How to 3D Print a Photo

The four methods differ in one way that matters more than any other: how much of a three dimensional object they actually contain. A photo holds one viewpoint. Everything past that viewpoint has to come from somewhere, and each method solves that differently, or declines to solve it at all.

Two of them do not try. A lithophane and a relief both keep your picture flat and vary its thickness, so what you print is a picture with depth rather than an object with volume. That is not a shortcoming when a picture is what you wanted, and both print beautifully with almost no setup.

The other two produce something with a back. A scan gets it by looking at the object from every side, which means you need the object. A reconstruction gets it by inferring the shape from the one view you have. Below is what each of them puts on your print bed.

A lithophane is an old idea that 3D printing revived. You print a thin plate whose thickness varies with the brightness of your image: thick where the photo is dark, thin where it is light. Held up to a window or backed with an LED panel, the light coming through renders the picture in greyscale.

It is the most forgiving option on this list. A phone snapshot works. There is nothing to model, nothing to clean up, and free generators have existed for years, so the whole job is upload, set a size, download the STL. Printed in white PLA with a lot of top and bottom layers, the result is genuinely striking.

What it is not is an object. A lithophane of your dog is a rectangle. Turn it edge on and it is four millimetres of plastic. If what you pictured was a small figure of the dog, this is not the method, and no amount of tuning gets you there from here.

A relief maps brightness to height in a different way: instead of a plate you look through, you get the shape raised off a flat backing, the way a face sits on a coin. Most tools that call themselves an image to STL converter are doing this, and it is why so many of them ask you to upload a logo rather than a photograph.

Line art, logos, lettering and silhouettes come out well, because the source already reads as shape rather than as a scene. Photographs come out as a lumpy landscape, since the converter has no way of knowing that a bright shirt is not closer to the camera than a dark face.

The giveaway is the back. Turn a relief over and it is flat, because the method never had any information about what is back there. That is fine for a plaque, a stamp, a keyring or anything else that hangs on a wall or sits against a surface.

Photogrammetry builds a model out of many photographs by finding the same feature in several of them and working out where it must be in space. It is what produces the scanned props you see in games and the terrain models in surveying.

It is also the most accurate method here, because it is not inferring anything: every surface in the result was actually photographed. If you need a copy of a specific real object, down to the dent in it, this is the method.

The catch is in the inputs. You need the object in front of you, forty or more shots that walk all the way around it, consistent lighting and a surface with enough texture for the software to find features on. A single photo from the internet cannot be turned into a photogrammetry scan, which rules it out for most people who arrive at this question.

The fourth method takes one image and produces a complete model: front, back, sides and depth, as a closed mesh, with textures. Instead of measuring the shape the way photogrammetry does, it recognises what the object is and builds the whole of it, using everything it has learned about how objects of that kind are shaped.

This is the only one of the four that turns a single photograph into something you can rotate. A phone snapshot of a toy gives you a model of the toy with a back on it, in a couple of minutes, with no modelling and no rig of forty photographs.

It is also the method that fits how most people arrive here. You have one picture, of something you do not physically own, and you want a small object of it. That is exactly the case the other three cannot serve.

STL is the format almost every slicer opens without argument, so whichever method you picked, STL is where you are heading. It stores one thing: a surface made of triangles. No colour, no materials, no units, no scene. That simplicity is why it has outlasted every format meant to replace it.

Which means the conversion to STL is rarely the hard part. What decides whether you get something printable is what went into the converter, and what the converter thought it was looking at.

One subject, clearly separated from its background. Every method on this page gets better with the same photo: the object roughly filling the frame, lit evenly, against a background that is not busy or the same colour as the subject.

Resolution matters less than people expect. A clean 1000 pixel image beats a noisy 4000 pixel one, because noise reads as texture and texture reads as shape. If the photo has heavy JPEG artefacts, motion blur or a strong shadow across the subject, fix that before you convert anything.

For reliefs and lithophanes the file type genuinely matters. PNG is lossless and, more usefully, can carry transparency: a PNG with the background cut out tells the converter exactly where the subject ends, which is the single biggest improvement you can make to a relief. JPG has no transparency and its compression softens exactly the edges the converter is looking for.

For reconstruction it matters much less, because the model is identifying an object rather than reading brightness values. JPG, PNG and WEBP all work. A cut out background still helps, but a clean photo on a plain surface is enough.

If you are converting a logo or line art, export it as PNG with a transparent background and hard edges, and skip any smoothing your editor offers. Everything you soften, the converter has to guess about.

Reconstruction tools usually produce GLB, because GLB carries the mesh and its textures in one file and STL cannot carry textures at all. If you are printing in a single colour that difference costs you nothing, and the conversion is mechanical: the triangles are already there, and going to STL just discards what the printer was never going to use.

OBJ sits in between. It holds geometry and can reference a material file, and most slicers read it, though STL remains the safer choice if you are handing the file to someone else or to a print service.

Convert once, at the end. Keep the GLB, since it is the version with the textures in it, and export STL from it when you are ready to slice. If you later want the model in a game engine rather than on a printer, you will want that GLB back.

An STL file has no colour field. Not a limited one, none: the format stores triangles and nothing else. Any colour you have seen described as being 'in an STL' is either a non standard extension that most software ignores, or the slicer showing you a preview colour it chose itself.

