Cross-dissolves that hide drift at the seam

There is a persistent idea in editing circles that reaching for a dissolve means you failed to find the cut. On live-action footage shot in one session with one camera, that is often fair. On a sequence built from four separate generations, it is bad advice, because the thing a dissolve does best is precisely the thing that goes wrong between generated clips: it blurs a boundary where two shots disagree about exposure, colour or the exact shape of a face.

The standard fix for temporal drift across a chained sequence is to break the piece into short segments and dissolve the joins. That is not laziness. It is the same reasoning that puts a fade on a tape splice.

What actually drifts, and why

Generated video accumulates error. Each frame is conditioned on what came before, so small deviations compound rather than cancel, and by the time a model is four or five clips into a chain the drift is visible without looking for it. Three things move most:

Exposure and white balance. The second clip is a quarter-stop brighter or a touch warmer than the first. Individually fine, side by side obvious.
Identity. A face reconstructed from a reference is close on clip one and noticeably a different person by clip five. This is the failure mode that ends most character-led sequences, and the underlying cause is error compounding across an autoregressive chain rather than any single bad generation.
Set dressing. A jacket loses a button. A mug moves along a table. The wall behind changes texture.

The first of those has a real fix. The other two have a mitigation. Knowing which is which decides how you spend your time — the full diagnostic is in diagnosing drift between shots, and the mechanism sits under temporal consistency in the glossary.

Colour-match first, then dissolve

Order matters here, and getting it backwards is the most common way a dissolve makes things worse.

Colour-match every clip at assembly before you decide anything about transitions. Matching exposure and white balance across the sequence is the direct counter to the drift a dissolve is otherwise being asked to hide, and it is non-optional on any multi-clip piece. If you dissolve first and match after, the blend region contains a mixture of two grades, and grading a mixture is a worse job than grading two clean clips.

Once the clips agree on exposure, you often find the dissolve you were planning is unnecessary — a lot of what reads as "these shots don't belong together" is a brightness difference, not a content difference. What a grade can and cannot rescue is worth reading separately, because there are limits and knowing them stops you burning an hour on something a reroll would have fixed.

What survives the colour match is the drift that actually needs hiding: identity shift, geometry changes, set dressing that moved. That is what the dissolve is for.

Dissolve lengths

Versely timelines default to 25 fps, so a frame is 40 milliseconds. Lengths below are in frames, because "half a second" means different numbers of frames on different timelines and frames are what the join is actually made of.

Length	At 25 fps	What it does	Use for
3–5 frames	120–200 ms	Barely perceptible. Softens the boundary without reading as a transition	Exposure or texture mismatch on an otherwise matched pair
6–10 frames	240–400 ms	Reads as a deliberate soft join	Identity or geometry drift between two shots of the same subject
12–18 frames	480–720 ms	Clearly a dissolve. Carries a sense of time passing	Scene changes, or covering a large mismatch
25+ frames	1s+	A statement. Slows the whole piece down	Openings, endings, montage sections

Two things constrain the top of that table. The first is shot length. Short-form pacing runs somewhere around 1.5 to 3 seconds a shot on TikTok, a little longer on Reels, longer again for tutorial content on Shorts. A 12-frame dissolve inside a 2-second shot is spending a quarter of that shot on a blend, which is why long dissolves and fast cut rates do not coexist. If your average shot is under two seconds, cap dissolves at about 6 frames.

The second is that a dissolve averages two images. During the blend, both shots are semi-transparent, so a long dissolve over two clips that disagree produces a stretch of footage where a face is genuinely two faces. Short dissolves work on drift because the eye does not have time to resolve the mixture. Long ones give it that time.

The practical default: 5 frames for a matched pair, 8 frames when the identity has shifted, and stop there unless the piece is deliberately slow.

When a hard cut is still the better call

A dissolve is not always the cheaper decision. Cut instead when:

The action is continuous across the join. If a hand is mid-reach in both clips, a dissolve turns one hand into two overlapping hands. Cut on the motion instead.
The cut is carrying rhythm. Anything cut to music. A dissolve has no attack, so it lands soft and the beat goes missing.
The subject is centred and identical in both shots. This is the worst case for a dissolve, because the two versions of the same face sit directly on top of each other and the morph is unmissable. Either cut, or change the framing so the two shots don't stack.
The drift is large. A dissolve hides a small disagreement. A big one just becomes a visible morph, and at that point you are better off regenerating the clip, re-anchoring to a locked reference image, or shortening the chain. The full set of options for continuing a shot is laid out in extend, chain, or first/last frame continuations.

There is also the case where neither works, and the answer is a cutaway: drop two seconds of B-roll between the mismatched shots and the viewer's memory of the first one degrades enough that the second stops disagreeing with it. This is an old trick and it is still the strongest one available, because it fixes the problem rather than blending it.

Building the sequence

The pass order that produces the fewest surprises:

Cap clip length at the source. Shorter generations drift less. Three to five seconds a clip is the range where most video models hold together, and it happens to match short-form cut rates, so the format hides the constraint.
Assemble on one timeline, not as a chain of exports. The editor is EDL-based, so the timeline is a re-renderable object rather than a one-shot render. Re-encoding footage once per join costs quality you cannot get back.
Colour-match every clip. Before transitions, before captions, before anything.
Set transitions and their durations explicitly. The editor takes both a transition and a transition duration as parameters, so the length is a number you choose rather than a default you inherit. A broader tour of what is available is in the transition effects guide.
Preview at 480p before exporting. preview: true renders a 480p pass at no credit cost with a short per-user cooldown, and the final export is charged once regardless of clip count. Both dissolve length and drift survive the downscale — a morphing face is a shape problem, and shape is exactly what a low-resolution pass preserves.

If all you need is several generations stitched with consistent joins, the merge path in the AI video editor covers it, and the preview and export cost breakdown has the billing shape. For the simple case of joining finished clips end to end, merging videos is the one-step version.

FAQ

Does a dissolve look like I couldn't get the shots to match?

Only if it is too long. Under about 6 frames at 25 fps, a dissolve reads as a soft join rather than as a transition, and most viewers will not register that a blend happened at all. The dissolves that announce themselves as apologies are the half-second-plus ones used on fast-cut material, where the pace makes them conspicuous. Length is the tell, not the technique.

Should I dissolve every join in a chained sequence?

No. Uniform dissolves flatten the piece and make everything feel like a montage. Dissolve the joins where the drift is actually visible after colour-matching, and hard-cut the rest. A sequence with two dissolves and six cuts reads as edited; a sequence with eight dissolves reads as a screensaver.

Will a dissolve fix a face that changed between clips?

It reduces how much you notice, it does not fix it. A short blend gives the eye less time to compare the two versions, which works when the change is small. When the identity has clearly shifted, the dissolve just makes the change visible as a morph instead of as a cut. At that point the real fixes are shorter clips, re-anchoring each generation to the same reference image, and not letting a chain run five links deep.

Where does the dissolve go relative to captions and music?

After the picture is assembled and colour-matched, before audio and captions are locked. Dissolves change where a shot's visual weight lands by a few frames, and captions timed to the edit beat need to follow the final boundary positions rather than the provisional ones. Music is last, because you want to place beats against a picture cut that is no longer moving.