📸 Photogrammetry: Step-by-Step Tutorial

From your first photo to a clean 3D model — practical, with concrete settings.

This tutorial guides you through a complete photogrammetry run on a small to medium-sized object. You need neither expensive equipment nor prior knowledge — a reasonably recent camera (or a good smartphone) and a PC are enough. All software details are current as of 2026.

Steps

Requirements & Equipment
Step 1: Prepare the object & setup
Step 2: Configure the camera
Step 3: Take the photos
Step 4: Choose the software
Step 5: Process (Align → Mesh → Texture)
Step 6: Cleanup in Blender
Step 7: Export
Quick Checklist

Requirements & Equipment

Camera: a mirrorless/DSLR camera with a manual mode is ideal. A modern smartphone also works to start with (use pro/RAW mode).
Tripod: recommended when the object stays fixed and you move around it; less necessary with a turntable.
Light: two diffuse light sources (softbox/lamp with diffuser) or an overcast day outside.
Surface: a turntable for small objects, plus a neutral, opaque background.
PC: for processing. An NVIDIA GPU speeds up many tools significantly (practically required for dense reconstruction in Meshroom).

💡 Choosing an object for the first time: matte, textured, immovable, handy. Good: a rock, a statuette, a brick, tree bark, old shoes. Bad: glass, chrome, single-colored/glossy surfaces, fur, water.

Step 1: Prepare the object & setup

Setup

Place the object on the turntable, with enough space around it.
Position both lights so that no hard shadows and no highlights appear. The goal is even, "flat" lighting.
Keep the background neutral and opaque. A slightly textured background can help with walk-around, but hurts with a turntable — more on that shortly.
Optional: include a color chart and a scale in the shot if correct colors/measurements matter to you.

⚠️ Turntable trap: When the object rotates and the background stays still, the software "sees" two contradictory motions. Either you rotate against a completely neutral, uniform background (which the software ignores) or you mask out the background later in the software. A patterned, stationary background with a turntable ruins the alignment.

Step 2: Configure the camera

Manual mode — fixed values for all photos

Setting	Starting value	Note
Mode	M (manual)	Keep all values constant across the whole series.
Aperture	f/8 – f/11	Full depth of field. Don't stop down further (diffraction).
ISO	100 – 200	Minimal noise.
Shutter speed	whatever exposes correctly (tripod: any; handheld ≤ 1/800 s)	Handheld short enough to prevent camera shake.
Focus	manual, on the object's center, then lock	Autofocus off — it must not drift between shots.
White balance	manual (fixed Kelvin)	Constant color.
Format	RAW	More headroom; only develop uniformly, don't edit creatively.

💡 Smartphone: Open pro/expert mode, ISO to minimum, set focus to "manual/locked", enable RAW (DNG) if possible. Turn off HDR and "photo beautification"/AI sharpening — they alter each image differently.

Step 3: Take the photos

Systematically all the way around

Overlap 70–80%: the image content of two adjacent photos must overlap heavily.
Angular spacing 10–15°: that yields 24–36 photos per full rotation. For complex objects, feel free to shoot 60+.
Several height rings: one pass slightly from above, one at eye level, one slightly from below.
Underside: at the end, flip the object and photograph the previously hidden side separately (merge as a second scan later).
Not too close, not too far: the object fills the frame well but stays completely sharp and within the frame.

✗ Typical mistakes: too few photos · angle jumps too large · autofocus left on · light/object moved in between · only one height photographed (→ object full of holes at the top or bottom).

Step 4: Choose the software

For processing in 2026 there are essentially three sensible options — depending on your operating system, budget and ambition.

Software	Cost	OS	For whom
RealityScan 2.1 (formerly RealityCapture)	free*	Windows	The fastest way to top quality. *Free for hobbyists/studios under $1M annual revenue (Epic Games).
Agisoft Metashape 2.2	~€179 Standard	Win/Mac/Linux	Cross-platform, a very good all-rounder, and since 2.2 also exports Gaussian Splats.
Meshroom 2025.1	free	Win/Linux	Open source, node-based, fully transparent. Practically requires an NVIDIA GPU.

I describe the following steps generically — the buttons are named slightly differently in each tool, but the process is the same everywhere. Detailed prices/strengths in the Software Comparison.

Step 5: Process (Align → Mesh → Texture)

Import the images

Load all photos into a new project. Develop RAWs uniformly to TIFF/JPEG beforehand (same values for all) or — where supported — import them directly.

Alignment / camera orientation (Structure-from-Motion)

The software looks for common features and computes where each image was taken from. Result: a sparse point cloud and a "camera swarm". This reveals whether the capture is good: are all cameras aligned? Do they sit logically in a circle?

⚠️ If cameras are missing or sit "off": usually too little overlap or textureless surfaces. Remedies: reshoot the missing angles, set masks (object vs. background), or raise the sensitivity/"High" preset in Metashape/RealityScan.

Set masks (with a turntable)

So that only the object counts, mask the background. Modern tools (RealityScan 2.x) offer AI-assisted mask generation for this; otherwise use a color/brightness threshold.

Dense reconstruction / depth maps

A dense point cloud is computed from the aligned images. Compute-intensive — this is where the GPU helps the most.

Meshing

The point cloud becomes a closed surface (mesh). High-resolution at first; it's reduced later during cleanup.

Texturing

The original photos are projected onto the mesh. Choose a texture resolution of e.g. 4096 or 8192 px, depending on object size.

💡 Merging two scans: Captured the top and underside separately? In Metashape/RealityScan, align both "components" via common points and join them into one model before you mesh.

Step 6: Cleanup in Blender

The raw scan is rarely usable straight away. Blender (free) is the standard tool for post-processing.

Import & remove excess

Import the mesh as .obj/.fbx. Select and delete the floor, turntable remnants and loose fragments.

Close holes

In Edit mode, select the open edges and close them with Fill Holes or F (Face). Remove small stray geometry with Mesh > Clean Up.

Decimate (for lightweight assets)

Add a Decimate modifier, lower the ratio step by step until the polygon count is manageable without losing visible detail.

Optional: Retopology & baking (Pro)

For games/animation, create a clean low-poly topology (e.g. QuadRemesher/Instant Meshes), then bake color and the Normal Map from the high-poly onto the low-poly. This preserves fine detail at a low polygon count.

Step 7: Export

Target	Format
Universal exchange	`.obj` (mesh + texture)
Game engine (Unity/Unreal)	`.fbx`
Web / AR	`.glb` / `.gltf` (compact, PBR)
Film/VFX pipeline, Apple AR	`.usd` / `.usdz`
Point cloud, surveying	`.ply`

When exporting textures, make sure the image files are written/embedded along with them (automatic with .glb, as a separate .mtl + image files with .obj).

Quick Checklist

Chose a matte, immovable, textured object
Diffuse, constant light — no hard shadows/highlights
Manual mode: f/8–f/11, ISO 100–200, locked focus & white balance, RAW
70–80% overlap, one photo every 10–15°, several height rings
Captured the underside as a second pass
Checked the alignment (all cameras aligned?), masked the background
Cleaned the mesh, filled holes, decimated
Exported in the right format (textures embedded)

💡 Next step: Want your object in real time and photorealistic instead of as a mesh? Then Gaussian Splatting is worth a look — the same capture technique, a different result.

← Workflow Overview Gaussian Splatting →