Leave Handles on Generated Clips
Models return exact-length clips. Pad 8-12 frames on each end so dissolves, J-cuts, and limiter look-ahead do not hit hard cuts.
You asked the model for five seconds. You got five seconds. There is no pre-roll, no camera run-out, and nothing to dissolve onto. A 12-frame dissolve needs 12 frames of overlapping media. A J-cut needs audio before the picture cut. A limiter with a few milliseconds of look-ahead needs samples before the transient. Exact-length clips turn all three into pops and hard cuts.
Pad 8–12 frames on each end of the file. Cut the timeline shorter. The extra media is not "more video." It is the only run-out a generated clip will ever have.
Exact-length clips have no run-out
Live-action handles exist because the camera was rolling before action and after cut. Even a sloppy take has frames you can use for a dissolve, a split edit, or a fade.
Generated clips are a duration you requested. Many models will not give you a frame more. Some will give you a frame less. Either way, the last frame of the file is the last frame of the action, and the first frame is the first. There is no pre-roll to hide a limiter attack. There is no tail to hold a line of dialogue over the next shot.
That is why a stitched sequence of generated clips sounds like a slideshow even when the picture cuts are good. Every edit is a hard cut on picture and on native audio. J-cuts and L-cuts are how you break that weld, and they are impossible if the file ends on the cut. Cross-dissolves have the same requirement on the picture side: the blend needs media from both shots in the overlap.
Eight to twelve frames, and why that range
Versely timelines default to 25 fps. At 25 fps:
| Pad per end | Time | What it buys |
|---|---|---|
| 8 frames | 0.32 s | A short dissolve, a brief J-cut, a fade that is not a pop |
| 12 frames | 0.48 s | A more visible dissolve, a dialogue lead of about half a second |
| 25 frames (1 s) | 1.00 s | Mix handles. This is what an AAF should carry, on top of the 8–12. |
At 24 fps, 8 frames is 0.333 s and 12 frames is 0.500 s. Same idea, different arithmetic. Confirm the frame rate of the file before you count.
Why not 2 frames? Because a 2-frame dissolve reads as a flash, and a 2-frame J-cut reads as a mistake. Why not 2 seconds on the generated file itself? Because you are paying for duration you will never see, and because many models degrade when you ask them for longer than the shot needs. 8–12 frames is the editorial pad. Mix handles (a second or two) are added at ingest transcode, when you rewrite the file into a finishing codec.
Limiter look-ahead is smaller than one frame. Typical look-ahead is a few milliseconds. One frame at 25 fps is 40 ms. If the only problem you had was the limiter, a single extra frame of media would cover it. You still pad 8–12 because the dissolve and the split edit are the things that actually fail, and the limiter is a free passenger on that pad.
Head and tail. A J-cut uses the incoming clip's head. An L-cut uses the outgoing clip's tail. A dissolve uses both. Padding only the tail is how your incoming audio still starts on a hard cut.
Pad the file, cut the timeline
The handle has to live in the media file. A timeline that is longer than the file is not a handle. It is a freeze, a black, or an offline.
Three ways to get the pad, in order of preference:
- Request a longer generation than the cut. If the shot should play for five seconds, ask for a duration that covers five seconds plus 8–12 frames on each end (at 25 fps, 5.00 s + 0.64 s ≈ 5.64 s, so request the next length the model will actually give you). Set in and out on the timeline. Do not file-trim the extra frames away.
- Extend a take you already have. Extend a video generates new footage that continues from the last frame. It is not a freeze, and it is not a head handle: it only adds a tail. Check the join. If the extend invents a new action, you do not have a handle. You have a new shot.
- Ingest transcode with handles. When the take is locked and you are rewriting to ProRes or DNxHR for mix, the transcode can include extra frames you already have. It cannot invent frames you deleted.
What not to do:
- Do not freeze-frame the first or last frame and call it a handle. A freeze is obvious on a moving shot and useless for audio. Native audio does not extend with a freeze. You get a held picture and a hard out on the track.
- Do not pad with black. Black is not overlap. A dissolve to black is a fade, and a limiter looking at silence then a transient will still pump.
- Do not trim the file to the cut "to keep things tidy." That is how the handles die. Tidiness is a timeline in/out. The file stays long.
- Do not generate extra length only at the tail. Models often put their worst frames at the start and the end. A little extra on both sides also gives you trim room when the first frame is a mess.
If you are assembling only to check pace, hard cuts are enough. Build handles before that assembly becomes the mix source.
Mix and limiter look-ahead
Once the files have 8–12 frames, the mixer still wants more. That is not a contradiction.
Editorial pad is there so you can cut dissolves and split edits without running out of media. Mix handles are there so they can fade, de-click, and let a limiter see a transient before it hits the output. EBU R 128 sets programme loudness at −23.0 LUFS and a true-peak ceiling of −1 dBTP. Hitting that ceiling without a pop at every picture cut requires the processor to have samples before the cut, not only after it.
If you bounce a mix against exact-length clips:
- Every cut is a discontinuity in the audio. Even a well-matched bed will click if the phase and the room tone do not continue.
- Look-ahead limiting on the master bus will duck after a transient that sits on frame 0 of a clip, which is how a line of dialogue starts with a bite out of the first syllable.
- You cannot build a J-cut in the mix that you did not leave media for in the picture. The mixer can slide audio they have. They cannot slide audio that was never in the file.
So the working stack is:
- Generate long enough for an 8–12 frame editorial pad.
- Cut picture using those frames for dissolves and splits.
- At ingest for the AAF, copy audio with a second or two of handle from the same files.
- Mix against that. Do not print a limiter onto a handle-less bounce and hope the platform's normalizer hides the pops. It will not. Loudness normalisation pulls hot mixes down. It does not repair a click.
Music and VO you add in the editor should be files that already have run-up. A spanning bed across the join is the cheap way to hide remaining hard cuts: the bed continues even if a clip's own audio does not. That spanning track still needs a handle at the start of the sequence, or the first limiter attack is the first note.
FAQ
Can I add handles by slowing the last frames?
No. Retiming the tail to stretch eight frames out of two gives you a slow-motion hitch and pitch-shifted audio, or a freeze by another name. Generate or extend. Do not retime a handle.
What if the model will not render the extra length?
Render the length it will give you, then extend, or generate a second clip that continues the action and keep both files. Two files with a join in the handle region is uglier than one long file, and better than no handle. If the shot is a hard lock-off with no motion, you still need audio run-out; picture freeze is only acceptable when nothing in the frame is moving and the sound is replaced.
Do 8–12 frames apply to stills I animated?
Yes, if they will be dissolved or split. A still that is exactly the cut length has the same hard edge. Hold extra frames of the still in the file. That is the one case where a freeze is the handle, because the source never moved.
Should the client see the handles?
Not in the review movie. The reference they approve is the cut, in/out on the timeline. The files you keep and the AAF you send still contain the pad. A 480p preview pass (preview: true, short per-user cooldown) is the right way to check that a dissolve or a J-cut is using real media rather than a freeze, before you pay for the export.