Image to 3D Model: Three Honest Ways to Do It

Image to 3D model: three different jobs under one name

Search for "image to 3D model", "picture to 3D model" or "photo to 3D model" and you get tools that look alike on the surface but do very different things. Some take one picture and return a model quickly. Others need dozens of photos and a longer wait. The results are not interchangeable.

There are three main routes today:

• One image in, a generated model out. An AI model guesses the full shape from a single 2D image.
• Many photos in, a measured mesh out. This is photogrammetry.
• Many photos or a video in, a realistic viewable scene out. This is Gaussian splatting.

The first route invents the parts of the object the camera never saw. The other two only reconstruct what was actually photographed. Which one you want depends on what you plan to do with the result, so this guide goes through each route and then sorts them by goal.

Route A: 2D image to 3D model with generative AI

Single-image tools are what most people picture when they think of turning a 2D image into a 3D model. You give them one photo, sketch or piece of concept art, and a trained model outputs a 3D shape with a texture.

A few examples, described only as their own pages describe them:

• TripoSR, developed by Tripo AI and Stability AI, is an open model for fast feedforward 3D reconstruction from a single image, released under the MIT license.
• TRELLIS, published in Microsoft's GitHub organization, takes text or image prompts and can output radiance fields, 3D Gaussians and meshes. Its README asks for an NVIDIA GPU with at least 16 GB of memory and says the code has only been tested on Linux.
• Hunyuan3D 2.0 from Tencent works in two stages: it first creates a bare mesh from the reference image, then paints a texture onto it. It comes with its own community license, which excludes some regions, so read it before use.
• Meshy is a browser-based service. Its image-to-3D page accepts a PNG, JPG or WebP image and offers downloads in FBX, OBJ, GLB, USDZ, STL, BLEND, 3MF and DXF.

The important thing to understand is what one image can and cannot contain. A single photo shows one side. The back, the underside and anything hidden are filled in by the model based on what similar objects usually look like. That is fine for a game prop, a character from concept art or a quick placeholder. It is not a record of your actual object. If the back of your vase has a chip or a maker's mark, a single-image tool will not know.

Route B: photos to a 3D mesh with photogrammetry

Photogrammetry works from many overlapping photos of the same real subject. Software finds the same points in different photos, works out where each camera stood, and builds a surface from that geometry. The output is usually a textured mesh: triangles with a color image wrapped over them.

Some options, again limited to what their own pages state:

• Meshroom is a free, open-source 3D reconstruction program based on the AliceVision framework.
• RealityScan from Epic Games is described as a photogrammetry solution that creates 3D models from images and laser scans. Its license page says it is free for individuals and small businesses with under US$1 million in revenue in the past 12 months, and for students and educational institutions.
• Agisoft Metashape is a commercial package that offers a demo mode and a 30-day trial license.
• Apple's Object Capture is a photogrammetry API on macOS that developers build apps on. Apple says the output includes a geometric mesh and material maps.

Because the mesh comes from real photos, it reflects the real object, including its flaws. The cost is effort: you need good coverage, and processing takes time. Shiny, transparent and featureless surfaces are difficult, because the software cannot find stable points to match.

Route C: photos or video to a Gaussian splatting scene

Gaussian splatting starts the same way as photogrammetry: many overlapping views, and a step that solves the camera positions. The difference is what gets built afterwards. Instead of a triangle surface, the scene is represented by a very large number of small, soft, colored blobs (3D Gaussians) that are adjusted until they reproduce the original photos. The method comes from the 2023 SIGGRAPH paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering" by Kerbl, Kopanas, Leimkühler and Drettakis, which describes starting from the sparse points produced during camera calibration and rendering novel views in real time.

What you get is a scene you can orbit and fly through that looks close to the photos, including soft things that meshes handle badly, such as foliage, hair and fine detail. What you do not get is a surface. A splat scene is not a mesh, so it is not something you can send to a 3D printer or drop into a workflow that expects triangles without further conversion.

Like photogrammetry, it only shows what you captured. Areas you did not photograph come out blurry or empty.

How to choose: start from what you need at the end

The quickest way to pick a route is to ignore the input for a moment and look at the destination.

• A game or animation asset from a drawing or idea: Route A. You want a clean, editable mesh, and it does not have to match a real object. Expect to clean up the result in a 3D editor.
• A 3D print of a real object: Route B. Slicer software turns a 3D model into printing instructions, so you need a mesh, and you want it to match the real thing. Single-image tools that export STL or 3MF can also work when accuracy does not matter, for example a decorative figure.
• A realistic showcase of a real object, room or place: Route C. If the goal is to let someone look around something that exists, and you care more about how it looks than about editing its surface, Gaussian splatting is a good fit.
• A copy of a real object for AR or a product page that needs a standard mesh file: Route B.
• Measurements or engineering drawings: none of the casual tools. That is surveying and CAD territory, with scale references and dedicated software.

If you only have one photo and cannot take more, Route A is your only option. Just be clear with yourself and your audience that part of the model is a guess.

Turn photos into a 3D model: capture tips for routes B and C

Both multi-photo routes depend on the same thing: the software must recognize the same points across many views. Most failed results trace back to the capture, not the software. The COLMAP documentation gives guidance that applies to both routes:

• Avoid texture-less subjects and backgrounds, such as a white wall or an empty desk. If the scene is plain, add textured objects around it, such as posters.
• Keep lighting similar across shots and avoid glare on shiny surfaces.
• Make sure every part of the subject is seen in at least three images. More overlap is safer.
• Move the camera. Do not stand in one spot and only rotate; the software needs different viewpoints to work out depth.
• Avoid harsh lighting contrast, such as shooting against the sun or through doors and windows.
• More images are not automatically better. Past a certain point they mostly slow things down.

Some practical habits that follow from these: walk a full circle around the subject, then do a second circle from higher up looking down. Keep the subject still, and keep people and pets out of the frame. Soft, even light, such as an overcast day, is easier than direct sun. If you shoot video, move slowly so frames stay sharp.

As a rough guide to quantity, Apple's Object Capture guidance says 20 to 200 close-up images are usually enough for a single object.

Common misunderstandings

A few expectations cause most of the disappointment with image-to-3D tools.

• "One photo is enough for an accurate copy." It is enough for a plausible model. Accuracy needs views of every side.
• "A splat scene is a 3D model I can print." It is a way of displaying a scene, not a surface. Printing needs a mesh.
• "A mesh will look as real as the photos." Meshes are solid and editable, but thin, fuzzy or semi-transparent things often look wrong.
• "Better software will fix a bad capture." Missing coverage cannot be recovered by routes B or C. Only a generative tool will fill gaps, and it fills them with guesses.

Where SplatLoom fits

SplatLoom belongs to Route C only. It is a desktop app for Mac (Apple Silicon, macOS 15 or later) and Windows 10/11 that turns overlapping photos (typically 30 to 150) or a short video into a Gaussian splatting scene on your own computer, with no upload and no account. You pick the material, press Generate, tidy the result with crop and selection tools, and export a Gaussian PLY, a PNG still, an MP4 orbit video or a self-contained offline HTML viewer. It costs US$19.90 once.

It does not work from a single image, and it does not export a mesh, so it is the wrong tool for 3D printing or game-ready assets; use a photogrammetry program or a single-image generator for those. It also shares the usual limits of this route: reflective, transparent and moving subjects are hard, and results depend on how well you covered the subject. If what you want is a realistic, shareable view of a real object or place, it is worth a look.