motion-background

Motion Background → the bed

A background that draws attention has failed. Slow, low contrast, looping with no visible join, and cheap enough that the picture in front of it still composites.

timeline-edit-ops carries the op contract for the calls named here. color-motion chooses the palette the bed is built from, and motion-curves writes the loop when no wrapping preset holds the seam.

Pick the cheapest thing that reads

Look	Build it from	Cost
Flat or gradient wash	One full-frame rect shape with fillStyle linear or radial	Shape clips do not count against the video layer cap
Soft mesh or conic wash	One full-frame shape with a generator effect, mode: "meshGradient" or "conicGradient"	One procedural effect, no model call
Shaped colour blobs	Radial-filled ellipse clips, low opacity, blurred	More control over each blob's motion
Noise, fractal, particles, light leak, grid	One full-frame shape with the matching generator mode	Procedural and seedable
Drifting texture	An image clip with a kenBurns animation	Outside the video layer cap
Smoke, ink, or clouds	A generated video bed, looped	One video layer and a model call
Colour cycle	Any of the above plus a hueShift loop	An additional grade

At most eight video layers composite at once, resolved from the lowest track index. Text, shape and image clips are not counted, so a bed built from shapes costs nothing in that budget. Nine overlapping video clips means the bottom one silently does not draw, reported as layer_cap_exceeded.

A procedural bed

Add a full-frame shape and apply a generator effect to it. Modes are noise, fractal, conicGradient, meshGradient, gradientField, particles, lightLeak, and gridPattern. colorA and colorB set the two colours; scale, amount, angle, time, and seed shape the field. For movement, set animate: true, which samples the current timeline time into the effect. Keep a fixed seed for repeatable frames. This clock keeps moving; it does not make the procedural field wrap at the clip boundary. For a seamless loop, use a wrapping clip animation or match the endpoints in a rendered loop.

{"op": "set_effects", "target": "Bed", "effects": [
  {"type": "generator", "mode": "meshGradient",
   "colorA": "#10233c", "colorB": "#ec6b45", "scale": 6,
   "amount": 0.6, "seed": 9, "animate": true}]}

For a dark wash that bands, put a fixed-seed stylize dither after the generator or gradient in the effect list. Dither breaks up 8-bit steps but cannot restore precision lost in the source. color-motion covers this look.

A mesh from shapes

For individual moving blobs, stack radial fills over a base. Each ellipse is its own clip on an overlay track, at 30–60% opacity, with a blur effect wide enough that no edge reads:

{"op": "add_shape_clip", "shape": {"kind": "ellipse",
 "x": 0.05, "y": 0.1, "width": 0.6, "height": 0.7,
 "fillStyle": {"type": "radial", "stops": [
   {"offset": 0, "color": "#5b8cff"}, {"offset": 1, "color": "#5b8cff00"}]}},
 "durationMs": 12000}

{"op": "set_effects", "target": "Ellipse", "effects": [{"type": "blur", "radius": 120}]}

Then give each blob a different loop: float with amplitude 0.01–0.015 (a fraction of frame height each way, so at most 3% of travel), breathe at intensity 0.03–0.08, and cycle lengths that do not divide each other — 9000, 13000, 17000ms — so the composite never visibly repeats.

blendMode screen or add on a blob over a dark base builds light rather than covering it; multiply or soft-light deepens instead. That choice matters more than the fill colours.

Loops that actually loop

float, breathe, rotate and hueShift are built to wrap: they start and end at the same value, so durationMs is a period and the seam is invisible. A custom loop curve must do the same — the value at t: 1 has to equal the value at t: 0 or the bed jumps every cycle.

kenBurns is the exception: it is a one-shot across the whole clip, holding at its end, so it does not repeat and its durationMs is ignored. On a long bed a drift from kenBurns finishes and then sits still. For continuous drift over a long clip use float with a long period instead.

Periods that read as ambient: 8000–30000ms. Anything under about 5000ms reads as a screensaver, and text in front of it becomes hard to hold.

Keep it behind

• Background sits on the highest track index, because the lowest draws on top. A bed that covers the picture is a track-order mistake, not an opacity one.
• Contrast: pull it down with a color effect at saturation 0.5–0.7 and brightness slightly negative before touching opacity. A desaturated bed reads as depth; a faded one reads as a mistake.
• Amplitude: nothing in the bed should travel more than about 3% of the frame, or the eye starts tracking it.
• Give the foreground a scrim if the bed's bright regions cross the type: a black rect shape at 40–60% opacity on a track index between the bed and the text. validate_timeline reports text_illegible for what it can measure, and says nothing about gradients — look at the frame.

A generated bed

For smoke, ink, or clouds that the procedural modes cannot draw, generate video. Use find_model with text_to_video, then generate_video, then add_media_clip with the returned asset:// reference. The find_model result names a prompting_skill for the line it picked; load it, because each line wants the prompt shaped differently. Whatever the line, prompt for slow, even motion with no cuts and no subject; a bed with an event in it is a shot, not a bed. On a line that writes native audio (Seedance 2, Veo 3, MiniMax H3, Kling 2.6 and later) the bed arrives with a soundtrack: either brief it as room tone and keep it, or mute the clip under the track you meant to use — video-audio-continuity is the rule for which.

A generated clip does not loop. Cover a longer sequence by duplicating it and crossfading the joins: overlap the copies by 800–1500ms on one track, which dissolves with no call at all, or set_transition crossfade on the incoming copy to pin the length.

Reduced motion, and knowing when to stop

A bed exists to sit under something. Before adding a second idea to it, check whether the first one is doing anything at all — a drift too slow to see is budget spent on nothing, and one too fast is the reason nobody read the headline. One idea, one amplitude, one period family.

Check it

• preview_timeline_frame with a range across a full cycle and sheet: true, which is how you see whether a loop breathes evenly or lurches. Sampling three timecodes will not show a seam.
• The same sweep with the foreground on, to confirm the type still reads at every point in the cycle rather than only where you happened to look.
• render_timeline at a low preview_scale while iterating. A bed is judged over seconds, and a still cannot tell you whether it is calm.