Lower-Third Contrast on Busy Generated Plates

A name super that looks fine on a paused frame of a generated plate will often fail three frames later. Bokeh discs drift through the letters. Neon edges pulse. Speculars on wet streets slide under the type. Contrast for a lower third is a hold-duration problem, not a type-specimen problem. If you only checked the still, you have not checked the super.

This is not a recap of caption contrast, reading speed, or line breaks. Those belong to a transcript track that changes every second. A name super is a short, deliberate graphic: a person, a title, sometimes a company, sitting in the lower third for two to five seconds while the plate keeps moving. The job is to keep that graphic readable on the worst frame of that hold, on footage a model made busy on purpose.

Why generated plates eat name supers

Live-action plates give you some control over the background. You light a darker band, you flag a window, you put the subject against a wall. Generated plates do the opposite by default. Prompts that ask for cinematic bokeh, night-city neon, rain, particles, or shallow depth of field are asking for high-luminance events scattered across the frame. Those events do not stay put. A bokeh highlight that missed the type at 0:04 is inside the bowl of the "O" at 0:05.

WCAG 2.2 Success Criterion 1.4.3 sets the floor for text contrast: 4.5:1 for normal text, 3:1 for large-scale text (18 point, or 14 point bold). A name super is often large enough to use the 3:1 floor. That does not rescue you. The ratio is measured against the pixels behind the glyphs, and those pixels change every frame. A 6:1 ratio on a dark coat becomes 1.8:1 the moment a cyan sign reflection crosses the surname. WCAG's own note is that 4.499:1 does not pass 4.5:1. Averaging the hold and calling it fine is the same kind of rounding.

Two more traps that are specific to generated video:

• Local contrast, not average contrast. A plate can be dark overall and still put a bright streak through two letters. Measure the actual background of the name, not the frame histogram.
• Color-only type. White on a pale bokeh ball, or brand-accent cyan on a neon plate, fails even when the designer "knows" the type is high-key. Color is not a contrast strategy on footage that already used that hue as a light source.

Safe-zone placement is a separate failure. The lower third is also where platform chrome sits. A super that holds contrast and then disappears under a username is not a contrast win. Check current platform templates and keep the name clear of that furniture, as covered in where on-screen text survives platform UI. Contrast and placement both have to pass.

Three treatments, not one style

When a super fails on a busy plate, teams usually recast the type: bigger, bolder, all caps. That is the wrong first move. Bigger type on a moving highlight is just a larger region of failure. Run three treatments on the same clip and the same copy before you touch the wording.

Treatment	What you actually change	Holds when	Breaks when
Stroke	Fill plus a dark outline (and a soft shadow if you have it)	Moderate motion, mixed midtones, no neon through the letters	Saturated highlights, rain, particles, thin weights
Scrim	Semi-opaque or solid block behind the name and title	Any plate, including neon and bokeh	Almost never on contrast; only if the block covers a face or a product
Plate	The footage under the type: darker band, calmer lower third, or a reframe	You still control generation, or you can recrop	Locked live-action you cannot reshoot; a plate that must stay neon for the story

Stroke is the caption-preset construction: fill, outlineColor, outlineWidth. Versely's caption styles expose those fields directly. It is the lightest visual change and the weakest defence. Use it as the control in the test, not as the default for night-city plates.

Scrim is the reliable one. Versely's text overlay path (add_video_captions) takes an optional background block. The agent's own example for a hook line is white bold text with a dark background, which is the same construction a name super needs. You give up some of the plate in a strip along the bottom. You gain a background whose luminance you chose, so WCAG 1.4.3 becomes a one-time check instead of a per-frame gamble. Pick the scrim from the plate rather than from the brand deck: a near-black at 70–85% opacity usually holds, a brand red on a red neon street does not. The in-browser color palette extractor runs on your machine and never calls a model, which is the right way to steal a dark from a frame without inventing one.

