The screws and hinges are fascinating.
It hints that the model may know both a pool of fasteners and individual types. It wasn't just trained on what fences are or board constructed elements as those would have defacto defined what a fastener means in the context of a fence. This looks more like a mix of hardware types, so on some level it knows about these specific details. There is probably a trained use case where the company is generating media for product, architectural, and/or automotive design.
The latch/hinge/tool looking monstrosity with a pipe/nail, looks like it was in the mid evolutionary stage of iteration. If anyone reading this has never had advanced control of a model it will be hard to understand, but there are steps to the generation process. In some sampling techniques a similar type of object is first generated quickly out of the static noise. Then the model focuses on small areas of the image specifically, evolving these almost in isolation for several iterations before moving to another small area. With this kind of sampler, some part of the image is always in transition. If you have control over the stepping process, it is possible to dial in the steps to a point that the attention is on the background or in a intermediate state that still looks natural.
There are also sampling techniques that make almost video like changes that look more like claymation in an animated stop motion like film where stuff slides into frame and moves with lots of things happening at once.
Muad'Dib rides the Shai Hulud of Catadan
