HeyGen Launches Video Model With Built-In Sound

HeyGen, best known for AI avatars that read a script on camera, now wants to generate the rest of the shot as well. On 21 October the company announced HeyGen Video, which it calls “our general-purpose video model”, and said a single prompt produces “the whole scene: subject, setting, light and sound, in a single call”.

The clips are short. HeyGen says the model, listed as heygen-video-1, makes videos of 5 to 15 seconds “from text, images or your own references, in styles from live action to stop-motion, at up to 2K”. It is available through the HeyGen API and the company’s Creative Studio, and runs on paid API keys. In a help article, HeyGen calls it “our first general-purpose video model”, saying that until now it has been home to its avatar and voice models.

HeyGen’s pitch is sound made in the same pass

The company’s argument is that most AI video is assembled from parts. “One model makes the picture, another voices the lines, lip-sync stitches them together and a sound library fills the gaps,” the announcement says. “The result: room tone that doesn’t match the room, footsteps that land a beat late.”

HeyGen says its model avoids that. Every clip “ships with its own audio track: dialogue, ambience and sound effects, generated alongside the image. There is no separate speech pass, no lip-sync step and no avatar to pick first.” Its example is a train pulling into an empty platform, which it says comes out as “near silence, then the horn, the wheels on the rail joints, the rush of air and the brakes, in that order”.

It draws a clear line between the new model and Avatar V, its avatar engine. “Avatar V animates a look you already own from a script you already wrote. HeyGen Video invents what isn’t there yet.” The company suggests combining the two, opening on a generated scene, handing over to an avatar to deliver the message and cutting back to a product shot.

The model can be called in three ways. Text to video builds a scene from a prompt alone. Image to video takes a still and uses it as the first frame, which HeyGen suggests for bringing a storyboard frame or the last frame of a previous shot to life. Reference to video accepts “up to 9 images, 3 videos and 3 audio files (12 in total)”, which users name in the prompt. HeyGen says a product photographed this way keeps “its colour, form and markings” in the scene built around it.

The technical details are specific. Output comes as MP4 at 24 frames a second with stereo audio, in aspect ratios from 21:9 to 9:16 and resolutions of 480p, 768p, 1080p or 2K. Prompts can run to 32,000 characters, and HeyGen says prompts of “a few hundred to a few thousand characters produce materially better output”. Setting a seed returns the same clip for the same prompt, so users can change one clause at a time.

Pricing is per second of output. According to HeyGen’s help article, from November text-to-video and image-to-video clips will cost $0.02 a second at 480p and $0.03 at 768p, while reference-to-video costs $0.04 and $0.06, with the seconds of any reference video also counted. The blog adds that 1080p and 2K “bill at 3× the 768p rate”. Until the end of October, HeyGen is running a 50% launch discount.

The quality claims are HeyGen’s own

HeyGen makes strong claims about consistency. It says the model was “post-trained for the shot as a whole”, and that “objects keep their size and shape, stay where they were placed and obey physics. Nothing duplicates, vanishes or reappears halfway through.” It also says the model follows a brief literally: “If you didn’t ask for dialogue, you don’t get dialogue.” In its help article, HeyGen goes further, saying that in blind side-by-side evaluations against Seedance 2.5, Seedance 2.0, H3 Max and Kling 3.0, HeyGen Video “ranked #1 in quality”, based on 4,800 blind votes on the Artificial Analysis Arena query set. HeyGen does not say who ran that comparison, and it cites no independent tests, so those claims should probably be treated as the company’s description until outside users have put the model through its paces.

The company is also candid about where the model works best. “It shines brightest on short, contained shots: one subject, one place, one action and a locked or gently moving camera,” it says. Short titles can be rendered inside the shot, but for longer copy HeyGen advises adding real type in the edit. That is a narrower promise than “the whole scene”, though HeyGen has an answer ready: such shots, it says, describe “most of the shots brands need”.