Silhouette extrusion, lithophane height maps and AI image-to-3D all end in the same file extension and almost nothing else. Here is what each route controls, which photos survive it, and the checks that happen before the first layer.
Type image to STL into a search box and the results quietly mix three unrelated workflows: tracing an outline and extruding it, mapping brightness to thickness for a lithophane, and handing the picture to a generative model that invents a shape the camera never saw. They share a file extension and almost nothing else, including the ways they fail.
This tutorial walks all three, gives the settings that change the outcome, and says when to abandon a route. It also covers where Open3DAI fits, and what it leaves to Blender and the slicer.
STL stores triangle vertices and normals and nothing else: no colour, no material, no units, no way to mark a surface as glass. Turning an image into an STL is therefore never a format conversion; it is a decision about which 3D shape the flat picture stands for. Three answers cover most of the work: the outline becomes a wall, the brightness becomes height, or a model guesses the geometry the photo only hints at.
Vectorise the bitmap, save an SVG, import it into Blender, Fusion 360, Tinkercad or OpenSCAD, and extrude. Inkscape's Trace Bitmap and Illustrator's Image Trace handle the first step. Typical depths are 2–4 mm for a flat sign, 3–5 mm for a keyring, and 6–10 mm where the part takes a screw. Every edge is a vertical wall, so the slicer has nothing to interpret and the first slice is usually clean.
A lithophane encodes the image in material thickness: dark pixels print thick, bright pixels print thin, and a backlight turns the thickness map back into a picture. Typical numbers are 0.8 mm at the thinnest point, 2.8–3.2 mm at the thickest, a 100 × 150 mm panel for a portrait, 0.1 mm layers with a 0.4 mm nozzle, and 100% infill so the light path stays solid. Print it standing on its long edge, so the layer lines run vertically and act as the image's horizontal resolution. itslitho, Lithophane Maker and the Cura lithophane script do the mapping; a subdivided plane with a Displace modifier does the same in Blender.
| Route | What the image controls | Best source image | Typical tools | Starting settings |
|---|---|---|---|---|
| Outline extrude | The silhouette path only | Logos, line art, stencils, high-contrast cut-outs | Inkscape, Illustrator, Blender, Tinkercad, Fusion 360 | 2–4 mm extrusion, 0.2 mm layers |
| Lithophane height map | Greyscale turns into wall thickness | Portraits and landscapes with even lighting | itslitho, Lithophane Maker, Cura script, Blender Displace | 0.8–3 mm thickness, 0.1 mm layers, 100% infill |
| Depth-map relief | Estimated depth becomes surface height | Single-subject photos with clean edges | Depth Anything, MiDaS, Blender Displace, Meshmixer | Relief depth 2–6 mm, 2 mm backing plate |
| AI image-to-3D | A generated full 3D shape | One clear object, plain background, 1–4 views | Open3DAI, Hunyuan3D, TRELLIS 2, hosted services | Export GLB, decimate, check manifold, scale to mm |
Some pictures are worth dropping before you spend an hour on them: transparent or translucent objects, polished metal with blown highlights, two subjects at different depths, strong wide-angle distortion. Thin hair strands are the usual culprit: at 0.4 mm nozzle resolution they sit below the smallest printable feature anyway. If the only photo you have is a family portrait, clean it first: cut the background, lift contrast, correct perspective.
Image-to-3D models return a triangulated surface with a baked texture, often between 20,000 and 200,000 triangles depending on the model and export settings. STL keeps the triangles and drops the texture, so a colourful preview and a grey STL are the same geometry. Expect triangle soup instead of quad topology, walls thinner than one nozzle width, internal shells and small floating fragments. Blender's 3D-Print Toolbox flags non-manifold edges and self-intersections, and Meshmixer's Inspector fills holes. A 30–80 mm figure slices fine under 150k triangles; a 200 mm statue is happier below 300k.
Open3DAI covers the generative route rather than the vector or lithophane ones. It accepts a single image or a multiview set, runs text-to-3D and image-to-3D generation through one local procedural engine plus optional third-party engine adapters you connect with your own API keys, and lets you compare the same input across those engines side by side before exporting. It also does AI texturing, and exports GLB, glTF, OBJ, STL, PLY and 3MF, alongside 48 format converters, 12 browser tools and a REST API.
