Guide
How to keep the same face across AI photos and video
Train the face instead of describing it. Prompts and reference images make the model redraw the face every time, so it drifts from shot to shot. A character LoRA, a small add-on trained on one face, makes the model draw that face the same way in every photo, and image-to-video carries it into clips.
Three ways to hold a face, and where each breaks
| Method | How it works | Where it breaks |
|---|---|---|
| A detailed prompt | Describe the face in words every time | Words fit thousands of faces; it drifts on every new image |
| A reference image | Feed one picture of the face with each job | Holds on similar angles, slips on new poses, light and profiles |
| A trained character LoRA | Train the model on a set of shots of the face once | Needs a training run first, and a good, varied set to learn from |
The workflow that holds
- Cast the face. Describe the character and generate takes until one is right. That take is the reference.
- Build a training set. Turn the one reference into a set of varied shots: front, three-quarter and profile, different expressions, outfits and light. Leave out any shot where the face slipped.
- Train the LoRA. Train on the kept shots, then compare checkpoints with the same test prompts and keep the one that looks most like the reference.
- Refine once. Use the first LoRA to draw a fresh set, keep the best and train again. The second LoRA learns from cleaner shots.
- Shoot photos, then move them. Make every photo with the LoRA, and make video from those photos, so clips keep the trained face.
castingcard.ai runs these steps as one flow in the browser, on your own GPU or a rented one. Once the character is trained, a 5-second clip costs about 39¢ of GPU time. Read what a character LoRA is for the details of training.
Common mistakes
- Training on near-identical shots. The LoRA learns the pose and the light along with the face, and fights you when you change them.
- Keeping shots where the face slipped. The LoRA learns the slip too. Leave them out.
- Making video straight from text. Text-to-video draws a new face. Start from a photo made with the trained character.
Questions
- What is the most reliable way to keep an AI character consistent?
- Train a LoRA on the character. A prompt or a single reference image only describes the face, so the model redraws it each time; a LoRA has learned it, so it comes back the same across poses, outfits, light and video.
- Does a consistent character work in video too?
- Yes, if the video starts from a photo made with the trained character. Image-to-video keeps the face of the frame it starts from, so the clip matches the photos.
- How many photos does a character need?
- Enough varied, good shots of the same face: different angles, expressions, outfits and light. castingcard.ai plans a 24-shot set from one reference photo, you keep the shots that look right, and it trains on those.
- Can I make a model of a real person?
- No. Every character is a fictional adult. Real people, celebrities and lookalikes aren't allowed.
Cast your first model
Cast a fictional model and train it once, in the browser.