A consistent AI presenter for explainers and courses

Decide what the presenter actually does

Most explainer and training videos are mostly screen content: slides, product UI, diagrams. The presenter appears at the start, between sections and at the end. That is good news, because short presenter shots are exactly where reference-based video is strongest.

Split your presenter needs into two kinds. Scene shots: the presenter walking into frame, gesturing at a screen, standing in a workshop or office, reacting. These are short, visual and suit reference-to-video models.

Long talking segments: the presenter speaking a 60-second script to camera with accurate lip sync. Dedicated talking-avatar tools are still more reliable for those.

Choosing and building the presenter

Real or fictional. If the presenter is based on a real employee or trainer, get written consent that covers AI-generated video, the courses it will appear in, and what happens when they leave the company. A fictional presenter avoids that question and never ages out of the material.

Look. Pick an appearance that suits the audience and will not date quickly: a plain outfit in a solid color, no busy patterns, which also hold up better in generation, and a hairstyle that is easy for the model to keep.

References. Use five photos: a front view at chest height, a three-quarter view from each side, a waist-up shot with hands visible, and a full-body shot standing. Presenter shots often include gestures, which is why the hands matter. Shoot or generate all of them in the same outfit and the same light.

Set. Give the presenter a home: a studio corner, a classroom, a warehouse floor for safety training. Keep one photo of it and use it as the scene photo for every opener, so they look like they were filmed in one place.

A shot set for one course

Generate a reusable library rather than new shots for every module:

1. Course intro: presenter walks into the set and stops at center, 6 seconds.

2. Module opener, three or four variants: presenter turns to camera, nods, gestures toward an empty area where the module title will appear, 4 seconds each.

3. Transition: presenter steps aside or points off-screen to hand over to screen content, 3 seconds.

4. Emphasis beats: leaning in, counting on fingers, a short pause with a smile, 3 seconds each.

5. Outro: presenter waves or nods and the camera pulls back, 5 seconds.

About fifteen clips cover a twenty-module course. Titles, captions and the voice-over are added in your editor, which keeps every module consistent even when the script changes.

Script and voice

Write for cuts. Keep on-camera lines to one or two sentences, then cut to screen content while the voice-over continues. Viewers accept a voice that continues over slides; they notice a mouth slightly out of sync on a long shot.

Use one voice for the whole series, whether a recorded narrator or a synthetic voice, and keep its settings with the course files. For a short opening greeting, the video model can generate the line itself. For anything longer, generate the shot without speech and add the voice-over in your editor, or use a separate lip-sync tool on the finished clip.

Label the presenter as AI-generated in the course credits or description. For workplace training, learners should not be left thinking a real colleague recorded it.

Keeping a long series consistent

Freeze the reference set when production starts. Version it if you must change the outfit, and do not mix versions inside one course.

Fix the framing too: the same camera height, distance and aspect ratio for every opener. Consistency of framing reads as consistency of person.

Generate updates against the original references, not against earlier clips, and review new shots next to the course intro before publishing.

Where RefVid fits

RefVid keeps the presenter as a saved character with 1 to 5 photos, and a photo of the set can go in as the scene photo, so a module made six months later starts from the same references. Every video of the presenter is listed on its page in the studio for review. Clips run 3 to 15 seconds, in 16:9, 9:16 or 1:1, on Kling 3.0 Omni.

With sound on, the model generates the audio with the picture, and a line written in quotes is said as written; English and Chinese speech sound best. RefVid does not do voice cloning, lip-sync or editing. Long voice-overs, captions and lip-synced segments are done in your editor or an avatar tool.