Points To Image (Projection)

Points To Image (Orthographic) renders your point cloud like a scanner - no perspective, nothing changes size with distance. This one is the other half of the pair: it renders with a real perspective projection and gives you a fov dial, which is what makes a rotated cloud actually feel like it's in space rather than just sliding sideways.

The practical difference shows up the moment you animate. Rotate a head a few degrees and re-render through the orthographic node and you get a believable parallax flip, sure. Run the same rotation through the projection renderer and the near side of the face gets slightly bigger while the far side recedes - that's the "camera moved around it" illusion that the demo gif in the README is showing off. If you're building the parallax or pseudo-3D effect where a flat image gains motion, this is the renderer you'll reach for.

How it works

The node takes the cloud, centers it, places an imaginary camera in front of it, and projects every point through a perspective transform - X and Y divided by distance from the camera, the standard pinhole-camera math. The fov input sets how wide that camera's view is, and the camera distance is derived from it. Points behind the camera are culled rather than drawn, which is why rotating the cloud too far produces the correct "you've turned past the object" behavior. As with the ortho sibling, depth becomes brightness and closer points are drawn on top of farther ones.

The inputs that matter

• images - the size template. The node reads width and height from it and renders at that resolution. Feed it your original photo so the re-render lines up with the source frame.
• points - the Points3D cloud.
• color (default OFF) - OFF renders depth-as-grayscale; ON colors each point with the RGB it carried from the original image.
• fov (1–2000, default 35) - field of view in degrees. This is the parameter you'll actually touch. Low values flatten the perspective and zoom in; high values widen it into a fisheye. For a natural look keep it around 30–50 and animate the Transform node's rotation instead.

Output: IMAGE - a regular ComfyUI image, so it wires into anything downstream: save, preview, img2img, mask.

Install

Standard for the pack - Manager, search "ComfyUI_depthMapOperation", or:

cd ComfyUI/custom_nodes
git clone https://github.com/chri002/ComfyUI_depthMapOperation

Restart. No models to download, no keys; torch, numpy, opencv-python, scipy and pandas are all things a stock ComfyUI already has.

Where people get burned

The fov range is huge (up to 2000) and the default 35 is sensible - resist the urge to crank it. Extreme values are where the README's "artifacts with some extreme values" warning bites: a tiny fov can make the cloud clip through the near plane, and a huge one makes everything smear toward the center. Sparse clouds also render with holes here; if yours does, either bump quality in Image To Points (Torch) to thicken the cloud, or use the "advance" version of this renderer, which adds a hole-filling pass. And as always with this pack: single frame only, no batches.