Generated objects and the space around them

Two different things, both called 3D

Meshes. Vertices, edges and faces, with a texture map. This is what every game engine, animation package and 3D printer consumes. It can be rigged, animated, cut, measured and manufactured.

Radiance representations — neural radiance fields and, more usefully now, Gaussian splatting. These store a scene as a cloud of coloured, semi-transparent blobs optimised to reproduce a set of photographs from any viewpoint. They render beautifully and photorealistically, in real time, and they are not geometry. There is no surface to attach anything to, no clean silhouette, nothing to 3D print.

The single most common disappointment in this area is someone capturing a splat of an object, finding it looks perfect, and discovering it cannot be imported into anything that expects a mesh. Conversion exists and loses much of what made the splat good.

Choose by what happens next. Viewing, walkthroughs, product presentation, virtual tours: splats. Anything to be edited, animated, simulated or manufactured: meshes.

How a mesh gets generated

Three approaches are in use.

Multi-view then reconstruct. Generate several consistent views of the object with an image model, then reconstruct geometry from them. Cheap and it inherits the consistency problems of this whole module — the back of the object frequently disagrees with the front.

Native 3D generation. A diffusion model trained directly on 3D representations. Better geometry, limited by the scarcity of 3D training data, which is orders of magnitude smaller than image data. This scarcity is the field's binding constraint and it is why 3D generation lags image generation by several years.

Photogrammetry from real photographs. Not generative at all: dozens of overlapping photographs reconstructed into geometry. Free software does this well. For a real object you can photograph, this remains the highest-quality route and it is often forgotten in the enthusiasm for generation.

What generated meshes are actually like

Honest description of the current state, because the promotional images do not show the wireframe:

• Topology is poor. Generated meshes are dense, irregular triangle soup rather than clean quad flow. They render acceptably and they deform badly, which means anything intended to be animated needs retopology.
• Scale and orientation are arbitrary. Everything arrives needing to be scaled and rotated to fit a scene.
• Textures are baked and low resolution, frequently with the lighting of the generated views burned into the colour map, which makes them look wrong under different lighting.
• Hollow interiors and non-manifold geometry are common, which matters enormously for 3D printing and not at all for a background prop.

The realistic use today is as greyboxing and background assets: blocking out a scene, filling a shelf, prototyping a shape before modelling it properly. As a hero asset for animation, it is a starting point that a modeller cleans up.

Where it fits with everything else in this course

There is a workflow worth knowing that runs the other way, and it is the most practical use of 3D for most people here.

Build a rough scene in Blender — free, on any platform — with grey blocks in the right positions and the camera where you want it. Render a depth map. Use that as structural conditioning for an image or video generation. You now have exact composition, exact camera, exact perspective and repeatable framing, supplied by software designed for precisely that, while the generative model does the surfaces.

This is the answer to nearly every spatial problem in this course. It requires learning a small amount of Blender — moving objects, placing a camera, rendering — which is a weekend rather than a career, and it converts an unreliable prompt into a reliable pipeline.

A note on the rights position, which differs from images in a way people miss. A generated mesh of a recognisable product, vehicle or building carries the same trademark and design-right questions as a photograph of it would, and in some cases more, because a 3D model can be manufactured. Printing a generated model of a branded object is not the same act as generating a picture of it, and the relevant law is design right and trademark rather than copyright in an image. The final module covers the shape of these questions; the practical note here is that "the model generated it" changes nothing about the object you then make from it.

The honest limitation: 3D generation is the least mature part of this field, the training-data scarcity is structural rather than temporary, and the gap between a viewable result and a usable asset is wider than in any other medium covered here. Expect it to improve and do not plan a production pipeline around it yet.