What Is Meshroom and How Do You Turn Photos Into a 3D Model?
Meshroom is the free, open-source photogrammetry application built on the AliceVision framework. It turns a set of overlapping photographs into a 3D reconstruction — a textured mesh, a point cloud, and a solved camera path — by running a node-based pipeline you can inspect and modify. It runs on Windows and Linux and requires an NVIDIA CUDA-capable GPU for the default pipeline; without one, you can still open the software and edit graphs, but the reconstruction nodes will not execute.
What Meshroom actually is
AliceVision is a photogrammetric computer vision framework: a collection of algorithms for 3D reconstruction, camera tracking, HDR and panorama work, packaged as a library. Meshroom is the graphical front end for that library. It gives you a visual node graph where each node is one processing step, and the connections between nodes define the order of operations.
That distinction matters in practice. When people say "Meshroom can't do X," they often mean the underlying AliceVision pipeline does not expose X as a ready-made node. Because the graph is editable, you can rewire it, disable stages you don't need, or substitute parameters — but you are working with the framework's building blocks, not a black box.
The photogrammetry workflow
The default graph in Meshroom is a complete structure-from-motion-plus-multi-view-stereo pipeline. The practical sequence is:
- Import photos. Drag a folder of images onto the Image node (or into the workspace). The images must overlap — typically 60–80% overlap between neighbours — and should be sharp, evenly lit, and free of motion blur.
- Run the pipeline. Start the computation. Meshroom executes the graph from left to right: feature extraction, image matching, structure from motion (camera poses and a sparse point cloud), depth map estimation, meshing, and texturing.
- Watch for failures. Nodes turn green on success, red on failure. A red StructureFromMotion node almost always means the photos did not match — too little overlap, too much blur, or a reflective/transparent subject.
- Export. Use the Export node or the file menu to write out the mesh (OBJ, PLY), the point cloud, and the camera data. You can also send the result onward to a viewer or a 3D tool.
The output you care about depends on the task. For a visual model, the textured mesh is the deliverable. For measuring or matching camera positions, the sparse point cloud and camera poses are what you need.
Hardware requirements
The default pipeline's depth map computation is CUDA-based, so an NVIDIA GPU is effectively mandatory for a normal run. This is the single most common blocker for new users on laptops with integrated graphics or AMD cards.
| Component | Practical requirement |
|---|---|
| GPU | NVIDIA with CUDA support; more VRAM lets you process larger image sets and higher resolutions |
| RAM | Scales with image count and resolution; large sets can exhaust memory before the GPU does |
| Storage | Depth maps and cached intermediates are large — budget several times the size of your photo set |
| OS | Windows or Linux |
If you don't have a CUDA GPU, the realistic options are to run the pipeline on a machine that does, or to use a different photogrammetry tool. Meshroom will install and open regardless; it just won't complete the reconstruction.
Typical use cases
- 3D reconstruction from photos — objects, buildings, terrain, or interiors, producing a mesh you can view, edit, or 3D print.
- Camera tracking — recovering camera positions and orientations from a photo set, useful as a starting point for matchmoving or for documenting a shoot.
- Sparse reconstruction only — if you only need camera poses and a point cloud, you can stop the graph before the dense stages and save significant time.
- HDR and panorama work — these come from the broader AliceVision feature set rather than the standard reconstruction graph, so expect to build or adapt a graph for them.
Common problems and workarounds
Structure from motion fails or produces a partial reconstruction. Usually a photo problem, not a software problem. Add more overlap, remove blurry or duplicate frames, and avoid scenes with large uniform surfaces (blank walls, water, sky) that give the matcher nothing to lock onto.
Runs out of GPU memory. Lower the image resolution or the depth map downscale factor, or split a large capture into smaller chunks and merge later. Processing fewer, larger images is not always better than more, smaller ones.
The mesh looks noisy or has holes. Dense reconstruction struggles with thin structures, shiny surfaces, and areas seen from only one angle. More viewpoints around the subject generally help more than higher resolution.
Everything is slow. The dense stages dominate runtime. If you only need camera poses, disable the dense nodes. If you need the mesh, accept that a few hundred photos can take hours.
Textures look stretched or smeared. This often traces back to inaccurate camera poses or a mesh with bad topology. Check the sparse reconstruction visually before committing to the full run.
Where to start
Install Meshroom, load a small, well-lit set of 30–60 overlapping photos of a single object, and run the default graph end to end before attempting anything larger. A successful small run teaches you what the node colours mean, how long each stage takes on your hardware, and what a good reconstruction looks like — which makes every later failure much easier to diagnose.