Why Text On Video Suffers More Than Faces

Titles, captions and screen recordings often look noticeably worse than the footage surrounding them. Compression's assumptions about natural images do not hold for lettering.

Encoders assume photographic content

Compression is built around statistics drawn from natural scenes, where brightness varies gradually and sharp discontinuities are rare.

Text is the opposite: hard transitions between two flat values, repeated across a frame. Describing such an edge requires many coefficients rather than few.

The encoder spends heavily and still rounds some of it away. The result is faint ripples radiating from letterforms.

Ringing is most visible against flat backgrounds

Artifacts around edges show up strongly where nothing masks them. White text on a solid color is the worst case.

The same artifacts exist around edges in ordinary footage and are hidden by texture. Nothing is different except visibility.

This is why lower-third graphics can look shimmering and unstable while the person behind them looks fine. The instability is worst on thin strokes, where the ripple is comparable in width to the letterform itself.

Color subsampling damages colored lettering

Most delivery formats store color at reduced resolution because vision is less sensitive to it. On photographic content this is nearly invisible.

Fine colored text is exactly the case where it is not. Edges between saturated colors are reconstructed from fewer samples and become blurred or fringed.

Red text on a dark background is the classic failure, because red carries little luminance contrast and depends heavily on the reduced color channels.

Static graphics are cheap, moving ones are not

Text that stays still between frames costs almost nothing after the first frame. The encoder simply states that nothing changed.

Scrolling or animated text changes across its whole area every frame, and sub-pixel movement means the edges never repeat exactly. Motion matching fails.

Scrolling credits over a moving background are among the most demanding content a general encoder faces, which is why they so often break down. Moving the text in whole-pixel steps rather than fractions gives the encoder repeatable edges and helps considerably.

Where the text is added changes the outcome

Titles rendered at full resolution and encoded once survive best. Every subsequent scaling or re-encoding softens them further.

Text added to already-compressed footage and re-exported gets encoded on top of existing artifacts, and both are then compressed again by the delivery platform.

Larger, heavier lettering with strong contrast against its background survives the chain far better than thin, fine type. The design decision has technical consequences.