Adjust Speed

Change video playback speed (0.05x slow-mo to 100x hyper-speed), reverse, time-remap with piecewise segments, with optional motion-compensated frame interpolation.

Overview

The Adjust Speed (Speed Ramp) node is a pure-FFmpeg post-processor. It changes the playback speed of any input video, optionally reverses it, and optionally interpolates new in-between frames for cinematic slow-motion. For shots that need varying speed across the clip (the classic “normal → slow-mo → normal” action-cam beat), it supports a piecewise time remap defined by start/end/speed segments.

It is distinct from the Temporal parameter picker — Temporal nudges the AI generator’s prompt with motion-character text (“slow motion, reverse playback”), Adjust Speed actually warps the rendered video frames.

Configuration

Field	Type	Default	Description
Speed	number	`1.0`	Constant speed factor, 0.05× to 100×. Ignored when `ramps` is non-empty.
Reverse	bool	`false`	Reverse playback (applied after speed change). Audio is also reversed in pitch-preserve / pitch-shift modes.
Audio mode	enum	`pitch-preserve`	`pitch-preserve` (natural voice, `atempo` chain), `pitch-shift` (chipmunk/giant, `asetrate`), or `drop`.
Frame quality	enum	`fast`	`fast` (frame-duplicate via `setpts`, 2 credits) or `smooth` (motion-compensated `minterpolate`, 5 credits, ~5-20× slower to render).
Ramps	array	`[]`	Piecewise segments — `{ start, end, speed }` tuples in input seconds. Sorted ascending, non-overlapping. Audio is forced to `drop` while ramps are active.
adjustAudio	bool	-	Deprecated. Legacy field — when present and `audioMode` is unset, `true` maps to `pitch-preserve`, `false` to `drop`. New workflows should use `audioMode`.

Inputs & Outputs

Inputs:

in (required) — Video to speed-adjust.

Outputs:

video-out — Speed-adjusted (and optionally reversed / interpolated / ramped) MP4.

How it works

Setting	FFmpeg primitive
Constant speed	`setpts=PTS/speed` (video) + chained `atempo` (audio)
Reverse	`reverse` + `areverse`
Pitch-shift audio	`asetrate=44100*speed,aresample=44100`
Smooth frames	`minterpolate=fps=60:mi_mode=mci:mc_mode=aobmc:me_mode=bidir`
Ramps	Piecewise `setpts` expression — output PTS at input time T is the cumulative output duration of all earlier segments plus `(T - start[k]) / speed[k]` for the segment containing T

The setpts expression for ramps is built in backend/src/providers/video/speed-ramp.ts::buildRampSetptsExpression. Test coverage lives in __tests__/speed-ramp-expr.test.ts.

Credit pricing

Mode	Credits
`fast` (default)	2
`smooth`	5

Composite credit identifier: speed-ramp:smooth when quality === "smooth", otherwise speed-ramp. Both are seeded in STATIC_CREDIT_COSTS and the route’s creditGuard dispatches based on the request body.

The 5 cr smooth-mode price is a flat ~2.5× of fast. Real CPU time is closer to 5-20×, so heavy use should be monitored via the credit-anomaly audit (/admin/credit-anomalies).

Best Practices

Slow-motion: 0.5× with quality: smooth looks the closest to “240 fps shoot” — frame-duplicate at the same speed looks choppy.
Time-lapse / hyperlapse: 4-10× with quality: fast, audioMode: drop.
Reverse: pair with a non-zero speed change for dynamic moments (e.g. speed: 2, reverse: true = fast rewind).
Pitch-preserved audio: keeps dialogue intelligible at moderate speed changes (0.7× – 1.5×). For voiceover narration, prefer pitch-preserve.
Pitch-shift audio: chipmunk-voice effect at high speed, giant-voice at low — useful for comedic/character moments.
Ramps: use 2-3 segments for a classic action-cam beat — e.g. {0,1,1.0}, {1,3,0.25}, {3,5,1.0} slows down the middle two seconds of a 5-second clip.

Limitations

No per-segment audio time-stretch. When ramps is set, audio is dropped. Cinematic speed-ramp shots typically swap in music post-hoc; the worker enforces this in resolveAudioMode.
Smooth-mode cost. minterpolate=mi_mode=mci is CPU-expensive. The flat 5-credit price is a deliberate undercharge to encourage experimentation; revisit when usage data is available.
No GPU acceleration. FFmpeg minterpolate runs on CPU only. For batch slow-motion at scale, an external service (e.g. Topaz, RIFE) would be required — those would be separate node types.
No rubberband audio time-stretch. Higher-quality audio time-stretch via librubberband would require a custom FFmpeg build.

Common Use Cases

Slow-motion highlights from action footage (e.g. 0.5× + smooth).
Time-lapse / hyperlapse compression (e.g. 8× + fast + drop audio).
Reverse shots for psychological / sci-fi sequences.
Speed ramps for music-video transitions (normal → slow-mo at the drop → normal).
Pacing adjustment for AI-generated video clips that feel too fast or too slow.

Tips

Below 0.5×, fast mode shows visible frame holds; smooth is strongly recommended for slow-motion below 0.7×.
Above 10×, the file gets very short — pair with Loop Video if you need it longer.
Chain with Trim Video when you want to speed-adjust only a section.
For audio-only speed changes, this node still works — just wire a video input that has the audio track on it. Audio-only speed changes via a dedicated audio node may come later.