Upscaling is interpolation with judgement: the tool studies what is there and guesses what a higher-resolution version would have contained. That sentence already contains the map - it wins on dense organic detail like fabric, foliage and scans, and it loses on anything exact: text, architecture, and the direction a surface normal points.
One word, three jobs
| What you have | The honest move | Why it holds |
|---|---|---|
| Colour map, fabric scan, photograph | Upscale 2x to 4x | Dense organic detail is exactly what upscalers were trained on. |
| A render you can still open | Re-render at a higher resolution | Real samples beat invented pixels, every time. |
| A normal map | Re-bake from an upscaled height map | Direction data breaks the moment an invented pixel nudges it. |
| A render whose scene is gone | 2x upscale, then inspect at 100 percent | The only option left - price it as a rescue, not a workflow. |
| A low-poly mesh | Retopology and displacement | The silhouette is a decision, not an interpolation. |
Renders: denoise is not upscale
The denoiser and the upscaler get confused constantly, so keep them apart. Denoising removes noise that a low-sample render left behind, at the resolution you rendered. Upscaling invents pixels the render never had. One cleans the image; the other guesses at it - and a render made at twice the target size and downscaled still beats both, because every pixel in it started as real image data.
So the order for any final frame: render larger or longer if the scene is alive, denoise to clean what you have, and upscale only when the first two are off the table. When compositing, upscale before the grade and bloom passes - sharpening halos baked in early only grow by the end of the chain.
Textures: the strongest case
Materials are where upscaling genuinely earns its keep. A 2K colour map upscaled to 4K keeps its character, and the memory bill stays sane when you upscale only the maps the camera actually reads. Three rules cover almost every job: upscale colour and roughness, re-bake normals from an upscaled height map, and re-check tiling afterwards.
The tiling check is the one people skip. An upscaler treats the image as a picture, not as a pattern, so the wrap edge can fall out of alignment. The five-minute fix: offset the texture by half its width, patch the seam with the clone brush, then offset back.
Meshes: detail in, decisions out
A mesh has no pixels to upscale, so the question becomes who gets to decide the new detail. Displacement from a real height map is authored detail - you own every bump and can argue with each one. Tools that infer detail are suggesting it; for hero geometry the silhouette is a design decision, and decisions come from humans or from a dedicated retopology pass, not from an inference engine.
The printable test settles most arguments: if invented detail survives a 28mm print, keep it; if it collapses into mush, it was never really there.
The rescue checklist
01 - Name the job. Render, texture or mesh. The tool choice follows from the table above, not from whatever upscaler advertisement found you first.
02 - Keep the original. The upscaled file is a derivative; archiving only it is how detail quietly disappears forever.
03 - Compare at 100 percent. Fitted-to-window previews hide halos, warped edges and smeared texture - the three ways upscaling announces itself.
04 - Up-res only what reads on screen. A roughness map at ten percent strength does not need 8K. Spend the memory where the camera looks.
05 - For print, proof on paper. Screens flatter invented detail; paper does not.
Used this way, upscaling stops being a magic button and becomes a rescue tool with a narrow, well-lit use case - and knowing its edges is what keeps the rest of a pipeline honest. When a mesh and its maps both need help, fix the order first: retopology, then a re-bake, and only then upscale whatever the bake still leaves soft.