How to Use HeyGen Video: A Step-by-Step Guide to Prompts, References and Voice

Key Takeaways

• Pick the mode first: text to video to explore, image to video to animate a still you own, reference to video to keep a character, product or place consistent.
• References are the skill. Make a finished character image (one full-body view, one face close-up, plain background) and the model will hold it from clip to clip.
• A prompt that works has the same parts every time: camera, subject, one action with a time, place, light, sound, and a no-text rule.
• Keep clips short. Reference to video accepts 5 to 15 seconds, and a longer film is several clips cut together.
• Add the voice after the picture works, and check every frame for stray text before anything goes out.

Step 1. Pick the mode

HeyGen Video has three modes, and the choice decides everything after it.

• Text to video when you are exploring ideas and do not need a particular look.
• Image to video when you already own the picture. The image becomes the literal first frame.
• Reference to video when something has to stay the same: a mascot, a founder look, a product, a street. This is the mode we use for anything that appears in more than one clip.

Step 2. Build the reference images

Reference to video accepts up to 9 images, 3 videos and 3 audio files, with 12 references in total, and reads only the first 5 seconds of each video. Image and audio references are free to send; a reference video is billed as input seconds on top of the output.

What worked in our tests:

• A finished character image, not a rough stand-in. The references are copied almost literally. When we used a rough 3D mannequin as a placeholder, the mannequin came back in the clip.
• Two images are enough. One full-body view and one face close-up on a plain background held the character across places.
• Keep place references free of people. If you add a place, send it without a figure in it.
• No text on anything. Letters in a reference are copied too.

Step 3. Write the prompt with the same parts every time

This is the formula we settled on. Fill every slot, in this order:

1. Camera. "Shot on a 35mm cinema camera, locked off on a tripod with one very slow push-in. No pan, no zoom, no handheld."
2. Subject. "The person from the reference images, with exactly their face and clothing."
3. One action, with a time. "Folds a white sheet, bringing the corners together in one slow clean motion at about two seconds."
4. Place. Name an ordinary, specific place and who is in the background.
5. Light. Say what the light really is: "warm yellow-orange fluorescent tubes through steamed glass".
6. Sound. What you would hear, and "no music" if you will add your own.
7. The no-text rule. "There is no text, no letters, no logos and no brand names anywhere in the frame: every sign, label and package is blank."

Prompts can run up to 32,000 characters, but long is not the goal. One action per clip is.

Step 4. Set the numbers

• Duration. Start at 5 seconds, the default. Reference to video accepts 5 to 15; text to video and image to video accept 5 to 30 on the standard enhancement settings.
• Resolution. 768p is the default. 480p is cheaper for tests. 1080p and 2K bill at three times the 768p rate.
• Aspect ratio. Choose it for where the clip will run: 9:16 for stories and short video, 16:9 for a site or a spot, 1:1 for a feed. Reference to video can also follow the shape of your reference (adaptive).
• Seed. A seed is a number that makes a result repeatable. Keep it when you re-run a clip with a small change, so the only difference is your change.

Step 5. Run one clip and read the frames

Run a single clip first. Image-only reference clips finished in about one to three minutes in our tests, and several can run in parallel once the first one is right.

Then read the frames, not just the first second. Three things went wrong for us, and all are avoidable:

• Camera brand names turn into text. Writing a professional camera maker's name in the prompt put that name on a sign and on a cereal box. "35mm cinema camera" gave the same look without it.
• "No readable brand names" is not enough. Gibberish lettering still appeared on packaging. Ask for plain, unmarked, blank props, and check every frame.
• The model copies what you show it. If a reference carries something you do not want, remove it from the reference.

Step 6. Add the voice

HeyGen Voice is a separate step when you need narration. It has three versions: the full HeyGen Voice model, Voice Turbo (faster, public and instant voices) and a free Turbo that HeyGen lists as free until November 30, 2026, on shared capacity. A speech request takes 1 to 5,000 characters of plain text and a voice, and returns a WAV file.

Pick a public voice that matches the tone of the script, keep the script short, and read it aloud once before you generate. Here is a seven-second sample made on HeyGen Voice with the voice Sterling:

HeyGen Voice sample, voice Sterling: "Your launch is tomorrow, and the video is already done. One script, one voice, ready before your coffee gets cold."

A speech request in HeyGen's API needs only the model name, a voice id and the text:

{
  "model": "heygen-voice-1-turbo-free",
  "voice_id": "<a public voice id>",
  "text": "Your launch is tomorrow, and the video is already done."
}

Step 7. Cut it together and check once more

Cut the clips on the held moment of each action, add the voice line, and watch the whole film once with the sound off, then once with it on. Look for stray text, a changed face between clips and any logo you did not ask for. For anything that will run as an ad, check the platform's rules on AI-generated media before you publish.

What it costs, in one place

• HeyGen bills video per second of output, and a reference video adds its own seconds. HeyGen is running a 50% launch promotion through the end of October.
• Voice is billed per character: $0 (Turbo Free, until November 30), $4 (Turbo) and $15 (HeyGen Voice) per million characters, according to HeyGen's own page.
• In our own tests on a HeyGen Pro account, a five-second image-only reference clip cost 5 credits and one with a reference video cost 10. That is one plan, not a price list.

Where MITPO fits

MITPO's Creative Studio lets you pick a video model, attach reference images and keep your results in one place, and HeyGen Video and HeyGen Voice are being added to its pickers. For the bigger picture, read HeyGen Video and Voice for marketers.

FAQ

How do I keep the same character in every HeyGen Video clip?

Use reference to video with a finished character image: one full-body view and one face close-up on a plain background. Send the same images with every clip and describe the clothing the way the images show it.

Why does text or a logo appear in my clip?

Naming a brand in the prompt, or showing letters in a reference, can put them on screen. Leave brand names out, ask for blank, unmarked props and signs, and check every frame.

How long can a HeyGen Video clip be?

Reference to video accepts 5 to 15 seconds. Text to video and image to video accept 5 to 30 seconds on the standard enhancement settings. Longer films are several clips cut together.

Does HeyGen Video make sound?

Yes. Each clip includes dialogue, ambience and sound effects generated with the picture. Add a separate voice with HeyGen Voice when you need narration.

Can I use HeyGen Video through an API?

Yes. HeyGen documents a video endpoint on its API (POST /v3/models/videos) with an API key and a speech endpoint (POST /v3/models/audio/tts). Check HeyGen's developer documentation for the current fields before you build.