AI Viral Clip Generator – Long Video to Shorts

Turn one long video into several distinct vertical clips for TikTok, Instagram Reels, YouTube Shorts, or another vertical-video workflow.

The Actor transcribes the audio, selects self-contained passages, prevents overlapping or duplicate clips, cuts on speech boundaries, reframes each clip to 9:16, optionally translates and publishes subtitles, validates the rendered files, and returns the videos with metadata.

Important: This is an automatic first edit, not a human editorial or linguistic review. Watch every delivered clip before publishing it.

Which Actor should I use?

• Create translated subtitle files from audio or video, without rendering a final video: Video & Audio to Translated Subtitles
• Check or conservatively correct an existing SRT or VTT file: Professional Subtitle QC
• Run the complete transcription, translation, subtitle QC, and optional subtitled-video workflow: Professional Subtitle Translator & Video Captioner
• Convert an existing subtitle file between SRT, VTT, ASS, and JSON, or adjust its timing: Subtitle Converter & Resync
• Select several distinct moments from a long video and turn them into vertical short videos: use this Actor.

Each Actor covers a different stage of the media and subtitle workflow, so you only pay for the processing you actually need.

1. Read this before your first run

Question	Answer
What source can I provide?	One uploaded MP4, MOV, MKV, or WebM, or one direct URL to such a file.
Can I paste a YouTube, TikTok, or Instagram page?	No. Platform page URLs are refused. Upload a file you own or provide a direct media-file URL you control.
Are the requested clip count and length guaranteed?	No. The Actor returns fewer clips when the source does not contain enough distinct, complete passages.
Does Follow the main face remove camera shake?	No. It stabilises the virtual crop around a subject; it does not stabilise shake already present in the source footage.
Are subtitles compulsory?	No. Burned subtitles, SRT files, and VTT files are three independent options.
Can it translate?	Yes, into ten validated output languages. Translation is automatic and unreviewed.
What format is rendered?	Vertical 9:16 only, at 720 × 1280 or 1080 × 1920.
Does it publish to social networks?	No. It produces files and metadata; you review and publish them yourself.

The Actor needs a video and an audible speech track. Weak, distorted, overlapping, or music-covered speech can reduce transcription, selection, and active-speaker accuracy.

2. Quick start in the Apify Console

1. Open the Actor's Input tab.
2. Under Video file, click Upload new files and select one video. Alternatively, enter one direct media URL under Direct video URL. Do not fill both fields.
3. Tick I own this content or am authorised to transform it.
4. For a first run, use:
   • 3 clips;
   • 30–60 seconds;
   • 9:16 at 720 × 1280;
   • Follow the main face;
   • Centre crop if no reliable face is found;
   • original language;
   • maxCostUsd: 5.
5. Choose the subtitle result you want using the recipes below, then run.

The restorable example intentionally contains no media. Add your own video before pressing Start. Starting with no video performs a free readiness check and therefore produces no clips.

3. Choose the exact subtitle result

These options are independent. burnSubtitles controls visible text inside the MP4. generateSrt and generateVtt control separate subtitle files.

Desired result	outputLanguage	outputLanguageMode	burnSubtitles	generateSrt / generateVtt
Video only, no subtitle artifacts	empty	original	false	both false
Original-language subtitles	empty	original	as desired	as desired
Translated subtitles only	choose a target	translations	as desired	as desired
Original + translation	choose a target	both	as desired	as desired

In both mode, two subtitle levels in the video are intentional: the translation is the primary lower track and the original is a smaller secondary track above it. Choose translations, not both, when you want French-only, Spanish-only, or another single translated language.

Translated-only example:

{
  "outputLanguage": "FR",
  "outputLanguageMode": "translations",
  "burnSubtitles": true,
  "generateSrt": true,
  "generateVtt": false
}

Video-only example:

{
  "outputLanguageMode": "original",
  "burnSubtitles": false,
  "generateSrt": false,
  "generateVtt": false
}

If subtitles are already burned into the source image, the Actor cannot remove them. Burning a new track on top can look like duplicate subtitles. Embedded text or bitmap subtitle streams are not used as transcription input: this version transcribes the audio.

