Speed ramps on generated clips

A speed ramp is the cheapest shot you never generated. Take one clip, run the first half at normal speed, ramp into slow motion over a few frames, hold, then snap back. You now have a beat with a rise and a landing where before you had a plain piece of movement, and the credit cost is unchanged because there was no second generation.

That is the pitch, and it holds. The catch is that slowing a clip down does not create frames, it spreads the frames you have across more time, and the point where that stops working is arithmetic rather than taste. Get the arithmetic wrong and the ramp does not look slow, it looks broken.

The frame budget, stated plainly

Versely timelines default to 25 fps. A one-second clip at 25 fps contains 25 real frames. Slow that second to half speed and it now has to fill two seconds of timeline, which needs 50 frames. Twenty-five of them do not exist.

Something has to produce them, and there are only two options: hold each frame for two slots, or synthesise the missing ones. Holding gives you a visible strobe. Synthesising is frame interpolation, and it is inference rather than recovery — anything that happened entirely between two source frames was never captured and cannot be reconstructed, only guessed at.

So the honest way to think about a ramp is as a ratio of invented frames to real ones.

Slow-down	Real frames per output second	Invented per output second	Verdict
0.75×	~19	~6	Safe on almost anything
0.5×	12.5	12.5	Safe on smooth motion, risky on contact
0.33×	~8	~17	Needs clean, moderate motion
0.25×	~6	~19	Only from natively higher-rate footage
Below 0.25×	<6	>19	Regenerate the shot instead

The practical line is at 0.5×. Down to half speed, half of what you are watching was actually computed by the video model, and a competent interpolator has enough neighbouring information to fill the rest. Below about 0.33×, three out of every four frames are guesses, and the clip starts to look like the smooth, slightly waxy motion that gives away an over-retimed shot.

Two smaller rules that matter more than they sound:

Two 2× passes beat one 4× pass. Each pass has closer neighbours to work from, so the guesses are shorter.
Interpolation never adds detail. If the clip is soft, more frames make it soft more often. Sharpness is a separate problem for a separate pass.

Where the extra frames come from

A minority of video models let you pick the render rate at all. Across the 146 video models in Versely's catalog, 48 declare a selectable frame rate: 29 of those offer 24 and 30 only, 18 add 60, and exactly one offers 24 and 48. The rest render at whatever their pipeline settled on.

That matters for ramps because it changes the arithmetic completely. If the model exposes 60 and the shot has real fast motion in it, generate at 60 and your 0.25× ramp becomes a straight retime rather than a synthesis job — those frames were decided, not inferred. If the model does not expose the choice, you are interpolating, and the table above is your budget.

There is a caveat worth knowing before you trust a number in a file header: some generators produce a modest count of real frames and interpolate up to the rate they advertise. A clip labelled 60 fps has not necessarily had sixty moments per second decided for it. Slowing footage that was already interpolated means running a second pass over guesses, which is where the mush comes from. The full comparison of interpolation against native frame rate covers how to tell.

Where interpolation smears

Interpolation does not degrade evenly across the frame. It fails at three specific things, and all three are places where a motion vector stops behaving like a straight line.

Contact points. A foot landing, a hand closing on a mug, a bat meeting a ball. At contact the motion reverses or stops dead, but an interpolator extending the motion it can see puts the foot slightly through the floor, and the next real frame snaps it back. The result is a half-frame ghost at exactly the moment the viewer is looking hardest. Generated footage already struggles with weight and contact — physics failures around gravity and contact is the standing version of the problem — so a ramp over a contact point compounds an existing error rather than smoothing one.

Occlusion. Anything passing behind anything else. The interpolator has no information about the hidden portion and invents it, which is where detached limbs come from.

Objects entering frame. A thing present in frame two but not frame one has no motion history, so the invented in-between pops it in whole or ghosts it.

The design consequence is simple: put the ramp where the motion is boring. Ramp into slow motion across an empty stretch of movement, hold the slow section through the part you want the viewer to dwell on, and if there is a contact moment in the shot, either keep it at full speed or accept that it will be the weak frame.

Ramp shapes that work

A ramp is two decisions: the ratio, covered above, and the length of the transition between the two speeds.

Short ramp, 4 to 6 frames. 160 to 240 ms at 25 fps. Reads as a snap into slow motion. Good for impact beats and product reveals.
Medium ramp, 10 to 15 frames. Half a second or so. Reads as deliberate and cinematic. The default for most brand work.
Long ramp, 25 frames or more. A full second of easing. Only works when the shot is long enough to give the slow section somewhere to live afterwards.

Ramping up is nearly free, and it is underused. Speeding a clip from 1× to 2× discards frames rather than inventing them, so there is no interpolation cost and no smear risk. Speeding up a dull middle section is often a better fix than cutting it, because it keeps the shot continuous. The one thing that breaks is audio: any speed change on a clip with speech in it has to decide what happens to pitch, and if picture and audio are retimed separately, sync goes with it.

If what you actually want is a fast move rather than a fast clip, that is a generation-side problem, not an edit-side one. Crash zooms and whip pans are better prompted than retimed.

Running it

Decide the delivery frame rate before you generate, and check whether the model you want exposes the rate at all. Most do not.
Generate the shot long. A ramp eats source. If you want a two-second slow section at 0.5×, you need one second of source for it plus whatever the surrounding full-speed footage needs. Pick a duration with headroom rather than trimming to length first.
Place the ramp away from contacts and occlusions. Step through the clip and note where they are before choosing the ramp position.
Describe the speed change rather than keyframing it. The editor handles speed as part of a full edit — changing playback speed walks through the instruction form. State the range, the factor, and what should happen to the audio.
Preview at 480p first. preview: true renders a 480p pass at no credit cost with a short per-user cooldown, and the final export is charged once no matter how many clips are on the timeline. Retiming artifacts are temporal, not spatial, so 480p shows them perfectly.
Watch the contact frames specifically. Not the clip. Step the boundary. That is where a bad ramp hides.

The whole assembly lives in the AI video editor, and the billing shape for iterating this way is in the preview and export cost breakdown.

FAQ

What is the slowest I can go without it looking wrong?

On moderate, smooth motion with no contact points, 0.33× holds up. On anything with impacts, hands, or objects crossing in front of each other, 0.5× is the sensible floor. Below 0.25× from 25 fps source you are watching four invented frames for every real one, and no interpolator makes that look like footage. If the shot genuinely needs quarter speed, generate it at a higher rate if the model offers one, or generate the slow version as its own clip.

Does a speed ramp cost extra credits?

The retime itself is part of the edit rather than a new generation, so you are not paying per ramp. The editor charges a single price for the final export regardless of how many clips and effects are on the timeline, and the 480p preview pass carries no credit cost with a short per-user cooldown between passes. What does cost more is generating the shot long enough to ramp inside, since most video models price by duration.

Should I ramp in the edit or generate slow motion directly?

Ramp in the edit when the slow section is short, the motion is smooth, and you want the speed change itself to be part of the beat. Generate directly when the whole shot is slow motion, because a model rendering slow movement natively is computing those moments rather than inferring them. The ramp is an editing device; native slow motion is a shot.

Why does my ramp look fine at full speed but bad when I step through it?

Because that is how interpolation fails. Individual invented frames near a contact point or an occlusion are visibly wrong, but at 25 frames a second the eye integrates them into an impression of motion rather than resolving them. If the clip reads fine at speed, it reads fine — the frame-by-frame check is for finding out where to move the ramp if it does not, not a standard the footage has to pass.