Relief reconstruction is real and useful, but it only knows what the image shows. Here is what comes out and how to check it before you slice.
Convert an image to STLYes, AI can convert an image to STL, but only in one direction: it turns brightness and contrast into height. A photograph, a logo or a scanned drawing becomes a relief model you can print, not a full object with a hidden backside. The output is real STL geometry, watertight when the settings are right, and it opens in any slicer. The caveat is that the software is guessing depth from shading and edges, so the result is only as good as what the flat image shows. Nothing appears that was never visible.
What that guessing produces falls into two groups. Flat, high-contrast artwork with clear shapes converts cleanly: silhouettes, coins, stamps, trail maps, line art. Photographs also work, but as a smooth height field, so a face becomes a shallow mask rather than a bust. Where it fails is anything the image does not describe: the back of a subject, an undercut, a hole behind a bar, or text smaller than about 1 mm tall after scaling. You get a plausible surface, not a reconstruction of the original object.
Judging the file is a printing question, not a looking question. Open the STL in your slicer, confirm the mesh is manifold with no open edges, then slice it at the layer height you actually plan to use. A relief that looks crisp in the preview can still have walls thinner than two extrusion widths, and anything under roughly 0.8 mm will not survive a 0.4 mm nozzle. Height range matters too: 2 mm of relief spread across a 150 mm base reads as almost flat.
The pipeline starts by turning colour into one number per pixel. Most converters use luminance, a weighted mix of red, green and blue, so a dark pixel becomes a low point and a bright pixel becomes a high point. That mapping is usually invertible, which matters for engraved text and for images with white backgrounds. The grayscale grid that comes out is the height map, and its resolution is whatever you uploaded. Downsampling here is normal: it keeps the mesh light enough to slice.
The height map is then meshed. Each sample becomes a vertex, or a small block of vertices if the tool reduces the grid, and triangles are stitched between neighbours. Vertical walls drop from the edges down to a flat base and get capped, which is what makes the solid watertight and printable. File size is decided here. A 2000 by 2000 image holds four million height samples, and a mesh at that density can be too heavy for a slicer to open comfortably.
Where AI enters is the depth estimate, not the meshing. A plain converter treats brightness as height and nothing more. A learned model has seen a large number of images and can separate surfaces using shading, focus and learned shape priors, so a sphere reads as round instead of flat grey. That is a real gain on photographs and a real risk on line art, where the model may invent a soft slope where the drawing showed a hard edge. Switching to brightness-driven mode avoids it.
The same tool that produces a clean coin can produce an unusable portrait, so know which job you are giving it.
Logos, silhouettes, stamps, coins and single-colour maps convert to crisp reliefs with near-vertical walls. There is little for the model to infer, because every edge in the mesh matches an edge in the image. Print them at 0.1 mm layers and the outline stays sharp.
Portraits, terrain and texture shots become smooth reliefs that read naturally against raking light, and they work well as lithophanes. The rule that keeps them legible is a sensible total height: enough rise to catch light, not so much that the model turns into spikes.
Anything narrower than about 0.8 mm disappears once the nozzle gets involved, and small text turns to mush. Text inside an image needs to be embossed large, or treated as a separate solid model. Scaling the whole relief up helps only until the base outgrows your build plate.
The back of a head, the underside of a car and the inside of a cup are not in the pixel data, and no model can recover them. The converter returns a shell of what the camera saw. If you need the full object, this is a modelling job, not a conversion job.
Image to STL conversion pays off in specific situations and wastes time in others.
Upload a JPG, PNG, SVG or HEIC up to 20 MB and get an STL back in the browser, on any desktop or phone. No account, nothing to install.
Yes. A picture can be read as a height map and meshed into a watertight STL that slices normally. What you get is a relief of what the picture shows, so it is not a full 3D model of the subject with a back, an underside or an interior.
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