Splat best practices

Splat quality depends on both the scene and the views you provide. Start with a bounded, textured environment, cover the areas you want to explore, and check that overlapping views agree before submitting posedImages2Splats.

Where splats work well

With good input, bounded indoor spaces with rich texture are a strong starting point. Successful examples include furnished living rooms, home offices, retail shops, wood-paneled rooms, and stylized science-fiction interiors. Furniture, shelving, floor patterns, and wall detail give these scenes visible structure across viewpoints.

A good overall room reconstruction does not guarantee accurate detail in every object. Inspect the features that matter to your use case from the viewpoints you plan to use.

Features that need closer inspection

Scene feature	What can go wrong	What to inspect
Large, open spaces	The foreground can look good while distant areas become fuzzy or lose structure.	Check the background as well as nearby objects.
Sky and clouds	Large sky regions can contain streaks, blobs, or floating artifacts.	Look at the sky while moving through the scene.
Trees and foliage	Leaves and small branches can blur or merge together.	Inspect fine detail where vegetation is important.
Large plain walls	Uncertain depth can produce bulges or uneven surfaces.	View walls from an angle to check their shape.
Human faces	Facial detail may be inaccurate.	Avoid relying on faithful facial reconstruction.
Thin structures	Chair legs, hanging wires, and similar parts can have missing segments.	Check that important thin parts remain continuous across views.
Windows	Glass and the scene beyond it can reconstruct poorly.	Inspect the view through the window, as well as its frame.

Consistent inputs help, but sky, foliage, and fine detail can remain difficult when the views agree. These are quality limitations to evaluate for your scene, rather than a rule that every scene containing these features will fail.

Prepare a consistent set of views

Keep the scene mostly static, use sharp images, and keep exposure and lighting consistent. For generated views, compare overlapping regions before splatting: objects should retain their position, orientation, shape, and appearance. Check again when a sequence returns to an earlier part of the room.

For example, in a cabin scene, generated views placed and shaped the same dining table differently. The reconstruction combined those conflicting views into overlapping, ghosted surfaces. In another living room, a coffee table became two tables on a return view, and the duplication carried into the splat. Resolve these disagreements in the input set before generating the splat.

Input view 1: note the dining table's shape and its position relative to the kitchen.

Input view 2: the same room contains a different table shape, orientation, and leg appearance.

Result from the full input set: inconsistent table geometry produces ghosting. The two images above are selected views from that set.

Cover the whole area you need

For a single room, start with 8–16 well-distributed views. Four views can miss too much of a complex room. Add views where they reveal an unseen area or a new side of an object; repeating an already covered viewpoint adds less value. Larger or more occluded spaces may need more views. The count is a starting point, not an API limit or a guarantee of quality.

Include the floor, lower furniture, walls, and areas behind obstructions that you intend to explore. In one lobby example, the input sequence looked mostly up at the ceiling and missed the floor. The resulting lower scene contained blurry, invented furniture. More views of the ceiling would not fill that gap.

Neighboring views should share recognizable scene content. Aim for roughly 50% shared image content, with at least about 10% as a practical lower bound when planning a capture. These are approximate capture recommendations, not request validation thresholds. Sparse overlap can make camera estimation unreliable, and camera errors can degrade the resulting splat.

Keep the image, camera, and depth for each view together. Use the returned frames from images2PosedRGBD when estimating cameras and depth from photographs; their intrinsics describe the returned image grid. See Cameras and posed images for coordinate and depth conventions.

Check the result before using it

Move through the intended viewing area, rather than judging only the first view. Look for ghosting, duplicated objects, missing floor coverage, wall bulges, and the difficult features above. When a problem corresponds to an inconsistent or uncovered input region, improve those views before rerunning. For persistent fine-detail or sky artifacts, assess whether the result meets your use case even with a consistent, well-covered input set.

Continue with the API task guide for submission, waiting for completion, and using the SPZ output.