How to Make a 3D Zoom Effect for Video With AI

The four zooms people mean by a 3D zoom effect

When people search for a 3D zoom effect video, they usually mean one of four moves, and each one does a different job. Name the one you want before you prompt for it.

• Digital zoom, or punch-in. The frame jumps or eases closer to your face on a key word. Nothing moves in front of the camera. Wikipedia describes digital zoom as cropping to an area with the same aspect ratio as the original and scaling it back up, with no optical resolution gained.
• Zoom-through cut. A zoom in transition: the outgoing shot rushes forward, usually with a blur, and the next shot opens out of it. It swaps a plain cut for a sense of travel.
• Dolly zoom, or vertigo effect. You hold your size in the frame while the room behind you seems to stretch towards you or drop away. This is the one with a genuine 3D quality, because you and the background change at different rates.
• Pull-back reveal. The reverse of a punch-in. The view widens to show context the viewer did not have, such as the room, a screen or the edit itself.

The “3D” part comes from depth: two layers moving at different speeds. Wikipedia’s article on the Ken Burns effect, the pan-and-zoom technique used on still photos, notes that simulated parallax can make a two-dimensional image appear three-dimensional. Video works the same way once you sit on a separate layer from your background.

How a dolly zoom is shot, and how to fake one in post

In camera, a dolly zoom pairs a camera move with an opposite zoom so the subject stays the same size. Wikipedia describes it as zooming a zoom lens to adjust the angle of view while the camera dollies towards or away from the subject. The most noticeable result is that the background appears to change size relative to the subject. It is also called the Hitchcock shot, Vertigo shot, Jaws effect or zolly shot, and the article credits Irmin Roberts, a Paramount second-unit cameraman, with creating it for Hitchcock’s Vertigo. Filmmakers use it to suggest vertigo, unreality or a character’s sudden realisation.

Doing it for real takes a dolly or slider, a zoom lens and practice. From a static talking-head clip, you fake it with layers:

1. Cut yourself out. A person mask separates you from the room in every frame. A video matting model such as Robust Video Matting uses a recurrent network with memory across frames and needs no extra inputs. How AI removes a video background explains where masks come from and why their edges flicker.
2. Fill the hole. Lifting you out leaves a person-shaped gap in the background layer. It has to be filled from surrounding pixels or from earlier frames, or it shows as the background grows past your outline.
3. Scale the layers apart. The background scales up around your face while you stay almost the same size. A slight extra stretch towards the edges sells the warp.
4. Put yourself back on top. ffmpeg’s overlay filter places one video on top of another; a Python compositing step does the same job frame by frame.

The fake has a ceiling. A real dolly zoom changes perspective because the camera physically moves. In the fake, the room is one flat picture, so a lamp beside you and a wall far behind you grow together.

How punch-ins and zoom transitions are built with ffmpeg

A punch-in is a crop that changes over time, resized back to the full frame. Three ffmpeg filters do this, and each has a catch worth knowing before you ask Claude to use it.

• zoompan applies a zoom and pan from expressions, with zoom between 1 and 10. Its d option sets how many output frames each input image produces and defaults to 90; the output size defaults to hd720 and the frame rate to 25. The documentation’s video example sets d=1, so each input frame gives one output frame.
• crop evaluates its x and y position for every frame, but its width and height are set when the filter is configured, not per frame. A crop alone can pan; it cannot animate a zoom.
• scale evaluates its width and height once by default. With eval=frame it does so for every frame and can read the frame’s timestamp, so a growing scale followed by a fixed crop gives a push-in.

The filter documentation’s own example zooms 2× into the centre of the picture for only the first second of the input video:

zoompan=z='if(between(in_time,0,1),2,1)':d=1:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)'

That is a hard punch-in, on and off, aimed at the centre. Add the s option with your frame size, or the output comes out at hd720. A punch-in that follows your face and eases in needs a face position and an easing value for each frame, which is why the free prompts on this site do the zoom in Python. A detector such as MediaPipe Face Detector finds face locations in each frame, and the script crops around that point. Resampling decides how sharp the enlarged crop looks. ffmpeg’s scaler offers Lanczos among its algorithms, and the free prompts ask for Lanczos to keep enlarged edges crisp.

For a zoom transition effect between two clips, ffmpeg’s xfade filter includes a built-in zoomin among its transitions. Both inputs must have the same resolution, pixel format, frame rate and timebase, at a constant frame rate. A zoom-through anchored on your face, with a radial blur, is again a per-frame job. AI video transitions built this way are ordinary image operations, scale, blur and blend, calculated for each frame and written as code.

When a zoom helps and when it hurts

A zoom helps when it marks a moment the viewer should notice. Every zoom says “this matters”, so its value depends on how rarely it fires.

• Punch-ins help on the word that carries the point: the number, the turn, the instruction. On every sentence they stop meaning anything. They also enlarge flaws, so soft focus and compression become easier to see.
• Zoom-through cuts help where something changes: a new section, a new shot, the move from problem to fix. On every jump cut they turn into noise, and in a calm, reflective video they can feel loud.
• Dolly zooms help on one line that should land hard, such as an unexpected fact or a realisation. Repeated, they lose the surprise. Against a plain wall there is little to warp, so the move reads as an ordinary zoom.
• Pull-back reveals help when the context is the point: the room, the screen, the process. They hurt if your face shrinks while you say something important.

Cinematic video transitions follow the same rule as any emphasis. Watch the whole video once before deciding how many to keep; that call is taste, and no formula replaces it.

Ways to make a zoom effect, compared

Each route trades realism, control and effort differently. Rows are categories, not products.

Make each zoom with the free prompts

Each of the four zooms has a free prompt for Claude Code, and all of them start from the same one-time setup.

• Punch-in. Digital Zoom zooms towards your face on six to eight key words, from 1.00 to 1.18 in 120 ms, eases back to 1.06 and alternates close and medium pushes. It is rated Easy, which makes it the natural first effect for learning the preview-and-notes loop.
• Zoom transition. Zoom-Through Cut is a 10-frame move. The outgoing shot scales from 1.0 to 2.2 with a radial blur, a faint white flash marks the middle, and the next shot opens from 0.6 to 1.0. Your audio is left untouched.
• Dolly zoom. Vertigo scales the background from 1.0 to 1.35 around your face while you go from 1.0 to only 0.97, over about 1.2 seconds, with a bold word in the centre. The prompt also includes the reverse.
• Pull-back reveal. Pull-Back Reveal shrinks your video to 58% inside a cream desktop editing window, with a waveform drawn from your real audio and edit steps ticking off beside it.

Pick one zoom for one moment in one clip, and render it before you add another.