Local vs cloud AI video processing: privacy, speed, and cost

Which jobs to run on your own machine

Local AI processing suits repeat batches, NDA footage, and family video, because the file is processed on your machine and no queue or upload sits between you and the result. A cloud service can still make sense for a one-off short. KwaFlux runs on your GPU; license checks and updates use the network. This page compares privacy, speed, and cost.

Where the GPU boundary actually sits

Local means the file is processed on this machine. The shipped modules run on the GPU in the machine: enhancer, repair, cutout, AI Beauty, AI Audio Denoise, converter, and file repair. The clip does not leave to finish a preview or an export. The network is still in the product for a different job. Sign-in, license checks, and updates use it; a 1s/3s/5s preview after download needs no account. Unplug the cable and the model can still run on frames already on disk. A license check for a new export cannot.

The boundary you feel is the card you already bought. A 2016 or later Intel, AMD, or NVIDIA GPU is the floor we state, integrated or discrete, with a recent RTX card as the comfortable recommendation. We do not publish a VRAM figure you can treat as a contract. Integrated graphics can run this class of model. A short preview on a thin laptop is a fair test of that. A long 4K timeline with several modules stacked is a scheduling problem: the same weights, more pixels, more minutes. That pace is the local cost. It is not a hidden quality rank, and this app has no cloud queue you can pay to skip.

• Preview the harshest beat at 1s before you commit a full timeline. Guests can do that with no account
• One module at a time. Cutout, beauty, and an upscale each ask the GPU to touch the frames again
• A weaker card wants the faster tier where a tool offers one. The tier buys time. The pass is still local
• Disk and memory cap the file before any product line does. We do not sell a size tier with no ceiling
• Two shorts a year on a laptop that thermal-throttles can still belong in someone else's queue. That is an honest fit

If you keep producing video, the payoff is the one described in the previous section. The footage isn't uploaded to finish the job, each attempt doesn't wait in a vendor's queue, and the bill is the license on pricing, not a per-minute meter on a card you already own. A datacenter GPU may win a single-frame race, but if you chose local so the master doesn't sit on someone else's server overnight, that's the wrong place for it. Keep the boundary where it really is: on your GPU, at your card's pace, with the network used for the account and the license, not for the frames. One practical check before you blame the model: watch the GPU during a 3-second preview. If the card is busy and the frame arrives late, you're at the limit this section describes. If the card is idle and the preview is still waiting, look at the disk, and at whether another module is queued ahead. Local doesn't mean instant. It means you can see the wait, on hardware you know, instead of a place in a vendor queue you can't inspect.

Ledger 1 — privacy and compliance

• A cloud job starts with an upload. Travel vlogs may not care. Client work under NDA, family faces, and internal screen recordings often do — upload is a veto, not a preference
• Local processing removes the "in transit / stored by a vendor / what is the retention policy" layer. You do not have to parse a privacy page to learn where the master sat overnight
• Most serious desktop enhancers share this shape. It is a desktop-versus-cloud difference, not a KwaFlux-only claim

Ledger 2 — what "speed" actually includes

• Cloud wall-clock is upload + queue + process + download. A multi-gigabyte 4K master on a home uplink is minutes even when the remote GPU is fast
• The expensive loop is iterate: tweak, preview, tweak again. Locally that is a 1s/3s/5s preview after download, with no account. In the cloud each loop often re-enters the transfer queue
• Local throughput is capped by your GPU. An older card is slower — that is honest — but it is predictable, parallel, and not waiting in someone else's line. We do not publish a VRAM floor here; see download for the current hardware note

Ledger 3 — how you pay

• Cloud enhancers usually meter minutes or credits. More footage means more spend, month after month
• A local tool is a license plus hardware you already own plus electricity. Ten times the volume does not ten-times the software bill
• KwaFlux lists monthly $35.99, yearly $99.99, or a Perpetual License at $199.99 on pricing. There is no per-minute meter inside the desktop suite. We are not quoting unnamed cloud rate cards here; those go stale

When the cloud is still the right tool

• A GPU from 2016 or later can run the model, Intel, AMD, or NVIDIA, integrated or discrete. Integrated graphics are slower on long clips and 4K. Older machines are the cloud case. Mac is Apple Silicon.
• Two shorts a year: a metered bill can be cheaper than any annual license
• Phone-only workflows: there is no realistic heavy local AI path on a handset today
• A one-shot job that truly needs more compute than a consumer GPU — film-out reconstruction, not a weekend vlog

How to test the local side tonight

1. Install from download. Windows 10 64-bit 1809+ / Windows 11, and macOS 12+ Apple Silicon, are shipping
2. Import a real clip into the AI Video Enhancer (or the module you actually need) and run 1s/3s/5s previews — guests can do this
3. Export only if a paid plan is worth it for your volume. Signing in is required to export, convert, repair a file, or download a file from a link, not for those previews