PNG to STL: turn a height map into a mesh without guessing scale

A PNG can store 8-bit or 16-bit grayscale, RGB, RGBA, palette, or transparency. A height map usually uses one channel, where black is the low end and white is the high end, but the convention is not guaranteed. Some tools write the opposite direction, some store a normalized range that does not use the full 0-255 or 0-65535 span, and some attach an alpha channel that means nothing to the geometry.

Before conversion, inspect the file in an image editor that reports bit depth and the actual minimum and maximum values. If the useful range occupies only the middle third of the histogram, the relief will look flat until you remap the black and white points. If the file is 8-bit, do not put it into a 16-bit container and call it higher precision; the missing levels cannot be recreated.

A height map is already a measurement or an artistic decision: each pixel says how high the surface should be. A photograph is a record of reflected light. Bright skin and a bright wall can share a value while representing completely different depths. If the PNG is a photograph, the first job is depth estimation, not direct conversion. If it is a height map, skip the model call and convert the data you already own.

DepthCarve accepts both paths. The free normal-map converter and the relief workbench can read a grayscale PNG as height, while the AI path estimates relative depth from a photograph. The important part is to know which one you have before you judge the result.

8-bit grayscale gives 256 height levels. On a 6 mm relief, one level is about 0.023 mm, which is below many CNC depth steps but can become visible on a smooth gradient or a large print. 16-bit gives 65,536 levels and more headroom for a gentle slope. The advantage is real only when the source actually contains that precision and the downstream process can resolve it.

A common failure is an 8-bit image saved into a 16-bit file. The container has more levels, but the values still land on the original 256 steps. DepthCarve labels the free compatibility file honestly and encodes the true 16-bit pair from raw floating-point depth only in the approved production bundle.

Decide whether white means near or far. If the map was authored for a different renderer or CAM package, the direction may be reversed. Use the invert control and compare the 2D and 3D previews before export. A reversed field can look plausible in grayscale and become a hole where a face should be.

Transparency is not height. If the PNG has an alpha channel, either flatten it onto a known background or use it as a mask to remove unwanted geometry. Holes, floating islands, and disconnected pieces are a manufacturing problem even when the preview looks correct. Flatten the background, add a stable border, and inspect the mesh for closed topology.

A 1024-pixel-wide map says nothing about millimeters. Set the width, relief height, base thickness, and mesh grid deliberately. A 60 mm test tile and a 600 mm wall panel can use the same PNG and need very different relief heights, base thicknesses, and mesh densities. The base is what makes the mesh a solid; it is not padding.

The free preview STL is intentionally low resolution: it answers scale and orientation questions quickly. The production mesh is denser and validated for closed topology. If the preview already shows a flat subject, a broken border, or an impossible slope, a denser mesh will only reproduce the same problem with more triangles.

Import the STL into the target program at real size. Confirm the bounding box, the base, and the manifold status. In a slicer, look for islands, unsupported overhangs, and walls thinner than the nozzle. In CAM, compare the deepest relief with the cutter's reach and the stock thickness, then simulate the toolpath.

Keep the PNG and the manifest with the finished part. If a later revision needs a different width or a deeper relief, the source data and settings explain what changed. A mesh alone does not record the decisions that produced it.