AI Video Editing Flows

Camera moves on talking footage: micro-zoom, slow push, and when to hold still

Three small camera habits for footage of people talking: tighten the frame after a jump cut, push in very slowly during a long shot, and never magnify a detailed object. We ran the FFmpeg commands and measured how smooth three ways of zooming really are.

The habits

  • Jump cuts inside one speaker's answer: show the next piece slightly tighter, then alternate.
  • Long single shots: a very slow push, so slow you feel it and do not see it.
  • Detailed objects (a table, a ruled grid, small interface text): never magnify them with the camera. Build them larger and film them at scale 1.0.

The numbers (6% tighter, 28 px, 0.0055 zoom a second) are the ones we use. They are our taste, not a standard.

Micro-zoom on a jump cut

When a cut removes a filler word or a pause but the framing stays the same, the picture appears to jump. We show the next piece about 6% tighter, shifted about 28 px sideways on a 1920-pixel-wide frame, and swap the side on each further cut so the move never drifts. The command below cuts an eight-second clip into four two-second pieces and applies the move to pieces two and four. Our test clip is 1280 pixels wide, so we scaled the shift to the clip's width in the picture and kept the 28 px figure for 1920-wide frames.

Four pieces, the second and fourth tighter
ffmpeg -y -i clip.mp4 -filter_complex "[0:v]trim=0:2,setpts=PTS-STARTPTS[a];[0:v]trim=2:4,setpts=PTS-STARTPTS,crop=w=iw/1.06:h=ih/1.06:x=(iw-iw/1.06)/2+28:y=(ih-ih/1.06)/2,scale=1280:534,setsar=1[b];[0:v]trim=4:6,setpts=PTS-STARTPTS[c];[0:v]trim=6:8,setpts=PTS-STARTPTS,crop=w=iw/1.06:h=ih/1.06:x=(iw-iw/1.06)/2-28:y=(ih-ih/1.06)/2,scale=1280:534,setsar=1[d];[a][b][c][d]concat=n=4:v=1:a=0[v]" -map "[v]" -c:v libx264 -crf 18 microzoom.mp4

It ran without an error and wrote 192 frames at 1280x534, which is eight seconds at 24 frames a second. We looked at a frame from piece one and one from piece two: the two people sit larger in the second, and the whole picture still fills the frame. setsar=1 after the scale matters: without it the concat step refused to join the pieces.

A slow push that stays smooth

We push in by 0.0055 of the picture size for each second of the shot, capped at 3% in total, and not at all on shots under six seconds. Long shots alternate push in and pull out. The cap means a twelve-second shot ends 3% larger than it began.

How you zoom matters as much as how much. We rendered the same six-second 3% push of a still three ways and tracked three soft dots through every frame to see how smoothly each moved. A perfectly smooth push would move the dot about 0.09 px each frame.

MethodWobble: spread around a smooth path (px, RMS)Largest single-frame step (px)
zoompan filter0.270.92
perspective filter, evaluated every frame0.0110.12
scale with a changing size, then crop0.3171.66

The perspective filter with cubic interpolation gave the smoothest move by a wide margin. The command is independent of the picture size because the filter exposes the frame width and height as W and H, and in is the input frame number. Replace 144 with the number of frames in your shot (6 s at 24 fps) and 0.03 with your zoom.

Slow push with the perspective filter (as tested on a grey still)
ffmpeg -y -loop 1 -framerate 24 -t 6 -i dots.pgm -vf "format=gray,perspective=x0=W/2-W/(2*(1+0.03*in/144)):y0=H/2-H/(2*(1+0.03*in/144)):x1=W/2+W/(2*(1+0.03*in/144)):y1=H/2-H/(2*(1+0.03*in/144)):x2=W/2-W/(2*(1+0.03*in/144)):y2=H/2+H/(2*(1+0.03*in/144)):x3=W/2+W/(2*(1+0.03*in/144)):y3=H/2+H/(2*(1+0.03*in/144)):interpolation=cubic:eval=frame,format=yuv444p" -c:v libx264 -crf 10 -pix_fmt yuv444p push_perspective.mp4

For colour footage drop the two format filters at the start and end of the chain.

Do not scale up a detailed object

Never magnify a detailed object with camera scale. A hairline menu, a ruled grid or small interface text shimmers or shakes when it is resampled at fractional scales. Build the object larger in the first place and shoot it at scale 1.0. This is a rule from our own work. We did not measure it on this page, so treat it as advice and test it on your object.

What failed or surprised us

  • The scale-then-crop method zoomed far more than we asked. A dot that should have travelled about 12.2 px travelled 32.8 px. We did not find out why, so we do not recommend that recipe and we show its numbers only as a warning. Its wobble and step figures describe a zoom of the wrong size.
  • zoompan is the filter most guides name for a push, and it was measurably less smooth than the perspective filter: its largest step was 0.92 px where a smooth push asks for 0.09. Whether you can see that depends on the screen and the footage. We would use perspective for slow moves on large screens.
  • The test uses three soft dots on a still, so it measures the filter and not what a real shot does.

Method and limits

Every command on this page was run with FFmpeg 8.1.2-full_build-www.gyan.dev on a Windows laptop, from one folder, with the file names shown. The clip is eight seconds of Tears of Steel, (CC) Blender Foundation, mango.blender.org, CC-BY 3.0; other inputs were generated by FFmpeg or by a short script. The measurements are on our test clips, not a survey of footage. Where a rule is our habit and not something we measured, the page says so. Our method is on {link:VID-trust-method}.

  • Wobble is the root-mean-square distance of each dot's tracked position from a smooth curve fitted through its path, over 144 frames.
  • Largest step is the biggest change in position of any dot between two frames.

Page checked on Nov 4, 2026.