4. Inputs

Field	Meaning
videoFile	One real file upload. In JSON it is an array containing the URL created by Apify's upload widget. Do not add characters to that URL manually.
videoUrl	One publicly reachable direct media-file URL. It must return the video bytes, not an HTML page.
confirmContentRights	Required. Processing does not start until this is true.
clipCount	1 to 10; default 3. It is a target, not permission to duplicate material.
targetClipDuration	15-30, 30-60, or 60-90 seconds. The Actor may adjust a boundary to preserve a complete sentence.
platform	TikTok, Instagram Reels, YouTube Shorts, or Generic vertical. This affects suggested metadata only. JSON values are case-sensitive.
sourceLanguage	Leave empty for automatic detection. If set through JSON, use a language code such as en, fr, es, de, it, pt, ar, zh, ja, or ko.
outputLanguage	Optional translation target: EN-US, ES, FR, DE, IT, PT-BR, AR, ZH-HANS, JA, or KO.
outputLanguageMode	original, translations, or both. A translated mode requires outputLanguage.
aspectRatio	9:16. Other ratios are refused rather than silently rendered incorrectly.
resolution	720x1280 or 1080x1920. HD has a higher render price.
cropMode	face_tracking, center_crop, or blurred_background.
fallbackCropMode	center_crop or blurred_background, used when tracking has no reliable subject.
burnSubtitles	Whether subtitles are visibly rendered into the MP4.
subtitleStyle	Clean, Bold, or Minimal; relevant only when subtitles are burned.
generateSrt, generateVtt	Whether separate subtitle files are produced.
createZip	Whether all produced artifacts are also packaged into ai-viral-clips.zip.
maxCostUsd	Internal cost ceiling checked before any paid provider call. Minimum $1.
minimumViralScore	Optional editorial-score threshold. A high value can reduce the number of delivered clips.
avoidOverlaps	Keep enabled to prevent overlapping or reused source passages.
maxMediaMinutes	Refuse longer sources. Default 90 minutes; maximum 180.

The form supplies the allowed values. When using JSON, copy their spelling and capitalisation exactly. For example, use "platform": "TikTok", not "tiktok".

5. How the clips remain different

The analysis model proposes candidate passages from the timed transcript. The Actor then validates the candidates independently of the model:

• source intervals cannot significantly overlap;
• textual duplicates and near-duplicates are rejected;
• neighbouring slices of the same exchange are rejected;
• delivered clips normally keep a 20-second editorial gap;
• cuts are moved to word, phrase, and sentence boundaries where possible;
• if the model concentrates all proposals in one section, a deterministic pass searches other complete transcript regions.

clipCount: 3 means “find up to three honest clips,” not “produce three files at any cost.” A short or repetitive source may correctly return only one or two clips. Reducing the target duration may reveal more distinct passages; turning off overlap protection is not recommended.

6. What the viral score means

The 0–100 score is an editorial estimate based on hook strength, standalone quality, clarity, informative value, curiosity, emotional intensity, pacing, and conclusion, with penalties for context dependency and artificial cuts. The full breakdown is returned with every clip.

It is not a performance prediction or a promise that a clip will go viral.

7. Framing and active-speaker tracking

Mode	Behaviour	Best for
face_tracking	Detects persistent faces, compares visible speech with the audio, follows the likely active speaker, and smooths the virtual camera.	Interviews, podcasts, panels, and talking-head footage.
center_crop	Uses a fixed vertical crop through the centre of the landscape image.	A subject that remains centred, or a completely static result.
blurred_background	Preserves the full landscape frame over a blurred vertical background.	Slides, demonstrations, wide group shots, or footage where cropping would remove important content.

Face tracking does not simply choose the largest face. YuNet detects faces and Light-ASD helps associate visible speech with the audio. The track is then median-filtered, dead-zoned, velocity-limited, and smoothed. A speaker change must remain dominant before the crop switches, and scene cuts are reacquired without a slow pan.

When active-speaker evidence is unavailable, the Actor may retain a stable visual face track. If no face is reliable, it uses fallbackCropMode. During B-roll or voice-over, there may be no visible person who corresponds to the voice; in that situation no system can infer the narrator from lip movement.

Profiles, covered mouths, heavy backlight, small faces, overlapping speech, and crowded scenes remain difficult. For footage where every part of the wide frame matters, choose blurred_background instead of tracking.

Stabilised face tracking means stabilised crop coordinates. It does not apply a general camera-shake stabiliser to the source video.

8. Languages and translation

Leave Input language empty unless automatic detection is wrong. The Actor normalises common two-letter and three-letter source codes before translation; you do not need to know the translation provider's internal code format.

