Lego Technic AI Models general discussion topic

[Auroralampinen]

Ok, soon the 42232 will be hopefully revealed. But im now serious, why there is a massive craze for 42232 AI slop pictures. This is insane. Pls note, This topic is used for general discussion/roasting for the AI slop pictures for the 42232. But after the 42232 offical reveal i will update this topic for the 42232. So this will be temporary "junk" topic. Because the AI sloppification is getting out of hand and i wanted in the meantime before the launch create this topic for roasting the AI slop pictures for the 42232. Have a fun time in here before the reveal of the 42232:).

[I really hope this will not be taked down. I wont create any more AI roast topics and only reason why i made this was because the 42232 has the most created AI slop pictures and this is getting out of hand and this topic hopefully will prevent anyone making new "leak" topics for 42232. Because the upper picture is circulating online as a "real" 42232 leak picture which its not. So this will be the only one before the launch and hopefully this will help to stop misinformation spread and i will update this when the 42232 will be revealed offically and this topic is only used for now roasting these AI pictures so hopefully mods will allow this because i don't want a massive 42232 misinformation spread to happen thank you nice mods if you don't take this down and this topic can stay here under special premissions:).].

Edited 2 hours ago by Auroralampinen

[Jockos]

I don't know what car the 2nd picture tries to resemble but there's nothing Sadair's Spear about it. IIRC the first pic has been created way before in brown by a member of EB. I could accept this dark blue colour and overall aesthetics.

[Paul B Technic]

Sometimes it is really easy to spot AI images, other times they are harder to pick from real images.
In this case, until we get official photos, anything we see "should" be marked as a mock up / AI / best guess.

[mahjqa]

It's kind of nuts how fast these things have become surface-level convincing. Many AFOLs will be able to spot inconsistencies with panels and construction, but your average lego enthusiast likely won't.

[Jurss]

Yes, AI is pretty good at coming up with new parts 

[Anio]

Which IA made this Twingo ?

[Jurss]

gemini

[Anio]

That is interesting. I got that IA render :

It is impossible to build it because of curved non-existing parts, but the render is impressive nonetheless (note the thickness of the stick to make the Renault logo !).

Using an AI tool might also be interesting to get some ideas on how to design some details.

There is no doubt that, some day, IA will be able to generate full model with existing Lego parts and optimised building process. And this time might come earlier than we think.

[howitzer]

Those AI images are indeed pretty impressive at first glance, and have gotten a lot better in last couple of years compared to the AI generated images of Lego when they first started to appear. Of course someone with actual knowledge on the parts and construction techniques will quickly notice the inconsistencies with real Lego, as is seen with the Twingo pictures above such as bent liftarms, wrong length/width of panels and completely nonexistent parts. For images depicting System creations it's probably a bit more difficult to separate AI images from traditional CGI and photos due to wider parts selection and it also seems to me that AI is slightly better at creating System builds than Technic builds.

When it comes to official sets, I wouldn't be too worried. The official reveal will happen when it will happen, and before that AI images are no more useful than the blurry "leaked catalog photos" of olden days or whatever, which often were nothing like the actual, final product. I find however that AI images could be used to make fake Lego content revolving around MOCs, which could be used to get ad revenue or something. I don't see even this as a huge problem though, as I don't think there's _that_ much money to be made this way. Maybe it could be a problem whenever there's some contest with photo submissions?

[mahjqa]

25 minutes ago, Anio said:

There is no doubt that, some day, IA will be able to generate full model with existing Lego parts and optimised building process. And this time might come earlier than we think.

I'm not convinced. This is a 2D image. An impressive one, but still 2D. Realising a fully functional 3D model with very specific constraints about parts use, buildability, functionality, structural integrity and so on is an entirely different game. It's not even the same league. It's not even the same sport. 

[Anio]

14 minutes ago, mahjqa said:

I'm not convinced. This is a 2D image. An impressive one, but still 2D. Realising a fully functional 3D model with very specific constraints about parts use, buildability, functionality, structural integrity and so on is an entirely different game. It's not even the same league. It's not even the same sport. 

I totally see your point, and I agree with you when you say that a picture and a 3D model are not the same sport. A bit like video games from the 80's and after 2010.

But the thing is that we are at the very early stages of AI. Models properly designed in 3D by an AI are bound to happen.

The question is rather how long it will take to get there. 5 or 20 years ?

[msk6003]

Gemini generated image of Escort RS1800 few month ago.

New ChatGPT generated image of Escort RS1800.

Honestly, in terms of exterior, it seems to have already reached a level where it is worth looking at and referencing. Internal mechanism is still bad.

Edited 21 hours ago by msk6003

[gyenesvi]

8 minutes ago, msk6003 said:

Gemini generated image of Escort RS1800 few month ago.

Well, that's really smart, because it invented some proper fender parts that should exist in technic, to be able to build these kind of cars, with less curved bodies :D

27 minutes ago, Anio said:

The question is rather how long it will take to get there. 5 or 20 years ?

Jokes aside, I'm in the camp that guesses it's more like 10-20 years.. (I've been working with AI for 20 years now)

I think the way these images are generated is that it starts from known mental images of the given car, and kind of applies a lego-like took to them. It's a visual effect mostly, and it has little to do with proper spatial relations. In order to build proper lego models, having proper understanding of spatial relations would be key, and then on top of that it would have to layer part-shape and connectivity / collision constraints. Currently, it not only invents non-existent parts, but even in case of existent ones, their connectivity is fake. And that's a big part of designing lego models. Unfortunately, the way I see it is that spatial relations are kind of an under-researched area of AI. It would be very much required in the field where I work (self driving cars), but even there, it's only used in basic ways. 10 years ago, I thought that self driving car developments will give a big boost to that direction of research, but it did not happen as much as I expected. I think we'll need some major breakthrough to get there, not just incremental improvements.

