ReprojectDepth

ReprojectImage handles the color; this is its twin for the depth channel. ReprojectDepth takes a depth tensor and remaps it between the pack's three projections (PINHOLE, FISHEYE, EQUIRECTANGULAR), optionally through a 4×4 camera transform, and - the part people actually appreciate - hands you back a mask of which pixels have valid depth in the new view.

Why does that mask matter? Because depth is how this pack lifts images into 3D. When you reproject a depth map into a new projection, the edges of the source view become holes in the target, and if you don't know where those holes are, you'll happily lift garbage geometry into a point cloud. The reprojected_mask output (1.0 = valid) is exactly the coverage signal you need to feed downstream nodes or to composite reprojected depth with something else.

The inputs

• depth - a depth TENSOR; the node accepts 2D [H,W], 3D, or 4D batch layouts and normalizes them internally.
• input_projection / output_projection and the two fov fields - same projection language as ReprojectImage. Fov matters just as much here; a wrong fov warps geometry even if the image looked fine.
• output_width / output_height - 0 (default) keeps the input resolution; set them to match a reprojected color image so the pair lines up pixel-for-pixel.
• transform_matrix - optional MAT_4X4; rotate or move the camera before sampling.
• inverse - invert the matrix when your poses use the other convention.

What comes out

• reprojected_depth - the depth map in the new projection.
• reprojected_mask - 1.0 where the new view actually has data, 0.0 where it sampled outside the source.

The classic usage is keeping color and depth in lockstep: run the same image through ReprojectImage and the corresponding depth through ReprojectDepth with identical projection/fov/matrix settings, and you end up with an RGB-D pair in the new view that you can hand to DepthToPointCloud.

One caveat to internalize

The pack's docs hammer on the difference between Z-depth and ray depth (radial distance). VideoPoseEstimator and the Depth-Anything metric models output Z-depth; the lifting and reprojection nodes expect ray depth. If geometry bows outward at wide fovs, you skipped the ZDepthToRayDepthNode step. ReprojectDepth itself is just resampling whatever it's given - garbage depth in, garbage depth out - so make sure the source is in the convention your downstream nodes want.

Installing it

Part of camera-comfyUI. Manager → search camera-comfyUI → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Alexankharin/camera-comfyUI.git
cd camera-comfyUI && python install.py

No extra dependencies for this node.

Common issues

• Depth and color don't line up - almost always mismatched projection, fov, or matrix settings between the ReprojectImage and ReprojectDepth instances. Mirror them exactly.
• Mask is all zeros at the edges - correct behavior; that's the boundary of your source view.
• Bow-shaped geometry at wide fov - Z-depth vs ray-depth mismatch upstream. Insert ZDepthToRayDepthNode.