Subtitles are not a decoration on short-form video

Every platform will auto-caption for you, at no cost and no effort. It is still the wrong choice for short-form, for four reasons that compound.

• You do not control whether they are on. Platform captions are a viewer setting on some surfaces and off by default on others. A video whose meaning depends on text the viewer has switched off is a video that fails silently for a large share of its audience.
• You do not control the timing. Auto-captioning groups words into lines and displays them in blocks. That is fine for accessibility and poor for pacing, because the emphasis lands on a line boundary rather than on a word.
• You do not control the position. Platform captions sit where the platform puts them, which is inside the region its own interface occupies. You cannot move them out of the way of a caption block you did not place.
• They do not survive cross-posting. Each platform generates its own from scratch, so a video published to three destinations gets three different caption treatments with three different error patterns. Burned-in text is identical everywhere because it is part of the frame.

Line-level captions show a phrase for its whole duration. Word-level captions reveal or highlight each word as it is spoken, which is the pop-in and karaoke styling you see on high-performing short-form.

The difference in effect is larger than it sounds. A revealed word pulls the eye and re-anchors attention at the exact moment the audio delivers it, so the text stops being a transcript and starts being pacing. It is one of the few purely presentational choices that measurably moves retention.

The reason many tools do not offer it is a pipeline ordering problem. To time a word you need to know when it was spoken, and a script does not contain that information. Synthetic narration does not reliably return per-word timings either, and the same sentence read at a different speed produces entirely different boundaries. So the timings have to be recovered from the audio after it exists, by transcribing the generated narration back with a speech recognition model and reading the word boundaries off that transcription.

That is an extra generation pass per video, which costs time and money on every single render. Tools that skip it ship line-level captions and describe them as subtitles, which is true and not the same thing. AutoVidGen runs the narration back through a transcription pass for exactly this, which is why the highlight lands on the word rather than the line.

It is worth noticing what that makes subtitles, strategically. They are one of the very few levers in short-form that sit entirely inside your control and are not subject to a feed's opinion. The algorithm decides who is shown the video. You decide whether the people it shows can read it, and on a series that decision is made once at setup and then holds for every episode you publish afterwards.

The ordering in full, including why narration has to exist before subtitles can be timed: where transcription sits in the render pipeline

Higher than film convention suggests, and this is the single most common mistake in vertical video.

The lower third is correct for landscape because nothing else is there. In vertical, the bottom of the frame is where every platform stacks its own caption text, account name and controls, and TikTok occupies roughly the bottom 20 percent. Subtitles placed at the traditional lower third land underneath all of it.

The safe region on a 1080 by 1920 frame is the central band: below the top 15 percent, above the bottom 20 percent, and clear of the right 15 percent where the interaction buttons sit. Practically that puts subtitles somewhere around the vertical middle, often slightly below centre, which looks wrong in an editor and correct on a phone.

Legibility first, personality second, and the ordering is not negotiable because unreadable text is worth less than no text.

• Heavy weight, generous size. Short-form is watched on a phone at arm's length. Text that looks oversized on a monitor is usually about right on the device.
• A hard outline or a solid backing. Generated imagery changes brightness constantly between and within scenes. White text with no outline will be invisible over some frames in almost every video.
• One highlight colour, used consistently. Highlighting the key word in a single accent colour is effective. Highlighting several words in several colours is noise, and it stops reading as emphasis.
• Two lines maximum, three or four words each. Long lines force horizontal eye movement, which is what breaks reading at speed. Short-form subtitles want to be read in one fixation.
• Never all-caps for the whole video. It scans as shouting, it is measurably slower to read, and it removes the ability to emphasise anything, since everything is already at maximum.

At the point where someone relies on them. Subtitles are, before anything else, how deaf and hard-of-hearing viewers get access to your video, and treating them purely as a retention lever leads to decisions that look good and exclude people.

The practical version is a short list. Keep contrast genuinely high rather than fashionably low. Do not animate individual words so aggressively that the sentence never sits still long enough to be read. Do not compress the text to fit a design. And if you burn subtitles in, consider still supplying a caption track where the platform allows one, because burned-in text cannot be resized, restyled or read by assistive technology.

There is a commercial version of this argument as well, if it helps: the styling that makes text readable for someone with low vision is the same styling that makes it readable for someone glancing at a phone in bright sunlight. The two goals point the same direction almost all of the time.