Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
600
1.21k

16-bit vs 8-bit Depth Map Comparison

A small, measurable example of why bit depth in a depth map shows up in a laser engrave, a lithophane, or a CNC relief.

The package is built from the public lion-medallion sample published by Depth-Map.com. The genuine 16-bit manufacturing PNG and the closed STL were downloaded from the site on 15 November 2026. The 8-bit file and the "fake 16-bit" file were derived from that PNG in this repository, so the three depth maps are the same height field.

This is a file-format comparison, not a trained model and not a benchmark of Depth Anything V3.

What is in the repo

File What it is Measured on 2026-11-15
lion-source.jpg Public sample photograph (/demo/medallion-source.jpg) 1024×1024 RGB JPEG
lion-depth-16bit.png Genuine 16-bit grayscale depth map (/demo/lion-depth-16.png) PNG bit depth 16; 19,662 unique values; min 0, max 19,661
lion-depth-8bit.png Same field quantized to 8-bit on the same 0–65,535 scale (value // 257) PNG bit depth 8; 77 unique values; min 0, max 76
lion-depth-8bit-upscaled-to-16bit.png That 8-bit file written back into a 16-bit PNG (value * 257) PNG bit depth 16, but only 77 unique codes, all multiples of 257
lion-relief.stl Closed relief mesh published with the sample (/demo/lion-relief.stl) Binary STL; header starts with depth-map.com relief
previews/ 8-bit RGB / JPEG shaded illustrations from the site, useful for seeing terraces. These are not manufacturing depth maps. See folder
verification.json Hashes, header fields, and unique-value counts —
verify.py Re-runs the Pillow counts —

Why the 16-bit file does not peak at 65,535

Depth-Map.com writes depth on a 10 mm scale: 65,535 maps to 10 mm. The site's 16-bit vs 8-bit guide describes this lion export as a 3 mm relief on a 1.5 mm base (Invert on). A 3 mm relief therefore uses about a third of the gray range, and the brightest pixel is 19,661 — which is what we measured.

An 8-bit PNG on that same 10 mm scale only has 77 height steps for a 3 mm relief. Stretching those 77 codes into a 16-bit container does not invent the missing steps. The upscaled file in this repo is an example of that failure mode: the header says 16-bit, the unique-value count does not.

The guide's expected ImageMagick reading for this manufacturing PNG is:

Depth: 16-bit
Channel statistics:
  Gray:
    min: 0 (0)
    max: 19661

A stretched 8-bit file is the one whose values are multiples of 257 and whose distinct count is 256 or fewer (here, 77).

How to verify

python verify.py

Or with Pillow:

from PIL import Image

im = Image.open("lion-depth-16bit.png")
print(im.mode)  # I;16
print(len(set(im.getdata())))  # 19662
print(min(im.getdata()), max(im.getdata()))  # 0 19661

ImageMagick (identify -verbose) is the method described on the site guide. These counts were measured with Pillow, not ImageMagick, because this package was prepared in an environment without identify.

Source files

The 8-bit and upscaled-16-bit PNGs were produced here from lion-depth-16bit.png. They are not extra downloads from the site.

License

CC BY 4.0. Attribution: Depth-Map.com, https://depth-map.com/. The sample is the site's public lion-medallion demo. The depth field was estimated with a Depth Anything V3 model; this repository does not redistribute model weights.

Limitations

  • One subject, one export (3 mm relief, 1.5 mm base, invert on, as described on the guide). This is not a dataset for training.
  • Preview images in previews/ are 8-bit illustrations. Do not treat them as 16-bit depth maps.
  • Unique-value counts come from Pillow on 2026-11-15. Re-run verify.py if you need to confirm a local copy.
Downloads last month
81

Spaces using depthmap/16bit-vs-8bit-depth-maps 2