Prepare Inputs for 3D Generation

Every image-based 3D pipeline — photogrammetry, neural reconstruction, multi-view generation — obeys one rule: a surface that no image observed will be invented or missing. Most disappointing results trace back to the capture rather than to the model, and capture is the cheapest thing in the pipeline to get right.

Twenty minutes of careful photography routinely beats an hour of mesh repair.

Coverage is geometry, not photography

Think of coverage as observations of surfaces rather than as pictures of an object.

Shoot three horizontal rings at different heights — roughly at the object's midline, one above looking down, one below looking up. Each ring should have enough positions that consecutive frames overlap substantially; a useful rule is that any point on the surface appears in at least three images.

Add top-down frames, which people always forget and which are the reason so many reconstructions have a mushy crown.

Add detail passes: closer frames of anything small or intricate. Resolution on a detail is a function of how much of the frame it occupied, not of the camera's megapixels.

Then find the concavities. Undersides, inside corners, the gap between an object and its base. These are where reconstruction fails, and they fail silently — the mesh closes over them with invented geometry that looks plausible until you rotate the model.

For a subject that can be moved, capture it twice in different orientations. The second pass observes what the first one was resting on, which is the only way to get a complete object.

The capture properties that matter

Ranked by how much damage they do when wrong:

Sharpness first. Motion blur destroys the feature matching that reconstruction depends on. Use a tripod or fast shutter, and check frames at full zoom before you leave. A blurred image is worse than a missing one, because it contributes wrong information rather than none.

Flat, diffuse lighting second. Reconstruction assumes a surface looks the same from different angles, and hard light with strong specular highlights breaks that assumption — the highlight moves with the camera and the algorithm reads it as surface detail. Overcast daylight or a large diffuse source is ideal. Baked shadows in the texture are the second cost of bad lighting, and they are covered in the texturing lesson.

Consistent exposure and white balance. Auto modes change between frames and produce a patchwork texture. Lock both.

Depth of field. Enough that the whole object stays sharp. This usually means a smaller aperture and more light.

A static subject and scene. Anything that moves between frames — a person shifting, leaves, changing light — becomes noise or ghosting.

Background. A neutral, non-repeating background helps segmentation. Featureless white can actually hurt, since some methods use background features for camera alignment; a lightly textured surface is a good compromise.

When you are generating rather than reconstructing

Text-to-3D and single-image-to-3D are generating, not measuring, so the input job changes: you are constraining a space rather than observing a surface.

Assemble a small reference set that fixes the properties you care about. One image for silhouette, which is what determines whether the result reads as the right object at a glance. One or two for proportion, ideally orthographic-ish views. One for material and colour. One for the level of detail you expect, which manages your own expectations as much as the model's.

Then write the constraints that images cannot show: real-world scale in units, which way is up and which way is forward, whether the object must be a closed watertight solid, and whether it will be seen from all angles or only from the front. That last question changes how much work the back of the object deserves, and it is worth asking before rather than after.

Some subjects will not work from photographs at all: transparent glass, mirrors, fine mesh or hair, dark matte objects with no visible features, and anything shiny enough that its appearance is mostly reflection. For these, generation from references or manual modelling is the realistic path, and knowing that in advance saves an afternoon.

Practice

Capture the same object twice: once quickly, freehand, in whatever light is available, and once with a tripod, diffuse light, locked exposure, three rings plus top-down, and a second pass with the object turned over.

Run both through the same pipeline and compare where they fail. The differences will be concentrated in the concavities and the underside, which tells you exactly which part of the capture protocol is doing the work. That comparison is worth more than any amount of reading about capture, and it takes an afternoon once.

Common mistakes

Too few frames. Storage is cheap; a return trip is not. Shoot more than you think.

Hard directional light. Highlights move with the camera and get read as geometry.

Auto exposure and white balance. A patchwork texture that is tedious to fix afterwards.

Never checking sharpness on site. Blur is invisible on a camera's rear screen and fatal to reconstruction.

Forgetting the underside. The single most common gap, and the one that requires a second capture rather than a repair.