AI Video for Instagram Reels: Specs, Labels, and Settings That Matter
The export settings Reels actually wants, the aspect-ratio default that ruins generated clips, and how Meta's AI labelling works. Everything specific to publishing generated video on Instagram.
Generated video meets Instagram at three points: the frame it comes out in, the file you upload, and the label Meta attaches. Each has a specific answer.
The frame
1080 × 1920, 9:16. This is the only ratio that fills the Reels tab without cropping or letterboxing. Instagram's minimum is 720p and its practical maximum is 1080p — exporting higher does not help, because Instagram compresses above that anyway.
The generation-side trap: HappyHorse defaults to 16:9. So does a good deal of the category. If you generate at defaults and post to Reels, you get a landscape clip in a vertical slot, and you have paid for it. Set aspect_ratio to 9:16 explicitly — or 4:5 if the clip is destined for feed rather than the Reels tab.
Keep the important content in the central safe zone, away from the caption block at the bottom and the UI down the right. And design the cover so the key elements survive the central 1:1 crop that appears on your profile grid.
The file
MP4 or MOV, H.264, 30fps, AAC audio, exported at a high bitrate — around 10–15 Mbps — so detail survives Instagram's compression. Generated video is often softer than filmed footage to begin with; giving the encoder headroom matters more here than with camera output.
The length
Reels run to 90 seconds in the standard flow, with longer uploads possible in-app and via third-party tools. For reach, 7 to 30 seconds is the band that works.
This aligns unusually well with what generation models are good at. Identity drift compounds past about eight seconds, so a 20-second Reel is best built as three short generations cut together — which is also better editing.
The label
Meta has applied AI labels on Instagram and Facebook since 2024, using C2PA — provenance metadata written into the file recording which tools touched it. Detection is automatic; the label does not wait for you to disclose.
For organic posts a label is not a penalty and there is no credible evidence of suppression for being labelled. For ads, disclosure of AI-generated visuals, text or audio is an enforced requirement, and enforcement can reach retroactively into a campaign already running — an ad can be labelled, rejected, or flagged after approval.
Practical advice: do not strip metadata to avoid detection. It is the behaviour the policy is written against, and the upside is nil.
A publishing checklist
- Generate at
9:16— never at the model's default. - 5–8 seconds per generation; cut several together for anything longer.
- Export H.264 MP4, 1080 × 1920, 30fps, 10–15 Mbps.
- Check the central 1:1 crop for your cover.
- Use in-platform audio for trending sounds — licensing is handled there and nowhere else.
- Leave the AI label alone. Add a manual disclosure on anything sponsored.
What still gets you actioned
Not "used AI". Someone else's face, someone else's voice, or someone else's music. Those three are the enforcement surface on this platform, and the label is not what triggers any of them.