CRF 18 vs 23 vs 28 for a Vertical AI Short: File Size and VMAF on One 10-Second Clip

Disclosure: this post has no affiliate links. The shop page linked at the end is our own product list and may contain affiliate links; we may earn a commission if you buy through those. The test clip is an AI-generated video we made ourselves, and this post was drafted with AI assistance, then checked line by line against our measurement file from October 29, 2026 (KST). No tool vendor provided or reviewed it.

Short answer from one test: on our 10-second vertical AI clip, each 5-step rise in CRF cut the file by 41-51% (about half from CRF 23 upward), and CRF 23 on the x264 "veryfast" preset was not the same quality as CRF 23 on "medium". Veryfast made a file 30% smaller, but its VMAF score (91.93) landed next to medium at CRF 28 (92.18), not next to medium at CRF 23 (95.97). If you change the preset to save time, do not assume the CRF number still means the same quality.

The question

You export a 1080×1920 short and have to pick a CRF value. Lower CRF means a bigger file and fewer compression losses. How big is the jump between common values, how much quality does the score say you lose, and does the encoder preset change the answer?

What we ran

  • Source: one finished 10.03-second AI-generated clip from our own short (local image-to-video plus one generated paper insert, no paid API calls). 1080×1920, 30 fps, 301 frames, H.264 at about 6,456 kbps, 8.35 MB.
  • Encodes: ffmpeg 9.0.1 with libx264, video only (audio removed), yuv420p. Preset medium at CRF 18, 23, 28 and 33, plus CRF 23 on the veryfast and slow presets. One encode per setting.
  • Score: VMAF with ffmpeg's libvmaf filter and its default model (vmaf_v0.6.1), every frame compared with the source file, plus PSNR on the luma channel. "Worst frame" below is the lowest single-frame VMAF.
  • Machine: one Apple M5 Pro laptop, 8 encoder threads. Times are wall-clock seconds for the encode only.

Results

Preset / CRFFile sizeBitrateVMAF meanWorst frameEncode time
medium / 186.83 MB5,444 kbps97.5295.853.2 s
medium / 234.00 MB3,189 kbps95.9793.732.8 s
medium / 282.00 MB1,593 kbps92.1889.352.2 s
medium / 330.99 MB787 kbps85.5481.321.9 s
veryfast / 232.81 MB2,240 kbps91.9388.331.0 s
slow / 233.82 MB3,049 kbps95.8693.684.6 s

What the numbers say

  • Size dropped 41-51% every 5 CRF steps here. 6.83 → 4.00 → 2.00 → 0.99 MB (18→23 cut 41%; each later step about half). The score drop grows as you go: 1.6 points from 18 to 23, 3.8 from 23 to 28, 6.6 from 28 to 33.
  • Preset changes what CRF means. Veryfast at CRF 23 was smaller and faster, but it scored like medium at CRF 28, which was smaller still (2.00 MB vs 2.81 MB). On this clip, medium at a higher CRF gave a similar score in a smaller file than veryfast at CRF 23.
  • Slow bought little here. Slow at CRF 23 was 4% smaller than medium with almost the same score (95.86 vs 95.97) and took about 1.7 times as long.
  • What we saw by eye. We pulled the frame at 5 seconds from the source and the CRF 33 file and cropped the printed bill on the table. The digits were still readable at CRF 33; the marble texture around it looked flatter. That is one still frame, not a viewing test.

What we decided

For our own uploads we keep medium at CRF 18 to 23, and we will not switch to veryfast at the same CRF to save time. If a smaller file matters more than the last few points, raising CRF on medium looked like the better trade on this clip than changing the preset.

Limits

  • One 10-second clip with mostly calm motion. Fast motion, grain or text overlays can behave differently.
  • The reference is our already-compressed export, so VMAF measures loss against that file, not against the original render.
  • The default VMAF model was built for viewing 1080p video at a set distance on a larger screen, not for a phone held in hand. We did not run a viewing test with people.
  • Platforms re-encode what you upload. We did not measure what YouTube, Instagram or TikTok delivered after upload.
  • One encode per setting on one laptop; times will differ on other machines.

Try it yourself

Encode: ffmpeg -i source.mp4 -an -c:v libx264 -preset medium -crf 23 -pix_fmt yuv420p out.mp4. Score: ffmpeg -i out.mp4 -i source.mp4 -lavfi libvmaf -f null - (encoded file first, reference second).

Sources checked

Related: our earlier test on fitting an AI voiceover to an exact length.

Our product list: Label Picks shop page (may contain affiliate links).

Comments

Post a Comment

Popular posts from this blog

FFmpeg Shows −1.5 dBTP. Did You Actually Fix the Audio?

An output_tp of −1.50 in an FFmpeg log does not mean your original audio has been fixed. A command can show that number without saving any corrected audio. Before uploading, measure the file you actually exported. This follow-up to our AI song quality checklist tackles one practical mistake: confusing filter output statistics with a saved deliverable. You need FFmpeg installed and a local audio file. These commands process files on your computer; they do not upload your recording. One recording, two different stages On September 18, 2026, we measured one newly generated, roughly 2-minute-20-second MP3. Its integrated loudness was −11.41 LUFS and its true peak was +0.84 dBTP . The same analysis log showed output_tp: -1.50 . input_i and input_tp describe the file supplied to that command. output_i and output_tp describe the signal after the filter in that execution. They do not establish the measured properties of a separately encoded final file. We then reduced gain by...
Read more

VoxCPM2 vs ElevenLabs for a 30-Second Science Short: We Timed Both on the Same Script

Image
Disclosure: DigitalHarvest (NSST) is an independent affiliate of ElevenLabs. We may earn a commission if you subscribe through the referral link in this article. ElevenLabs has not sponsored, reviewed, or approved this article. This relationship does not change the numbers below, and the free local option wins on cost in our own test. We make 30-second science Shorts from public-domain NOAA ocean footage, and every one needs about six lines of narration. On October 4, 2026 we took one finished script and voiced the same six lines twice: once with VoxCPM2, an open-source model running on our own Apple Silicon Mac, and once with ElevenLabs on the paid Creator plan we already use. Then we ran both outputs through the same Whisper check our pipeline uses before anything is published. Here is what came out. The script we used The episode is "Why doesn't hot water on the seafloor boil?" (a small black smoker vent filmed at 2,947 m by NOAA Ocean Exploration). Six lines, 428...
Read more

Before You Upload an AI Song or Video: The Checks We Run (Free Commands)

The FFmpeg and file-hash commands below run locally. No purchase is needed to follow them. In September 2026 we published a 140-second "music video" for one of our AI songs. It was one still picture with a small moving waveform along the bottom. Our pre-upload checks had measured the length, the resolution and the loudness. Nobody had checked whether the picture actually moved. We deleted it from our channel the next morning. This post shows the checks we run now. Every command below runs on your own computer with ffmpeg , and every number comes from files we measured ourselves. 1. Loudness and true peak (free, one command) ffmpeg -hide_banner -nostats -i song.mp4 -af ebur128=peak=true -f null - Read the Summary block at the end: I: is the integrated loudness in LUFS. Peak: under "True peak" is the true peak. ffmpeg prints the unit as dBFS. A test file we deliberately pushed too loud read I: -8.4 LUFS and Peak: 3.5 . Our own working baseline is ...
Read more