My AI video character's face changes between shots

Any AI video tool Last checked 2026-10-14

What you saw

You generate shot 1 and shot 2 from near-identical prompts — same character description, same wardrobe — and the person in shot 2 is visibly a different human. Different jawline, different age, different eye colour.

Or identity holds for a few seconds and then drifts mid-clip, usually after a camera move or a cut-like transition inside the generation.

Or the face is right but the clothing, hair length or scar/tattoo detail changes between shots.

What it actually means

A text-to-video model samples from a distribution. "A woman in her thirties, dark hair, grey coat" describes a region of that distribution containing millions of faces. Two generations from the same prompt are two draws from the same bucket, not two shots of the same person.

There is no persistent character memory between API calls unless the model is explicitly given one. Seeds help reproducibility of a *whole generation*, not identity across different prompts.

Mid-clip drift is a different problem: the model is maintaining coherence frame to frame and loses the thread when the scene changes enough. Longer clips and bigger camera moves both increase drift.

This is the single most common reason AI video series look amateur. Audiences forgive weird physics. They do not forgive a protagonist whose face changes every cut.

The fix

1. Stop describing the character and start showing it. Use image-based direction: Veo 3.1 accepts up to three reference images to guide the content of a generated video. Feed the same reference set to every shot in the sequence.

2. Lock the first frame. Veo 3.1 supports frame-specific generation — you specify the first and last frames and the model generates between them. Generating a still of your character first, then using it as the first frame of every shot, is far more reliable than any prompt wording.

   # conceptually, per shot:
   # 1. generate ONE canonical still of the character
   # 2. reuse that still as the first frame / reference image for every shot
   # 3. vary only the action and camera in the text prompt

3. For continuous action, use video extension rather than generating a new clip. Veo 3.1 can extend a video it previously generated, which preserves identity far better than starting fresh.

4. Consider a model built for this. Google now positions Gemini Omni Flash as the default video model specifically citing superior video coherence and character consistency, with Veo 3.1 reserved for scene extension, last-frame control and existing pipelines. If identity is your bottleneck, test the consistency-focused model before you spend another week on prompt wording.

5. Keep shots short. Drift compounds with duration. Three 4-second shots that hold identity beat one 12-second shot that loses it.

6. Fix identity in post when generation will not cooperate — a face-swap or re-render pass over a clip that is otherwise correct is cheaper than regenerating the shot forty times.

How to stop it happening again

Build a character bible before you generate anything: one canonical reference still per character, from the same angle, saved in the project. Every shot references it. This is the AI-video equivalent of a continuity folder, and it is the single highest-leverage habit in the whole workflow.

Freeze the wording of the character description and store it as a constant in your pipeline, not retyped per shot. Small wording changes move the sample.

Keep a per-character seed where the tool exposes one, and record which seed produced your canonical still.

Review continuity on a contact sheet, not in the timeline. Export one frame per shot, lay them side by side, and identity breaks become obvious in seconds.

Budget for rejects. A realistic shot acceptance rate for character-consistent AI video is well under 50%; plan generations accordingly rather than treating each one as final.

Related questions

Does using the same seed give me the same character?

Only if the prompt and settings are also identical. A seed makes one generation reproducible. Change the action or camera in the prompt and the same seed will give you a different person.

Why does the face drift partway through a single clip?

Frame-to-frame coherence degrades over time and across large camera moves. Shorter clips and smaller moves reduce it. For longer continuous action, extend a previous clip instead of generating a long one from scratch.

Is a LoRA or fine-tune the answer?

It can be, on open models where you control training, and it is the strongest form of identity lock available. On hosted APIs you usually cannot fine-tune, so reference images plus first-frame control are the practical equivalent.