HeyGen's new video model makes 15 second scenes with sound. Here is how a 30 minute episode actually gets built.

The first question we asked when HeyGen put out documentation for its new general purpose video model was whether it could produce a full 30 minute episode, the kind of weekly market show a brokerage or an owner might want to run. The short answer is that each generation is a clip between 5 and 15 seconds long, so an episode is something you assemble from scenes, and HeyGen already has a separate tool that handles the long talking segments.

What HeyGen Video 1 is

HeyGen’s documentation lists the model as heygen-video-1 and describes it as built on MiniMax H3 and post-trained by HeyGen. It generates a whole scene, with the subject, the setting and the sound, from a single request. The specs from the docs:

• Clips run 5 to 15 seconds each, at 480p or 768p, 24 frames per second, delivered as MP4
• Aspect ratios run from 21:9 widescreen down to 9:16 vertical, plus an adaptive setting
• Audio comes out in the same pass, with generated dialogue and sound effects
• Three modes: text to video from a prompt alone, image to video where your image becomes the first frame, and reference to video where you pass in up to 9 images, 3 videos and 3 audio clips, with a limit of 12 references in total
• Access is through HeyGen’s API on paid API keys

The model underneath, MiniMax H3, was released by MiniMax on July 31. MiniMax describes it as a general purpose model that reads text, images, video and audio and generates video with native stereo sound, up to 15 seconds at 2K. The cover image on this article is a frame from one of MiniMax’s own H3 sample clips.

Where the 30 minute episode comes from

HeyGen’s avatar videos are the long form piece. Per HeyGen’s usage limits, an avatar or image video renders as a single scene with a maximum of 30 minutes, and multi-scene studio videos and templates can run as long as the sum of their scenes, with each scene staying under 30 minutes.

So a half hour show on HeyGen is built in layers. The host segments, where your principal or your analyst walks through the numbers, come from an avatar video that can hold the full length. Video 1 supplies the short scenes that cut in between, like an establishing shot of the submarket, a slow push across a lobby, or a cutaway with its own ambient sound. Studio stitches them into one piece.

What it means for a commercial real estate team

The reference mode is the part worth paying attention to. Nine reference images is enough to hand the model the actual photography of a building, so the clip it returns starts from the real property. That makes it useful for listing teasers, the opening seconds of a market update, and vertical cuts for LinkedIn and Instagram where a 15 second clip is the right length anyway.

The resolution tops out at 768p, which reads well on a phone and on a website and looks soft on a conference room screen, so a flagship property film still wants a higher resolution engine or a real shoot. Anything generated from a real building also gets checked against the building before it goes out, because a model will happily add a window or a balcony that is not there, and a buyer or a tenant who tours the space will notice.

We already build our own video across several engines, and our Seedance 2.5 write-up covers why we keep the reference library and the workflow in our own hands and swap the model as better ones ship. Video 1 is one more engine for that bench, with sound built in and a price we will confirm on HeyGen’s pricing page before we budget a campaign around it, since the API pricing page did not list it yet when we checked on September 30.

If you want a weekly market show or a set of listing films running for your team, that is a 30 minute conversation.

Sources: HeyGen Video 1 documentation, HeyGen API usage limits, MiniMax H3 announcement. Cover image courtesy of MiniMax, a frame from an H3 sample clip.