HeyGen: HeyGen Video

HeyGen Video is HeyGen’s general-purpose scene-generation model, identified by the manufacturer as heygen-video-1. It creates the subject, setting, light and sound together from a creative brief. Each clip is 5–15 seconds long and contains its generated audio track, including dialogue, ambience and effects. Useful production inputs include a scene description, a product photograph or material that establishes a person, place or movement. The resulting asset is a short scene for a product presentation, an onboarding insert, a property view or other edited video work.

Three modes give the inputs different roles. Text-to-video creates a scene from the prompt. Image-to-video starts literally on the supplied image, then follows the brief for the action. Reference-to-video uses supplied images, video and optional audio to guide a newly composed scene; it requires at least one reference image or video. Choose a first-frame image when the opening composition must match a still, or references when a known object should appear in a different composition. This scene model is distinct from HeyGen’s avatar rendering engines and Video Agent composition workflow.

The maker’s guide favours detailed, concrete shot directions: one subject, one location and one contained action, with a locked or gently moving camera. Describe the camera position, lens or depth-of-field intent, light, surface texture and colour treatment. Explain where objects touch when an action depends on contact. For an original product example, use a reference photograph of a travel flask and ask for it to stand on a wooden counter while a presenter speaks beside it, with a small camera move toward the flask. Inspect the flask’s form, placement and visible markings in the result.

Reference lists support up to nine images, three videos and three audio items, with a maximum of twelve references in total. Images, videos and audio are numbered separately in the manufacturer’s prompt format: the first image is <Picture 1>, and the first video is <Video 1>. Give each reference a role and identify that role explicitly in the brief. The model reads the first five seconds of each reference video, so prepare the relevant section at the start. The API reference also specifies reference-video proportions between 1:4 and 4:1. A long clip containing the desired movement only at its end is a poor preparation for that five-second input window.

The documented reference interface accepts a directly accessible HTTPS resource, an uploaded asset identifier or inline base64. The guide lists maximum image sizes of 16 MB by URL and 5 MB inline; video or audio may be up to 32 MB by URL and 16 MB inline. Prepare readable source material with the important subject easy to distinguish and without competing instructions in different references. Keep the intended orientation in the first-frame image. These input details concern the maker’s reference workflow and help organize the creative materials for the selected mode.

Direct sound in the same brief as the picture. State the dialogue, vocal delivery, room tone and specific effects, including whether a sound should be nearby or distant. Specify the intended music treatment when the scene should contain only speech and ambience. The guide suggests allowing roughly 2.5 spoken words per second as a planning estimate, so leave room for pauses and the visible action. A short introduction can be planned as one line followed by a beat that reveals the product. Listen for the actual words and whether the generated sound fits the space and action.

Output duration is selected in whole seconds from five to fifteen, with a five-second default. Resolution is a size class: 480p, 768p, 1080p or 2k. The documented framing choices are 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. For text- and reference-to-video at 1080p or 2k, select 16:9 or 9:16; adaptive reference framing is documented at 480p and 768p. Image-to-video follows the first image’s proportions, so crop before generation when the required shape changes. Review the completed result’s actual dimensions rather than assuming a size-class label implies one conventional pixel pair.

The model guide specifies MP4 video with H.264 picture encoding and AAC audio at 32 kHz stereo. A generation is processed asynchronously, and the completed job supplies the playable video location, duration and dimensions. Its signed video location can be refreshed by retrieving the job again. Preserve the actual clip, the reference materials and the brief for an editing review. A pending or processing record tracks the generation; the completed video is the production artifact to assess for picture, speech and sound.

The manufacturer documents prompt enhancement choices of turbo, quality and disabled; disabled keeps the prepared wording as written. A fixed seed is useful when comparing a change to one clause while retaining the other inputs. Record the chosen seed and settings with the shot brief. For on-screen lettering, keep an intended phrase short and spell it precisely; the guide recommends compositing paragraphs, small print and labelled diagrams afterward. Check every visible character in a generated title rather than assuming the wording was reproduced correctly.

The guide identifies long on-screen text, soft organic movement such as petals, paper or hair, and close hand manipulation as less reliable areas. For a procedural explanation, let a clearly written line convey the steps while the image shows a stable person or object; use separately verified visuals when the physical operation itself must be exact. Review the whole shot for object continuity, contact, face and hand detail, dialogue and sound. A concise list of what should change in the next take keeps a detailed production brief practical.