On this page
AI creates a video in four stages: a language model writes the script, a video model generates (or software selects) the visuals for each scene, a text-to-speech voice reads the script, and an editor syncs captions and music and renders the file. Tools like Revid run all four from a single prompt, link or script.
The model is rarely what decides whether anyone watches. Most tools send prompts to the same rented models (Veo, Seedance, Kling, Wan, MiniMax Hailuo), so what changes the result is the input: the hook, the structure, the pacing. In Revid's study of AI videos versus real footage, AI-generated TikToks averaged twice the views of real footage. On X, Instagram and Threads, though, AI visuals clustered in the lowest engagement tier. Same technology, opposite outcomes.
How does AI create a video, step by step?
- Input. You give the tool a prompt, a script, a link (a blog post, a YouTube video, a product page), a song or a reference video.
- Script. A large language model turns that input into a scene-by-scene script: a hook, the body, an ending. Better tools ground this step in real examples. Revid's script writer, for instance, can first pull transcripts of related TikToks with more than 10,000 likes and use them as style references.
- Visuals. For each scene, one of three things happens: a video model generates a clip from a text description, an image model creates a still that is then animated, or the system picks matching stock footage. Generated clips are short (Google's Veo 3.1 makes 4, 6 or 8-second clips), so a 30-second video is several clips cut together.
- Voice, captions and music. A text-to-speech voice reads the script, speech recognition times the captions to each word, and a music track goes underneath.
- Edit and render. You fix weak scenes in an editor, export (720p, 1080p or 4K are the usual options) and publish.
Steps 1 and 2 decide most of the outcome. A weak hook stays weak no matter which model renders it.
What technology powers AI video generation?
| Technology | What it does in the video | Example |
|---|---|---|
| Large language models (NLP) | Read the input, write the script, split it into scenes, write a visual prompt for each scene | Turning a blog post into a short narrated script |
| Diffusion video models | Generate moving footage from a text description or a starting image | A product on a kitchen counter with a slow camera push-in |
| Image models | Create stills that get animated or used as scene backgrounds | A cartoon character for a story series |
| Text-to-speech | Produces the voice-over in many voices and languages | The same script read in Spanish and French |
| Speech recognition | Times captions to the spoken words | Word-by-word animated captions |
| Computer vision | Understands existing footage: picks stock clips, reframes landscape video to vertical, describes scenes | Cropping a webinar to 9:16 around the speaker |
Older guides credit Generative Adversarial Networks (GANs) for AI video. GANs powered earlier face and image synthesis, but today's video models are mostly diffusion models. These start from random noise and refine it step by step into frames that match the prompt. For the deep learning background behind all of this, Andrew Ng's conversation with Lex Fridman is a solid primer:
Text-to-video, image-to-video or footage-based: which approach fits?
| Approach | You provide | Best for | Watch out for |
|---|---|---|---|
| Text-to-video | A prompt or a script | Explainers, faceless channels, stories, product concepts | Generic visuals when the prompt is vague |
| Image-to-video | A photo or an illustration | Product shots, animating artwork, consistent characters | Motion that drifts away from the original image |
| Footage-based | Your own recordings or a long video | Clips from podcasts and webinars, captioned talking heads | You need footage worth cutting |
| Reference-based | A video that already performed | Adapting a proven format to your topic | Copying instead of adapting (see the rights section below) |
If you have no footage, start with text-to-video. If you have strong product photos, use image-to-video so the product looks like the real thing. If you record podcasts or webinars, footage-based clipping gets you more videos per hour of work than generating from scratch.
Why some AI videos get watched and most get skipped
AI footage is now good enough for the feed. What limits results is the idea and the opening seconds, and that is where audience data helps.
In Revid's analysis of 176,367 TikToks with 1,000+ likes posted between October 2025 and October 2026, faceless videos (no person on camera, which is how most AI-made videos look) had a median of 114K views, and 17.4% passed 1 million views. Videos with one person on camera had a 67K median and 12.7% past 1 million. The full TikTok virality study explains the method. The sample only includes videos that already reached 1,000 likes, so it shows what successful videos have in common, not what causes success.
Faceless videos earned a median of 114K views, against 67K for videos with one person on camera, and passed 1M views 38% more often.
Source: Revid analysis of 176,367 TikToks with 1,000+ likes posted between Oct 1, 2025 and Oct 16, 2026 (61,203 creators)
Show the data
| Median views | Share that passed 1M views | |
|---|---|---|
| Animals | 147.4K | 24.1% |
| Faceless | 113.9K | 17.4% |
| Groups of people | 96.1K | 16% |
| Objects or products | 70.8K | 13.2% |
| One person on camera | 67.4K | 12.7% |
The same TikTok data has a second lesson for AI scripts: less narration did better. Videos with under 40 characters of speech had a 123K median (20.2% past 1 million), against 68K for videos with talking throughout (300+ characters of speech). AI tools default to wall-to-wall voice-over because a script is the easiest thing to type. That default is worth fighting.
Rules that follow from this:
- Write the first line before anything else. Read it as on-screen text with the sound off. If it does not stop you, rewrite it.
