Reference to Video
Drop up to 9 references — keep characters, styles and products consistent across your video.
Reference to Video AI
Upload the face, outfit, product or location you need, tag each one in your prompt, and AI Hug turns them into a single shot that stays on-model. Reference to video is the fastest way to get a consistent character AI video without retraining anything or re-rolling dozens of times.
- Up to 9 reference images
- Reference video & audio
- Seedance 2.0 · Wan 3.0
What References Change in a Shot
Each clip below started from one or two still images. The prompt describes the action and camera; the references lock in what the subject actually looks like.
One reference, a whole scene
A single creature design became a backlit desert siege. The silhouette, leg count and body shape come straight from the reference image, so the creature reads the same in every frame instead of mutating between cuts.
Face plus outfit, kept together
A close-up of the face and a full-body shot of the costume were tagged as Image 1 and Image 2. Splitting identity and wardrobe across two references is the most reliable way to get character consistency AI video when the camera moves fast.
Stylized characters stay stylized
References are not only for realistic people. A plush, cartoon-proportioned creature kept its fur color and rounded shape through a high-speed chase, because the model was copying a design rather than guessing one from text.
Most drift problems start with the input, not the prompt
How to Pick References That Actually Hold
- 01
One job per image
Give each reference a single purpose: a face, an outfit, a product, a room. A busy collage forces the model to decide what matters, and it often decides wrong.
- 02
Clean, well-lit faces
For people, use a sharp, front-facing photo with even light and nothing covering the eyes. Heavy filters, sunglasses or extreme angles are the usual cause of a face that shifts mid-clip.
- 03
Separate identity from wardrobe
If the costume matters, add a full-body shot as its own reference next to the face close-up. Tag both in the prompt so the model knows they belong to the same person.
- 04
Products on plain backgrounds
Packshots on white or neutral backgrounds transfer labels, colors and proportions far better than lifestyle photos where the item is small or partly hidden.
- 05
Use video for motion, audio for voice
A short reference clip is the clearest way to show a dance move, a camera path or a gesture. A reference audio clip guides voice and rhythm. Keep both brief and on-topic.
- 06
Tag every reference in the prompt
Type @ in the prompt box to insert Image 1, Video 1 or Audio 1 exactly where it applies, e.g. "Image 1 walks into the café from Image 3, holding the bottle from Image 2." Untagged uploads are still sent, but the model has to guess their role.
Make an AI Video from Reference Images in Three Steps
From uploads to a finished clip
- 1
Upload your references
Add up to 9 images, plus optional reference videos and audio clips, in the upload bar above the prompt. Each upload appears in the prompt as a tagged chip you can drag into place.
- 2
Describe the shot and tag the references
Write what happens, where, and how the camera moves, then press @ to point at the right reference for each subject. Pick Seedance 2.0 for refined 4–15 second shots or Wan 3.0 for clips up to 30 seconds and 1080p.
- 3
Generate, review, refine
Choose duration, resolution and sound, then generate. If a detail drifts, swap in a clearer reference or make the tag placement more specific instead of rewriting the whole prompt.
FAQ
Reference to Video FAQ
Questions about reference to video, multi reference video AI and keeping characters consistent.
What is reference to video AI?
It is video generation guided by your own media. Instead of describing a character in words and hoping the model draws the same person every time, you upload images (and optionally video or audio) and the model treats them as the ground truth for how subjects look, move or sound.
How is this different from image to video?
Image to video animates one picture as the first frame, so the shot starts exactly where the photo is. Reference to video uses several inputs as a guide and builds a brand-new scene around them, which means your character can appear anywhere, from any angle. Use image to video when the opening frame matters and this tool when identity matters.
How many references can I use?
Up to 9 images, plus up to 3 reference videos and 3 audio clips. With Wan 3.0, reference videos can total up to 15 seconds and reference audio up to 15 seconds, and the reference video length plus the output duration must stay within 30 seconds. The tool checks these limits before submitting.
Which model should I choose?
Seedance 2.0 is the pick for polished, physically convincing 4–15 second shots at 480p or 720p; Seedance 2.0 Mini is the lower-cost version. Wan 3.0 handles longer clips from 2 to 30 seconds, goes up to 1080p and generates native audio. Credits are charged per second of output, and the exact cost is shown on the Generate button before you start.
Can I use photos of real people as references?
Only photos you have the right to use. Get consent from anyone recognizable before turning their likeness into video, and do not upload images of celebrities or other people to impersonate them. Prompts and reference images are checked by content moderation.
Why does my character still change between shots?
Usually because the reference is ambiguous: low resolution, a partly hidden face, or several people in one image. Crop to a single subject, add a second angle, and tag the reference directly next to the action it should perform. Very fast motion and extreme close-ups are also harder to keep consistent than medium shots.
Can I use the videos commercially?
Videos generated on a paid plan can be used in your own projects, ads and social content, as long as your references are yours to use. Avoid copyrighted characters, logos and brands you do not own, since a reference does not transfer any rights to the generated clip.