Runway states shapes as pixel pairs, not as ratios
The input reference gives shape as seven pixel pairs rather than as named ratios: three landscape, three portrait and one square. Every other entry in this column publishes names, so this is the one row where a placement check means dividing. As of 2026-09-22.
| Pair | Orientation | What it corresponds to |
|---|---|---|
| 1280:720 and 720:1280 | Landscape and portrait | Sixteen by nine, both ways round |
| 1584:672 and 672:1584 | Landscape and portrait | Close to twenty-one by nine |
| 1104:832, 832:1104, 960:960 | Landscape, portrait, square | Four by three and one by one |
Inclusion rule. Pixel pairs the image-to-video ratio field accepts on this reference page. The correspondence column is arithmetic on the printed pairs and is labelled as such rather than presented as a vendor statement. Order. Grouped by the ratio the pair works out to, widest first.
1A pair is more precise and harder to check
Named ratios are convenient and lossy: sixteen by nine says nothing about how many pixels arrive. A pair says both, which means a placement asking for a minimum pixel size can be checked directly rather than through a tier name and an assumption.
The cost is comparison. This row cannot be lined up beside the other fourteen in a shape column without the reader doing arithmetic, and these sheets do not do it for them, because a derived ratio would look identical to a published one.
2The seventh pair is square, and squares are rare here
960 by 960 is one of three square options in the register. The platform in this register that publishes minimums accepts square from 640x640 upward, so this pair clears it and the route is confirmable from the page rather than assumed.
That is worth noting because the same check is not possible on the eleven rows whose shape cell is a name or is empty. A pixel pair is the only form of this cell that settles a pixel floor without a second source.
3The pairs bound the output more than a tier name would
The largest long side on the page is 1584 pixels. Nothing here reaches 1080p on its long axis, which a resolution ladder expressed in tier names would have obscured, and which matters for reuse beyond a feed placement.
Read with the file ceilings on the same reference, the row describes a service tuned for short conditioned pieces rather than for large masters. Both columns say the same thing in different units, which is the advantage of reading them separately.
4Where this cell sits in the column
| Generator | Aspect ratio |
|---|---|
| Amazon Nova Reel | Not stated |
| CogVideoX | Not stated |
| Google Veo | 16:9 and 9:16 |
| Grok Imagine | Seven, 16:9 the default |
| Hedra | Seven, including 9:21 and 21:9 |
| HeyGen | 16:9 or 9:16 |
| Kling AI | Not stated |
| LTX-Video | Not stated |
| MiniMax | Six, required for text-to-video |
| Pika | Not stated |
| PixVerse | Eight on version 6 |
| Runway | Seven pixel pairs |
| SceneMixer | 9:16 and 16:9 |
| Vidu | Not stated |
| Wan 3.0 | Six, adaptive by default |
Inclusion rule. Every generator in the register, carrying whatever its own page says about output shape. Six rows say nothing and are marked hollow rather than dropped. Order. Alphabetical by generator name.
- Gen-4.5 image-to-videoLandscape 1280:720, 1584:672 and 1104:832, portrait 720:1280, 832:1104 and 672:1584, square 960:960shape given as a pixel pair rather than a ratio
- Ceilings on what goes in16MB for an image and 32MB for a video by URL, 5MB and 16MB as a data URI, 200MB for an ephemeral uploadnot on what comes out
5Sources
Read from the Runway API input reference, checked 2026-09-22. The column itself is described on aspect ratio, and the rule that keeps surfaces and generators apart is on method. Next along this column: Amazon Nova Reel, TikTok.