Open3DAI generates and exports meshes. It does not scan objects, measure real dimensions, rig or animate, retopologise, unwrap UVs or produce CAD-quality parametric solids, and it does not promise print-ready output with no cleanup. Treat every STL as a starting mesh: a few minutes in Blender, Meshmixer or Windows 3D Builder is normal print work.
Wall thickness of 0.8–1.2 mm is the usual safe band with a 0.4 mm nozzle, two or three perimeters; features below roughly 0.4–0.6 mm tend to vanish or print as strings. Anything overhanging more than about 45° from vertical needs supports, which is why a relief prints best flat on its back and a lithophane standing upright. Check the scale every time: STL has no units inside it and most slicers assume millimetres, so a 2.5× error at export turns a 60 mm figure into 150 mm. A 1 kg spool of PLA typically costs about 15–25 USD at the time of writing, and a 100 × 150 mm lithophane uses a few tens of grams of it.
Engine availability, credit pricing and third-party terms change, and slicer defaults change with them. Confirm the current numbers on the vendor page and in your slicer profile before committing a spool or a budget.
PNG, JPEG, WebP, BMP and TIFF all work as inputs to the vector, lithophane and AI routes; SVG is the intermediate for the extrusion route. For AI generation, a JPEG or PNG with at least 1500 px on the long edge and a plain background gives the most predictable result. A transparent PNG alpha helps on the lithophane and relief routes because it removes the background before the mapping step.
A single subject filling roughly 60–80% of the frame, even diffuse lighting, a plain contrasting background, and at least 1500 px on the long edge. Transparent, mirror-like, hairy or thin-stranded subjects are the usual failures, because the depth information simply is not present in the pixels, and no generator or height map can recover it.
Yes, twice over. Trace the silhouette and extrude it — Inkscape or Illustrator to SVG, then Blender, Tinkercad, Fusion 360 or OpenSCAD — or map greyscale to thickness for a lithophane with itslitho, Lithophane Maker or the Cura script. Both routes are deterministic: you set every number and nothing is guessed.
Image-to-3D output is usually a triangle soup with thin walls, internal shells and small disconnected fragments, and the STL format itself makes no promise of a closed surface. Run Blender's 3D-Print Toolbox or Meshmixer's Inspector, merge vertices by distance, fill holes, delete loose parts, and re-check for non-manifold edges before slicing.
STL stores coordinates without units, and most slicers read those numbers as millimetres. Measure the imported model against a dimension you know, scale by the ratio you need, apply the scale, and only then export. A 2.5× mistake is easy to make at export time, and it turns a 60 mm figure into a 150 mm one.
0.2 mm layers with a 0.4 mm nozzle is the common default. Lithophanes are the exception at 0.1 mm layers, because the layer lines are the image resolution. Keep walls in the 0.8–1.2 mm band, treat anything under roughly 0.5 mm as decorative, and add supports wherever a surface overhangs more than about 45° from vertical.
No. STL has no colour or texture channel, so the photo's colours are discarded at export. Keep the GLB, OBJ or 3MF file for anything coloured, and plan on 3MF or a multi-material printer if the colour has to appear in the printed part itself rather than in a render.
No. Generation invents the surfaces the photo never showed, so dimensions come out plausible rather than measured, which is fine for shape and look and wrong for fit. Photogrammetry and structured-light scanning measure those surfaces instead, with their own limits on reflective, transparent and moving subjects. Use generation for appearance, measurement for interfaces.
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.
Four ways to produce an STL: browser primitives, parametric CAD, Blender sculpting and AI image-to-3D. What each route is good at, the export settings that decide printability, and the checks worth running before you slice.
One picture can become a screen-only model, a game character or a printed figure, and each route needs different work. Here is the five-step photo-to-STL pipeline, the print settings a small figure needs, and how to convert an existing AI action-figure image into a real object.
A photograph carries no depth, so the printed result depends on how much the software is allowed to guess. This guide covers capture, three routes to a printable solid, the mesh checks that matter, and the scale and wall thickness a small figure needs.