TikTok Video Size Ratio: A Practical AI Video Workflow

Start With the Frame, Not the Prompt

Most AI-generated video that underperforms in vertical feeds fails for reasons decided before a single word of prompt is typed. A clip that looks cinematic in a wide desktop preview can arrive on a phone screen cropped, soft, or oddly empty — not because the model was weak, but because the frame was designed for the wrong canvas.

The correction is unglamorous: choose the delivery format first, translate that format into prompt language, then generate. Everything downstream — composition, captions, compression, even the words you put on screen — inherits from that decision.

This guide walks through the whole chain for short-form vertical video: the specs that genuinely change how a clip reads, how to write them into prompts, where interface overlays will land, how to encode without losing detail, and the recurring mistakes that quietly cost watch time.

The Delivery Specs Worth Memorizing

Platform documentation changes, but the working numbers for vertical short-form have been stable for years. Treat these as defaults and verify against current documentation before a large campaign.

Aspect ratio and resolution

Vertical 9:16 is the native format for full-screen mobile feeds. The practical resolutions:

Target	Pixels	When to use it
Standard vertical	1080 × 1920	Safe default, fast to encode
Lower-bandwidth vertical	720 × 1280	Drafts, previews, internal review
High-detail vertical	1440 × 2560	When you plan to crop or punch in later

Square 1:1 and portrait 4:5 still matter for profile grids, thumbnails, and reposts to feed-based platforms. That is why you should compose with a center-safe core: keep the subject and any essential detail inside the middle band of the frame so a square crop still reads as intentional rather than accidental.

Frame rate and motion

24 fps reads cinematic and keeps file sizes modest. 30 fps is the general-purpose middle. 60 fps helps fast action, dance, sport, and rapid camera movement, but it roughly doubles the data and can make subtle generated motion look unnaturally smooth.

A practical rule for AI clips: generate at the frame rate you will deliver at. Converting 24 fps to 60 fps with interpolation tends to produce ghosting around hands, hair, and fabric edges — exactly the details a viewer's eye locks onto on a small screen.

Codec, bitrate, and file size

The upload chain is boring on purpose: MP4 container, H.264 video, AAC audio, 1080 × 1920, 8–15 Mbps for the video track. Platforms re-encode everything you send them, which means every compression pass removes detail. If your file was already compressed twice, the third pass is where skies turn into blocks and gradients band.

So export at a higher bitrate than the platform's own target, and avoid stacking compression. Generate, edit, export once.

Turning Specs Into Prompt Language

Here is where most creators lose the plot. They know they need 9:16 and 1080p, then write a prompt that describes a scene with no framing instruction at all — and are surprised when the model returns something composed for a widescreen.

Build a reusable spec block

Write a short block of framing language you paste in front of every prompt. It does not need to be poetic:

vertical 9:16 frame, subject centered in the middle third, headroom above the eyes, full-body shot, camera at chest height, shallow depth of field, no text overlays

That block does three jobs: it sets the canvas, it constrains where the subject sits, and it reserves visual space where captions and interface elements will land. Reuse it, edit one clause at a time, and you get consistent output instead of lottery results.

Framing language models actually respond to

Broad instructions like "cinematic" do very little for composition. Specific ones do a lot:

• Distance: close-up, medium shot, full body, wide establishing shot.
• Height: camera at eye level, low angle, overhead.
• Placement: centered, offset left, negative space on the right.
• Depth: foreground blur, layered background, flat backdrop.
• Lens feel: 24 mm wide, 50 mm neutral, 85 mm compressed.

Pairs of these give you a usable shot description. "Medium shot, camera at eye level, subject offset left, negative space on the right, soft window light" is a frame you can build a caption into. "Cinematic vertical video" is a wish.

Motion language for narrow frames

Vertical frames are tall and thin, so lateral movement eats screen space fast. Prefer motion that reads inside a rectangle shaped like a door: slow push-ins, subject-led walks toward camera, vertical reveals such as curtains opening or a drone rising, and gentle handheld drift.

If you need a pan, describe its limits — "slow pan from left edge to center, subject stays in frame" — and keep it short. Fast pans in vertical video become a smear, and the platform's thumbnail picker will find the worst possible frame of it. The prompt library has reusable blocks for this kind of framing language if you would rather start from examples than a blank field.

A worked example

Weak prompt: "cinematic vertical video of a barista making coffee, moody lighting."

Working prompt: "vertical 9:16 frame, medium close-up, camera at chest height slightly below eye level, barista centered in the middle third with headroom, hands and portafilter in sharp focus, steam drifting through backlight, shallow depth of field, slow push-in, no text overlays."

Read the second version clause by clause and notice what each one prevents. The aspect ratio stops the model from composing wide. The height keeps the subject's eyes away from the top band. The centered placement protects the square crop. The hand focus gives the small screen a single detail to land on. The closing clause saves you from generated text you cannot spell-check.

Safe Zones, Captions, and Readability After Compression

A vertical frame is not fully yours. Interface elements occupy predictable regions: a caption and username cluster near the bottom, action buttons down the right edge, a top band for tabs and search.

A simple safe-zone discipline:

• Keep faces and key objects in the upper-middle and center of the frame.
• Reserve roughly the bottom quarter for platform text and your own captions.
• Avoid placing critical detail in the rightmost strip.
• Leave margin above the subject's head so the top band does not clip it.

Caption styling also has to survive re-encoding. Thin fonts, low contrast, and subtle gradients are the first casualties. Use bold or heavy weights, strong contrast, and a solid background or stroke behind the text. Test at 50% zoom on a phone in daylight; if it is unreadable there, it is unreadable in the feed.

A Repeatable Production Workflow

This is the loop that keeps a batch consistent rather than producing one lucky hit.

