As part of the "Lunchabuild This" campaign, Lunchables encourages kids to play with their food and build with their products. And they released blueprints to help kids on this. So we are wondering, can we use AI to empower kids to build anything they want, in a Lunchabuilder style?

The challenge here is, AI can’t quite nail the style of Lunchabuilder - there’s simply not enough data for AI to pick up. And if we just prompt AI to make “a helicopter made out of ham, crackers and pretzels”, AI doesn’t know exactly what ham and crackers are in Lunchables, so the results are still far from real Lunchabuilder.

​

So this is where LoRA came in. Since we already have a decent dataset of Lunchabuilder sculptures made by food stylist, we picked out 32 images and labeled them, which is a good size for LoRA training.