The most common complaint about AI video is not quality. It is that clip two shows a different person from clip one. You describe "a woman with a copper-red bob and an olive jacket", and the model draws a new woman who fits that description every time you press Generate.
The fix that works is reference images: you show the model the same few photos of your character on every clip, and you stop describing her in words. This tutorial walks through how to do that, which photos to pick, how to write the prompt, and how the main reference models differ: Kling 3.0 Omni, MiniMax H3 (Hailuo), Wan 3.0 and Seedance 2.5. Every image below is a frame from a real, unedited clip we made in RefVid, the tool I work on.
Step 1: Make one invented character and photograph her four ways
Our test character is Mia. She is invented, not a real person, which matters later (some models refuse real-looking faces, and you should only ever use photos of people who agreed to it).
She has four reference images, and the set is the whole trick:
- Front, head and shoulders
- Three-quarter view
- Profile
- Full length, so the model knows her jacket, jeans and build
Reference 1 of 4: the front view. The sharpest, most evenly lit photo goes first.
Reference 3 of 4: the profile. Without it, a model has to guess her nose and ear when she turns.
Rules that held up across every model we tried:
- Three to five photos beat one. One photo is enough to start, but the model has to invent every angle the photo does not show. More angles hold better when the character turns her head.
- Plain background, even light, no filters, no text. Anything in the photo can leak into the clip.
- Never change the set between clips. A new haircut photo or a different outfit in clip two is, to the model, a new character. Save the set once and reuse exactly those files.
Step 2: Write the scene, not the person
Once the photos are attached, the prompt should only say where she is, what she does and what the camera does. Do not repeat her hair colour or her jacket. When the words and the photos disagree even slightly, the model blends them, and the blend is the drift you were trying to avoid.
Here is the prompt for the first clip, word for word:
She walks through a neon-lit Tokyo side street at night in light rain, wet asphalt reflecting pink and blue neon light. The camera tracks alongside her at shoulder height in a smooth lateral dolly as she walks.
Kling 3.0 Omni, 5 seconds, from the four photos. Not edited.
Then the same four photos with completely different scenes:
Same photos, new prompt: a windy beach at golden hour, slow push-in to a close-up.
Same photos: a Scandinavian kitchen in morning light, pouring coffee, then smiling at the camera.
Same photos: a misty pine trail, turning back toward the camera.
Four scenes, four lighting setups, one face. For comparison, the prompt-only version of the kitchen scene (no photos, Mia described in as much detail as we could fit) gave a woman who matched the words but not the other clips:
Words only, no reference images. Close, but not Mia.
All of the Mia clips above ran on Kling 3.0 Omni, which takes up to 10 reference images and makes 3 to 15 seconds. If you want to try it without the Kling app or an API key, our Kling 3.0 Omni free online page runs a first 5-second clip on it for free.
Step 3: Keep the photos, switch the model
Reference images are not a Kling feature; most current video models have a reference mode. The useful part is that the same saved photos work on all of them, so you can pick a model per shot and the character stays put. What changes is how many photos each takes, how long a clip can be, and what kind of subject it accepts:
| Model | Photos it takes (in RefVid) | Clip length | Good for |
|---|---|---|---|
| Kling 3.0 Omni | up to 10 | 3–15 s | People, products, pets |
| MiniMax H3 (Hailuo) | up to 5 | 5–15 s, with sound | People, short talking shots |
| Wan 3.0 | up to 10 | 3–30 s | People and products, long single takes |
| Seedance 2.5 | up to 10 | 4–30 s | Products, pets, illustrated characters (refuses real-looking people) |
A new account's free credits cover one short first clip from your photos on several of these: MiniMax H3 (5 seconds at 768p, with sound), Wan 3.0 (4 seconds at 720p) and Vidu Q4, another reference model that takes 1 to 10 photos.
MiniMax H3: Hailuo's subject reference, now with several photos
Hailuo's original subject reference model, S2V-01, read a single face photo. MiniMax H3 reads several. We sent it the same four Mia photos:
MiniMax H3, 5 seconds at 768p, same four photos, with street sound and her line "Hi, I'm Mia." Back in 33 seconds.
The differences between S2V-01 and H3 are laid out in our Hailuo subject reference guide.
Wan 3.0: the long take
Most Wan reference-to-video tutorials are ComfyUI workflows written for Wan 2.6 or 2.7. Wan 3.0 takes up to 10 reference images and runs up to 30 seconds in one pass, which is the easiest way to keep a character through a long continuous action instead of stitching 5-second clips. Write the actions in order, one per sentence, so the model knows the sequence. More in Wan reference to video online.
Seedance 2.5: for products and drawn characters
Seedance 2.5 refuses photos of real-looking people before the clip runs. For a product, a pet or an illustrated character it holds detail very well. Here is AERO 1, an invented running shoe, from three product photos:
Seedance 2.5, 5 seconds at 720p, from three photos. The knit upper, heel tab and white sole match the photos.
For an illustrated character, a reference sheet that shows it from several sides works as one of the photos. Details in Seedance reference to video.
Checklist
- 3–5 photos of one subject: front, three-quarter, profile, full length.
- Plain background, even light, sharpest photo first.
- Save the set once; never swap photos between clips of the same story.
- The prompt says place, action and camera. Never appearance.
- Pick the model per shot: Kling 3.0 Omni or MiniMax H3 for short character shots, Wan 3.0 for long takes, Seedance 2.5 for products and drawn characters.
If you try this and a model still drifts on your character, tell me in the comments which model and how many photos you used. That is usually where the answer is.
Top comments (0)