A depth map can look fine as a thumbnail and still fail on a Blender Displacement modifier: the relief comes out inverted, the background ripples, or a shallow face turns into a staircase. That is usually the height field, or a mesh too coarse to sample it.
Before I render or export, I confirm the file is real 16-bit grayscale, look at the height field, clean what I can without a new generate, and only then set up the modifier.
What Blender is sampling
In this workflow a depth map, a height map, and a displacement map are the same thing: a single-channel grayscale image whose brightness is a height per pixel. Bright raises the surface; dark stays low. Blender's Displacement modifier reads that image as a texture and moves vertices. A Displace or Bump node in a shader only fakes lighting and leaves the silhouette flat. For a portrait that casts a real shadow, or geometry you can export to USD, FBX, or a slicer, use the modifier.
The modifier can only move vertices that already exist. A plane with four corners gives you four samples, and the relief reads as a flat tile. In Edit Mode, subdivide the plane six times, or leave it low-poly and add a Subdivision Surface modifier (viewport 3 / render 6 is a workable starting point). That geometry is what the texture is sampled against.
Check the file, not the browser preview
Browsers are a poor place to judge a manufacturing PNG. When depth is written on a 10 mm full scale, a 3 mm relief peaks well below the brightest possible gray, so the real file can look almost black even when the data is fine.
First confirm bit depth. Some tools write an 8-bit map and upscale it into a 16-bit container, which does not invent the missing steps. ImageMagick is enough:
identify -verbose your-depth-map.png | grep -E "Depth|Statistics"
You want Depth: 16-bit. On a 10 mm scale, a 3 mm export should peak around 19,661, not 65,535 — a lower max is expected. A file with 256 or fewer distinct grays, or values on multiples of 257, is an 8-bit field in a 16-bit wrapper.
The difference shows up once a process samples the height. In an 8-bit PNG, a 3 mm relief on that same 10 mm scale has 77 height steps; in a 16-bit PNG it has 19,661. Shaded, the 8-bit steps read as terraces, and in Blender they show up on a dense mesh even if a preview looked smooth. Keep the import as 16-bit grayscale: do not save a JPEG, and do not let the texture slot treat it as sRGB.
Look at the height field itself
Before you touch Strength, flip between the photo and the grayscale.
Is the subject one clear object? Single faces, emblems, and coin-style objects usually give a clean occlusion boundary. Group shots tend to flatten. Glass, water, and polished chrome often invent false near-far edges. Front light helps; side light can etch a cheek shadow as fake depth, and backlight can collapse the subject into mid-gray.
Is high actually high? If a coin or medallion comes back hollow, the model has read the recessed field as closer than the rim — swap high and low before you displace. Watch the background too: a busy wall or textured table will rise and fall with the subject, and that noise later looks like a rippling floor around the portrait.
I used a depth map generator to get a 16-bit PNG from one photograph, then judged the field with those questions before I opened Blender. Generate is one model call; invert, gamma, flatten, and crop should be decided against the preview, not a first render.
Clean what you can without a new generate
If the estimate is roughly right, you usually do not need another photograph through the model. These sliders rewrite the same field:
- Flatten background (0 to 0.80) pulls a textured backdrop down toward a constant height.
- Crop (0 to 35%) trims an edge that is stealing attention.
- Gamma (0.30 to 3.00) compresses highlights or lifts shadows when the relief feels too harsh or too soft.
- Invert swaps near and far when the direction is wrong.
A new photo spends a credit; moving these controls does not. If the subject is still a group shot, a backlight silhouette, or a mirrored surface, no slider will invent a depth the model did not see. Reshoot with front light and one subject, then generate again.
Then set Strength and modifier order
With a clean 16-bit PNG, add a Displacement modifier, create a new Image or Movie texture, and open the file. Strength is a multiplier on the gray range, not a millimetre setting. A useful rule of thumb is Strength ≈ maximum relief depth divided by the plane's width in Blender units: on a 1-metre plane, 0.03–0.08 reads as subtle relief, 0.1–0.2 as a printed bas-relief, and 0.3 or more as a stylised exaggeration. Start at 0.1 and adjust after the first render.
Optionally stack a Subdivision Surface modifier above Displacement so it smooths already-displaced geometry and keeps portrait detail from stair-stepping on a sparse mesh. Subsurf below densifies the plane first, but uses more memory. Cycles and Eevee both render the displaced mesh. If you need the geometry outside Blender, Apply the modifier first; a modifier stack does not travel out in a USD or FBX.
This photo-to-height workflow is not a procedural terrain generator, and it does not export a normal map or a vector-displacement EXR. Blender reads the 16-bit PNG natively; if you need EXR or a tangent-space VDM, use a different tool.
Do not debug Strength until you have checked bit depth, invert, background, and mesh density. Most bad displacements were already visible in the grayscale.
Top comments (0)