What reference-to-video AI is, how it differs from text-to-video and image-to-video, how many references each model accepts, and how to get consistent results.
Text-to-video is great until you need the same product, person or place to look identical in every shot. A prompt that says "white sneaker with a red heel tab" produces a different sneaker each time.
Reference-to-video AI fixes that. Instead of describing your subject, you show the model images, clips or audio of it, and the model generates new footage that keeps those details.
This guide explains how reference-to-video works, how it differs from image-to-video, how many references each leading model accepts, and how to prepare references that hold up. Specs were checked in October 2026.
What reference-to-video AI means
Reference-to-video is a generation mode where the model conditions new footage on one or more reference files. The references define what something looks or sounds like; the prompt defines what happens.
It sits between the two older modes:
- Text-to-video generates everything from a description, so subjects vary between generations.
- Image-to-video uses one image as the literal first frame and animates it.
- Reference-to-video uses images, clips or audio as guides, so the subject can appear in new angles, scenes and actions while keeping its identity.
That makes it the mode to use for product shots, recurring characters, brand mascots and locations that appear across a campaign.
How many references each model accepts
Seedance 2.5 accepts the most references, which suits complex scenes with several products or characters. Vidu Q3 added its Reference-to-Video mode on April 13, 2026, and Grok Imagine added image and voice references on July 31, 2026.
Where reference-to-video works best
- Product videos: several angles of one product keep the logo, stitching and color consistent across shots.
- Recurring characters: a mascot or presenter keeps the same face and outfit in every scene.
- Locations: a store, office or listing stays recognizable as the camera moves.
- Voice consistency: audio references, where supported, keep a character's voice the same across clips.
It works less well for long on-screen text, close-up hand interactions and fast, soft motion such as hair or fabric, where most models still drift.
How to prepare references that hold up
- Use clean, well-lit images. Plain backgrounds help the model separate the subject from its surroundings.
- Cover the angles you need. Front, side, back and a close-up of key details beat ten near-identical photos.
- Keep references consistent. Mixing two versions of a product teaches the model to blend them.
- Describe the action, not the subject. The references carry the look; the prompt should carry the camera move and the action.
- Fix the seed when iterating. Models with deterministic seeds, such as HeyGen Video, let you change one variable at a time.
Reference-to-video in HeyGen
HeyGen Video, launched September 30, 2026, is HeyGen's own scene model, built on MiniMax H3. It accepts up to 12 references, made up of 9 images, 3 videos and 3 audio clips, and generates 5 to 15 second clips with sound. At launch it is available through the HeyGen API and partners including OpenRouter, Runware and ComfyUI, with the web app to follow.
Inside the HeyGen web app today, reference-driven shots pair with presenters. A product still becomes a talking performance with image to video, and the product video workflow wraps a presenter and captions around product footage. Avatar V keeps the same presenter consistent for up to 30 minutes.
FAQs
What is reference-to-video AI?
Reference-to-video AI generates new video using reference images, clips or audio to keep a subject's look or voice consistent, while a text prompt describes the action. It is used for products, characters and locations that must match across shots.
How is reference-to-video different from image-to-video?
Image-to-video animates one image as the first frame. Reference-to-video uses one or more references as guides, so the subject can appear in new angles and scenes.
Which AI model accepts the most references?
Seedance 2.5, with up to 50 reference files per project. Seedance 2.0 and HeyGen Video accept up to 12, and Vidu Q3 accepts 1 to 4 images.
Can reference-to-video keep a person's face consistent?
Partly, for short clips. For a person speaking a script at length, a presenter model such as HeyGen's Avatar V holds one identity for up to 30 minutes, which clip models cannot.
Is HeyGen Video available in the HeyGen app?
Not yet. HeyGen Video launched on September 30, 2026 through the HeyGen API and partner platforms, and HeyGen has announced web app access for later.
Greetings! My name is Ayesha Shaheryar. My words have helped millions over the past two years. As a HeyGen expert and a writer, I am here to introduce tips and tricks to edit your next video in no time.
