How to upscale AI-generated video (Veo, Sora, Runway…) on a Mac

Generator specifications checked on the vendors' own documentation on November 9, 2026. Local Upscale Studio is in development.

AI video generators usually give you a short clip at 720p or 1080p. If you want it bigger, for a 4K edit, a large screen or a client, you need to upscale it. AI clips aren't like old home videos or phone footage, though, and the usual advice doesn't all apply. This guide covers what the generators output, what is different about their clips, and what to try first.

What the generators give you

These services change often, so check the current documentation for the one you use. As of today:

Google Veo 3.1 (on Google Cloud's Vertex AI): the main model, veo-3.1-generate-001, lists 720p, 1080p and 4K output at 24 fps, in 16:9 or 9:16, for clips of 4, 6 or 8 seconds. The Fast and Lite versions list 720p and 1080p. Google's Gemini API documentation says its Veo 3.1 preview models were shut down on October 22, 2026.
Runway: its help centre says generative video outputs are created in 720p and can be upscaled to 4K with an Upscale option beneath the finished generation. It also has an Upscale app for uploaded videos of up to 40 seconds. Exporting a 4K project from its video editor does not upscale footage that isn't already 4K.
OpenAI Sora: OpenAI's developer documentation says the Sora 2 models and Videos API were shut down on September 24, 2026. When it ran, sora-2-pro was the option for 1080p exports (1920×1080 or 1080×1920). If you have Sora clips saved, they are most likely 1080p or smaller.

If your generator offers a higher resolution or its own upscale, try that first. It works on the clip at the source, before any download or re-sharing. Compare it with a Mac upscaler on the same clip if you have the time.

Why AI clips behave differently

Most upscalers are trained to repair damage: blur, noise, compression blocks, low light. Old footage has plenty of that. A fresh AI clip usually doesn't.

It's already clean. There is little noise or blocking to remove, so a restoration-style upscaler has less to do. The visible gain from upscaling is often smaller than on old footage.
The faults are generation faults. Warped hands, melting background details, garbled text and objects that change shape between frames come from the generator. An upscaler doesn't know they are wrong, and may make them sharper and easier to notice.
The texture is already synthetic. Generators often produce very smooth skin and surfaces. An upscaler can add fine texture, which some people like and others find overdone.
Re-shared copies are different. A clip downloaded from social media has been compressed again. That copy does have compression damage, and a restoration model has more to work with. Always start from the original download if you can.

What our own test found

While choosing models for Studio, we ran nine items through Real-ESRGAN, SeedVR2 and FlashVSR and compared the results blind, without knowing which model made which. One item was a clean clip generated with Veo 3, upscaled from 1280×720 to 1920×1080.

On that clip, no model was a clear winner. On every degraded video in the test (a 1993 camcorder tape, low-light phone footage, a 1955 film transfer and a badly compressed web clip), the newer video model was picked. On the clean AI clip the differences were too small to call.

That is one person's judgement on one clip, so treat it as a hint, not a result. But it matches the reasoning above: when the source is already clean, a heavier model has less to fix. You can see the degraded examples on our results page.

A practical approach

Download the original file from the generator, not a copy from a feed or a chat app.
Check what you have. This prints the codec, width, height and frame rate:
ffprobe -v error -select_streams v:0 -show_entries stream=codec_name,width,height,avg_frame_rate -of csv=p=0 clip.mp4
Edit first, upscale second. Trim and assemble the clips you will actually use, then upscale only those. It saves a lot of time.
Pick the target. 720p to 1080p is a 1.5× step and usually looks natural. 720p to 4K is 3×; 1080p to 4K is 2×. See how to upscale video to 1080p or 4K for scale factors and codecs.
Try plain resizing. For a clean clip, ordinary Lanczos resizing may be all you need:
ffmpeg -i clip.mp4 -vf "scale=1920:1080:flags=lanczos" -c:v libx265 -crf 20 -preset medium -tag:v hvc1 -c:a copy clip-1080p.mp4
Then try a fast AI upscale on the same few seconds, and compare both on the screen you will use. Pick the one you prefer, not the one that looks sharpest at first glance.
Use a slow, heavy model only if you can see the gain. On clean clips it may not be worth hours of processing.
Keep the frame rate. Veo outputs 24 fps. Leave it at 24 unless you have a reason to convert, and keep the generator's audio as it is.
Check faces, hands and text at full size before you deliver. Upscaling makes them more visible, good or bad.

We tested both commands with ffmpeg 9.0 on macOS 27, on a short test clip. For a vertical 9:16 clip, use scale=1080:1920 instead.

Where Studio fits

Local Upscale Studio is a pay-once Mac app for Apple silicon that upscales photos and video in one app: $49 at launch and $79 from version 1.0.

Fast mode uses Real-ESRGAN and runs on every Apple silicon Mac. Given our test, it is a sensible first try on clean AI clips.
Quality mode uses FlashVSR for video. It is slow: 1.0–1.6 seconds per frame at 540p–720p output on an M5 Max with 64 GB, using 13–17 GB of memory. It will likely be offered on Macs with 24 GB or more, for short clips, which suits 4–8 second generations.
A test on first launch picks the models your Mac can run, and you can override it. Nothing is uploaded, which matters if the clip is unreleased client work.

Studio isn't released yet, and we haven't published results from the finished app. For other options, see choosing an AI upscaler for Mac and AI upscaler prices.

Questions

Should I upscale before or after editing? After trimming, so you only process what you keep. If you add titles or graphics, add them after upscaling, at the final size, so they stay crisp.

Will upscaling fix warped hands or flickering? No. Those come from the generator. Regenerating the shot is usually quicker than trying to repair it.

Is 4K worth it for a short AI clip? If it goes into a 4K timeline, yes, because something has to scale it. Upscaling it yourself lets you choose how. For social media, 1080p is usually enough.

Can I upscale on a Mac without uploading the clip? Yes. ffmpeg, HandBrake and Studio all work on your Mac. See offline upscalers for Mac.

Get the launch price. Studio is $49 until version 1.0. See pricing.

Sources

Veo 3.1 output specifications: Google Cloud, Veo 3.1 models
Veo 3.1 preview shutdown in the Gemini API: Google AI Studio video documentation
Runway resolution and upscaling: "Can I generate videos in 4K?"
Sora 2 output sizes and shutdown: OpenAI video generation guide
ffmpeg scale filter: ffmpeg filter documentation

Veo and Gemini are trademarks of Google LLC. Sora is a trademark of OpenAI. Mac, macOS and Apple silicon are trademarks of Apple Inc. All other names are trademarks of their owners. Local Upscale Studio is not affiliated with any of them.