Plate is the generated-video option live-action editors do not have. If you are prompting the shot, ask for a calmer lower third: subject upper-centre, dark clothing or a dim floor, no practicals in the bottom band. Depth of field can help if the blur sits behind the type, not as bright discs in it; that is a prompting choice, not a grade. Depth of field and rack focus in prompts is the lever. If the clip already exists, a reframe that lifts the subject and leaves a darker band is still a plate treatment. Regenerating the whole shot to save a super is justified when stroke and scrim both look wrong on a hero interview.

Do not bake the name into the generation prompt. Baked type cannot be restyled, retimed, or A/B tested, and it will still lose a contrast fight with neon. Keep the super as an overlay.

The three-treatment test

Run this on staging footage, not in your head. Use the editor's 480p preview (edit_video with preview: true). That pass is free and is gated by a short per-user cooldown, so space the three renders instead of firing them back to back. The final export is charged once. Free 480p previews exist so you can do this kind of check without paying for three full-resolution versions of a two-line graphic.

1. Lock the copy and the hold. Name on line one, title on line two. Two to five seconds, timed to the shot where the person is actually on screen. Do not test on a title card.
2. Render treatment A, stroke only. White or near-white fill, dark outline, no background block, position bottom. Pull the typeface from the font registry rather than a random default. Thin display faces fail this treatment first; that is useful information.
3. Render treatment B, scrim. Same type, same position, dark background block behind both lines. Keep the block tight to the text. A full-width bar is easier to see and harder to defend in a design review; a rounded card behind the two lines is usually enough.
4. Render treatment C, plate. Either a regenerate with a calmer lower third, or a reframe of the same clip so the type sits on darker pixels. Keep the type from A or B, not both, so you can tell whether the plate did the work.
5. Watch all three at 1x on a phone. Not paused, not on a calibrated monitor. Scrub only to mark the worst frame of the hold. The super that survives that frame is the one you ship. If A fails and B holds, you do not owe the plate a regenerate. If B covers a face, C is the real fix.
6. Check placement against captions and chrome. If a caption track also sits bottom, the name super and the transcript will collide. Move captions, or time the super only while the person is identified and the transcript is elsewhere. The overlay tool itself is add a text overlay for a line that holds the whole clip, or timed text overlays when the name should appear and leave. Position is top, center, or bottom: there is no free-form pixel placement, so "a bit higher" is not a setting.

A pass is not "I can read it if I already know the name." A pass is a colleague who does not know the speaker reading both lines on the worst frame, on a phone, without pausing.

What the test is not deciding

This test does not pick your caption style, your hook line, or your brand typeface. It answers one question: on this generated plate, which of stroke, scrim, or plate choice keeps the name super legal and readable for the whole hold.

If all three fail, the copy is too long or the hold is too short, and that is a different craft. Text overlays and hierarchy on video covers word count and timing. Do not solve a six-word title on a two-second neon whip-pan by adding a thicker stroke.

If the super is actually a title card (centered, designed as an intro, no moving face underneath), use a background block on purpose and stop pretending it is a lower third. Adding a title card is that construction.

Ship the treatment that held. Save it. The next generated interview on a night-city plate should not restart from white type and hope.

FAQ

Does a 3:1 ratio count if the name is large?

WCAG 2.2 SC 1.4.3 allows 3:1 for large-scale text. Measure it on the pixels behind the letters on the brightest frame of the hold, not on a design-file swatch. A large super that drops below 3:1 when a bokeh disc crosses it still fails.

Why not just use the caption preset I already like?

Caption presets are built for a transcript that moves. A name super is a graphic with a job: identify a person. You can borrow a preset's outline or background, and the caption styles page is the right place to see those fields. You still need the three-treatment test, because a style that survives talking-head midtones can vanish on generated neon.

Can I skip the scrim if I regenerate the plate?

Yes, if treatment C holds on the phone check. Plate choice is the cleanest look when you still control the prompt. If the shot is already locked, or the neon is the story, scrim is the honest fix. Stroke-only is the control, not the hero, on busy generated plates.

Do I need a new export for each treatment?

Use the free 480p preview pass with preview: true. It has a short per-user cooldown, so do not slam three requests at once. Pay for the final export once, on the winner.