AI Upscaling: When It Recovers Detail and When It Invents It

Traditional resizing interpolates. It looks at neighbouring pixels and averages them, which is why a small photo scaled up the old way looks blurry rather than wrong. AI upscalers do something different: they were trained on many pairs of small and large images, so they have learned what edges, skin, fabric and foliage tend to look like at high magnification. Given a soft face, the model does not average, it proposes. That is the whole trade, and it explains nearly every result you will see.

When upscaling helps

Upscaling earns its keep when the information is present but under-sampled. A slightly soft photo from an older camera, a screenshot that was compressed on the way to you, a print scanned at a modest resolution, a frame pulled from a video: in each case the structure is real and the file simply does not have enough pixels to describe it cleanly. The model is being asked to render an edge it can already see, and that is a task it does well.

It also helps before any generative step. Every image model, including image-to-video, reads your file at a fixed size and works from whatever detail is there. A clean, larger input gives it an unambiguous subject, which is the single biggest factor in whether a face stays stable across a clip.

When it invents detail

The rule underneath the table: upscaling restores what the file almost contains and fabricates what it does not. A thumbnail with two hundred pixels of face has almost no real structure, so the model supplies a face from its training. It will look convincing. It will not be your subject. This is the failure that matters most in practice, because it is invisible in the output and only obvious when someone recognises the person.

Image upscaling versus video upscaling

Still images and video are different jobs. A still is one frame, so the model has time and no consistency problem. Video adds a second requirement: every frame has to agree with the frame before it, or the picture boils. That is why video upscalers are more conservative, and why they are usually better at removing compression and noise than at inventing fine texture.

In the studio the crystal upscaler handles stills from 40 credits per job, priced by output size. Video upscaling starts at 75 credits, and the pro tier at 300, both multiplied by the output settings you pick. On a long clip, treat the first pass as a test: upscale a short section, watch it at normal speed, and only then commit to the whole thing.

Upscale first, then animate

If the destination is an image-to-video clip, order matters. Animating a soft photo and then upscaling the video means paying a video model to work from bad input and inventing movement on top of invented detail. Upscaling the still first means the video model sees a clean subject, and the resulting clip is usually more stable for the same number of credits.

A workable order for a portrait: upscale the still, check it at normal size, then animate with a small motion and a static camera. If the clip still drifts, the cause is nearly always motion strength rather than resolution, so reduce the movement before you spend credits on another upscale pass. Keep the original file untouched throughout, because every judgement you make should be against the photo as it was, not against the last version you generated.

Print sizing, honestly

Printers want roughly 300 pixels per inch at normal viewing distance. An 8 by 10 inch print therefore wants about 2400 by 3000 pixels, and a 16 by 20 inch print wants about twice that in each direction. Upscaling can produce those numbers from a smaller file, but producing the pixels is not the same as having the detail. A print made from a heavily upscaled thumbnail is a print of an invented image at large size, and it will look soft or strange the moment someone stands close to it.

Failure modes worth recognising

• Over-sharpening: halos along every edge and a brittle, crunchy look. Usually the result of pushing the strength slider to maximum.
• Plastic skin: pores and fine lines erased into a smooth surface. A sign the model decided skin should be clean rather than textured.
• Hallucinated text: signs, labels and documents come back with letterforms that are nearly right and completely wrong. Never upscale a document you intend to read.
• Crowded detail: foliage, gravel and fabric weave turn into repeating patterns because the model tiles what it learned.
• Identity drift: a face that looks better but is no longer quite the person. Compare against the original, never against the previous upscale.

The practical habit that prevents most of this is comparison. Keep the original open beside the result, look at both at normal viewing size, and judge the upscale by whether it looks like the same photograph rather than by whether it looks sharper. Sharper is cheap. Same is the thing you actually wanted.