Skip to the content
Agent Tools

fframes vs Remotion: fframes won all 26 timed renders, Remotion won the edit loop

On one simple 1080p scene, fframes was faster than Remotion in all 26 paired renders, by 1.8 to 8.1 times on the median across three sessions. After a one-line edit, Remotion had a new 6 second clip ready in a median 9.1 seconds against 16.5 for fframes, which must recompile first. Pick by what you do more often: render a fixed scene many times, or change the scene between renders.

By OpenKrill. Published 2026-10-10. Every time here comes from our own runs on 2026-10-10 (the intro run on 2026-10-09) on an Apple M5 Pro with 18 cores and 48 GB of memory. fframes 1.2.0 drawing on the GPU through Apple's Metal; Remotion 4.0.531 with default settings except the quality setting below.

Why we ran this

Dmitriy Kovalenko, who makes fframes, quoted a user whose heavy 81 second video went from 700 seconds in Remotion to under 21 in fframes. We pick a renderer for short product videos, so we wanted our own number on a plain clip built in both tools. We did not rebuild that heavy video, so this post does not test the 700 to 21 claim.

How the two tools draw a frame

Remotion renders a React page in headless Chrome (a browser with no window), screenshots each frame and hands the frames to FFmpeg. By default it ran 8 browser tabs in parallel on our Mac. fframes draws each frame as an SVG straight to the GPU with Skia, the graphics library inside Chrome, with no browser and no screenshot step.

The clip

We first made a small probe scene with fframes, then ported it element by element to Remotion: a 1920 by 1080 card at 30 frames a second with a drifting glow (an SVG gradient, no blur), a headline that springs up and fades in, a growing underline, a subtitle, a frame counter and the DM Sans font. Both encode with x264 at CRF 20, its quality dial, and the medium preset. fframes also sets x264's animation option, which Remotion does not.

To check that both tools drew the same thing, we compared the two 6 second files frame by frame with FFmpeg's SSIM score, where 1 means identical. They scored 0.997.

Render times: three sessions

Each number is the elapsed time of the whole render command, from start to a finished MP4. Session 1 ran each tool 3 times, one tool after the other, through cargo run (Rust's build-and-run command, which adds 2 to 4 seconds) and npx remotion render. Its machine load was not recorded. Sessions 2 and 3 ran the built fframes program directly and alternated the tools, 5 runs each. The Mac was shared with other jobs: load average 65 to 78 in session 2 and 17 to 41 in session 3, on 18 cores, where 18 means every core busy. We had no idle machine that day, so every number below comes from a busy Mac.

Clip and sessionfframes, fastest / median / slowestRemotion, fastest / median / slowestMedian gap
6 s, session 12.3 / 2.3 / 3.4 s9.8 / 13.0 / 14.7 s5.6 times
6 s, session 21.0 / 1.3 / 5.9 s8.1 / 10.2 / 21.1 s7.6 times
6 s, session 30.85 / 0.89 / 0.96 s6.9 / 7.3 / 8.2 s8.1 times
60 s, session 112.3 / 21.5 / 22.1 s37.3 / 38.1 / 40.2 s1.8 times
60 s, session 26.6 / 9.0 / 23.9 s37.1 / 50.0 / 60.1 s5.6 times
60 s, session 35.8 / 8.8 / 11.3 s30.9 / 36.0 / 40.2 s4.1 times
  • No fframes run was slower than any Remotion run of the same clip in the same session. That holds in all 6 rows, which is a firmer claim than any one ratio.
  • Noise was large. The same command varied by up to 5.6 times between runs in the busy session 2. Even session 3, the quietest, ran at load 17 to 41, and the 60 second gap moved from 1.8 to 5.6 times between sessions. Read the ratios as a range, not one number.
  • We tried more Remotion tabs once. One session 1 run of the 60 second clip with every core given to Remotion (--concurrency 100%) took 47.2 seconds, against 37 to 40 at the default. One run is not enough to conclude anything.

A straight line through session 3's medians at 6 and 60 seconds gives Remotion about 4 seconds of start-up plus 18 ms a frame, and fframes almost no start-up plus 5 ms a frame. So the gap has two parts: Remotion's browser start-up, which dominates a short clip, and a cheaper frame in fframes. Session 1's noisier medians give a smaller per-frame gap (15 against 12 ms), so treat the split as an estimate.

The edit-to-video loop

Render speed is only half of daily use. In a fourth session we made 5 one-line text edits in each tool, alternating, and timed each edit through to a finished 6 second MP4 (load 22 to 38).

Edit then render, 6 s clipFastest / median / slowest
fframes: release rebuild with cargo, then render14.1 / 16.5 / 22.8 s
Remotion: render, no compile step8.1 / 9.1 / 10.8 s

Remotion was quicker in all 5 pairs. The render itself takes fframes under a second, and nearly all of its time is the Rust compile. The first fframes release build, before any of this, took 2 minutes 8 seconds.

What we read from it

If you render many videos from a fixed scene, such as one template with new text per customer, fframes rendered this card 1.8 to 8.1 times faster on the median and was never slower in our runs. If you change the scene between short renders, Remotion gets you to the next video sooner, and React may be a language your team already knows.

For a heavy scene we have one fframes-only number. Its own 128 second intro example took 124 seconds on this Mac on 2026-10-09, about 16 ms a frame against 5 for our card. The maker reports 36 seconds for that video. 124 is 3.4 times that, and we do not know which hardware difference explains it. We did not run the intro in Remotion.

What we did not test

  • A heavy scene in Remotion, or any scene other than this one card.
  • A fully idle machine, any machine other than this Mac, Linux or a CI server.
  • A scene written for Remotion first: ours was written for fframes and ported, which may favour fframes.
  • Remotion tuning beyond one concurrency run, and fframes on its CPU backend.
  • An fframes debug build, which compiles faster, or Remotion Studio's live preview. Either could change the edit-loop result.

How to repeat it

Our scene code for both tools is linked under Sources. Time each command 5 times with /usr/bin/time -p, alternating the tools:

cargo build --release && ./probe render
npx remotion render src/index.ts Probe out/remotion.mp4 --crf 20 --x264-preset medium --overwrite
ffmpeg -i fframes.mp4 -i remotion.mp4 -lavfi ssim -f null -

If the SSIM score is well under 0.99, the scenes differ and the times do not compare. Log uptime before each run. On a busy machine, alternate the tools so both see the same load.

Our own tool

If you only need a still share card rather than a video, our Live Card renders one from HTML in headless Chrome, the same capture model as Remotion. We did not benchmark it here.

Slides

The same numbers as 10 slides: fframes vs Remotion in 10 slides.

Sources