Topaz Labs Wonder 3.5

by Topaz LabsAugust 11, 2026

Wonder 3.5 is Topaz Labs' updated generative image enhancement and upscaling model for recovering detail and realism from both high-quality and degraded source images. It builds on Wonder 3 with finer detail generation, fewer repetitive patterns, improved handling of visual noise, and better precision on structured content such as tables, charts, and text-heavy graphics. It is well suited to photo restoration, compressed-image cleanup, creative enlargement, and enhancement workflows that need stronger texture recovery without drifting into overly artificial output.

Upscaling text and structured graphics

How to sharpen text, charts, tables, and packaging with Topaz Wonder 3.5, recovering legible type and clean structured graphics from soft or compressed sources.

Text is where most upscalers fall apart. They treat letters as texture rather than language, so a compressed screenshot comes back with smeared type and numbers that dissolve into mush. Wonder 3.5 was tuned for the opposite, with sharper handling of structured content, so labels, figures, and layout come through the upscale legible. The coffee bag below pairs a soft 640 px source with its 4× rebuild, each with the label zoomed in:

A soft, low-resolution photo of a kraft coffee bag with unreadable small label text — Source

Wonder 3.5

The brand name survives most upscalers because it is large, but the small print under it is the real test. In the zoom, the descriptor lines go from an illegible blur to readable type.

Wonder rebuilds text it can still make out. If the source is so degraded the letters are already gone, it can confidently render the wrong characters rather than the right ones. Feed it type that is soft, not type that is destroyed, and proof-read anything that has to be exact.

Charts and dashboards

Analytics UI is full of the details a generic upscale destroys: thin axis rules, small numbers, and legend labels. Wonder holds the chart's geometry and reconstructs the figures, so a soft dashboard export comes back sharp enough to read the values off. The zoom sits on the bar labels:

A soft, compressed screenshot of an analytics dashboard with unreadable numbers — Source

The same dashboard, sharp, with crisp bars, axis labels, and value figures — Wonder 3.5

The axis numbers and the small value labels on the bars are the parts that decide whether the chart is usable, and they are the first thing a plain resize loses.

Tables and documents

Tables punish an upscaler twice, on the grid lines and on the dense small text inside the cells. Wonder keeps the columns aligned while it rebuilds the type, so a photographed report page or a scanned document stays a table rather than collapsing into a grey field. The zoom lands on the data rows:

A soft, low-resolution photo of a report page with an unreadable data table — Source

The same report page, sharp, with a clean aligned table and readable rows — Wonder 3.5

Posters and layouts

Design work mixes display type that survives anything with small print that does not. The event poster below keeps its bold title through the compression, but the date and venue lines only come back once Wonder rebuilds them, so the zoom sits on the fine print at the foot of the poster:

A soft, compressed event poster with a readable title but unreadable small lines — Source

The same poster, sharp, with clean display type and readable date and venue lines — Wonder 3.5

Layouts also carry hard edges and flat colour fields that a generative pass can smudge. Wonder keeps them crisp, so a rebuilt poster or infographic reads as a clean graphic rather than a softened photo of one.

Where it holds and where it slips

Structured content is Wonder's strong suit, but it has limits worth knowing. It recovers text that is soft but still legible, and it holds chart lines, table grids, and flat colour. It is far less reliable where the source has already lost the information: type too small or too crushed to read can be rebuilt as plausible but wrong lettering, and an exact logo or wordmark is not guaranteed to reconstruct faithfully.

For anything where the copy must be exact, a label price, a legal line, a brand name, treat the upscale as a legibility boost, not a source of truth. Rebuild the layout with Wonder, then confirm the critical strings against the original.

Tips

Upscale soft text, not destroyed text. Wonder rebuilds letters it can still read. If the words are already gone in the source, it may invent the wrong ones, so proof-read exact copy.
Feed the raw file. As with photos, a compressed screenshot rebuilds cleaner from the original than from a pre-sharpened copy that has thrown detail away.
Lean on it for charts and tables. Thin rules, small numbers, and aligned columns are exactly what Wonder 3.5 improved over earlier versions, so structured graphics are a strong fit.
Verify critical strings. Prices, names, and legal text should be checked against the source after the upscale, since legibility is not the same as accuracy.
Keep logos in mind. An exact wordmark or brand mark may not reconstruct perfectly. Composite a clean logo back in when the source was too small to hold it.