Upscale an image without inventing detail

Upscale takes one image and returns a larger, sharper version of it: more pixels, crisper edges, cleaner texture. It is the last step of a pipeline rather than a rescue for a bad one.

No prompt, no model, nothing to set. One image in, one bigger image back.

The rule that makes it worth having

Iterate small. Enlarge once, at the end.

Every decision you make about a picture, the composition and the light, the wardrobe and the colour, is visible at a size that costs a fraction of the full one. Deciding those at full resolution means paying full price for every attempt at a question the small version already answered.

So the shape of the work is:

1. Generate at the cheap setting until the picture is right.
2. Fix anything structural: cut it out, clean it up, crop it.
3. Upscale, once, when nothing is going to change again.

This is also why upscaling belongs after Remove background rather than before it: a cut-out is decided on the edge between subject and ground, and enlarging first just means the model resolves that edge over four times as many pixels for no better answer.

The same rule, in video

Video is where the arithmetic gets loud, because clips are priced by the second and resolution multiplies that.

The standard move on a model that offers 480p and 720p is to work at 720p and enlarge afterwards if the delivery needs it. A thirty-second take is not a thing you re-roll casually, and paying a resolution premium on every attempt of a shot you have not settled is the most expensive habit available.

Two things it is worth being straight about here:

• This tool takes one image, not a clip. Upscaling a video means extracting the frames, and that is a different job from the one described on this page.
• What it genuinely fixes is the still end of a video pipeline: the first frame you generated, the reference you are about to bind, the poster frame you will publish beside the clip. Those are images, and they are where sharpness is actually looked at.

See the Seedance 2.5 prompting guide for where this sits in a full clip pipeline.

What it does not do

It raises resolution and clarity. It does not invent what was never recorded.

• A face too small to read stays unreadable. Enlarging gives you a larger version of an ambiguity, and the plausible features that appear belong to nobody.
• Text that was mush stays mush. If the letterforms were not resolved in the source, there is nothing to sharpen, so it will produce confident, crisp, incorrect glyphs. Set text at the image stage, at a size that survives.
• Compression artefacts get sharper too. A heavily-saved JPEG has blocking and ringing baked into it, and clarity is applied to those as faithfully as to the subject.
• It is not a restoration. Scratches, tears and colour that has drifted are a different problem with a different tool, Restore photo.

The one-line test: is the detail present but soft, or absent? Soft is what this fixes. Absent is a re-shoot.

When it goes wrong

It came back over-sharp, or plasticky

Skin, fabric and foliage are the three places clarity reads as artificial first, because their real texture is fine and irregular and sharpening regularises it.

Fix: ask whether you needed the resolution at all. An image that is already large enough for its destination gains nothing from this step and can only lose texture to it.

The edges have a faint outline

Halos around high-contrast boundaries are the signature of sharpening applied to an edge that was already crisp.

Fix: upscale the version before whatever else sharpened it. Running two sharpening passes over one picture compounds, and the second one has nothing left to recover.

The file is enormous

That is what the extra pixels are. A lossless format at four times the area is a large file by construction.

Fix: convert at the point of delivery rather than upstream. The full-resolution copy is the one worth keeping in the library; the web-sized derivative is cheap to make again and expensive to have made too early.