HeyGen, best known for AI avatars, has released a general-purpose video model. HeyGen Video (API model id heygen-video-1) was announced on October 21 and the blog post was updated on October 22. Instead of animating a presenter, it generates the subject, the setting, the light and the sound from a text description.

Built on MiniMax H3

HeyGen's developer docs describe the model as "built on MiniMax H3 and post-trained by HeyGen." In other words, the base model comes from MiniMax, and HeyGen did the post-training. For more on MiniMax's recent releases, see our coverage of MiniMax M3.1 Flash.

What HeyGen says it does

According to HeyGen, each clip runs 5 to 15 seconds and can be rendered at 480p, 768p, 1080p or 2K. HeyGen says picture and audio come from the same pass: each clip ships with its own track of dialogue, ambience and sound effects, with "no separate speech pass, lip-sync step or avatar to pick." Output is MP4 (H.264) at 24 fps with AAC stereo audio. A duration of "auto" lets the model pick a length from 5 to 15 seconds to fit the dialogue.

There are three modes:

  • Text to video: a prompt only.
  • Image to video: a prompt plus one image, which becomes the literal first frame.
  • Reference to video (the API default): a prompt plus images, videos or audio, producing a new scene that keeps the products, people or places you supplied.

How it differs from Avatar V

The docs draw the line plainly. The avatar engines "animate a look you already own from a script you already wrote," while HeyGen Video "invents the subject, the setting, the light and the sound at once, from a description." HeyGen pitches the two as a pair: Avatar V delivers the message, and HeyGen Video supplies the cold opens, product shots and B-roll around it.

How references control people and scenes

In reference to video, you can supply up to 9 images, 3 videos and 3 audio files, with a cap of 12 in total. Each list's order becomes a label you use in the prompt: the first image is Picture 1, the first video is Video 1, and so on, numbered separately per type. The model reads only the first 5 seconds of each reference video, so HeyGen advises trimming clips to the part you want followed. HeyGen says a product photo supplied this way keeps its colour, form and markings in the generated shot. If you want to compare this approach with another reference-driven tool, see our guide to directing a shot with Kling 4.0 keyframes and references.

Limits HeyGen itself lists

The docs say the model is at its best on short, contained shots with one subject, one place, one action and a locked or barely moving camera. They also say it is least reliable on long on-screen text, soft organic motion such as petals, paper and hair, and close hand work like assembling or operating equipment. The clips shown so far are HeyGen's official demos. We have not tested the model, and there is no guarantee that people or objects will stay consistent in every clip.

Access and pricing

HeyGen says the model is available through the HeyGen API and Creative Studio. The API runs on paid API keys only. The docs say 1080p and 2K bill at 3x the 768p rate, but neither the blog nor the docs list a per-second or per-clip price. Which Creative Studio plans include it, and in which regions, has not yet been announced.

Sources

  • HeyGen Blog: Announcing HeyGen Video
  • HeyGen Developer Docs: HeyGen Video