Upscaling: restoring detail versus inventing it

The difference

Restorative processing stays within what the original pixels mathematically imply. Enlarged four times, it does not invent pores or letters that were not captured. Generative processing draws from learned priors — 'there should be detail like this here'. For a poster, the second looks better. For a news photograph, the second is unusable.

Where generative upscaling must not be used

Generative upscaling makes things appear that were not in the frame: facial detail, text on a sign, digits on a plate — replaced by something plausible and different. When that happens the image stops being a record.

• News and journalistic photography
• Evidence for accidents, insurance or litigation
• Product photography showing the actual item
• Medical and diagnostic imaging
• Photographs of people used for identification

Product photography is the one people miss. A generatively enhanced texture makes the delivered item look different from the listing, which in many jurisdictions is a misleading-representation problem. For e-commerce images, stay restorative.

Topaz products and pricing

Topaz Labs is the main restorative family, and it runs on the desktop. Products are split by use, with a bundle available.

• Topaz Photo: $199/yr — denoise, sharpen and enlarge photographs
• Topaz Gigapixel: $149/yr — enlargement specifically
• Topaz Video: $299/yr — video enhancement, enlargement, stabilisation
• Topaz Studio: $399/yr — the bundle, with 300 monthly cloud video credits
• Topaz for Web: $109/yr images, $199/yr video — browser versions

Running locally is the practical advantage: material is never uploaded, so the tools can be used on work under NDA. The cost is hardware — processing 4K video takes several times longer without a capable GPU.

A decision procedure

• 1. Decide whether the image is a record or an expression
• 2. If it is a record, use restorative processing only — generative is off the table
• 3. If it is an expression, generative is fine, but treat faces carefully
• 4. Tell the client that AI processing was applied

The fourth is routinely skipped and routinely becomes a problem later. Being unable to answer 'is this part generated?' after delivery is a bad position. Keep the unprocessed source and a record of what was applied.

Frequently asked questions