An AI model reads depth and shape cues from a single flat image, estimates the geometry behind them, and writes an STL you can slice and print.
Convert an image to STLThe image to STL AI route is the one you take when a single photograph is all you have and the shape matters more than the surface. A heightmap can only push pixels up and down a flat plane, which is fine for a coin relief and useless for a mug handle. A reconstruction model reads the whole silhouette, estimates the volume behind it, and hands you a mesh that stands up in three axes. It is an estimate, not a measurement, and this page says exactly where that estimate holds.
Most people arrive with a JPEG and no scanner, no calipers, and no CAD session ahead of them. Photogrammetry wants many overlapping frames taken from a steady orbit; a structured-light scanner wants hardware and a dark room. Reconstruction from a single image exists because the alternative is redrawing the object by hand in a modeller, guessing every radius. It trades accuracy for speed, and it is honest about that trade. If you need a dimension to land on 0.2 mm, measure the real part. If you need a recognisable bracket blank to refine, this gets you there in one upload.
Each stage is a judgement call, and each one is a place where a weak input turns into a weak mesh.
The model first decides which pixels belong to the object and which belong to the background. A plain, evenly lit backdrop makes that call easy, so a white wall or a sheet of grey card beats a cluttered kitchen counter. Anything it labels background is discarded before geometry is built.
A trained network predicts how far each pixel of the subject sits from the camera, and which way the surface faces there. That gives a dense depth map rather than a set of measured points. Soft shadows and glossy reflections fool it, because they look like shape changes and are not.
The depth map becomes a surface, then the unseen rear of the object is closed off with a plausible shell. Triangles are merged and simplified so the file stays a workable size. What comes out is a single closed body, though thin walls may still need a repair pass.
Four steps, in the order that saves the most re-printing.
One subject, filling most of the frame, against a background that differs in tone from the object. Even light from a diffused source helps more than resolution does. JPG, PNG, GIF, BMP, TIFF, WebP, SVG and HEIC are accepted, up to 20 MB.
Trim away clutter around the subject so the segmentation has less to guess at. If hard shadows stretch across the object, pick a different frame instead of pushing through. The preview shows the reconstructed shape before you commit to a file.
Download the mesh and open it in your slicer's 3D view, not just the print preview. Look for holes, floating shells and spikes that follow a shadow. Most of what needs fixing is visible in the first few seconds of orbiting.
STL stores no units, so set the size in millimetres yourself and check wall thickness against a 0.4 mm nozzle. Cut a flat base where the object will sit so it adheres. Print a small version at 0.2 mm layers before committing filament to the full size.
These are the cases where a different route, or a ruler, will serve you better.
Upload a file up to 20 MB and get an STL back in the same browser tab. It runs on Windows, macOS, Linux, iOS and Android, needs no install and no account, and a basic conversion is free.
It is a trained network that predicts depth and surface direction from the pixels, then builds geometry from those predictions. A filter would only remap brightness into height. The output here has volume, so you can orbit it and see the back of the object, even though that back was inferred rather than seen.
The pages that answer the next question after this one.
Skip the complex modeling software. Upload your image now and get a print-ready STL file instantly.
Start Converting Free