI can determine the perspective and vanishing point, it can deduce the length/width/height more accurately. That functionality is in Photoshop also, called “perspective warp”.  That’s how MTL converted photos of weathered freight cars into printable artwork.

[Jesse6669]

I'm going to be semi-scratch building (3D printing plus some styrene and commercial parts) the sand tower used at the shops in McCullough Yard and it's a pretty unique design--at least I don't know of anything commercial like it. 

I want "blueprints" for making parts.  I fed two photos into Gemini and asked it to make "architectural line drawings" of it;  It came up with decent initial results but I had to go back and tweak it several times, telling it to move things, etc  It came up with what has decent proportions, perhaps a foot short on the bottom--but I can adjust the model with a 20' clearance to the sanding platform.  It initially had dimensions on the drawing but they were all wrong--sometimes just gibberish, with letters and numbers being backwards, upside down, etc.  You can see it calls elevation B "rear/side view".  No.. just rear.
 
Here are the inputs:





Ultimately I came up with this, although I cut/pasted the ladder/cage from an earlier rendering because it omitted it on this one (another quirk..)  So, it required work but saves me a bunch of time drafting/CAD and I'm "mostly" there. 

[Chris333]

I still think that the graininess (the halftone dots of the low-res photo) might have something to do with the inability to produce good results.  I could be wrong  . . .

Out of curiosity, try using this version.
(Attachment Link)

I tried that photo. It was slightly better. Finally made a 0-4-0 instead of a 2-4-0. But it did not make a tank loco.

[Chris333]

This is the same locomotive, but later on after the cab was rebuilt and handrails added to the tank.


Pretty clear photo, but it gave me this:


I think it is just giving me a "steam locomotive" and not looking at the photo. Again it gave it a cow catcher pilot and a pilot wheel. It did give it a tank, but also an air pump. The problem is I can do better by my self:

[nickelplate759]

It's really fond of 54" as a dimension.  Also it has no idea what a diameter is.

[Jesse6669]

Playing around with the sand tower drawing I came to realize after some frustration that the AI thinks in terms of pixels, less so logic.  So if you set a dimension on a picture and tell it to calculate how many pixels per inch (or whatever unit) then establish that standard it starts to more accurately depict length, etc.  It took me many revisions to get this drawing with accurate dimensions on it. 

Basically, you have to figure out how the AI thinks to give it the correct instructions.  Reading through it's step-by-step of the output helps, but it takes time. 

I'm sure there will be specialized AI catering to tasks like doing dimensional drawing/engineering drawings (if there isn't already) with more industry-specific vocabulary.  We are perhaps asking Gemini to do a specialty when it's really a generalist.

[CR4100]

I wouldn't have thought about using AI to create a drawing before this thread. Gemini did a great job removing perspective and converting this https://www.railcarphotos.com/PhotoDetails.php?PhotoID=48174 into this:



Now, I need AI to find a way to make the 3D printed model heavy enough to run at the front of a train!

[Iain]

[ Guests cannot view attachments ]

[ Guests cannot view attachments ]

Well, at least this is a starting point

[Dave V]

(Attachment Link)

(Attachment Link)

Well, at least this is a starting point

The floor plan seems a little sketchy. By 1940 there's a good chance there'd be a toilet (although on the RGS, not a single depot had indoor plumbing for the public at any point). But more importantly, I feel like the ticket office would include the bay window. The station agent would use the bay window to watch up and down the line.

[nickelplate759]

What I find interesting is how much of the drawing the AI gets wrong.   Even though it looks plausible at first, there's a lot that's clearly not right.

A few things jump out at me right away:
1. the end window is too short, and missing half of its panes.
2. More generally, the building appears to be either too short or too wide.  Probably too short.
3. One chimney got lost, and the other is in an unlikely place.
4. The floor plan shows two fireplaces on the outside walls, neither under a chimney.  I'd bet that one chimney went to a stove in the office, and the other to the waiting room.  Probably neither went to a proper fireplace.
5.  The eave brackets got simplified.


As @Iain said - "... at least this is a starting point"

[Jesse6669]

I tried Gemini out for creating backdrops from Google Map images (not Streetview, the aerial ones in 3D where trees and such are extremely blocky with artifact ).  It was useful but definitely required work.  The initial command was "make trees and structures realistic.  Keep buildings in the same architectural style.  Make sky overcast gray."  I had to go back and forth quite a lot, and do some manual editing in PhotoScape then re-upload and let Gemini try again. 

See pics in the Layout Engineering thread here:  https://www.therailwire.net/forum/index.php?topic=58640.msg832615#new

Incidentally, I got the "free" trial month so I have the Pro version, and cancelled right away, but I might actually pay for a month or more if my backdrops aren't done by then (iirc it's about $15-20).

[Chris333]

What sort of prompts do you guys use?

I have 2 photos of a locomotive:
https://photos.app.goo.gl/wgd2NGBBCERSo3Ay6
https://photos.app.goo.gl/zvGmeW29HiKGXgQZ9
I know it was built on 4/1886. Is a Porter 0-4-2T. Is 30" gauge. Has 6x10 cylinders. Has 24" drivers. Has 16" trailing wheel. Has a 48" driver wheelbase. Has a 102" total wheelbase. And is 14' overall length.

Edit: I think it is able to Google that info because it gave me this.
https://photos.app.goo.gl/uQw5jL8fT9j62mm9A
And when I scale it so the wheels are 24" the wheelbase turns out correct. But notice it insists on adding a cow catcher.