Reframe to vertical and upscale, without re-rendering

The render is finished, approved — and the wrong shape. A 16:9 cut does not go on Reels, a 9:16 phone clip does not go on a landing page, and a print vendor wants more pixels than any generator produced. Re-rendering to get there throws away work that is already done, and usually produces a slightly different picture into the bargain: a re-run is a new take, not the take everyone signed off.

Prepare exists for exactly this moment, the step between rendered and posted. Two jobs share the one surface: change the shape, or make it bigger. Both take one file from your asset library and file one transformed file back into it, and neither touches the original.

Two jobs on one page, because they are the same moment

Reframing and upscaling are the same shape of operation — one library file in, one transformed library file out — and they arrive at the same point in the work, which is why they share a surface instead of being two rows on the rail. Both jobs take minutes, so the rail, the drawer and the credit pill stay live while one runs; and the page cannot be double-driven, because a second Run while the first is still going would charge twice and file two copies of a file you asked for once.

The one requirement is that the file is in the library. Both jobs run server-side and read the source by id, so a file picked from disk needs the picker's save-to-library box ticked — which is also what makes the result part of a record instead of a download. That is the same logic as the rest of a workflow with no export step: the handoff is the product.

Reframing is deciding what to lose

Going from 16:9 to 9:16 discards 68% of the width. There is no setting that avoids that arithmetic — only three genuinely different answers to it, and Prepare makes you pick one rather than pretending the question away.

Fill the frame crops: the edges go, the subject stays, and the result is what make this vertical normally means. Keep everything letterboxes instead — nothing is lost and nothing is invented, which is what an archivist actually wants. Blurred surround fills the frame with a blurred enlargement of the same footage and sits the whole picture on top: the look most vertical reposts already have, because it loses nothing and still fills the screen.

Before anything runs, the page says in one sentence what the chosen combination will do — Cuts 68% of the picture away, keeping the subject in frame, or Fits the whole picture inside 9:16 and letterboxes the rest — because that is the thing worth knowing while you can still pick a different mode, and it is not something you can work out from “16:9 → 9:16” at a glance. If the file is already the shape you asked for, it says that too, before you pay for a no-op.

The crop follows the subject

A centred crop is only right when the subject happens to be centred, and in the shots people actually want reframed, it usually is not. With subject detection on, Prepare finds the salient objects in the frame first and places the crop window against them; with nothing detected, it falls back to centre rather than guessing.

The window is pulled toward the subject rather than snapped onto it. A window locked exactly to a detection jumps every time the detection moves a few pixels, and on a shot with two subjects it sits on whichever one the model happened to rank first. Weighting toward the middle keeps the framing stable — and keeps the second subject in shot more often than not.

The shapes themselves come from the film vocabulary, served to the page rather than hard-coded in it: widescreen 16:9, anamorphic 2.39:1, theatrical 1.85:1, portrait 4:5, vertical 9:16 and square — the same aspects a film project delivers in, so a shape that exists anywhere in Grafy exists here. For video, only the picture is touched: the output is a clean H.264 encode and the audio track is copied through untouched, because a reframe changes the picture and nothing else, and a needless re-encode would cost quality for no reason.

Make it bigger means a model, not a resample

The other job is enlargement, and the word matters: this is not the bicubic resize your image viewer does, which invents no detail and simply spreads the same pixels thinner. Prepare runs a trained super-resolution model — a local ESRGAN-family model where one is installed, otherwise hosted upscalers with automatic fallback across models — which reconstructs texture and edge detail as it enlarges.

You ask for it in one of two ways. A multiplier, 2× or 4×, is how people think. An exact longest side is what a print vendor or a platform actually demands, and it is the option that saves the calculator: ask for 4,096 pixels on the long edge and the aspect ratio is preserved around it. Under the hood the model runs at its own native factor and the result is then fitted to your target — it is never resampled up beyond what the model produced, so the number you asked for is real detail, not stretched pixels. The ceilings are generous and stated: up to a 16,000-pixel long edge, about 96 megapixels, factors from 1 to 8.

Video upscaling promises less, on purpose

Video can be upscaled too, by factor only. The exact-size option disappears for a video source rather than being greyed out or quietly approximated, because the video upscalers take a factor and the output dimensions are whatever the model produces — offering a pixel target the server cannot honour would be a promise it has to break. Video work is priced by the clip's duration, and the estimate is shown before you run, like every other paid act in Grafy.

Whether upscaling is available at all is asked, not assumed: the page probes the server on load, and a deployment with no upscale model disables that half only. Reframing needs ffmpeg alone, so it keeps working everywhere.

Everything lands as a new file

Neither job modifies its source. The result is a new asset with its own name — the original title plus what happened to it, unless you name it yourself — and the completion message says where it went and at what size, because the result lives in the library rather than on the page, and done would leave you hunting for it. An upscale that overwrote the original would be unrecoverable; a reframe that did would destroy the landscape master the moment you made the vertical cut.

Working at the library level is also why Prepare exists as its own surface instead of a button inside one studio. Everything in Grafy files its output into the asset library — a Creator image, a film frame, a Video Edit render, an upload — so one surface pointed at the library can post-process the output of every app, including the ones that have no upscale button of their own. The result card names the model that did the work, and the usage dashboard itemises the charge afterwards like any other generation.

Compared with the alternatives

Re-rendering at the new shape or size costs a full generation and changes the content: a new take might be better, but it is not the approved one, and for a client-signed frame that is disqualifying. A converter website handles the arithmetic but breaks the record — the file leaves the library, the result comes back with a filename like final-vertical-2.mp4, and nothing connects the two. And doing it on the Video Edit timeline is the right answer when the work is an edit — keyframes, layers, a grade — but wrong for the ninety-second errand of making one finished file fit one platform. Prepare is the errand tool: one file in, one decision about what to lose or how much to gain, one file out.

Frequently asked questions

How do I convert 16:9 video to 9:16 without cutting off the subject?

Pick the file in Prepare, choose the vertical shape, and leave subject detection on. The crop window is placed against the detected subject rather than the centre of the frame, and the page tells you before running how much of the picture the crop discards. If nothing may be lost, switch the mode to the letterbox or the blurred surround instead — both keep every pixel.

What is the difference between AI upscaling and just resizing?

A plain resize spreads the same information across more pixels, which is why it looks soft. A learned upscaler is a model trained to reconstruct plausible detail — texture, edges, hair — as it enlarges. Prepare uses a real model for every upscale, runs it at the model's native factor, and only fits the result down to your target, never stretches it up past what the model made.

Can I upscale a video as well as an image?

Yes, by a factor of 2× or 4×. The exact-pixel target is images-only, deliberately: video upscalers produce whatever dimensions the model produces, and Prepare will not offer a promise the server cannot keep. Video jobs are priced by the clip's duration, with the estimate shown before you run.

Does reframing or upscaling change my original file?

No. Both jobs read the source and file their result as a new asset in your library, next to the original. The landscape master survives the vertical crop, and the native-resolution image survives its enlargement. You can run both jobs, or the same job with a different mode, from the same untouched source as many times as you need.

How large can I upscale an image?

Up to a 16,000-pixel long edge or roughly 96 megapixels, whichever comes first, from factors between 1 and 8. Exact targets keep the aspect ratio: ask for a 4,096-pixel long edge and the short side follows. Past the ceiling the request is clamped rather than refused, so you always get the largest result the budget allows.

Open Prepare, pick the file that is finished but wrong-shaped, and decide what to lose — the page will tell you exactly what that decision costs before you make it.