Validated output languages:

Language	Code
English (US)	EN-US
Spanish	ES
French	FR
German	DE
Italian	IT
Portuguese (Brazil)	PT-BR
Arabic	AR
Simplified Chinese	ZH-HANS
Japanese	JA
Korean	KO

Input detection covers more languages, but only these ten have been validated end to end for translation, wrapping, direction, and rendering. Translation is machine-produced and marked machine-translated-unreviewed. When source and target are the same language, including EN to EN-US, no translation is performed or charged.

9. Subtitle behaviour

All ten output languages have validated language profiles. English, French, Spanish, German, Italian, Portuguese, and Arabic avoid awkward line endings on frequent grammatical function words. Simplified Chinese and Japanese use punctuation-aware kinsoku boundaries. Korean prefers spaces and safely falls back to Hangul-syllable boundaries. Arabic uses right-to-left shaping through libass. Unicode grapheme measurement prevents internal cuts in combining marks and emoji sequences.

Subtitles use at most two lines and remain inside a phone-safe region. The reading-speed target is 23 characters per second, with script-specific limits where appropriate. When fast speech makes readability and synchronisation incompatible, synchronisation wins and the subtitle track is labelled needs_manual_review; text is not silently removed or desynchronised.

There is no word-by-word karaoke or dynamic highlight style. The transcription provider does not guarantee the word timings required for an honest result.

10. Outputs and statuses

One Dataset row is created per selected clip. It includes the source interval, duration, score and breakdown, hook, title, description, hashtags, selectionMethod, cropMethod, language codes, compliance statuses, warnings, and links to files that actually exist.

The run's Key-Value Store can contain:

• one MP4 and one metadata JSON per delivered clip;
• SRT and VTT tracks only when their toggles are enabled;
• original and translated tracks with language suffixes when both are requested;
• manifest.json, including counters, billing reconciliation, warnings, artifact sizes, and SHA-256 checksums;
• ai-viral-clips.zip when createZip is enabled.

An absent artifact is represented by null, never by an empty fake link.

Status	Meaning
success	Every selected clip was rendered and delivered.
partial_success	At least one clip was delivered and at least one failed. Inspect each row.
failed	No clip could be delivered. Inspect error and the logs.
needs_manual_review	A subtitle track exists but violates a linguistic reading constraint, usually because speech is too fast.

A subtitle warning does not mean the MP4 failed to decode. Technical media validation and linguistic subtitle review are separate facts.

11. Pricing

This Actor uses pay-per-event pricing. A started minute is a billed minute.

Event	Price	Charged when
actor-start	$0.05	Once, after input validation and the cost check.
transcribed-minute	$0.03	Per source minute, after a usable transcript is produced.
analysed-minute	$0.02	Per source minute when the model selects passages.
fallback-analysed-minute	$0.01	Per source minute when deterministic selection is used instead.
translated-minute	$0.04	Per delivered clip minute, only when translation is produced.
rendered-clip-minute	$0.08	Per validated clip minute at 720 × 1280.
rendered-hd-clip-minute	$0.13	Per validated clip minute at 1080 × 1920.

Exactly one of the two analysis event types is used in a run. For each delivered clip, exactly one of the two render event types is used. Failed renders, rejected candidates, and unnecessary translations are not charged.

Example for a one-hour source producing five clips of 30–60 seconds, using model-based analysis:

• 720 × 1280, no translation: $3.45;
• 720 × 1280, translated: $3.65;
• 1080 × 1920, no translation: $3.70;
• 1080 × 1920, translated: $3.90.

If deterministic fallback selection is used instead, the analysis portion is charged at the lower fallback rate.

maxCostUsd is the Actor's preflight ceiling. Apify's separate Maximum cost per run setting must also be high enough. If the Actor estimate exceeds maxCostUsd, the run stops with cost_limit_exceeded before a paid provider is contacted. If the Apify charge ceiling is exhausted later, the run reports billing_limit_reached and keeps already produced artifacts.

12. Complete JSON example

For a direct file URL, three distinct clips with French subtitles:

{
  "videoUrl": "https://example.com/interview.mp4",
  "confirmContentRights": true,
  "clipCount": 3,
  "targetClipDuration": "15-30",
  "platform": "TikTok",
  "sourceLanguage": "en",
  "outputLanguage": "FR",
  "outputLanguageMode": "translations",
  "aspectRatio": "9:16",
  "resolution": "720x1280",
  "cropMode": "face_tracking",
  "fallbackCropMode": "center_crop",
  "burnSubtitles": true,
  "subtitleStyle": "Clean",
  "generateSrt": true,
  "generateVtt": true,
  "createZip": true,
  "maxCostUsd": 5,
  "minimumViralScore": 0,
  "avoidOverlaps": true,
  "maxMediaMinutes": 90
}

The videoFile value generated by the Console is normally an array containing one Apify storage URL. Use the upload widget instead of typing that structure by hand whenever possible.

