A video upscaler for Mac that needs only a browser

Apple silicon Macs are built around unified memory rather than a discrete card with its own pool. Wikipedia's Apple silicon article lists the M-series chips with unified RAM and per-chip memory bandwidth — for example the M1 Pro at up to 32 GB unified RAM with up to 200 GB/s — and notes that the Mac line completed its transition to Apple silicon in June 2023. That architecture is efficient, but it also means a local upscaling job shares memory and GPU time with everything else the machine is doing.

A hosted route sidesteps that entirely. The browser uploads the file, a remote GPU runs the model, and the Mac does nothing heavier than a download while the job runs. Nothing is installed, no Python environment or model weights are fetched, no driver compatibility is involved, and closing the lid mid-job does not affect work already submitted. For a single clip or an occasional batch, that is usually the lower-friction choice.

What it costs you is control and locality. The file leaves the machine, it has to be within the size and duration limits, and the request carries only four things: the file or URL, the model, the output tier and the billed duration. If a project requires the footage never to leave your hardware, or the same pipeline to run inside an existing local toolchain, the local route is the right one, and the sections below say so rather than talking you out of it.

A Mac with a current browser, a file, and a network connection. That is the whole list. You upload a supported video file — MP4, MOV, MKV, WebM, AVI and similar — of at most 500 MB, or paste a publicly reachable HTTPS video URL for larger material. The URL route is the practical one for a long recording, because sourcing by link avoids the upload and keeps the browser tab light.

Then the limits decide the shape of the job. One source is capped at 10 minutes and billing is rounded up from a five-second minimum, so a feature-length recording becomes a series of segments rather than one submission. The pipeline exposes no aspect-ratio, crop, frame-rate or audio option, so anything structural — deinterlacing, framing, soundtrack — has to be resolved on the Mac before the upload or after the download.

Two habits keep a Mac workflow tidy. Export the segment you intend to test straight from your editor rather than screen-recording a preview, because a screen capture adds a scaling step and a second encode to material that is already the weak link. And keep the segments in a folder named after the source, so the untouched files and the processed ones travel together.

Choose five to ten seconds that carry the hardest material: a dark scene, a fast pan, fine texture, on-screen text, a face in motion. Upload that segment, pick the model and the output tier, and read the credit quote before submitting. At a five-second billing minimum the test costs a handful of credits, and it answers the question that matters most on a Mac workflow — whether the hosted result is good enough that you can stop thinking about a local pipeline.

Set your pass and fail rules before you look. Reject a setting when skin goes waxy, when fine texture becomes a repeating pattern, when lettering changes shape, when flat areas crawl, or when edges acquire haloes. Accept it when motion is stable at your target tier, when the result survives a 100 percent crop, and when it looks the same on the Mac display you will actually deliver to.

Then process the rest in 10-minute segments, keeping notes per segment: source export, scene, model, tier, credits and verdict. Because the untouched exports stay on the Mac, any segment can be re-run at a heavier model later, which is cheaper than discovering after a full batch that one scene needed different treatment.

All figures come from the published credit table. At the Standard model 1080p bills 9 credits a second, which is 540 credits a minute and 32,400 an hour of source footage; 720p bills the same, 2K bills 17 a second and 4K bills 33. FlashVSR bills 30 credits a second at 1080p, and SeedVR2, Pro and Ultimate bill 49 credits a second at 1080p.

Against the yearly plans, Creator is $238.80 for 28,000 credits, about 10 hours of 1080p Standard work; Starter is $118.80 for 9,600 credits and Studio is $598.80 for 80,000 credits. The smallest public general pack starts at $29.99 for 1,500 credits when you only want to check one scene.

The comparison to make on a Mac is not against a price tag but against your own time. A local pipeline costs no credits per minute but occupies the machine, needs its own setup and competes for unified memory; the hosted route costs credits and leaves the Mac free. For occasional clips the credits usually win; for a standing workflow that already runs locally, they usually do not.

Three situations argue for local work. The first is confidentiality: if a client contract or a medical, legal or personal-privacy constraint says the footage must not leave your hardware, a hosted upload is the wrong tool regardless of quality, and no amount of convenience changes that. The second is volume: hundreds of minutes a month of routine work is cheaper and more predictable on hardware you already own. The third is integration: if the upscale is one step inside a local editing or compositing pipeline, moving single steps to a website breaks the automation.

This page will not pretend to be an installation guide for that route. The site publishes its own material on local workflows and memory requirements — the ComfyUI workflow, settings and node pages, the VRAM pages for 8, 12 and 16 GB, and the RTX-specific pages — and those describe parameters and trade-offs that apply to the local toolchain rather than to a Mac browser session. Read them for the local route and this page for the hosted one.

A practical pattern is to use both: run the hosted route for the jobs where a quick, supervised test is what you need, and keep a local pipeline for bulk or confidential work. Decide per project rather than per tool, and write the decision down with the project so the next batch does not re-litigate it.

Retina scaling makes previews flattering. A player that scales a 1080p frame to a 4.5K display is resampling it before you see it, and a scaled preview can hide reshaped lettering and invented texture that a 100 percent crop would expose. Set the player to a 1:1 view, or use a viewer that reports the zoom level, and compare the processed frame with the untouched export at the same crop on the same screen.

Then check it at normal size and speed, because delivery is watched in motion. Look for three faults that matter most: faces changing identity, small on-screen text shifting shape between frames, and flat backgrounds that crawl or band. Test on the display you actually deliver for, and if the work goes to a phone screen, look at it there too, because a reconstruction that reads well on a 27-inch Mac panel can look over-smoothed on a small one.

Keep one written line per segment — source, timestamp, model, tier, credits, verdict, reason — and keep the completed output beside the untouched export in the Finder. That record is what turns the next Mac project from an experiment into a decision, and it is also what proves the credits were spent on a result you can defend.