For a single extruder printer that is irrelevant, since the model comes out in whatever filament is loaded. For multi colour printing you either paint the model in the slicer, which works well for flat regions and badly for photographic detail, or you keep the textured GLB and use a workflow built around it.

The practical rule: choose your filament for the object, not for the photo. A reconstruction printed in one colour reads as a sculpture of the thing, and that is usually better than a rough approximation of its real colours.

Two of the four methods have been free for years, and they are free in the real sense: no account, no watermark, no credit counter.

Lithophane generators run in the browser and have done since long before any of this was fashionable. Upload, choose a shape, set the dimensions, download the STL. For a picture you intend to light, this is not a compromise, it is simply the right tool, and paying for it would be strange.

Relief conversion is free too, either through one of the many image to STL converters online or through Blender, where a displacement modifier driven by your image does the same job with more control. Blender costs nothing, runs everywhere and will do a great deal more than this once you know it. The price is an afternoon of learning where the buttons are.

Photogrammetry has free options as well, though the real cost there is never the software. It is owning the object and taking forty usable photographs of it.

Reconstruction is the method where free gets complicated, because unlike the other three it costs real money to run. Every reconstruction is time on a GPU, so a tool offering it free is either giving you a limited number of them or recovering the cost somewhere else.

The place to look is not the credit allowance, which is easy to compare, but the licence and the export. A free tier commonly caps the resolution, waters the mesh down, keeps the high quality textures behind the paid plan, publishes your results to a public gallery, or grants you a licence that does not cover selling what you make. None of that shows up until you have already picked a tool and made something with it.

Read the export formats and the commercial terms before the credit count. A free model you cannot sell, or cannot download at full quality, is not a saving if either of those is the reason you started.

For lithophanes and reliefs the free tools are the answer and there is nothing to weigh up. The question only becomes interesting when you want the object rather than the picture.

A reconstruction is only as good as what you hand it, and the difference between a great result and a mediocre one is almost always the photograph rather than the settings.

Give the model as much of the object as you can in one frame. A three quarter view, where you can see the front and one side at once, carries far more shape information than a straight on shot, and the results show it immediately.

Keep the whole subject in frame with a little room around it. A cropped edge is a surface the model has to complete, and completing it is guesswork you could have avoided by stepping back half a metre.

Even, soft light beats dramatic light. A hard shadow across the subject reads as a change in shape, so an object lit by a window on an overcast day converts better than the same object under a bright spotlight.

Put the subject against a plain background of a different colour. The model separates subject from background before it does anything else, and a patterned tablecloth or a wall the same shade as the object makes that harder for no reason.

Avoid reflections and glass. Anything the camera cannot see the surface of is a surface with no information behind it, which is the one situation where a reconstruction has to work hardest.

If the result is wrong in a way you can name, and the reason is visible in the photo, take another photo. A missing handle, a face in shadow, a limb cropped at the frame edge: none of those are fixed by running the same picture again with different settings.

Change settings when the shape is right and the finish is not. Texture quality, polygon budget and topology all change how the model is built out; they do not change what the model understood the object to be.

Slicers are stricter than viewers. A model can look perfect on screen and still refuse to slice, because the slicer has to answer a question the viewer never asks: which side of this surface is inside?

A watertight mesh is one with no holes in its surface, so it encloses a definite volume. That is what lets a slicer decide where to put walls and infill. A mesh with a gap in it has no inside, and the slicer either produces nothing or produces something with a wall missing.

Reconstruction produces closed meshes, which is the practical reason it slices without a repair step. Reliefs are usually closed too, since the flat backing seals them. Photogrammetry scans very often are not: any surface the camera never saw comes back as a hole, and the underside is almost always one of them.

If a file will not slice, run it through a mesh repair tool before you start changing print settings. Most slicers have one built in, and it fixes the majority of cases without you needing to know what it did.

A standard nozzle lays down a line about 0.4 mm wide. Any detail thinner than that does not get printed small, it gets skipped, which is why fine engraving and thin whiskers vanish while everything around them comes out fine.

Scale up, or simplify. A model that is unprintable at 40 mm is often perfect at 100 mm, and enlarging costs nothing but filament and time. Where you cannot scale up, thicken the thin parts before slicing rather than hoping the printer manages.

Watch for thin walls in reliefs in particular. A relief generated from a dark photograph can end up under a millimetre thick across large areas, which prints as something you can see through and break with a thumb.

The format stores numbers with no unit attached. Your slicer assumes millimetres, and most tools export on that assumption, but a file that came out of a program working in inches or metres arrives either tiny or enormous.

Set the size once, in the slicer, against a dimension you know: the height of the object, the width of the base, whatever you can picture. Then check it against the bed before you slice, because a model 30 times too large is easy to miss when the preview scales to fit.

Your printer does not read STL. It reads gcode, a list of instructions telling it where to move and how much to extrude. Turning one into the other is slicing, and it is the last step between the file and the object.

Layer height decides how much of the surface detail survives. For a reconstruction of anything organic, 0.12 mm to 0.16 mm is a good balance; going finer mostly adds hours. For a lithophane the layer height is the image resolution, so go as fine as your printer manages.

Infill barely matters for decorative prints. Ten to fifteen percent is plenty, and the walls do the visible work. Lithophanes are the exception again: they want zero infill and a lot of top and bottom layers, since the whole point is controlled thickness.

Print speed costs you more quality than any other single setting on detailed models. Slowing the outer wall alone, which most slicers expose as its own setting, gets you most of the improvement without doubling the print time.