Buy video training data for AI models

What is video training data?

Video training data is footage licensed for model training, together with the metadata and rights records that let a buyer use it safely. Versos, which runs a video archive licensing network, reports buyers seeking unpublished footage, film and TV libraries, corporate video and niche or large-scale libraries.

Footage from robotics and first-person cameras is a separate category, used to teach models how people and objects move. Its rights questions center on the people and places in shot rather than on music or performances.

How much does video training data cost?

Video training data is priced per minute, and the public figures run from about $1 to about $7.25 per minute.

Bloomberg reported that Adobe asked contributors for footage of people in everyday actions such as walking or expressing emotions, simple anatomy shots of hands, feet or eyes, and people interacting with objects such as smartphones or fitness equipment. Briefs like these show that buyers often specify content in advance.

What rights does video training data need?

Video training data needs clearance for every layer inside the frame, not only the footage itself. Getphyllo's guidance on training data licenses calls this per-layer clearance for complex content, since video can contain music, images, performances and trademarks owned by others.

• Copyright in the footage, confirmed by chain of title
• Music and sound recordings in the soundtrack
• Performers' and contributors' consent
• Model releases for identifiable people
• Trademarks, artworks and other copyright works in shot

Versos makes the same point for sellers: confirm clear, transferable rights, obtain model releases and third-party clearances where needed, and use anonymization or consent protocols for personal or sensitive material.

Under US copyright law, an exclusive license counts as a transfer of copyright ownership under 17 USC 101, and 17 USC 204(a) requires a transfer to be in writing and signed by the rights owner. Ask for the signed grant behind any exclusive rights the seller claims. Illinois' Biometric Information Privacy Act lists a scan of face geometry as a biometric identifier, which matters if faces in the footage will be analyzed.

Video data is more useful when it comes as original masters with shot-level descriptions that let you filter by content. Versos notes that without a semantic, searchable structure even valuable assets can stay hidden.

Ask for resolution, frame rate and codec in the data manifest, and check whether clips were cut from longer programs, since that affects both rights and duplication.

Who sells video for AI training?

Video for AI training is sold by broadcasters, film and TV libraries, stock footage agencies, creators and companies with corporate video archives. Versos lists Google, Meta, NVIDIA, Runway, Moonvalley, Synthesia and Getty Images among the places where creators can license video.

How should video training data be delivered?

Video training data should be delivered as original masters, with a manifest listing each clip's duration, resolution, frame rate, codec and rights record. Per-clip descriptions help you filter footage and make duplicate detection easier.

Ask how the minutes were counted. A per-minute price on a catalog padded with near-duplicates, slates or black frames costs more per useful minute than the headline rate suggests.

What should you check before buying video data?

Before buying video data, sample at random across the catalog, check resolution and duplicates, and audit the rights records for a subset of clips. Agree withdrawal terms that say what happens to models already trained if a rights holder objects, which Getphyllo lists as one of six things a training data license should cover.

Media owners can read about selling media and video archives, and our guide to licensing video and audio archives explains the seller's side. See the Versos profile for one licensing network. To source vetted footage, book a free consultation.