Reframe an AI video to 9:16: regenerate, crop or fit?
Luma lists enhanced reframe for Ray3.2. On Sume, a 9:16 frame comes from asking for it, cropping with video filter, or fitting in Timeline.

Luma's Ray3.2 announcement, dated June 9, 2026, lists enhanced reframe as a feature. Sume has no tool called reframe; to get 9:16 from a clip you have three options: generate at 9:16 in the first place, crop a 16:9 clip with video filter, or fit it into a 1080×1920 Timeline render. The crop is a $0.02 encode job.
What does Luma mean by reframe?
Luma's announcement, read 2026-09-29, names "enhanced reframe" in a feature list and gives no detail in the text saved for this page. This page therefore does not describe how it works.
Should I regenerate at 9:16 instead?
If the clip is not made yet, ask for the ratio. 9:16 is a listed aspect ratio, and each model advertises the subset it accepts in supported_aspect_ratios; gemini-omni-flash-1.1 is documented as 16:9 or 9:16. A new generation composes the shot for the vertical frame, so it is the option that changes the picture, and it is a new paid job.
When is cropping the right call?
When the subject already sits in a vertical slice of the 16:9 frame. The crop op takes fractions of the source frame. A full-height 9:16 slice of a 16:9 frame is 81/256 of its width, about 0.3164, and centered it starts at about 0.3418. That is arithmetic, not a docs rule. The unbilled check route validates the program without creating a job:
curl -X POST https://api.sume.com/v1/video-filter/check \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"video_url": "https://media.sume.com/artifacts/artf_demo/talk.mp4",
"ops": [{ "op": "crop", "x": 0.3418, "y": 0, "width": 0.3164, "height": 1 }]
}'How do the three options compare?
Pick by what you can afford to lose: picture, edges or a rerun. The full crop and fit walkthrough is in converting horizontal video to vertical.
| Option | Endpoint | Cost and limit |
|---|---|---|
| Regenerate | POST /v1/videos with aspect_ratio | A new generation job; limits per model |
| Crop | POST /v1/video-filter | $0.02 per encode job; source up to 300 seconds |
| Fit | POST /v1/timeline-1.0/render | Default 1080×1920; $0.10 per output minute, rounded up |
What must I do before either tool sees my clip?
Video filter and Timeline read only this workspace's media.sume.com files, so import the clip first with POST /v1/media-imports; both encode routes require an Idempotency-Key. The filter's check route does not. Read the video filter docs for the refusal codes.
What size is a cropped result?
The crop keeps the source's pixels, so it is smaller than the source: a 1920×1080 clip cropped to the 0.3164 slice is roughly 606×1080 (arithmetic), with sides rounded to even numbers for yuv420p. That is a 9:16 shape but not a 1080×1920 frame. If you need the full vertical frame, the Timeline route defaults to 1080×1920 and offers fit modes that place the clip in it.
A crop can fail with video_filter_crop_out_of_bounds when the rectangle leaves the frame or a side is under 0.05, and a source longer than 300 seconds fails with output_duration_exceeded. The check route surfaces the first kind before you pay.
Sources
Related posts
More in Media tools
- Amazon online video ad specs: OLV size, length, bitrate
Amazon online video (OLV) ads run 6–120 s in 16:9, at least 1920×1080 and 4 Mbps, with 192 kbps AAC on 2+ channels and up to 500 MB site-served.
- App Store screenshot size (1290×2796) and AI images
Apple lists 1290×2796 for 6.9-inch iPhone screenshots. Sume's ChatGPT Image 2.5 cannot output it exactly, so generate the same shape larger and resize.
- Audio ad specs: Spotify, Amazon, SiriusXM, and YouTube
Audio ad specs by seller: Spotify wants 192–320 kbps at -16 LUFS, Amazon a 10–30 s file up to 3 MB, SiriusXM a 44.1 kHz MP3, YouTube a video.
- Extract 16 kHz mono audio from a video for speech-to-text
Set sample_rate 16000 and channels mono on Sume's audio detach to get the speech-to-text shape from a video. Options, the 900 second cap and the price.
Written by Sume