1.3 Hours per 10 Seconds: What Topaz Starlight Really Costs to Upscale Your Archive

Topaz's Starlight diffusion models can do remarkable things with 320x240 archive footage. The cost is time and money, and those numbers get far less attention than the before-and-after clips. Our position is to use Starlight on the hero shots only and finish everything else with a cheaper upscaler. Run the numbers below before you trust Adobe's ownership to keep the economics as they are.

The measured numbers

The best public benchmark we have found is Nick Lear's in-depth review for ProVideo Coalition, published on 8 July 2026. He tested Topaz Video 1.6.1 on an RTX 4080, and these are his times for 10 seconds of 30 fps footage:

• Starlight Precise 2.5: 0.06 fps, about 1.3 hours
• Starlight Mini: 0.04 fps, about 2.0 hours
• Starlight HQ: 0.06 fps, about 1.4 hours
• Starlight Fast 2: 0.11 fps, about 0.8 hours
• Starlight Sharp: 0.41 fps, about 0.2 hours

For cloud rendering, Lear reports that a five-minute HD-to-UHD upscale with Starlight Precise 2.5 uses 750 credits. On a 9,000-credit top-up he works that out at about $167 for one export. One thing doesn't match: Topaz's product page lists extra cloud credits at $0.10 each, which would make the same export about $75. We couldn't work out which price applies to which plan or pack, so we give both. Topaz's pricing page lists Topaz Video at $299 a year. Topaz Studio comes with 300 video cloud credits a month and Studio Pro with 600.

Lear is fair to the newer models but not uncritical. On one of his test shots the generative Astra model "completely made up the statue on the clock", although he adds that you don't notice it at normal playback speed. He also found that the older Proteus model "again shows it can compete".

Worked example: a 60-minute archival documentary

The following is our own arithmetic, based on Lear's figures. Treat it as an order-of-magnitude guide, not a quote. Lear's local timings don't give the source resolution, and SD sources or a different GPU will give different numbers.

• Everything through Starlight Precise, locally: 60 minutes at 30 fps is 108,000 frames. At 0.06 fps that is about 1.8 million seconds, or roughly 500 hours: three weeks of one GPU running non-stop. At 25 fps it is about 417 hours. Either way, one workstation is out of action for most of a month.
• Everything through Starlight Precise, in the cloud: at 150 credits a minute, 60 minutes is 9,000 credits. That comes to about $2,000 at the rate implied in Lear's review, or about $900 at $0.10 a credit. This assumes the HD-to-UHD credit rate in his test applies to your source, which you should check with a short clip first.
• Selective approach: pick the three minutes that carry the film, such as the opening archive, the faces the story depends on and the shot the poster comes from. In the cloud that is 450 credits, within one month of a Studio Pro allowance. Locally it is about 23 hours, which is one long weekend. Put everything else through Resolve Studio's Super Scale, which Blackmagic describes as "high quality image uprezzing" from HD to 4K and 8K and which comes with a $295 perpetual licence. Alternatively, use Proteus, which Lear still considers competitive.

The selective approach costs a few per cent of the full-timeline price. It also keeps generative output to the shots where you have time to check it frame by frame.

Where invented detail shows up

Diffusion upscalers add detail that looks plausible, and in an archive documentary plausible detail can be a factual problem. On every shot that goes through Starlight, check these areas at 200%, frame by frame:

• Faces: eyes, teeth and ears can drift towards a generic face. That matters if the subject is a real, identifiable person.
• Text: signs, newspaper headlines, number plates and uniforms. A model that turns blurred letters into crisp wrong letters has changed the historical record.
• Repeating texture: foliage, brick, crowds and fabric, where invented detail can shimmer from frame to frame.
• Distinctive objects: Lear's clock statue is a good example. Anything a viewer might recognise or research should match the source.

Keep the un-upscaled original in the project, write down which shots were generatively enhanced, and tell the client or broadcaster. Some archive licensors and compliance teams will ask.

What Adobe ownership changes, and what it doesn't yet

Adobe announced the deal on 25 June 2026 and did not disclose the price. Camera Jabber reports that it closed on 23 September. Topaz stays a standalone brand and its apps are still sold separately. Adobe says image upscaling will reach Firefly and Photoshop first, followed later by video upscaling, archival restoration and low-light enhancement across Creative Cloud. PhotographyTalk reports a price of about $340 million, which it attributes to Adobe's quarterly SEC filing (unverified: we have not checked the filing ourselves).

As of 2 October, Adobe has announced no changes to pricing or licensing for existing Topaz customers. Our reading is that this will not stay the case indefinitely. Once video restoration is part of Creative Cloud, Starlight-class rendering could start to be paid in Adobe's generative credits rather than Topaz's. That is speculation, but it affects how you plan:

1. Budget on today's prices and put a date on the estimate.
2. Get large archive jobs rendered now if the budget is already approved, rather than waiting for an integration that doesn't exist yet.
3. Keep a working local install and a note of your licence terms, so a pricing change doesn't stop a project halfway through.

The decision rule

Starlight is worth it when the shot is very low resolution, central to the story and short. If footage is already SD or better, plays for a few seconds, or sits under graphics, use Super Scale or Proteus and spend the GPU time elsewhere. Run a ten-second test on your own hardware before you quote, because every number above changes with your source and your GPU.