Sharpen the still you will publish, not every tile in the grid

upscale_image makes an existing still larger and sharper. It does not invent a product that was never in the frame. Cost depends on the model you pick. estimate_cost with that model; check_credits if the batch looks big. The video cousin already has a dedicated rule — upscale after picture lock. Stills deserve the same habit: lock the tile, then spend the enhance meter.

Upscale an image's resolution is the job. Text-to-image is usually how the still was born. Print and hero crops are why you upscale. Browsing a 2×2 is not.

Grids are the trap

A variation sheet is four (or nine) maybe-tiles at a fraction of the parent resolution. Splitting then upscaling every cell is how a contact sheet becomes a credit event. Split if you need separate files. Pick the winner. Upscale that. Borders and gutters from the parent will upscale too if you never trimmed them — another reason not to enhance the whole sheet "to see detail."

People upscale first because a bigger thumbnail feels like a decision. It is a delay. Identity, lighting, and copy-fit are visible at the native size. If they are wrong, sharper is still wrong.

This enhances existing detail. Compression artifacts get sharper too. A smashed JPEG of a miss does not become a campaign by passing through an upscaler.

Order

1. Approve the still (split a grid, pick a cell).
2. Cut out or crop if the layout needs it, on that still.
3. Upscale the file you will actually print, post, or animate.
4. Image-to-video only after the still is the still.

If you will motion the image, upscale the parent you will drive, not a rejected sibling. Motion models do not inherit the discarded upscale.

FAQ

Is this the same meter as video upscale?

Different tool (upscale_image vs upscale_video), same discipline. Do not enhance a file you will replace.

Should I upscale before removing the background?

Only if this still is the still. Cutout-then-upscale or upscale-then-cutout on a keeper is a sequencing choice. Doing either on four rejects is the leak.

Can upscale recover a tiny face?

It cannot recover information that was never there. A 64px face becomes a bigger smear. Generate a proper still.

Why not upscale the whole archive overnight?

Because most of the archive is not shipping tomorrow. Enhance the next hero. The rest can stay native until a destination demands it.