[Lipko]

6 hours ago, mahjqa said:

It's kind of nuts how fast these things have become surface-level convincing. Many AFOLs will be able to spot inconsistencies with panels and construction, but your average lego enthusiast likely won't.

And this developed in about one year. I remember how crappy pictures AI generated a year ago.

1 hour ago, gyenesvi said:

Well, that's really smart, because it invented some proper fender parts that should exist in technic, to be able to build these kind of cars, with less curved bodies :D

Jokes aside, I'm in the camp that guesses it's more like 10-20 years.. (I've been working with AI for 20 years now)

I think the way these images are generated is that it starts from known mental images of the given car, and kind of applies a lego-like took to them. It's a visual effect mostly, and it has little to do with proper spatial relations. In order to build proper lego models, having proper understanding of spatial relations would be key, and then on top of that it would have to layer part-shape and connectivity / collision constraints. Currently, it not only invents non-existent parts, but even in case of existent ones, their connectivity is fake. And that's a big part of designing lego models. Unfortunately, the way I see it is that spatial relations are kind of an under-researched area of AI. It would be very much required in the field where I work (self driving cars), but even there, it's only used in basic ways. 10 years ago, I thought that self driving car developments will give a big boost to that direction of research, but it did not happen as much as I expected. I think we'll need some major breakthrough to get there, not just incremental improvements.

The spatial thing is developing fast. And the craziest thing that it just emerges by itself. As humans can feel spatial things are off withoug having an actual 3D model in their minds, this ability in AI starts to emerge. We now have videos of objects fully rotated, moving, and stuff, and you have too look closely to notice errors.

Edited 19 hours ago by Lipko

[gyenesvi]

33 minutes ago, Lipko said:

The spatial thing is developing fast. And the craziest thing that it just emerges by itself. As humans can feel spatial things are off withoug having an actual 3D model in their minds, this ability in AI starts to emerge. We now have videos of objects fully rotated, moving, and stuff, and you have too look closely to notice errors.

Well I can imagine some spatial consistency/reasoning emerging from things like training on lots of videos, which is indeed demonstrated by recent video generation capabilities, that's true. Though I wonder if it really understands spatial stuff or is it just really good at faking it. It could be similar to how arithmetic worked in some AI models when it was examined by researchers. They found that AI models can explain how to add numbers, but the way they actually did it inside was quite different; they basically just stitched together text strings in very tricky ways. So for me the jury is still out. But it's true that it's impressive how much stuff does emerge.

[Jurss]

The main issue is, that these are general AI models. If there would be somebody (who knows, maybe Lego already doing it) who feeds some specific model with all legal parts, their possible connections, all existing model instructions, engineering stuff etc., then after some year there would be something pretty OK ready, I suppose.

[aFrInaTi0n]

@Jurss we may not be far away there - blender is offering a so called MCP server already & there is the LDraw Loader to move all legit parts into Blender.. But that may not mean that this may directly allow for collision and other things to be understood by AI.

In my opinion most people are not really understanding the AI-products out there (ChatGPT, gEmini, CoPilot, whatnot), which all  have a huge portion of manual human work done to create needed logic & programming to allow those product to be felt versaitable as they seem to be "out of the box". It is funny that those product are mostly having multiple parts (LLM + RAG + KG), but people tend to think the LLM is the almighty AI - and with that they are not understanding that a model alone may not be worth much..

But I reckon this is another discussion :D

[Jurss]

1 hour ago, aFrInaTi0n said:

this is another discussion

@Auroralampinen can rename this just to AI topic. 

Maybe it also could be just not simply AI, but AI together with that LDraw or any other app. Just like Copilot is integrated in MS word, in which, f.i. I have IT system design,m and then I asked it to create test cases for all bussines logic chapter, and it did it pretty good. 

So, in LDraw You could put panels around, so they replicate some car, and then You ask AI to nicely connect them, if possible. And LDraw connection logic also could "fight" with AI, if it tries to connect something incorrect. Hybrid like this would be more usable, not to ruin all creativity.

[Auroralampinen]

5 minutes ago, Jurss said:

@Auroralampinen can rename this just to AI topic. 

Maybe it also could be just not simply AI, but AI together with that LDraw or any other app. Just like Copilot is integrated in MS word, in which, f.i. I have IT system design,m and then I asked it to create test cases for all bussines logic chapter, and it did it pretty good. 

So, in LDraw You could put panels around, so they replicate some car, and then You ask AI to nicely connect them, if possible. And LDraw connection logic also could "fight" with AI, if it tries to connect something incorrect. Hybrid like this would be more usable, not to ruin all creativity.

Ok then:). 

[aFrInaTi0n]

1 hour ago, aFrInaTi0n said:

But I reckon this is another discussion :D

Oh dang, hopefully I did not got understood wrongly here... I was only speaking about my own post - as I thought I would go too deep into IT topics. Sorry if that has been received in other ways.. 

[Auroralampinen]

1 minute ago, aFrInaTi0n said:

Oh dang, hopefully I did not got understood wrongly here... I was only speaking about my own post - as I thought I would go too deep into IT topics. Sorry if that has been received in other ways.. 

Yeah, for now i will keep this as AI "roast" topic but after the launch/reveal of the 42232 i will rename and rework this to the 42232 topic so in the mean time this will be temporary AI topic. But after the reveal this will be the 42232 topic and after that i will have to see if i will make another AI technic topic if this topic becamed popular;).