Consistent character AI video: keep a face, product or place the same
Make one clip of your founder and it looks like her; make the next and the model gives you a cousin. Reference images fix that. Here is what a reference is, which models accept one, how to shoot a set in ten minutes and how to build a month of clips around a mascot, a product or a clubhouse.
· 5 min read · 1.2K views · by the Cortex Digital Hub team
Drop two to four reference photos of your subject, pick a model with a reference slot, and describe only the scene and the motion. The face, product or building stays the same from clip to clip. From ₹24 a clip · no watermark · runs in your browser
Make a video → → Opens AI Video Generator · runs in your browser
Make one clip of your founder walking through the office and it looks like her. Make the next one and the model hands you a cousin: same hair, different jaw. The same happens with a product, a mascot and a building. Reference images stop it, and they are the feature that turns the AI Video Generator from a one-off toy into something you can plan a month of content around.
What a reference is
A reference is anything you drop, paste or click into the reference slot alongside your prompt: an image, a video or an audio file. What the model does with it depends on how much you give it.
With no reference, the model invents the picture from the prompt alone. Fine for a mood clip, useless for anything that has to look like a specific person or thing.
With one image, that image becomes the first frame and the prompt describes only the motion. This is image to video. It keeps your picture, but only while the subject stays roughly where it started.
With two or more images, a video or an audio file, the model treats them as references it keeps to. It learns what the subject looks like from several views and draws it fresh in whatever scene you describe. That is the mode this post is about. Two to four clear references work best: one is too few to learn from, and six or eight start to blur what matters.
Which models accept references
Not every model takes references, so choose the model first. As of October 2026, these are the ones with a reference slot.
Vela 2.5 takes references and also edits and extends, so it is the one for a clip that has to be long (up to 30 seconds) and cinematic. Nova 2K makes the sharpest single frame on the site, at 2K, which suits a product people will pause and zoom. Orbit 2.7 is listed as the consistent-characters model, at 1080p and up to 15 seconds, and is the sensible default for a person. Canter 1.1 takes references at 1080p for up to 15 seconds. Lumen Frames and Lumen Omni add first-and-last-frame control, which matters when a series has to stitch together.
Orbit 3 and Orbit Prime are the cheap, long-clip models: rough out the camera move and timing on them, then switch to a reference model for the clip you will publish. A 5-second 720p clip on Orbit 2.7 is 135 credits (₹67.50), on Nova 2K it is 175 (₹87.50), and on Vela 2.5 at 720p it is 855 (₹427.50).
How to shoot a reference set
The model copies what you show it, flaws included. Shoot the references like a passport photographer, not like an influencer.
Neutral light. Near a window in daytime with no direct sun, or under even indoor light. A hard shadow across a face gets baked into every clip, including the one set at night on a beach.
Three angles. Front, three-quarter and profile for a person; front, side and top-down for a product; the main elevation, a corner and the approach from the road for a building. The three-quarter view does most of the work; the other two tell the model what is around the corner.
Plain background. A white wall, a bedsheet, or run the shots through the background remover first. Anything behind the subject competes for attention.
Same session, same outfit, same version of the product. Crop tight in the photo editor so the subject fills the frame, and keep each file at least 1080 pixels on the long side.
Building a series around one subject
References pay off when you make more than one clip, because the shoot is a one-time cost and the consistency compounds.
A brand mascot. A D2C snack brand in Bangalore with a cartoon tiger on the pack can shoot the tiger from the artwork at three angles, then make a Reel a week: tiger at a cricket match, at a Diwali party, in the monsoon. Same tiger every time.
A founder. Kirana-to-D2C founders are the face of their brand on Instagram and rarely have time to film. Four reference photos and the founder can appear in clips that were never shot: unboxing in a studio, in the warehouse, walking into a Mumbai trade fair.
A signature product. A Meesho seller's bestselling kurta or a lamp from a Chennai decor store, shot on white, can be placed in a living room, on a balcony at golden hour or under studio lights without re-photographing. If the same shots go to marketplace listings, see the product photo requirements guide.
A project's clubhouse. For a Hyderabad developer, the clubhouse is the asset channel partners share most. Three views as references, and every walkthrough, festive greeting and WhatsApp status shows the same building. Starting from a plan? Make the 3D plan video first and pull stills from it.
Do and don't
- Do choose the model before the shoot; the reference slot only exists on the models above.
- Do keep the prompt about the scene and the motion; the references already say what the subject looks like.
- Do draft at the lowest resolution, then re-run at 1080p once the prompt is right.
- Do keep every reference from one shoot: same outfit, haircut and product version.
- Don't expect a first-frame image to be kept exactly once you add a second; with two or more, the model is referencing, not copying.
- Don't use group photos, sunglasses, heavy filters or a render from an earlier stage of the project.
- Don't expect small text on packaging to survive; add the logo afterwards with the free video tools.
- Don't upload someone's face without asking them.
What references cost
Nothing extra. Credits are charged per second of finished video at the model's rate, and references do not change the rate. A 10-second 1080p clip on Orbit 2.7 costs 400 credits (₹200) with no references and 400 with four. Credits are returned automatically if the model fails or declines, so a reference set a model turns down costs nothing.
What references change is how many clips you throw away. Without them, five founder clips might take ten generations to get five that match; with them, it is closer to five, and that saving is the point. Clips stay in your history for seven days, so download the keepers, then finish them in the free video tools: resize to 9:16, trim, add a logo and compress for WhatsApp, all on your own device.
Frequently asked questions
How many reference images do I need for a consistent character?
Two to four clear images work best. One image becomes the first frame rather than a reference, and more than four or five starts to confuse the model about what matters. Front, three-quarter and profile is the standard set.
Do reference images cost extra credits?
No. Credits are charged per second of finished video at the model's rate, and adding references does not change that rate. The generate button shows the exact total before you commit, and credits come back automatically if the model fails or declines.
Which models on the AI Video Generator accept references?
As of October 2026: Vela 2.5, Nova 2K, Orbit 2.7 (listed as the consistent-characters model), Canter 1.1, Lumen Frames and Lumen Omni. Pick the model first, then shoot the references to suit it.
Can I use a reference photo of a real person?
Only with their permission. The image is sent to the video model only to make the clip and nothing is used for training, but consent is yours to get, especially for a founder, an employee or a customer.