Depth Warp 3D

If you've tried the older ComfyUI Deforum 3D nodes and the camera felt subtly wrong - the in-plane axes mirrored, the pan rolling the wrong way - this is the node that fixes it. Depth Warp 3D is an exact forward port of original Deforum's transform_image_3d, verified pixel-identical to the real thing on all 9 camera axes by an oracle test. When the source moves right, your frame moves right, the same way the 2022 A1111-era Deforum did.

It's the camera-motion workhorse of the Stat-Imp pack: feed it a frame plus that frame's depth map, dial in translation and rotation, and out comes the warped frame - the per-frame "push" that turns a still image into a moving one inside a Deforum animation loop.

How it works

The mechanism is a forward projection, and it's pure torch - no pytorch3d, no kornia at runtime. It builds a world grid from your image (x, y in -1..1, z from the depth map), projects that grid through an old (identity) camera and a new (moved) camera, takes the screen-space offset, and resamples the image by it. The motion conventions are copied verbatim from the original's py3d_tools: row-vector transform, symmetric FoV projection, and a translate vector of [-tx, +ty, -tz] / 200. That last bit is why positive translation_x moves content right, translation_y moves it up, translation_z zooms in, and the rotations are a pan-right / tilt-up / roll-clockwise set that matches the original exactly.

Inputs that matter

• image and depth - the frame and its depth map. The node does not estimate depth; you bring it. The pack recommends comfyui_controlnet_aux: a MiDaS-DepthMapPreprocessor (auto-downloads Intel/dpt-hybrid-midas) or a Depth Anything V2 preprocessor, which is generally better than the MiDaS+AdaBins pipeline the 2023 original used. Wire two preprocessors into an AnySwitch and you can swap estimators at runtime.
• translation_x/y/z, rotation_3d_x/y/z - the camera move, usually driven by a ValueSchedule so the motion animates.
• fov (default 40) and translation_scale (default 0.005) - the two knobs that set how "wide" and how strong the move is.
• depth_invert - defaults to True, because MiDaS/Depth Anything output inverse depth (white = near). If your warp looks inside-out, this is the first thing to check.
• padding_mode / sampling_mode - edge and interpolation behavior; border / bilinear are the sane defaults.

Outputs

• image - the warped frame, straight into your loop.
• reveal_mask - optional but genuinely useful. It marks newly-revealed (disoccluded) pixels that have no source data. Wire it into Set Latent Noise Mask and you can inpaint only those regions - the standard fix for the smeared "holes" big camera moves leave behind. Safe to leave unconnected if your moves are gentle.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Statistical-Impossibility/comfyui-Stat-Imp-nodes

The core warp is pure torch with zero extra Python dependencies - no pytorch3d, no kornia. To run the shipped example workflow you also need the harness fork (https://github.com/Statistical-Impossibility/deforum-comfy-nodes), which widens the loop's carry slots and fixes the value-schedule parser. Restart ComfyUI, hard-refresh, and you'll find the node under Stat-Imp / Deforum / Depth3D. The depth preprocessor comes from comfyui_controlnet_aux, installed separately.

Common issues

The biggest misunderstanding is expecting depth to come free - no depth input, no warp; the node consumes a depth map, it doesn't make one. Next is polarity: with depth_invert left at its default and a non-inverse depth map, motion reads backwards, so flip the toggle and re-check. And one honest caveat from the pack's own docs: the geometry matches original Deforum exactly, but the depth values won't - the 2023 original blended MiDaS with AdaBins, and this pack leaves depth entirely to you. If your motion looks different from a vintage Deforum clip, that's almost always the depth source, not the camera math.