ComfyUI Stat-Imp Nodes

Clean, MIT-licensed nodes for Deforum-style animation loops in ComfyUI. This release ships three nodes: a depth-aware 3D camera warp and two color-coherence nodes (a fixed-anchor matcher and a rolling/EMA matcher) that stop the feedback loop's color from drifting into neon "acid".

Node: Depth Warp 3D

Menu group: Stat-Imp / Deforum / Depth3D.

An exact forward port of original Deforum's transform_image_3d (deforum-stable-diffusion, helpers/animation.py), reimplemented in pure torch — no pytorch3d, no kornia at runtime. It is verified pixel-identical to the real original on 9 camera axes by an oracle test (run in the development repo against the genuine Deforum/py3d code; not shipped here because it needs the original Deforum source to compare against).

Why it exists: older ComfyUI Deforum 3D nodes mirrored every in-plane axis (tx, ty, rx, ry, rz). This node reproduces the original's motion exactly.

Inputs

• image (IMAGE) — frame to warp.
• depth (IMAGE) — a depth map for that frame (see Depth below).
• translation_x / translation_y / translation_z — camera translation.
• rotation_3d_x / rotation_3d_y / rotation_3d_z — camera rotation (degrees).
• fov — field of view (degrees).
• translation_scale — global translation gain.
• depth_invert — True for MiDaS-style inverse depth (default True).
• depth_equalize — histogram-equalize the depth before use.
• padding_mode / sampling_mode — grid-sample edge / interpolation behaviour.

Outputs

• image (IMAGE) — the warped frame.
• reveal_mask (MASK) — optional. Marks newly-revealed (disoccluded) pixels that have no source data. Wire into Set Latent Noise Mask to inpaint only those regions. Safe to leave unconnected.

Motion directions (positive value)

Axis	On screen
translation_x	content moves right
translation_y	content moves up
translation_z	zoom in / approach
rotation_3d_x	tilt up
rotation_3d_y	pan right
rotation_3d_z	roll clockwise

Depth is external (bring your own)

This node does not estimate depth — it consumes a depth-map image from any upstream estimator. Recommended: comfyui_controlnet_aux:

• MiDaS-DepthMapPreprocessor (auto-downloads Intel/dpt-hybrid-midas), or
• a Depth Anything V2 preprocessor (recommended — generally better than the MiDaS + AdaBins pipeline the 2023 original used).

To switch estimators at runtime, wire two preprocessors into an AnySwitch node and select the active input.

Fidelity note: the camera geometry matches original Deforum exactly. The original blended MiDaS + AdaBins for depth; this pack leaves depth to external nodes, so depth values differ from the 2023 original. Geometry is exact; the depth source is your choice.

Node: Color Match LAB

Menu group: Stat-Imp / Deforum / Color.

Per-frame histogram color match to a reference image, in LAB / HSV / RGB. Unlike mean/std matchers, it remaps the full per-channel distribution, so it holds the whole palette instead of just the average — killing the slow color drift (the neon "acid" degeneration) of low-denoise img2img feedback loops. Pure torch, stateless.

Inputs

• image (IMAGE) — frame to recolor.
• reference (IMAGE) — palette source (e.g. the frame-0 init image for a locked palette, or any image).
• color_space — LAB (default; match color, keep detail natural), HSV, or RGB.
• strength (0–1) — how strongly to pull toward the reference palette (0 = off, 1 = full match). Schedulable.
• preserve_luminance — keep the frame's own brightness; match color only (LAB/HSV).

Outputs

• image (IMAGE) — the recolored frame.

It is a color stabilizer, not an injector: matching to a fixed reference holds a palette but cannot create new prompt colors. For an evolving palette, lower strength or use Rolling Color Match below.

Node: Rolling Color Match (EMA)

Menu group: Stat-Imp / Deforum / Color.

Same histogram match, but the reference is a moving anchor — an exponential moving average of past frames — instead of a fixed image. This lets the palette drift with the animation while staying coherent (no frame-0 jail, no acid). Stateless: the anchor is carried through the loop, not stored in the node.

Inputs

• image (IMAGE) — frame to recolor.
• anchor_in (IMAGE) — current anchor; thread from the loop's carry slot (ForLoopOpen value), bootstrapped at frame 0 with a reference image.
• alpha (0–1) — anchor update speed: 0 = frozen anchor (≡ Color Match LAB), 1 = follow the previous frame, middle = exponential moving average. Schedulable.
• color_space / strength / preserve_luminance — as in Color Match LAB.

Outputs

• matched (IMAGE) — the recolored frame.
• anchor_out (IMAGE) — updated anchor; wire to ForLoopClose to carry it to the next frame.

alpha × strength

strength = how much recolor (0 = none, 1 = full). alpha = what to match toward (0 = fixed origin, 1 = previous frame, middle = fading memory). They are independent; at strength 0 the node is a pass-through.

Practical guidance:

• To kill acid while keeping the original palette, use a very low alpha (≈0.02–0.1). Higher alpha lets the anchor chase recent frames, so a dark/dull source degenerates quickly. alpha of one frame is a fast memory — small is strong.
• Schedule strength (e.g. ramp 0→0.7 over the first frames) so early frames aren't slammed to the anchor before the scene forms.
• A poor/dark source palette can't be enriched by matching alone — feed a richer image as the bootstrap/reference and the whole loop inherits its colors.
• preserve_luminance keeps brightness, so this node controls color only; it will not stop a loop that is darkening on its own (that's a denoise/luminance issue).

Install

Manual: copy this folder into ComfyUI/custom_nodes/, restart ComfyUI, and hard-refresh the browser. No extra Python dependencies (pure torch, already in ComfyUI). The node appears under the Stat-Imp/Deforum menu group.

ComfyUI Manager: once published to the Comfy Registry, install by name.

Credits

• Camera math ported from Deforum (deforum-stable-diffusion, transform_image_3d).
• Projection conventions follow py3d_tools (BSD).
• Color coherence is a clean-room pure-torch reimplementation of Deforum's maintain_colors (helpers/colors.py) — quantile histogram match, no skimage/cv2.
• This wrapper and pure-torch reimplementation: MIT (see LICENSE).