Image to 3D: turning a photo into a 3D model with photogrammetry, AI and Blender
Short answer
Three routes take you from an image to a 3D model. Photogrammetry rebuilds a real object from dozens of photos, accurately. AI tools generate a volume from a single image, quickly but by inventing what cannot be seen. Finally, you can model it yourself in Blender over a reference photo. In all three cases, Blender is used to clean up, texture and render the result.
Turning a photo into a 3D object is an appealing idea: a single image, and the model is ready. The reality is more nuanced. Depending on the result you want (an accurate real object, a quick set piece or a clean character to animate), the right method is not the same. This guide compares the three routes, explains what to expect from each, and shows how to take the result further in Blender.
Three ways to go from an image to 3D
| Method | What you need | Result | What it is for |
|---|---|---|---|
| Photogrammetry | Dozens of photos of the real object | Accurate, textured, dense mesh | Real objects, heritage, realistic sets |
| AI from an image | One image, sometimes a few | Fast, hidden side invented | Background props, prototypes, ideas to test |
| Modelling over a reference | One or more photos, and time | Clean, light, under control | Characters, objects to animate or print |
Photogrammetry: rebuilding a real object
Photogrammetry starts from many photos of the same object, taken while moving around it. The software first works out the position of each camera, by spotting points shared from one image to the next, then reconstructs a point cloud, a mesh and its texture. Meshroom, the desktop application of the open source AliceVision project, released under the MPL-2.0 licence for Windows and Linux, goes in this way from raw photos to a textured mesh.
Quality depends above all on the shoot:
- Many photos with a wide overlap, from several heights, with no gap in the circuit around the object.
- Soft, constant light, without strong cast shadows or direct flash.
- A matte, textured object. Shiny, transparent or perfectly plain surfaces reconstruct badly, for lack of details to recognise.
- Sharp photos, without camera shake, with the same focus.
The result is realistic but heavy: hundreds of thousands or millions of triangles, and an irregular mesh. The tutorial on integrating photogrammetry into Blender (in French) shows how to lighten it and light it.
AI: a 3D model from a single image
Online services, often presented as AI 3D modelling tools, generate a textured object from a single image or a description. The principle: a model trained on large quantities of 3D objects guesses the most likely shape, including the parts that cannot be seen in the image. It is very fast, and sometimes free within certain usage limits.
What to know before using them:
- The hidden side is invented: the back of an object photographed from the front is a guess, often an approximate one.
- The mesh is rarely clean: dense, disorganised triangles, hard to animate or modify.
- The texture is built into the geometry, sometimes with lighting "baked" into it, which shows under different light.
- Terms of use vary from one service to another: check the right to commercial use and your rights to the source image.
These models work well for background props, quick mock-ups and exploring ideas. For an object in the foreground, they are better used as a starting point to rework.
Modelling it yourself over a reference photo
The safest method for a clean model is still to model in Blender over photos. Add an image with Shift+A, Image, as a reference or a background, aligned with a front or side view (Numpad 1, Numpad 3), then build the shape in polygons following its outlines. It takes longer, but the mesh is light, organised and ready to texture or animate. The 3D modelling guide presents this method and the others.
Taking the result further in Blender
- Import the model, as glTF, OBJ or FBX depending on what the tool exports, and check its scale.
- Lighten the mesh with the Decimate modifier, or rebuild it with Remesh, if the object does not need fine detail.
- Redo a clean topology if the object is to be animated or deformed: retopology (in French) redraws a light mesh over the dense model.
- Transfer the details from the dense model to the light one, as a normal map, by baking.
- Review the materials: remove the lighting baked into the texture if possible, adjust the roughness, and set the data maps to Non-Color.
An imported model can also be used for 3D printing, provided it is closed and cleaned up: the guide to 3D printing with Blender details these checks. For models that are already clean and under a clear licence, the guide to free 3D models for Blender lists sources.
Frequently asked questions
Can you create a 3D model from a single photo?
Yes, with artificial intelligence tools that guess the hidden volume from an image. The result is useful for a set prop or a quick mock-up, but the unseen side is invented and the mesh needs cleaning up. For an accurate copy of a real object, photogrammetry from many photos remains more precise.
What is photogrammetry?
It is the reconstruction of an object or a place in 3D from many photos taken from every angle. The software works out the position of each shot, then calculates a point cloud and a textured mesh. Meshroom, which is open source, is one example.
Can Blender turn an image into a 3D model?
Not automatically. Blender is used to import, clean up and texture the model produced by a photogrammetry or AI tool, or to model it yourself over a photo displayed as a reference image.
Can AI-generated models be used commercially?
It depends on the terms of the service used, which vary from one tool to another and between free and paid plans. Read them before delivering a model to a client, and also check that you have the right to use the source image.
Rendering the animation on TURNA
When a scene is already well optimised and the frame count is what takes the time, the remaining lever is rendering several frames at once. TURNA does this for Blender: an add-on (Blender 4.0 to 5.2, Cycles and Eevee) sends the scene in one click, and the frames are split across 1, 2, 4 or 8 GPUs in parallel, RTX 5090 first and RTX 4090 when none is free. The add-on shows the estimated cost before anything starts. Prices start at €2.90 per hour, billed on the time actually used, with a 30-second minimum, and new accounts get €25 of free credit. A single still image is still rendered by one machine, so the gain is on animations. Tiers and worked examples are on the Blender render farm page.
Models from scans or AI are often very dense and heavy to render. The guides to Blender textures and to rendering in Blender help you prepare them before rendering.
