HeyGen Video — Text, Image and Reference to Video with Native Audio
What is HeyGen Video?
HeyGen Video (heygen-video-1) is HeyGen's general-purpose AI video generator, built on MiniMax H3 and post-trained by HeyGen. From a single prompt it generates a whole scene — subject, setting, light and sound together — in one call, and every clip ships with its own audio track of dialogue, ambience and sound effects. There is no separate text-to-speech pass, no lip-sync step and no avatar to pick.
One model is called three ways, chosen by the inputs you send:
- Text to video — a prompt only. A new scene with its own sound (5–30 seconds).
- Image to video — a prompt plus one image. The clip opens on your image as its literal first frame (5–30 seconds).
- Reference to video — a prompt plus your images, videos or audio. A new scene that keeps the products, people or places you supply (5–15 seconds).
Key Features
Native synchronized audio generated with the picture; three generation modes in one model; reference-to-video that preserves a supplied object's colour, form and markings (up to 9 images, 3 videos and 3 audio files); first-frame image-to-video; durations up to 30 seconds; resolutions from 480p to 2k; seven aspect ratios plus auto; prompt enhancement (turbo, quality, ultraquality or disabled); reproducible seeds; and short, spelled-out in-shot lettering.
Best Use Cases
HeyGen Video is strongest on short, contained shots — one subject, one place, one action, a locked or barely-moving camera — where objects stay stable and obey physics and the model follows your brief literally. That makes it a fit for talking-to-camera clips, product-on-a-bench shots, brand and social spots, explainer beats, and reference-driven scenes that drop a real product or person into a new setting. Because audio is generated in the same pass, a vertical clip comes back finished rather than silent.
Prompt Tips and Output Quality
Write long, specific prompts: name the camera, lens and aperture instead of words like "cinematic"; quote dialogue exactly (roughly 2.5 words per second); and direct the sound by naming the room tone and effects. Ask for texture, then rule out the defaults with clauses such as "no retouching" and "no logos or printed words". A tested text-to-video run — a cheesemonger speaking a line to camera — returned a clean 8-second 768p clip with lip-synced dialogue and a layered soundscape. Fix the seed to iterate one clause at a time. The model is weakest at long on-screen text, soft organic motion (petals, paper, hair) and close hand work.

