Image to 3D: object generators and scene generators are not the same tool

Search “image to 3D” and you will meet two categories of tool wearing nearly identical marketing. Upload a photo to one and you get a model of the thing in the photo. Upload a photo to the other and you get the place the photo was taken. Both are accurate descriptions of “image to 3D.” Only one of them is what you wanted. Product names below are used only to identify the tools; this page describes what each category returns rather than benchmarking output.

Thirty seconds: which one do you need?

Three questions settle it almost every time.

1. Does the camera go inside it? Inside means scene. Around the outside means object.
2. Is there one subject, or a place with many things in it? One subject means object.
3. Does anything have to collide with it, be picked up, or be animated? That needs clean surface geometry, which points to object-level output — even if the surroundings come from a scene tool.

If the answers split — you need a room and a chair you can pick up — that is not a contradiction. It is the normal case, and the section on mixing both covers it.

Object generators: one thing, as a mesh

This is the larger and older category, and it owns most of the “image to 3D” search results. Tools such as Tripo and Meshy take a photo or a text prompt describing a single subject and return a mesh: triangles, UV coordinates and texture maps, usually as .glb, .obj or .fbx. Tencent’s Hunyuan3D and Meta’s SAM 3D work at the same level.

What makes this category coherent is the deliverable. A mesh is what game engines, DCC tools and 3D printers all expect. It has a pivot point you can rotate around, a surface you can attach a physics collider to, and a material you can retexture. You can drop it into a scene somebody else built and it behaves like every other asset there.

The boundary of the category is equally clear: these tools model one closed surface, seen from outside. That assumption is what makes them good at chairs, helmets, characters and product shots — and what makes them fail at rooms.

Scene generators: a place, usually as a splat

The other category generates or reconstructs a space. Marble from World Labs takes text, a single image, multiple images, a 360° panorama or video and returns a navigable world, exporting Gaussian splats as .spz or .ply alongside a mesh and a fly-through video (Radiance Fields). Atlas, the newer World Labs model, outputs point clouds and Gaussian splats but is limited to early-access partners (Dealroom).

Capture tools such as Polycam and Luma sit in the same category from the user’s point of view: photos in, a space out. The difference is that they reconstruct what was actually there, while generative scene models invent the parts you did not photograph. World Labs says Atlas can work from as few as two or three reference images (The Decoder), where classical splat capture normally wants dozens of overlapping shots. That claim is self-reported and has not been independently reproduced.

The signature of this category is that the camera belongs inside the result. You are not orbiting a subject; you are standing somewhere.

Side by side

	Object generators	Scene generators
Typical input	One photo or prompt of one subject	Photos, video, a panorama, or text describing a place
What comes back	One mesh	A whole space
Common formats	.glb, .obj, .fbx	.ply, .spz splats; sometimes a mesh
Camera position	Outside, looking in	Inside, looking around
Collision and physics	Straightforward	Needs a mesh export or a proxy
Retexture and animate	Designed for it	Not really
Photoreal environment	No	Yes, this is the strength
Examples	Tripo, Meshy, Hunyuan3D, SAM 3D	Marble, Atlas, Polycam, Luma

The format difference is the real difference

Everything above follows from what the two output formats actually are.

A mesh is a description of surfaces. Triangles have a front and a back, UVs map an image onto them, and a material says how light behaves at each point. Every game engine has been built around this representation for thirty years, so a mesh arrives ready to be lit, collided with, skinned and animated.

A Gaussian splat is a description of appearance. It stores millions of coloured, oriented, semi-transparent blobs positioned in space, which together reproduce what a camera would have seen. It can look startlingly photoreal, especially for a real place, because it never had to commit to where the surfaces are. That is also its limitation: there is no surface, so there is nothing to collide with, nothing to retexture, and no clean way to separate one object from its neighbour. The formats themselves are covered in PLY vs SPZ explained, and the trade-offs against the older NeRF approach in Gaussian splatting vs NeRF.

The practical summary: meshes are for interacting with, splats are for looking at. A splat also needs a renderer that understands it — a plugin, or a runtime such as the MIT-licensed Spark renderer, which is what powers the splat viewer on this site. A mesh needs nothing special.

This is why scene tools that also export a mesh are easier to fit into an existing pipeline, even when the mesh is rougher than the splat. It gives you something to stand on.

What going wrong looks like

Room from an object generator. You get a room-shaped object. From outside it reads correctly. Move the camera in and you are looking at the back of every surface: no interior detail, lighting that makes no sense, and geometry that was only ever solved for an exterior view. No amount of prompting fixes this, because the model was built for closed subjects seen from outside.

Prop from a scene generator. You get your prop embedded in its surroundings, in a representation with no boundaries between objects. Cutting the chair out of a splat of the room is not a selection operation; there is no chair, only the blobs that happen to look like one from where you stood.

Splat straight into an engine, expecting an asset. It renders — if you installed something that can render it — and then nothing else works. The player walks through walls, nothing casts a shadow onto it, and no material can be swapped. It was never geometry.

Mixing both is the normal answer

Most real pipelines use the two categories for what each is good at:

• Environment from a scene tool. Photoreal, fast to produce, and never touched by physics — a backdrop, a skybox replacement, a location plate.
• Interactive elements from an object generator. Anything picked up, opened, driven, shot at or animated needs a mesh with a collider, so it is authored or generated as an object.
• A collision proxy for the environment. A rough mesh — either the scene tool’s own mesh export or a hand-built box-and-plane approximation — sits invisibly inside the splat so the player has floors and walls. This is the step most often skipped, and it is what makes a splat environment playable rather than decorative.

The game development use case on this site works through how these pieces land in Unity and Unreal.

Where to go next

• What is a world model? — the category scene generators belong to, and the two other things that share the name
• Image to 3D world — this site’s page on the scene-level workflow
• Marble API guide — the scene-level API you can call today
• Free splat viewer — open a .ply or .spz in the browser and see what a splat is
• PLY vs SPZ explained — picking a splat format