The pipeline everyone gets backwards
Almost every guide shows upscaling as step one. Load image, pick 4x, done. Then people wonder why the result is a 4K file that still looks soft.
The problem is that upscaling is arithmetic, not restoration. It interpolates values that already exist. When a frame is noisy, those existing values are wrong -- and a 4x upscale does not average the error away, it amplifies it across sixteen times as many pixels.
What denoising actually does
Denoising is the step where you decide which detail is real. Grain from a high ISO shot, compression blocking from a phone video, sensor noise in a dark scene -- none of that is content. A good denoise pass removes it while protecting edges that matter.
Only after that does interpolation have clean material to work with.
Why the order cannot be swapped
Denoise then upscale | Clean edges, no amplified noise
Upscale then denoise | Noise is now large-scale structure; removing it destroys detail
Once noise has been scaled up it stops behaving like noise. It becomes texture. A denoiser cannot tell it apart from real detail, so it either leaves it or eats the detail with it.
Practical order of operations
- Inspect the source at 100% and identify the dominant defect
- Denoise with the mildest setting that removes the defect
- Only then upscale
- Sharpen last, lightly -- sharpening hides mistakes, it does not fix them
Where this matters most
Old family photos, scanned slides, phone footage from compact sensors, screen recordings that were compressed twice. These are the cases where the order is the whole difference between a usable file and a painted one.
UpgoLAB runs the pipeline in that order in the browser, so the denoise stage happens before any interpolation. It is a small thing that decides whether the output is worth keeping.