Upload one video and get it back reframed for the aspect ratio a platform actually needs, vertical for Reels and TikTok, square for the feed, widescreen for YouTube, with the subject kept in view instead of cropped off. Set where your footage sits in the new frame, add a short prompt if the new space needs filling, and download a clean file ready to post.
Every platform wants a different shape of the same video, and cutting it by hand means opening an editor, guessing at safe margins and re-exporting the same ten seconds four separate times. An AI reframing model looks at the footage, works out where the subject is, and rebuilds the frame at the ratio you need, so one upload turns into a vertical cut for TikTok, a square one for the feed and a widescreen one for YouTube without you touching a timeline.
Most footage is shot once, in whatever ratio the camera defaulted to, and then needs to live everywhere. A center crop handles that badly the moment the subject drifts to one side of the frame, which is most of the time in anything handheld or unscripted. Cropping blind clips a shoulder, cuts off a hand mid-gesture, or leaves the actual subject outside the safe zone a platform reserves for its own UI.
Reframing is different from cropping because it reasons about where the content is before deciding what to keep. Picasso IA runs luma/reframe-video for exactly this job, inside a catalog of 488 models covering image, video and audio generation. The model takes a short clip, a target ratio and, optionally, a prompt or a set of crop bounds, and returns a new video where your subject stays framed instead of guessed at. It will not replace a camera operator who reshoots for each platform on purpose, but for the ninety percent of clips nobody is reshooting anyway, it removes the manual crop-and-export cycle entirely.
Reframe Video supports seven output ratios, and picking the right one matters more than any other setting, since the wrong shape gets letterboxed, cropped again by the platform, or simply looks wrong in the feed. Match the ratio to where the clip is actually going before you touch anything else.
| Ratio | Fits | When to use it |
|---|---|---|
| 9:16 | TikTok, Reels, Shorts, Stories | Full-screen vertical for mobile-first feeds |
| 1:1 | Instagram feed, Facebook, LinkedIn posts | Square crop that holds its footprint while people scroll |
| 16:9 | YouTube, most websites, presentations | Native widescreen, no padding if the source already fits |
| 4:3 | Older embeds, some display ad slots | A boxier frame when the subject is centered and upright |
| 21:9 | Trailers, hero banners, cinematic teasers | Ultra-wide crop for a panoramic, film-like feel |
If a single clip needs to go to three platforms, run three separate generations rather than trying to find one ratio that half-works everywhere. Each run takes the same source and a different target, and the model handles the geometry differently for each one.
The model accepts a prompt alongside the ratio, and its job is narrow: it tells the model how to treat the space a new ratio adds, not what the scene is about. Going from 16:9 to 9:16 adds a strip above and below the subject, and the model has to decide what belongs there. A prompt like extend the plain studio background keeps that new space simple and unobtrusive; leaving the prompt empty lets the model guess, which works fine for plain backgrounds and less well for busy ones.
Crop bounds do more precise work than the prompt does. Setting x_start, x_end, y_start and y_end tells the model exactly which pixels of the source map to the new frame, and grid_position_x and grid_position_y control where that crop sits inside the output. For a talking-head clip, centering the crop on the face rather than the geometric middle of the frame is usually the single change that fixes a bad reframe.
Reframe Video caps input clips at ten seconds. For anything longer, cut the segment you actually need with Trim Video first, then reframe that shorter piece, rather than reframing the whole file and hoping the important part survives.
This is the full loop from a single source clip to a platform-ready export, and it runs the same way whether you need one new ratio or five.
A reframe only looks intentional when the subject lands where an eye expects it, and that comes down to a handful of habits rather than any single setting.
Reframing does the geometry well and nothing more than the geometry, so a few limits are worth knowing before you plan a whole export pipeline around it.
Output tops out at 720p, which is fine for social feeds but not for a client delivering broadcast-quality footage; run the reframed clip through video-increase-resolution or real-esrgan-video afterward if you need more pixels. The ten-second cap means longer scenes have to be split and reframed in pieces, then reassembled, rather than processed in one pass. Extending a frame significantly, such as going from a tight 16:9 shot to 9:16, means the model is inventing new background rather than revealing footage that was never there, and busy or detailed backgrounds invent less convincingly than plain ones. And a clip with two people who both matter cannot be perfectly centered on both at once; one crop decision per generation means picking who the frame favors.
None of this makes the tool unreliable, it makes it a geometry solver with known edges. Plan the source footage and the crop settings around those edges and the output holds up.
Yes. Upload the same source clip to Reframe Video twice, once targeting 9:16 for TikTok and once targeting 16:9 for YouTube, and the model recalculates the crop for each ratio separately rather than stretching one output to fit both. The subject stays framed in each version because the model repositions the crop per target rather than applying one fixed center point.
Seven: 1:1, 3:4, 4:3, 9:16, 16:9, 9:21 and 21:9. That covers square feed posts, vertical mobile formats, standard widescreen and the extra-wide and extra-tall ratios used for banners and cinematic crops. Pick the ratio that matches the destination platform rather than the ratio the source was originally shot in.
Yes, ten seconds per generation. Longer footage needs to be trimmed down first, either to the single moment you actually need reframed or into ten-second segments you reframe separately and stitch back together. Trim Video handles the cutting step with precise start and end timecodes before the clip ever reaches the reframing model.
Output is delivered at 720p regardless of the source resolution, so a 4K upload comes back smaller, not sharper. For social posting that ceiling rarely matters, since most feeds compress well past 720p anyway. If you need the reframed clip at a higher resolution afterward, running it through video-increase-resolution recovers sharpness at 2x or 4x scale once the geometry is already correct.
Yes, with the crop bounds and grid position settings. Setting x_start, x_end, y_start and y_end defines exactly which part of the source maps into the output, and grid_position_x and grid_position_y control where that crop sits inside the new frame. Centering these on a face or product is usually what separates a clean reframe from one that clips something important.
The model fills it based on the prompt you give it, or its own guess if you leave the prompt blank. A plain, simple background fills convincingly with almost any prompt; a busy or highly detailed background gives the model more to guess at, and the seams show more easily. Describing what belongs in that space, rather than describing the whole scene, gets the cleanest result.
Yes, reframing and captioning are two different steps on Picasso IA. Once a clip is reframed to the ratio you need, run it through autocaption to burn in styled subtitles sized for that format; a vertical Reel usually wants larger text and fewer characters per line than a widescreen upload does, and autocaption lets you set both independently per export.
Yes, and that is the normal workflow. Upload the source once, then run Reframe Video once per target ratio, since each generation only produces one output ratio at a time. Keep the crop bounds consistent across runs when the subject does not move much, and adjust them per ratio when a tighter or wider ratio changes where the subject needs to sit.
New Picasso IA accounts start with free credits, which is enough to test Reframe Video on a handful of clips before spending anything. Generation cost depends on clip length and the model you run, and current plans and credit allowances are listed on the pricing page rather than quoted here, since they change over time.
A single set of crop bounds works best for a subject that stays roughly in the same place for the length of the clip. For a subject that moves significantly, either trim the clip down to the segment where it holds still, or lean more on the prompt than on fixed crop bounds and let the model track the motion itself, checking the result afterward since motion-heavy reframes need more review than static ones.
A clip sitting in one aspect ratio is a clip half the platforms will never see properly. Open the Toolkit and turn it into the shape each one actually asks for.