13. Media validation

Every rendered clip is probed again before delivery. Validation checks include the required audio and video streams, dimensions, pixel format, duration, decodable beginning and end, and timestamp consistency. This catches damaged files that can report a plausible duration yet fail when their final frames are decoded.

The MP4 output uses H.264 video, AAC audio, and yuv420p for broad social and mobile compatibility. Rendering re-encodes the cut so boundaries remain frame-accurate instead of drifting to the nearest source keyframe.

14. Troubleshooting

Symptom or result	Explanation and action
The run succeeds in a few seconds but returns no clips	No source was supplied, so the free readiness check ran. Upload a video and start a new run.
The upload field is missing	The selected Actor build does not expose this project's .actor/input_schema.json. Select the deployed build/version that contains the complete project.
validation_error	Read the message and correct the named input. JSON enum values are case-sensitive.
A direct URL is refused	It is probably a platform page, private page, or redirect rather than a direct media file. Upload the video instead.
Fewer clips than requested	The source lacks enough distinct passages under the current duration or score constraints. Try a shorter duration or a longer source.
Two subtitle levels appear	outputLanguageMode is both, or subtitles were already burned into the source. Choose translations for translated-only output.
No visible subtitles, but SRT/VTT files exist	This is expected when burnSubtitles is false but generateSrt or generateVtt is true.
The wrong person is framed during a conversation	Active-speaker evidence may be ambiguous. Profiles, overlapping speech, and covered mouths are difficult. Try center_crop or blurred_background for that source.
The image shakes	Face tracking filters crop jitter but cannot remove shake present in the original camera footage.
cost_limit_exceeded	Increase maxCostUsd, or reduce source length, clip count, clip duration, or resolution. Nothing was charged.
billing_limit_reached	Increase Apify's Maximum cost per run and start a new run. Already produced files remain available.
configuration_error	A required provider is not configured on the Actor. Nothing was charged; report it to the Actor owner.
Subtitle needs_manual_review	Speech was too fast for the configured reading limit. Review or edit the external subtitle before publication.

15. Privacy and third-party processing

The media is downloaded into the run's container and temporary working files are deleted when the run ends. Audio is sent to the transcription provider; the timed transcript is sent to the selection provider; clip text is sent to the translation provider only when translation is requested. Face detection, active-speaker analysis, cropping, subtitle rendering, and media validation run inside the Actor container.

Output files remain in the run's Apify storage according to your Apify storage and retention settings.

16. Copyright and content rights

You must own the source or be authorised to transform it. Confirmation is required before any download or paid processing.

The Actor does not download from YouTube, TikTok, Instagram, or similar platform pages, and does not bypass authentication, access controls, or DRM. Download or obtain the source lawfully, then upload the file.

17. Known limitations

• Automatic selection, transcription, metadata, and translation require human review before publication.
• A requested clip count is not guaranteed when honest, non-overlapping passages do not exist.
• Active-speaker tracking can fail on profiles, small or obscured faces, overlapping speakers, heavy backlight, and crowded scenes.
• It stabilises the digital crop, not physical camera shake.
• Subtitle streams already embedded in the source are ignored; audio is transcribed. Text already burned into the image cannot be removed.
• Only vertical 9:16 output is rendered in this version.
• Source limits are 500 MB, 90 minutes by default, and 180 minutes maximum. The source must contain both video and audio.
• There is no automatic social publishing, B-roll insertion, voice cloning, face modification, AI-generated imagery, or word-by-word karaoke.

18. API call

curl -X POST "https://api.apify.com/v2/acts/lumaxys~ai-viral-clip-generator/runs?token=$APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "videoUrl": "https://example.com/interview.mp4",
    "confirmContentRights": true,
    "clipCount": 3,
    "targetClipDuration": "30-60",
    "platform": "TikTok",
    "resolution": "720x1280",
    "cropMode": "face_tracking",
    "outputLanguageMode": "original",
    "maxCostUsd": 5
  }'

Dataset rows and file URLs are available in the run output. The full artifact inventory, validation results, and billing reconciliation are recorded in manifest.json.

Support

Include the Apify run ID and a minimal reproducible input when reporting a problem. Do not include private API tokens or signed media URLs.