- Cut the script, then cut it again. If a scene works visually, let it play without narration.
- Start from a format that already works. A reference video with proven engagement gives the model a structure to follow. A blank prompt gives it nothing to aim at, so the output drifts toward a generic average.
- Generate several hooks, keep one. Regenerating an opening scene costs a few credits. A weak opening costs you the whole video.
- Check every fact. AI scripts state wrong things confidently, and a correction pinned in the comments hurts more than the time it takes to verify.
Prompt to try
Make a 30-second faceless vertical video about why octopuses have three hearts. Open with on-screen text: 'Octopuses have three hearts, and one stops when they swim.' Keep the narration under 40 words and let close-up underwater shots play without voice between lines. End with a question asking viewers which animal fact surprised them most.
How to write a prompt for an AI video
Write one prompt per scene, in this order: subject + action + setting + camera + light + mood.
- Weak: "a coffee shop video"
- Strong: "A barista pours latte art in a sunlit café, close-up, slow push-in, warm morning light, calm mood"
For the script itself, give the model the audience, the format and the constraint: "Write a vertical video script about [topic] for [audience]. Open with a surprising fact shown as on-screen text, give three quick tips, end with a question. Keep the narration short." If you need camera directions and examples, the free AI video prompt generator writes scene prompts you can paste into any tool.
Prompt to try
Create a 20-second vertical video for a neighborhood coffee shop. Open on a barista pouring latte art in a sunlit café, close-up, slow push-in, warm morning light, calm mood, with on-screen text 'The best first sip in town.' Follow with two short scenes of fresh pastries and regulars chatting, light acoustic music, minimal narration, and end on the shop's opening hours.
Which AI video tools should you use?
AI video products tend to do one job well. Match the tool to the job:
| Tool | What it does best | Starting price | Best for |
|---|---|---|---|
| Revid | Finished short-form videos from a prompt, link, script or song, grounded in a library of about 2.77 million viral videos; several models chosen per scene; timeline editor; direct publishing | Free sign-up with free credits to try; paid plans from $39/month | Founders, small businesses and creators publishing short-form video often |
| InVideo | Chat-style agent where you edit by typing commands, 16M+ stock library, avatars and voice clones | From $20/month; free plan with watermark | Light use, people who prefer editing by chat |
| Synthesia | Polished avatar presenters, 140+ languages, slide-based workflow, SCORM and LMS export | From $29/month; free plan with up to 10 minutes of video a month | Corporate training and internal communications |
| Pictory | Blog posts, scripts and URLs to video, auto-clipping of webinar recordings, 12M+ licensed stock clips from Getty Images and Storyblocks | From $29/month; free trial, no permanent free tier | Repurposing articles and recordings at volume |
| VEED | Browser editor with a multi-track timeline, auto-subtitles and translation, AI dubbing | Roughly $10 to $24/month; free plan with watermark and 720p export | Editing footage you filmed yourself |
Quick decision rules:
- If you have your own footage and mainly need to cut and caption it, use VEED or a clipping tool such as OpusClip.
- If you need the same human presenter in every training video, use Synthesia.
- If you want raw cinematic clips with shot-by-shot camera control, a model lab such as Runway fits better.
- If you want finished vertical videos from an idea or a link, published to TikTok, Instagram and YouTube from the same place, use Revid.
For fact-checked head-to-heads, see Revid's side-by-side tool comparisons.
What is Revid.ai and what are its features?
Revid is an AI video platform that turns an idea, a link or a proven viral format into finished short-form videos and publishes them to TikTok, Instagram and YouTube. It is built for founders and small business owners, and is also used by creators, agencies and AI agents. It started in Paris in 2023 as Typeframes and has been called Revid since April 2024. More than 100,000 videos are rendered on Revid every month (Revid's analytics, 2026).
| Feature | What it does |
|---|---|
| Viral video library | About 2.77 million short-form videos and posts with public engagement data. You can search them and remix any one with options such as "Same energy, new script" or "Make it about your topic". |
| Ways to create | Script, prompt, article or link, long YouTube video to Shorts, audio, song to music video, talking avatars, captions for any video: 100+ specialized tools |
| AI video models | Seedance 2.5, MiniMax Hailuo H3, Wan 3.0 and Google Veo 3 under one credit system, chosen per scene, up to native 4K |
| Voices and captions | 1,000+ AI voices in 70+ languages, 30 caption styles, automatic captions in 100+ languages |
| Editing and export | Full timeline editor; export in 720p, 1080p or 4K at 30 or 60 fps |
| Publishing | Direct publishing and scheduling to TikTok, Instagram and YouTube from one queue |
| Automation | Auto-Mode workers that create a fresh video every day from a prompt or a content source |
| For developers and AI agents | REST API, CLI and an MCP server for AI assistants such as Claude and ChatGPT, on every paid plan |
Paid plans start at $39/month (Growth, 2,000 credits), and a 30-second video with Pro AI visuals and a voice-over costs about 40 to 80 credits. See the full list of Revid features and the current Revid plans and prices.
How to make a video with AI in Revid
- Open the AI video generator and choose your input: a prompt, a script or a link. For an article, the blog to video converter reads the page and writes the script from it.
