Reframe a Horizontal Video to Vertical Automatically

Make Your Video Vertical

Upload real footage, type one sentence, watch the preview. 50 free credits, no credit card.

How to Auto-Reframe a Video to Vertical

1. 1
   Upload the horizontal video

   Drag in any MP4, MOV, MKV or WebM up to 14GB or 3 hours. Valmera runs a one-time analysis: a word-level transcript, shot detection, silence detection, and labeled frame tiles the agent reads directly. Your original file is never modified.
2. 2
   Type "make this vertical for TikTok"

   That single sentence is the whole job. "Convert this to 9:16", "make a square version for the feed" and "reframe it 4:5 for Instagram" all work the same way. The agent picks crop or fit by measurement, not by assumption.
3. 3
   Read what it measured, then correct it

   The reply names the focus point it found and how many of the sampled frames agreed. If the framing is wrong, say so in plain English — "frame the person on the left", "use blurred bars instead", "crop it anyway, I don't need the sides" — and it re-aims.
4. 4
   Export from the original file

   Previews render from a fast proxy; the final export renders from your original upload at source quality as an H.264 MP4. Change your mind later and "put it back to 16:9" restores the source ratio exactly — the reframe lives in the edit, not in the file.

The analysis runs once per upload. Reframing itself is one message, and every follow-up correction is another.

What a 9:16 Window Actually Takes

Start with the arithmetic, because it is the whole problem. A 16:9 frame has an aspect of 1.778. A 9:16 frame has an aspect of 0.5625. Fit the tallest possible 9:16 window inside a 16:9 picture and it is 31.6% as wide as the source — on a 1920×1080 file, a 608-pixel slice of a 1920-pixel frame. Converting horizontal to vertical always discards 68% of the width. The only question any auto-reframe tool answers is which 31.6% to keep.

How much a conversion costs depends on which two shapes you are moving between. These are the four common ones, computed from the ratios themselves:

Crop, Pad, or Blurred Pad — the Real Trade-off

There are exactly three honest ways to put a wide picture in a tall frame, and none of them is free.

Crop fills the screen and loses 68% of the width. It is the correct answer for footage with one subject in it — a vertical crop of a talking head is what vertical video looks like. Pad keeps every pixel and spends two thirds of a phone screen on black bars; it is honest and it looks like an upload nobody bothered with. Blurred pad keeps every pixel and still fills the screen, at the cost of the picture being small — which is why it is the compromise most people actually ship, and why it is the standard look of repurposed wide footage across Reels and Shorts.

Valmera's default is to measure which one your footage needs rather than pick a house style. An explicit instruction always wins: ask for a crop and you get a crop, sides and all.

How the Measurement Actually Works

Reframing samples five frames spread evenly across the footage you kept — not across the raw upload. If you cut the first four minutes, the framing follows what the viewer will see, not material that is no longer in the video.

Faces are measured in the pixels, not asked of a model. Each sampled frame is downscaled to 640px wide, converted to grayscale, histogram-equalised, and run through frontal and profile Haar cascades. This takes milliseconds, costs nothing, needs no vision provider, and on the footage people actually reframe — a person talking — it beats asking a multimodal model to estimate a coordinate. Two adjustments matter: detections within 2% of any frame edge are discarded as false positives, and the focus point is lifted by 12% of the face box's height, because portrait framing crops on the eyeline rather than the geometric middle of a head.

A face has to appear in at least two of the five samples to count. Cascade false positives are per-frame and independent; a real subject in a video someone wants reframed is in most of the frames. That quorum was learned the hard way. A wide gameplay recording containing no faces at all matched in exactly one of five frames and aimed a 9:16 crop at (0.39, 0.20) — the corner of a HUD. Requiring two agreeing detections costs a genuine subject nothing and costs a phantom everything. The five points are then reduced by median, not mean, so one wide establishing shot cannot drag the crop off every close-up.

Then it asks whether a crop is the right operation at all — the part almost no reframing tool does. Gradient energy is a good proxy for where a picture's content is: backgrounds are smooth (a wall, sky, bokeh, a blurred room) and subjects are not. So the renderer's exact crop-window geometry is reproduced in Python, the frame's gradient energy is integrated over that window, and the result is the share of the picture's detail the crop would keep. Below 55%, the crop is not reframing — it is truncation, and the whole frame is fitted into the new shape over a blurred backdrop instead. A centre-weighted talking head keeps well over 60% of its detail through a vertical crop. A wide game recording keeps about 35%, because the HUD, the minimap and the score all sit outside the window.

That test also gates the fallback path. When no face is found and a vision model is available, the model is asked where the subject is — but even when it confidently names a focal point, the same detail measurement runs against that point before a crop is allowed. Asked where to look in a game frame, a vision model will happily point at the character, and the HUD around it still gets cut off.

When there is no face and no vision model either, the last resort is the centroid of gradient energy, weighted 65% toward the measurement and 35% toward the frame centre — enough to move the crop meaningfully, not enough to slam it into a corner on a noisy frame. That is a weaker claim than a detected face, and the agent describes it as one instead of asserting it found your subject.

Everything above is written into a versioned edit decision list, not baked into pixels. The reframe is one line of that list; your upload is untouched, and "put it back to 16:9" removes the line.

