Upscale an image with AI, without inventing it.
Take a small, noisy, over-compressed photograph to 4K — and choose how much the model is allowed to add while it does. For archives, print, catalogues, and the tools inside your own product.
A 1024 × 768 snapshot with flash, noise and JPEG blocking, taken to 4K. Switch back to the original to see how little there was to work from.
This one has face recovery on, which is the largest single change an upscaler makes to a person. Measured against the source it moves the face further than the choice of model does — so it is a decision, not a default.
Recovering detail and inventing it are different jobs.
More pixels is the easy part.
This is the same photograph at 4K with nothing reconstructed: the colour noise is gone, the JPEG blocking is gone, the edges are clean. What has not happened is a miracle — her eyes are still soft, because they were soft in the original and no honest upscale can decide what they looked like. For an archive, a legal exhibit or a product shot, that restraint is the feature.
Then you decide how much to add.
Three faces, one upscale, three settings: recovery off, recovery on reconstructing, recovery on reimagining. Measured against the original, turning recovery on at all moves the face more than either later dial does — and the two recovered versions differ from each other about as much as either differs from the source. The dial is not a fidelity slider. It picks a different face.
When the whole frame has to hold up.
Faces are not the only thing in a photograph. The generative tier improves everything at once — the cabinet hardware, the weave of the shirt, the pendant on the chain — and when we measured it, it stayed closer to the source than a face-recovered pass did. It costs three times as much per megapixel, which is the right trade when the print is large and the background is not allowed to turn to mush.
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. Upscaling is priced by megapixels of output rather than by the second, so a back catalogue costs what it renders, and switching between the faithful tier and the generative one is a parameter rather than a migration.
Generate with API or Agent
Build with the API
One key, both tiers. Point it at an archive and set the recovery flags per image.
Upscale one now.
No code — drop in the photo, pick the tier, compare the two before you commit.
FAQs
If fidelity is the point — archives, evidence, product detail — use Gigapixel with face recovery off: it diverged least from the source when we measured it, and it is the cheaper tier. If a person needs to look good, turn recovery on and know that this is the biggest single change you are making. If the whole frame matters, especially for large print, use Wonder.
Image models →