Understand resampling
The original image is resized proportionally to a grid with at most 96 samples on either side. More samples preserve finer changes but increase triangles and processing cost. A very small feature may disappear when resampled even when it is visible in the original image.
Match detail to scale
A 100 mm wide relief with a 64-sample grid has roughly 1.6 mm between samples. Increasing the physical width without increasing samples makes each feature coarser. Compare that spacing with the detail you need and your printer’s achievable resolution; a dense image does not imply a dense output mesh.
Keep the source and settings
Download the processed height image alongside the STL. Save the width, base thickness, height range and interpretation in the report so that later versions are reproducible. If your slicer repairs or changes geometry, compare the repaired dimensions and layer preview before using it.
Calculate sample spacing from the finished size
For a square 100 mm model sampled at 64 points across, there are 63 intervals, so spacing is about 1.587 mm. At 96 points it becomes about 1.053 mm. Doubling the finished width doubles those spacings at the same grid setting. This explains why a large plaque can look less detailed even when its source image contains many pixels.
Compare a known narrow feature with this spacing before generating. A line narrower than one interval may be weakened or disappear during resampling. Raising the grid from 64 to 96 helps within the supported range, but does not create details absent from the image or recover geometry hidden behind the photographed subject.
Review a proportionally resized image
The longest side is bounded by the chosen 16–96 sample setting, and the other side is resized proportionally. Grid rounding can slightly change the final aspect ratio; use the reported model dimensions as the definitive output dimensions. Higher resolution increases mesh faces and validation work. Keep the smaller trial result if it already meets the physical detail requirement, and compare slicer toolpaths rather than judging quality solely from triangle count.