Compose test render (cut 2-6 s)

1 viewer

0:00.0 / 0:04.0

Look + overlayscontent-addressed viewer composition · source MP4 pixels stay unchanged

Timed text

Drag an overlay on the picture to place it; A− and A+ size its text.

No overlays yet

This manifest travels beside the video atom and is replayed by the share page. Downloading the MP4 returns the untouched source pixels; a future render/export connection must explicitly bake the composition.

♪ Audio shape fades, gain, peak normalise · the picture is never re-encoded

Fades are linear ramps. Normalise moves the loudest sample to −1 dBFS (peak, not loudness) and then the gain applies on top. Only the audio track is decoded, shaped and re-encoded — one AAC generation — and every video sample keeps its bytes and its stamp, so captions and overlays stay where they are. A clip with no audio is refused. The source is untouched.

♫ Music bed mixes under the clip's own sound · the picture is never re-encoded

The bed is stored as its own content-addressed atom, then mixed into a new clip: bed and clip each at their level, the bed pulled down wherever the clip's own sound is loud when ducking is on, cut at the clip's end or looped to fill it. WAV always works; MP3 only where the gateway was built with native-mp3-decode. Up to 15 MiB on this gateway. The source clip is untouched.

✂ Timeline drag the handles · I/O set in/out · space play · ← → nudge 0.1s (shift 1s) · , . one frame · J K L back, pause, play

Timeline scale 1×

frames appear as the clip loads

0:00.00:00.50:01.00:01.50:02.00:02.50:03.00:03.50:04.0

0:00.0

0:04.0

0:00.0

in 0:00.0 selection 0:04.0 out 0:04.0 of 0:04.0

The start snaps back to the previous keyframe (H.264 cannot open on a delta frame), so the saved clip can begin a little before the in point — the result reports the window actually cut.