When It Goes Wrong, and What to Say

Two people in the frame. The largest detected face wins, which is what a human editor does on a group shot — but if your two speakers sit at opposite sides of a 16:9 frame, no 9:16 window holds both, and no amount of cleverness changes that. The agent measures how far the subject points spread across the samples and, when that spread passes 0.18 of the frame, says so and offers the blurred-pad conversion. Say "frame the person on the right" if you want one of them, or "use blurred bars, keep both" if you want the exchange.

The subject moves. The focus is one fixed point for the entire video. If your speaker walks across the room, or you cut between two very differently framed shots, one point cannot serve both — and the agent tells you when it measured that, rather than shipping a crop that is right for half the video. Blurred pad is the reliable answer; a hand-aimed crop plus a punch-in on the moments that matter is the more interesting one.

It fitted when you wanted a crop. The detail measurement is deliberately conservative — it would rather show you the whole picture small than silently cut off a scoreboard. Override it in one sentence: "crop it anyway, I don't need the sides".

It cropped when you wanted everything. Say "don't crop anything out — use blurred bars". Pad modes discard nothing, so there is nothing to aim and nothing to measure.

The aim is simply off. Coordinates are the vocabulary here, not adjectives. "The crop is too far left" works; so does naming the person, the object, or the region you care about. The agent looks at real frames of what it produced before it answers, so a correction is checked against the picture rather than against its own memory of what it intended.

How People Do This Without Valmera

Adobe Premiere Pro — Auto Reframe. Apply it to a clip or to a whole sequence, choose a Motion Tracking preset (Slower Motion for talking heads, Default, Faster Motion for sports), and Premiere generates position keyframes that follow the action. The keyframes are fully editable afterwards, which is its real strength: when the tracker latches onto the wrong subject you fix that one moment instead of re-running the analysis. It is the most capable option here for footage where the subject genuinely moves. It also requires a Creative Cloud subscription and a machine that can scrub the timeline.

DaVinci Resolve — Smart Reframe. Object-aware reframing in the Inspector, with an automatic mode and a reference-point mode for telling it which subject to follow. Resolve's free version is famously generous, but Smart Reframe is not in it — it is a Studio feature, which is a one-time paid licence. If you already own Studio, it is excellent and free at the point of use.

Final Cut Pro — Smart Conform. Analyses the clip and writes a Transform you can then adjust by hand, applied per clip from the Inspector. Mac only, one-time purchase.

ffmpeg — free, exact, and blind. A centre crop to 9:16 is one filter: crop=ih*9/16:ih, and adding :x=... aims it wherever you like. The blurred-pad look is a handful more filters — split the input, scale one copy to cover and box-blur it, scale the other to fit, overlay the second on the first. ffmpeg will execute any of this perfectly and instantly. What it will never do is look at your footage and tell you the subject is not in the middle. You supply that judgement, frame by frame, from your own eyes. For a batch of shots framed identically it is the best tool on this list.

Browser tools. A large field of online resizers and "smart crop" tools will do a 16:9 to 9:16 conversion in a few clicks, and for a short clip with a centred subject they are a perfectly reasonable answer. Read the free tier carefully before you rely on one: watermarks, resolution caps, length caps and re-encoding on upload are the usual costs, and few of them will tell you what they measured or let you argue with the result.

Honest Limits

No tracking. One fixed focus point per video. Valmera does not keyframe a crop path along a moving subject the way Premiere and Resolve Studio do. For a skateboarding clip or a stage walk, use one of those, or accept blurred pad.

One frame setting for the whole program. One ratio, one mode, one focus point. You cannot give the wide establishing shot a different crop from the close-up in the same export. Two versions of one video is two exports.

Four ratios. 16:9, 9:16, 1:1 and 4:5, plus "source". There is no arbitrary ratio and no custom pixel dimensions. (An aspect shift — the frame changing shape partway through and back — additionally reaches 4:3, inside a fixed canvas.)

Nothing is upscaled, and cropping still costs sharpness. The output is capped at the source's pixel budget: a 1920×1080 file becomes 1080×1920 at 9:16, 1080×1080 at 1:1, 1080×1350 at 4:5. No resolution is invented. But a 608-pixel-wide crop window resampled up to a 1080-wide output is genuinely softer than footage shot vertical — that is true of every honest reframing tool, and worth knowing before you reframe a 720p source.

Pad bars are black. No custom bar colour, no branded background, no stored brand kit. Blurred pad or black.

Face detection is classical, not neural. Haar cascades are fast, free and reliable on frontal and profile faces at reasonable size. They are weaker on faces that are very small in frame, heavily backlit, or turned away. When they fail, the pipeline falls through to a vision model and then to detail energy — but the honest description of a fallback result is "an estimate", which is what you will be told.

Related Tools

Reframing is rarely the only thing a repurposed video needs. The same chat can cut the dead air, add word-accurate captions that rescale to the new vertical frame, pull short clips out of a long recording, blur a username or a face so the censor follows the footage through the reframe, and duck music under speech — from one message if you want: "make it 9:16, cut the silences, and caption every word".

If you already know the framing you want, resize video is the direct route to the same four ratios. To work out the numbers first, use the aspect ratio calculator. For the vendor-neutral definition of the technique, see the auto-reframe glossary entry and aspect ratio. The mechanics of the renderer are in the reframing docs, and everything Valmera can do is in the tools hub.

Reframe It and See What It Measured

Upload one horizontal video and ask for vertical. The reply names the focus point it found.

Frequently Asked Questions

Convert Horizontal to Vertical

The agentic AI video editor — describe the edit, read what it measured, download the export. 50 free credits, no credit card.