1. Fix the delivery spec. 1080 × 1920, 9:16, 30 fps, H.264, under the platform's file-size ceiling. Write it down once and stop relitigating it.
2. Write the hook first. One sentence describing what the viewer sees and learns in the first two seconds. If the hook depends on a text overlay, note where that overlay will sit.
3. Compose the shot. Choose distance, height, placement, and lens feel from your framing vocabulary. Put the subject in the middle third.
4. Generate variations, not one-offs. Three to five seeds per prompt with small deliberate changes.
5. Judge at phone size. Review on a handset, not a monitor. A clip that only works full-screen on a desktop is not finished.
6. Pick for the crop. The keeper survives a square center crop and a bottom-quarter caption block.
7. Assemble and pace. Cut on motion, keep the first frame busy, and keep runtime tight to the idea rather than to a target length.
8. Add captions and on-screen text. Bold, high contrast, inside the safe zone, one idea per caption.
9. Export once at a bitrate above the platform target, then upload the same file everywhere instead of re-exporting per platform.

When you are producing volume, templates shortcut steps three through six. The video templates page is built around that repeatable structure, and if you are weighing generation approaches across tools, the alternatives overview keeps the comparison focused on output quality rather than marketing claims.

Designing a series, not a single clip

The moment you plan more than one video, consistency becomes the product. Pick three fixed decisions — grade, lens feel, caption style — and let everything else vary. A series reads as a series because the same visual grammar repeats, not because every clip is identical. Keep a notes file with your spec block, your grade settings, and the exact export string you use. Six months from now, that file is worth more than any single prompt.

Keyword and Hook Planning Without Chasing Trends Blindly

Keyword work for short-form video is mostly about matching the words people already use to describe the outcome they want. Three layers work well.

Layer one — the searchable phrase. What would someone type to find this? "Vertical video framing tips," "AI video aspect ratio," "how to make 9:16 clips." Put that phrase in your on-screen text within the first two seconds and in the caption's opening line.

Layer two — the compliance vocabulary. Words that signal the clip is safe and standard: format, size, resolution, export settings, safe zone, caption placement. These are not exciting, and they are precisely what gets saved and rewatched by people solving a practical problem.

Layer three — the emotional hook. The reason to keep watching: a reveal, a transformation, a mistake avoided. Without this layer, the first two read like a manual.

The trap is treating trending audio and hashtags as strategy. A trending sound attached to a badly framed clip does not rescue the clip; it spends the trend on a weak asset. Treat trends as garnish, not as the recipe.

Also keep platform policy in mind when you plan hooks. Claims, before-and-after framing, and health or finance language all carry review risk that has nothing to do with your aspect ratio. A compliant frame does not compensate for a non-compliant claim.

Common Mistakes That Quietly Cost Reach

• Generating widescreen and cropping later. You lose resolution, and every composition decision was made for a different canvas.
• Ignoring the center-safe crop. The clip looks fine vertically and incoherent in the profile grid.
• Over-compressing before upload. Each export pass removes detail you cannot recover.
• Thin captions. Elegant typography that vanishes at phone brightness.
• Filling the whole frame. Dense compositions read as noise at small size; negative space is a feature, not waste.
• Fast lateral motion. Beautiful horizontally, unwatchable vertically.
• Inconsistent specs across a batch. Every clip then needs its own export settings and nothing matches visually.
• Judging on a monitor. The only honest review surface is a phone at arm's length.
• Skipping the first-frame check. A quiet opening frame is a scroll trigger for the wrong reason.

Pre-Upload Checklist

• Aspect ratio 9:16, resolution 1080 × 1920 or higher for punch-ins
• Frame rate matches delivery, with no interpolation artifacts on hands or hair
• Subject inside the middle third; nothing essential in the bottom quarter or right strip
• Captions bold, high contrast, inside safe zones, readable at 50% zoom
• Export once, H.264 MP4, bitrate above the platform target
• First two seconds carry both the visual hook and the searchable phrase
• Square center crop still reads as intentional
• File size under the platform ceiling

FAQ

Does resolution matter if the platform re-encodes everything?

Yes. Uploading 1080p or higher gives the encoder good material to work with. Uploading 540p and hoping the platform fixes it produces soft results no setting can recover.

Is 60 fps better for vertical video?

Only when the motion justifies it. For action, dance, and fast camera moves, yes. For talking-head or slow cinematic shots, 24–30 fps looks better and encodes smaller.

How do I keep a subject centered across different crops?

Compose for the middle third and keep essential detail there. If the frame works as a square center crop and as a full 9:16, you have a durable asset you can reuse across platforms.

Should I add text in the prompt or in editing?

In editing. Text generated inside video is unreliable across models and languages, and you cannot fix a typo without regenerating the clip. Generate clean plates and add text in post.

What breaks first when a vertical clip looks bad on a phone?

Usually contrast. Low-contrast grading that looks elegant on a calibrated monitor turns to mud in bright ambient light. Grade slightly punchier than feels necessary.

How many variations should I generate per shot?

Three to five with small deliberate changes. Beyond that you are browsing rather than directing, and the extra options rarely improve the final cut.

Do I need different files for every platform?

Often not. A clean 1080 × 1920 H.264 export at a generous bitrate travels well across short-form platforms. Only re-export when a platform imposes a hard limit or a different aspect ratio.

Make the Frame the First Decision

Vertical video rewards planning that feels almost boring: one spec, one framing vocabulary, one export, one checklist. Do that consistently and AI-generated clips stop looking like experiments and start looking like a channel with a point of view.

Start with a frame you have already decided on, then generate into it — the Orelon AI video generator is built for exactly that order of operations, from spec to shot to export. If you want to see how the output reads before committing to a style, browse the Orelon blog for breakdowns of framing, pacing, and prompts that survive real platforms.