A single photograph has no true depth. Software guesses depth from shading, edges, and perspective, or it triangulates the same points across many photos. Three practical routes: an AI generator from one picture, photogrammetry from a photo set, or modeling by hand in Blender using the image as a reference.
Pick the route from the job. A product mockup, a game mesh, a museum scan, and a CT-derived bone are not the same problem. Medical imaging is one use case at the end, not the method.
Which method
| Method | Best for | Input | Skill | Speed |
|---|---|---|---|---|
| AI generator | Mockups, concept art, a base mesh you will edit | One clear photo | Low | Seconds to a minute |
| Photogrammetry | A measurable digital copy of a real object | 20–200+ overlapping photos | Medium | Hours (capture + process) |
| Manual modeling | Stylized work, sketches, anything the photo cannot constrain | One or more reference images | High | Hours to days |
- Need a 3D product view on a site this afternoon → AI.
- Need a digital archive of an artifact, or a dimensionally honest scan → photogrammetry.
- Need a character from a drawing → Blender (or another DCC), image as blueprint.
AI from one image
Tools such as Luma, Meshy, and Kaedim take one picture and return a mesh. They are fast. They are not a survey of the object. The unseen side is invented.
Give the model a chance:
- One subject, uncluttered background (white/gray or transparent).
- Even light. Hard shadows and clipped highlights hide shape.
- A short prompt if the UI asks: material, style, use (“metallic,” “low-poly game asset”).
Treat the first export as a draft. Remesh if the polycount is insane. Then take it into Blender. Common exports: OBJ (works everywhere) and GLB (web / AR, textures in one file).
Photogrammetry
Many photos, known overlap, software triangulates. Free starting point: Meshroom. Paid and faster: RealityCapture.
Capture rules that actually matter:
- 60–80% overlap between neighboring shots. Too few photos is the usual failure.
- Soft, even light (overcast, or bounced indoor light). Specular hits move with the camera and break matching.
- Simple objects: ~30–50 frames. Busy ones: 100–200+.
- Shiny, glass, or featureless surfaces fail. Matte scanning spray (or a light dusting that you can clean off) and a few bits of tape give the matcher points to lock onto.
Pipeline: align cameras → sparse cloud → dense cloud → mesh → project textures from the source photos. That last step is why photogrammetry looks like the object and a single-image AI model often does not.
Clean it in Blender
Blender is the usual free finish tool, whatever produced the raw mesh.
- Cleanup. Edit Mode → Select All by Trait for loose verts, holes, non-manifold edges (an edge welded to more than two faces). Merge by Distance for stacked verts. Fill obvious holes.
- Retopology. Scans and AI dumps are dense, ugly triangles. Build a cleaner quad mesh on top (Shrinkwrap helps). You need this if the model will deform or run in real time.
- UVs. Unwrap so a 2D texture lies on the surface without stretching. Skip this only if you will never texture or bake.
- Export for the destination. Games: low-poly + normal map from the high-poly, often FBX. Print: watertight, STL. Web/AR: GLB / glTF. Archive / other DCCs: OBJ.
A hybrid that works: AI or photogrammetry for the shape, Blender for the mesh you actually ship.
Medical imaging is a different input
A CT or MRI is already a 3D volume (a stack of slices), not a photograph. You segment tissue, then meshing is the same last mile (repair, STL). Do not run a clinical DICOM series through a consumer “photo to 3D” site and expect a reliable bone. The medical path is CT scan to 3D model and, for how the MRI volume is built, 3D MRI reconstruction.
PYCAD builds custom web DICOM viewers when the 3D view has to live in a clinical or medtech workflow rather than a DCC. Case studies.