Upscale CapCut video without buying resolution you cannot get

An export dialog will happily let you choose a resolution larger than the clip you are exporting, and the encoder will write that larger frame. What it cannot do is put back detail the camera never recorded. Every pixel it has to invent is computed from pixels that already exist, through a resampling kernel that averages its neighbours — bilinear or bicubic, depending on the path the build takes. The file gets bigger, the upload gets slower, and the picture carries no more information than it did before. Magnification and reconstruction are different operations, and the export setting is the magnification half.

The disappointment arrives later and somewhere else. In the editor's small preview, and even on a phone held at arm's length, an interpolated 1080p export of a 720p clip looks acceptable. At full size on a laptop screen or a television, the giveaway is a specific kind of softness: edges that are smooth but hollow, fine texture such as hair, fabric or foliage flattened into a wash, and the source's compression blocks magnified into visible squares. None of that is damage the export added; the setting simply made an existing weakness larger and easier to see.

So export at the size the media actually is. If a clip was shot at 720p, export it at 720p and let the AI pass be the only step that invents pixels — that way there is exactly one place in the chain where prediction happens, and you can judge it on its own. Exporting at 1080p first is worse than doing nothing extra: the model receives a file whose invented pixels are indistinguishable from captured ones, and it will sharpen the interpolation along with the detail.

Two clips can both be labelled 720p and be nowhere near the same quality. A phone clip shot in good light with the subject in focus holds crisp edges and mild compression noise; the same phone in dim light, with digital stabilisation working hard or with digital zoom engaged, holds smeared detail and heavy blocking. The resolution label says nothing about any of that. Judge a source by opening one frame at full size, not by reading its properties panel.

Before paying for a pass, name what actually went wrong. If the problem is focus, motion blur or underexposure, a model will produce a confident-looking version of a blurred frame — sharper in texture, no more in focus. If the problem is that the frame is simply small — a 720p clip inside a 1080p delivery, a narrow crop out of a wider shot — that is exactly the case an upscaling model is built for. Matching the tool to the fault is most of the decision, and it is the part people skip.

Compression damage deserves its own paragraph, because it is the one case where sharpening too early makes the result worse. A model can turn block edges and ringing into crisp-looking artefacts that survive all the way to the platform. If the clip is heavily compressed, a denoise pass first and the upscale second usually beats upscaling twice or cranking the strength. Soft-but-clean footage is the easy input; blocky footage is where you test one scene before committing a whole video.

Export the shot once, from the original media, at the clip's native resolution, with the highest bitrate your build allows. Every build exposes slightly different export controls — mobile builds generally offer fewer knobs than desktop ones, and labels move between versions — so the useful instruction is the intent rather than a menu path: spend bitrate on detail, do not scale on the way out, and export once. A file that has already been exported and re-exported carries a second generation of compression into the upscale, before the model has done anything at all.

Think about what is baked into the frames before you export them. Stickers, burned-in captions, colour filters, grain overlays and speed ramps all become part of the image the model sees. A model estimating real texture will also sharpen a caption's lettering, and it can reshape strokes that were never crisp to begin with. Where your build lets you, export a clean plate and add text after the pass; where captions have to be burned in for the platform, check them at full size before publishing rather than assuming they survived.

Hand the file over as either an upload or a public HTTPS URL. The upload path takes a supported video file — MP4, MOV, MKV, WebM, AVI and similar — up to 500 MB; a bigger intermediate is better passed as a URL the service can fetch directly. Audio does not travel through the upscale request, so the soundtrack stays in your project and gets re-attached when the enhanced clip comes back.

A vertical delivery is not a rotated landscape one; it is a smaller frame. A 1080p landscape frame cropped to 9:16 keeps a region roughly 607 pixels wide and 1080 tall. That is the true source resolution for the vertical cut, whatever the timeline label says, and it is why a vertical version of a shoot often looks softer than the horizontal one. Cropping to 9:16 and then upscaling gives the model the exact frame you will publish. Upscaling first and cropping afterwards spends credits on pixels you delete and does nothing for the pixels that remain.

Settle the framing before the pass, not after. If the crop may still move — the subject drifts, or you are choosing between two framings — decide first, because re-cropping an enhanced clip means either a second paid pass or a visible fall back to the unenhanced pixels on one side of the frame. Leave a margin so that edge artefacts sit under the platform's interface overlays instead of on the subject's face, and check the safe areas of the app you are publishing to rather than the generic 9:16 rectangle, since the bottom strip is usually occupied.

When one shoot has to serve several aspect ratios, upscale the largest frame any delivery needs and derive the smaller ones from it. The other order — a separate pass per crop — costs more and produces inconsistent sharpness across the set, which becomes obvious the moment two versions of the same product appear in one feed.

Import the enhanced file and put it on a track above the original, aligned to the same start point, with the original muted or hidden. That gives you a clean A/B you can toggle at full size before committing to anything, which is the single most useful check in the workflow: model output can look plausible in isolation and wrong next to the frames around it. Only after that comparison should the enhanced clip replace the original inside the cut.

Speed ramps and retiming are the usual casualty. If a clip carried a speed change when you exported it, the model saw stretched or compressed frames and predicted detail for a cadence that never existed; apply the retime after the upscale, working from a full-speed intermediate. Transitions, captions and music live in the project and the pass does not touch them, but a dissolve that mixes an enhanced clip with an unenhanced one will show a sharpness jump mid-transition, so check any place where restored footage meets untouched footage.

Keep the project organised so the decision stays reversible. Name the enhanced file after the shot and its source resolution, keep the original media in place instead of overwriting it, and keep the upscale out of any nested sequence you render repeatedly. If the deliverable fails review, you want to swap back to the untouched clip without rebuilding the timeline around it.

Billing follows the duration you submit, not the length of your project. On the Standard model, 1080p footage bills 540 credits per minute of submitted source, which is 32,400 credits an hour, rounded up per submission from a five-second minimum, with a single source capped at 10 minutes. A three-minute vlog where thirty seconds is genuinely soft is a thirty-second job, and it is priced as one.

Convert that to money at the Creator yearly rate of $0.0093 per credit: a minute of 1080p Standard work is about $5.01, an hour about $300. Creator's 106,800 credits cover roughly 197 minutes of 1080p work, or about 53 minutes at the 4K tier. For a vertical feed watched on phones, the 1080p tier is normally the right purchase; 4K buys a larger frame for a delivery spec, not recovered detail.

The saving that actually matters is editorial rather than technical. Trim to the seconds that need the pass before uploading: an intro card, a static talking head in good light, or a screen recording that is already sharp does not need a model at all. Doing that trim in the editor you already have is free, and it is what keeps a weekly series affordable as the clip count grows.