- To start from a proven format instead, search Revid's viral library by topic, open a video that performed, and choose "Make it about your topic" to apply its hook style and pacing to your niche.
- Review the script and scenes, then pick a model for each scene, a voice and a caption style.
- Check the credit estimate, generate, and fix any weak scene in the timeline editor.
- Export the file, or publish and schedule it directly to TikTok, Instagram and YouTube.
How businesses use AI video, and how to measure it
Adoption is no longer early. Grand View Research's AI video generator market report valued the market at USD 554.9 million in 2023 and projected a 19.9% CAGR from 2024 to 2030. In Wyzowl's 2026 survey, 63% of video marketers said they had used AI video tools to create or edit marketing videos, up from 51% a year earlier (more in our short-form video statistics).
Where it pays off:
- Product videos. Paste a product page and the tool reads the product, keeps its real photos and writes the script. This matters most for stores with large catalogs, where filming every item is not realistic.
- Explainers and training. A script or a set of slides becomes a narrated video. Updating it means editing text and re-rendering, not reshooting.
- Ads. Generate several hook variants of the same ad and let the results pick the winner.
- Repurposing. Blog posts, long YouTube videos and podcasts become vertical clips for TikTok, Reels and Shorts.
- Personalized sales videos. A real estate agent can send each buyer a short tour of the listings that match their criteria.
To know whether it is working, track four numbers:
- Cost per finished video: tool credits plus your editing time, compared with what a video cost you before.
- Videos published per week: the output you actually ship, not what you generate.
- Median views per video: the typical result, not your one outlier.
- Clicks, sign-ups or sales attributed to video: use links or codes in the bio and the caption.
Third-party roundups such as 9cv9's AI video generator statistics and Visla's list of video marketing trends collect many more figures. Trace each one to its original source before you quote it. For film and streaming adoption specifically, Statista's topic page on AI and video streaming in the US tracks the data.
Do you have to label AI-generated videos?
Yes, when the footage or audio is realistic. Here is what each platform requires:
| Platform | When you must disclose | What happens |
|---|---|---|
| TikTok | Realistic AI-generated images, audio or video | Your label reads "creator labeled as AI-generated". TikTok may also auto-label uploads that carry C2PA Content Credentials, and an auto label cannot be removed. Labeling does not affect distribution if the video follows the Community Guidelines. |
| YouTube | Realistic altered or generated content, such as a real person appearing to say something they did not, or a realistic scene that never happened (the "AI use" setting in YouTube Studio) | Not required for AI-written scripts, captions, titles or cloning your own voice. Since May 2026, YouTube automatically labels undisclosed photorealistic AI. Disclosure alone does not limit audience or earnings. |
| Instagram and Facebook | Photorealistic video or realistic-sounding audio that was digitally created or altered | The label shows as "AI info". Meta reads C2PA and IPTC signals and may penalize people who do not disclose. |
In the EU, the AI Act's transparency rules, which require deepfakes to be clearly labelled, took effect in August 2026 according to the European Commission.
Monetization. YouTube allows AI tools in monetized videos when the result shows your creative vision and gives educational or entertainment value. It refuses to monetize three kinds of AI content: content made from generic templates that looks mass-produced, compilations of unrelated AI clips meant to shock, and AI personas posing as experts on health, legal, financial or political topics. Since March 2026, Facebook also deprioritizes reels that only narrate what is on screen or stitch clips together.
Copyright. You need the rights to any footage, music, script or likeness you reuse. When Revid adapts a reference video, it keeps the structure, timing and tone but writes new words, new visuals and an original music direction. Its "Reuse the script" option deliberately keeps the original spoken words, so the rights question is yours to settle before you publish.
Accuracy. Never use AI to show a real event or a real person in a way that did not happen. Treat every AI-written claim as a draft until you have checked it.
FAQ
Is Revid free?
Sign-up is free and needs no credit card, and new accounts get free credits to try video creation. A free account can also search the viral video library. Regular creation then uses the monthly credits of a paid plan, starting at $39/month for 2,000 credits. Revid's pricing page is the live source for prices.
Can you make money from AI-generated videos?
Yes, as long as the video shows your own creative input. YouTube monetizes AI-assisted videos with real educational or entertainment value, but not mass-produced template content or AI personas giving expert advice on sensitive topics.
Does labeling a video as AI-generated hurt its reach?
No. TikTok and YouTube both state that the disclosure label alone does not affect distribution or eligibility to earn. Skipping a required label is what creates risk.
Can AI make a whole video from just text?
Yes. Text-to-video tools write the script, generate or select the visuals, and add a voice-over, captions and music from a single prompt. Plan to edit afterward: tightening the hook and regenerating weak scenes is where the quality comes from.
Which AI video models are available in 2026?
The main ones are Google Veo, Seedance, Kling, Wan, MiniMax Hailuo and Runway's models. OpenAI's Sora is gone: the app closed on April 26, 2026, and Sora 2 left the API on September 24, 2026. Revid offers Seedance 2.5, MiniMax Hailuo H3, Wan 3.0 and Google Veo 3 under one credit system.
