--- title: Restoring old and degraded photos — Topaz Labs Wonder 3.5 | Runware Docs url: https://runware.ai/docs/models/topazlabs-wonder-3-5/guides/restoring-photos description: "How to restore old and degraded photos with Topaz Wonder 3.5: faded scans, over-compressed uploads, noisy low-light shots, and tiny legacy files rebuilt as they upscale." --- Real photos rarely arrive clean. A scan has faded and gone soft, an upload has been crushed by compression, a phone shot in bad light is buried in noise, or the only copy of an image is a tiny legacy file. Wonder 3.5 handles all of these the same way: it **clears the damage while it rebuilds detail**, so a degraded source comes back sharp enough to use. The faded print below is one pass: ![A faded, grainy, softly blurred vintage photo of a couple in a garden](https://runware.ai/docs/assets/source-archival.BqcbAprg_sE7jo.jpg) *A soft, colour-faded scan rebuilt at 3× with the grain and haze cleared. Drag to compare.* Restoration is not a separate mode. It is the **standard upscale from [upscaling images](https://runware.ai/docs/models/topazlabs-wonder-3-5/guides/upscaling-images)**, with a damaged image in `inputs.image` instead of a clean one. What changes is the input, not the request. The faces here came back at `medium` enhancement strength, which keeps skin natural on a subject where `high` can over-render, a trade covered in [enhancement strength](https://runware.ai/docs/models/topazlabs-wonder-3-5/guides/enhancement-strength). A crop of the faces shows how much Wonder reconstructs from a soft scan: ![A soft, grainy close-up crop of two faces from the faded photo](https://runware.ai/docs/assets/crop-source-archival.Cgymqo0a_Zahjct.jpg) *Faded scan · 100% crop* ![The same faces restored, with defined features and clean skin](https://runware.ai/docs/assets/crop-output-archival.JlA29yD6_ZNOVMM.jpg) *Restored · 100% crop* > [!NOTE] > Restoration is **generative**, so Wonder rebuilds a plausible version of a damaged region rather than the exact original. On faces from a badly degraded source, features can shift slightly. Feed the best copy you have, and check identity on anything that matters. ### [Cleaning up compressed images](https://runware.ai/docs/models/topazlabs-wonder-3-5/guides/restoring-photos#cleaning-up-compressed-images) Re-saved and re-shared images pick up **blocky JPEG artefacts**, the mosaic of squares that shows up around edges and in flat areas. Wonder reads through the blocking and rebuilds the underlying detail rather than smearing it, which is the difference between a clean result and a blurred one. ![A blocky, heavily compressed photo of a busy street market](https://runware.ai/docs/assets/source-travel.BvztNtXw_2pz7Xg.jpg) *A heavily compressed market photo with the JPEG blocking rebuilt into clean detail.* ![A close crop showing coarse JPEG blocking across the market stall](https://runware.ai/docs/assets/crop-source-travel.CX0HDlFg_Z1Gf7er.jpg) *Compressed · 100% crop* ![The same crop rebuilt into clean produce and fabric detail](https://runware.ai/docs/assets/crop-output-travel.BhhffKiq_Z1kJJXK.jpg) *Restored · 100% crop* Feed the compressed file as it is. Sharpening or denoising it first strips the faint signal Wonder needs to tell an edge from an artefact, so it rebuilds cleaner from the raw compressed image than from a pre-cleaned copy. ### [Rescuing noisy low-light shots](https://runware.ai/docs/models/topazlabs-wonder-3-5/guides/restoring-photos#rescuing-noisy-low-light-shots) Photos shot in poor light carry heavy sensor noise that a plain upscale would only enlarge. Wonder's **noise handling separates grain from real texture**, clearing the speckle while keeping skin and fine edges intact. ![A dark, grainy indoor photo of a woman cooking at a stove](https://runware.ai/docs/assets/source-lowlight.BXGIqLdk_ZWm6PG.jpg) *A grainy low-light kitchen shot with the sensor noise cleared and detail held.* As with compression, resist cleaning the shot first. Wonder does better on the raw noisy frame than on one you have already smoothed, since aggressive denoising takes the fine detail with the grain. ### [Enlarging tiny legacy files](https://runware.ai/docs/models/topazlabs-wonder-3-5/guides/restoring-photos#enlarging-tiny-legacy-files) Old catalogues and archives are full of images that are too small to use: a product thumbnail, a cropped headshot, a low-resolution export that is all that survives. A larger `upscaleFactor` gives Wonder room to rebuild one into something usable. The camera below is a 384 px thumbnail taken to 5×. ![A very small, soft thumbnail of a retro film camera](https://runware.ai/docs/assets/source-camera.DtYGuPyF_Z1BuG5X.jpg) *A 384 px legacy product thumbnail rebuilt at 5× into a usable packshot.* The smaller the source, the more Wonder has to invent, so a tiny file leans harder on generation than a merely soft one. It works well for product and catalogue images where the shape and materials are clear, and it is less reliable where the source is too small to read a face or fine text. ### [Tips](https://runware.ai/docs/models/topazlabs-wonder-3-5/guides/restoring-photos#tips) 1. **Feed the raw degraded file.** Wonder works from whatever signal survives, so pre-cleaning, sharpening, or denoising a photo before sending it usually removes detail it could have rebuilt. 2. **Lower strength for faces.** On restored portraits, `medium` keeps skin natural where the `high` default can over-render damaged features. Push higher only if the face comes back soft. 3. **Check identity on faces.** Generative restoration rebuilds a plausible face, not a guaranteed exact one. Verify anything where the person's likeness matters. 4. **Raise the factor for tiny sources.** A small legacy file needs a larger `upscaleFactor` to reach a usable size, and Wonder has the most latitude to reconstruct when the subject's shape is still clear. 5. **Restore before you edit.** Clean and enlarge first, then colour-grade or retouch the recovered image, so your edits land on real detail rather than on noise and blocking.