Difforum · Camera (expressions)

The Difforum pitch is "no camera expressions to write". They're right that most people shouldn't need them. But if you came from Deforum, this node is the comfort blanket: the same 0:(expr) syntax on seven axes, backed by a safe expression evaluator rather than eval().

The syntax

Each axis takes a Deforum-style schedule where the keys are frames and the values are expressions:

0:(0), 60:(2*sin(2*pi*t/30))

t is the current frame and s is seconds. pi, sin, cos, abs, min, max and friends are available. When audio is connected, so are the six bands - amp, low, mid, high, onset, beat.

Here's the part that trips people up, and it's the same trap as the original: values are per-frame increments, not positions. A constant translation_x = 0:(2) doesn't put the frame at x=2, it pans a bit every single frame, and the result accumulates through the feedback loop. That's the whole design - it's why a schedule of motions works at all - but it means "2" and "0.2" are wildly different moves. zoom follows the same logic: it's a per-frame scale factor, so 0:(1.0 + 0.004*sin(2*pi*t/48)) is a gentle breathing zoom, and 0:(1.01) is 1% closer every frame, which gets silly fast.

The default rotation_3d_z of 0:(0.3) is a nice illustration: a third of a degree of roll per frame, which over 120 frames is a full 36 degrees of rotation. Subtle constant, big cumulative result.

Inputs

Seven axes: translation_x, translation_y, translation_z (3D only), rotation_3d_x (tilt), rotation_3d_y (pan), rotation_3d_z (roll), and zoom. Plus:

• mode - 2d or 3d, matching the rest of the pack. In 3D you want a depth map downstream or you're back to pseudo-3D.
• fov (40) - lens field of view in degrees. 20–30 is long lens, 40 normal, 60+ wide. This matters more than people expect: it's what makes a dolly feel like a dolly rather than a zoom.
• audio - makes the bands available inside expressions.
• Seven *_curve sockets - a DIFFORUM_SCHEDULE (an Audio Curve, or a Schedule) replaces the text for that axis. This is the useful trick: write the shapes you like by hand, and hand the axis you want reactive over to a schedule you can tweak without touching the text.

Outputs

camera and info. camera overrides the Director's camera wherever it plugs in - Feedback Sampler, Guide Frames, Storyboard, Camera → Prompt, Camera Export. The info string is worth reading at least once per session; it reports the resulting motion range, which is how you catch "I typed 2 instead of 0.2" before rendering.

Install

Manager → search "Difforum", or:

cd ComfyUI/custom_nodes && git clone https://github.com/chillithebillis/Difforum.git difforum

Restart. numpy and a built-in AST evaluator - notably no eval(), which is a real difference from the 2022-era scripts people copy expressions out of.

Where people get burned

Motion through space needs depth. Pan, roll and zoom are in-plane and work on a flat image; dolly, orbit and tilt are camera-in-a-room moves, and without a depth map the engine falls back to a pseudo-3D affine - dolly becomes zoom. It works, it just won't have the parallax you're picturing, and the report says so.

The other one is continuity across transition. Expressions are per-frame increments, so a discontinuity in the expression is a jump in the image. If you build a move out of piecewise segments, ease them into each other or you'll get a visible snap - which is the same problem the Camera (keys) node solves with its own transition slider.