Depth Parallax Frames

This is the "living photo" node - the one that takes a flat image plus its depth map and animates the camera through it so the foreground drifts past the background. The effect has been internet-bait since LeiaPix and DepthFlow made it effortless, and Depth Parallax Frames is the in-graph version from gokayfem's ComfyUI-Depth-Visualization pack. What makes it genuinely better than the weekend depth-warp scripts is what comes out the other side: real frame batches and validity masks that tell you exactly which pixels were invented by the warp, so you can inpaint the holes instead of shipping a video with smeared edges.

How it works

Each frame, the node moves a virtual camera along your chosen path - horizontal sway, vertical sway, ellipse (a nice lazy dolly-around), or dolly (push in and out, which scales about the frame center). Displacement per pixel comes from (depth - depth_center) * amplitude_px: pixels at depth_center stay put, everything closer or farther shifts by an amount proportional to how far they are from that pivot. So depth_center is your "what stays fixed" dial - set it where your subject sits. Higher amplitude_px means more drama and, inevitably, more tearing.

Warping is bilinear sampling with clamp, and here's the honest part: pixels that would sample from outside the source image don't exist. They get flagged 0 in the validity mask. The frames output is the pretty video; the validity_masks output is the truth about where it's faking.

Inputs and outputs that matter

• image and depth_map - both in. Batches must match, or one of them can be a single image broadcast to the other's count.
• frames (24, 2–240) - how many frames per source image.
• amplitude_px (24) - total motion strength in pixels. Start here, not lower, and see the "burned" section.
• path - one of the four camera motions.
• loop (true) - seamless looping phases. Off if you want a one-way move.
• depth_center (0.5) - the depth that doesn't move.

Outputs:

• frames (IMAGE batch) - feed to a video save node (VHS/VideoHelperSuite) or an AnimateDiff-style batch consumer.
• validity_masks (MASK batch) - 1 where sampled inside the source, 0 at exposed edges. This is the inpainting driver.
• manifest_json (STRING) - counts, path, amplitude, and a note on mask semantics.

Install

ComfyUI Manager, search ComfyUI-Depth-Visualization, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-Depth-Visualization.git
python -m pip install -r ComfyUI-Depth-Visualization/requirements.txt

Restart, node under depth/toolkit. Deps are numpy + Pillow; no models, fully offline, CPU-warp.

Where people get burned

• Amplitude is the tearing dial. The hard limit of parallax is structural: the depth map only describes the original camera view, and when the camera moves you reveal regions with no data. Push amplitude_px past ~40–60 and the image starts tearing no matter how clean your input is. Keep the motion subtle, or budget for the fix below.
• The fix is inpainting, and the masks are the map. Run the frames through an inpaint pass masked by validity_masks (inverted) to fill exposed edges - the README's own framing is that these masks drive compositing/inpainting. This is the workflow that separates "demo clip" from "shippable video."
• Un-normalized depth breaks depth_center. If your depth isn't [0,1], depth_center 0.5 points somewhere meaningless. Depth Normalize first, like every other node in this pack wants.
• Wrong-sized batches error out. One image vs one depth is fine; mismatched multi-image batches raise a clear ValueError. Match them or broadcast a single input.