Reversed-Z Depth: Improve Precision With a Consistent Depth Convention

Diagnose the depth representation

Two nearby surfaces can flicker because their projected depths become indistinguishable. Perspective depth normally follows a remapping of reciprocal view distance, so equal stored increments do not correspond to equal distances. Moving the near plane unnecessarily close makes conventional depth precision worse. Inspect the projection and depth format before changing geometry to hide the symptom. [1]

Reverse the mapping, not just the comparison

With a suitable zero-to-one projection and floating-point depth buffer, reversed Z maps the near plane to 1 and the far plane to 0. Floating-point values become denser toward zero, counteracting the reciprocal mapping's loss of distant precision. Reversing a uniformly quantized integer buffer does not provide the same benefit. [1]

The complete convention includes the projection, a greater-depth comparison and a clear value of 0. Choose strict or inclusive comparison according to the passes that must share equal depths. A comparison change alone makes an inconsistent pipeline.

Worked example: read the value in its proper units

For an illustrative infinite-far projection, let positive view distance be z, near distance be n, and reversed depth be d = n / z. With n = 0.2 metres, a surface at 2 metres produces 0.1, while a surface at 2000 metres produces 0.0001. At the near plane the result is 1; increasing distance approaches 0.

Those values are encoded depth, not metres. A post effect that subtracts them and interprets the difference as physical thickness is using the wrong units. Reconstruct view-space position or use the matching linearization before applying a distance-based threshold. The numbers demonstrate the mapping, not measured precision or performance.

Migrate every reader and writer

Create a small scene with overlapping surfaces at several distances and preserve reference captures. Change projection construction, depth comparison and clearing together. DirectXMath documents reversed projection by exchanging the near and far arguments for its stated convention; both distances must remain positive and unequal. [3]

Audit custom depth writes, full-screen geometry, sky handling, depth reconstruction and raw depth comparisons. Also review reduction rules used by depth hierarchies. Godot's migration explanation demonstrates why shader assumptions about near and far values can break even when ordinary geometry renders correctly. [2]

Verify limits across the complete frame

Test occlusion, background pixels, equal-depth passes and post effects after camera changes. Confirm the API's clip-space range and the actual attachment format; merely selecting a floating-point format is insufficient. Reversed Z does not separate exactly coincident surfaces or repair inaccurate vertex positions. Keep a practical near plane and validate visual correctness on the intended rendering paths.

Sources

• [1] Visualizing Depth Precision — Nathan Reed, NVIDIA Technical Blog.
• [2] Introducing Reverse Z — Clay John, Godot Engine.
• [3] XMMatrixPerspectiveFovLH — Microsoft, DirectXMath documentation.