How to generate a 3D model from an image: the four decisions that shape the result

Image-to-3D tools are sold as a button that turns a photo into a model. What they really do is predict the geometry the camera could not see, and the quality depends far more on the picture you feed in and the engine you choose than on any slider in the interface.

This tutorial walks the four decisions that shape the outcome: what the model can infer, how to prepare the image, how to choose between single-view and multiview engines, and how to review, repair and export the mesh. The reasoning applies equally to hosted services such as Meshy, Tripo and Rodin, to local stacks such as Hunyuan3D 2.x in ComfyUI, and to Open3DAI.

What image-to-3D predicts, and what it never measures

A single-photo model has to invent everything behind the silhouette. Most systems lift the image into an intermediate 3D representation, a radiance field, a set of Gaussians or an implicit signed-distance field, then extract a surface with a differentiable isosurface method such as marching cubes, DMTet or FlexiCubes. The geometry is therefore plausible rather than observed: back panels are smooth guesses, and thin straps or fingers fuse into blobs. The output also carries no real-world units: a generated mesh usually arrives normalised to roughly a unit cube, so a 3 cm ring and a 3 m statue come out the same size.

Decision one: is a single photograph enough to define the shape?

The honest answer depends on how much of the object the camera saw. One view works when the visible half implies the hidden half: bottles, mugs, simple creatures, stylised props, anything close to rotationally symmetric. It struggles when the object's identity lives on an unseen surface, such as the back of a chair, a cape or a car's far flank. If only one picture exists, a three-quarter view beats a straight-on view, because it shows two faces and gives the model more to interpolate.

• Exposure: even, diffuse light and a sharp subject. Specular highlights read as holes, and motion blur becomes melted edges.
• Background: a plain background or a cut-out with clean alpha. Clutter behind the subject returns as extra geometry.
• Resolution: 1024 px on the short side or larger. Engines resize internally to 512 or 1024, and detail lost in a small crop never returns.
• Framing: the subject fills roughly 70–90% of the frame, centred, with a margin for the crop step.
• One object per image: a mug, a book and a plant give the model no way to know which you meant.

Preparing the file: square crop, clean alpha, no baked shadow

Background removal is the one preprocessing step that improves results almost every time. rembg with a u2net model, BRIA RMBG 1.4 and SAM 2 all produce usable alpha mattes, and Photopea or GIMP will do the same by hand. Feather the edge by a pixel or two; a hard aliased border becomes a jagged ridge in the mesh. Then crop to a square canvas with the subject centred: the common 512×512 and 1024×1024 inputs are square, and a non-square image is padded or stretched. Do not retouch the object itself; a painted highlight is a feature the model will reproduce.

Decision two: one view, or several calibrated views?

Multiview engines, including TRELLIS 2, the multiview mode of Hunyuan3D 2.x, and the options in Meshy, Tripo and Rodin, accept three to six images and reconstruct a consistent object instead of guessing a back. The price is calibration: the views should share lighting, camera distance and white balance, and you should enter each shot's approximate azimuth and elevation where the interface allows. Mixing a phone photo with a rendered turntable frame usually yields a lopsided mesh; two consistent views beat five contradictory ones.

Decision three: how to judge a mesh in the first minute

Open the result in Blender or a glTF viewer and check four things. Look at the silhouette with flat shading and no texture: if the outline wobbles against the reference, the surface is noisy. Orbit 180 degrees and ask whether the hidden side is plausible. Enable backface culling and look for sheets that disappear, because a zero-thickness surface will not survive a slicer. Then check scale and orientation: a glTF file is Y-up and measured in metres, while Blender is Z-up and unitless.

1. Read the bounding box first. A unit-cube mesh needs a deliberate scale factor, not a guess.
2. Recalculate normals and check for inverted faces with backface culling switched on.
3. Run a manifold check: Blender's 3D-Print Toolbox, the Meshmixer Inspector or a short trimesh report lists holes and non-manifold edges.
4. Decimate to budget for a game asset: 20k–50k triangles for a hero prop, 2k–8k for background clutter.
5. Compare a render against the source photo at the same angle, then decide between regenerating and repairing. A better crop or a second view is cheaper than an hour of sculpting.

Failure cases you will meet, and the fix for each

Repairs that belong in Blender or the slicer, not the generator

• Fused thin parts: straps, cables and fingers merge into blobs. Regenerate from a tighter crop, or cut them free and rebuild from swept geometry.
• Zero-thickness sheets: caps, cloth and fins come back single-sided. Solidify to 0.8–1.2 mm for FDM with a 0.4 mm nozzle, then re-check manifoldness.
• Hallucinated symmetry: a feature from one side gets mirrored. Fix it with a second view, or mirror the other half deliberately.
• Baked-in lighting: a hard shadow from the photo reappears as geometry and as a dark patch in the albedo. Start from a shadow-free cut-out, or repaint the texture in Substance Painter or Blender.
• Wrong scale and axis: the mesh arrives a thousand times too small or lying on its side. Set unit and axis at import, then re-export.

Decision four: export the format your next tool reads

GLB and glTF 2.0 are the default for game engines and the web: one file, embedded textures, metres and Y-up. Draco compression cuts file size several-fold at the cost of a runtime decoder, and KTX2 textures keep video memory down. OBJ with a separate MTL file remains the safest interchange format for older tools. STL carries triangles and no units; 3MF carries units, colours and multiple objects and suits Bambu Studio, OrcaSlicer, PrusaSlicer or Cura. PLY is useful when the engine returns vertex colours.

1. Decide units first: millimetres for printing, metres for engines. Scale in the modelling tool and apply the transform.
2. Decimate before exporting, not after. A 200k-triangle prop with a 2K texture is slow to iterate on and over most game budgets.
3. Keep the source photo, engine name and settings beside the exported file; a mesh with no provenance is hard to regenerate.
4. Round-trip once: re-import the exported file into an empty scene and check the bounding box, textures and axis.

Run it yourself in Open3DAI

Send the same prompt or image to several engines, compare the meshes side by side, then export GLB, glTF, OBJ, STL, PLY or 3MF. Creating an account grants free credits and needs no card.