Upscale a video with AI, and decide what it may add.
Some footage is fine and simply small. Some genuinely lacks the detail. Those want different models, and the difference is measurable — so this page measured it instead of describing it.
A 854 × 480 clip taken to 1080p by the most faithful model in the catalogue. Switch back and watch the grain return to the sign and the wrenches.
Both play here at 1080p, which is as far as a web player is worth pushing — the 4K renditions are larger than this page should serve. Measured against the source, this is the pass that moved least.
Upscaling is not a quality dial.
The model that added least was also the sharpest.
This is the faithful tier, and the hero toggle above is the test: switch to the 480p source and the grain comes back to the sign lettering and the wrenches. What the measurement adds is the part you cannot see by looking — against the source it moved about a third as far as the most expensive tier in the catalogue did, and between consecutive frames it moved less than an ordinary resize does. It is the cheapest path to 1080p here. Going in we assumed a generative tier would win this. It did not.
And sometimes you do want it to add.
This is the diffusion tier on the same 480p clip, and it is the honest reason to pay more: it resolves individual curls, skin texture and eyelashes that the source never carried, and it stays the same person. Measured, it strayed no further from the original than the faithful pass did. What it costs shows up somewhere else — the detail it invents is not quite the same detail on every frame, so it moves more between frames than doing nothing at all.
Check what comes back, not just how it looks.
All four models on this page were asked for the same thing — 1080p — and returned three different frame sizes and two different frame rates. This clip is one of them: 1922 × 1080 at 24fps, resampled from a 30fps source that nobody asked to be retimed. Nothing about it looks wrong until it lands in a timeline next to the footage it came from. It is the kind of thing worth knowing before a batch runs, which is why it is on the page rather than in a support article.
Upscale it — or render it large to begin with.
Built for visual inference.
Serving video is a different problem from serving text — a single request can saturate a GPU, and none of the tricks that made language models cheap apply. Hedra's engine was built for exactly that workload.
Whichever model you choose — ours or anyone else's — it runs on the same infrastructure, behind one key. Which is what makes the honest workflow affordable: generate a clip at 720p with one vendor's model, finish it with another vendor's upscaler in the same request chain, and compare two tiers on the same second of footage before committing a batch.
Generate with API or Agent
Build with the API
One key, every tier. Point it at a back catalogue and set the model per clip.
Upscale a clip now.
No code — drop in the footage, pick a tier, compare it against the source before you keep it.
FAQs
If the footage has the detail and is simply small or compressed, Topaz Proteus: when we measured it on a 480p clip it diverged least from the source, came back steadier frame to frame than the source itself, and it is the cheapest tier. If the footage genuinely lacks detail — old, soft or AI-generated — Starlight is built for that and added real detail without losing the face. Compare both on one clip before committing a batch; that costs a few seconds of footage.
Video models →
