Character Consistency Across Scenes.

Why Consistency Is Hard

Every video generation is independent. The model has no memory of previous generations. When you generate Shot A and Shot B separately, the model does not know they feature the same character. It samples from its probability distribution independently each time, producing subtle variations in facial structure, skin tone, hair, and proportions. This is fundamentally different from traditional filmmaking where the same actor physically appears in every shot. In AI cinema, you are effectively casting a new actor for every cut. The consistency gap manifests as: - **Facial drift**: Cheekbone height, nose width, eye spacing change between shots - **Color drift**: Skin tone, hair color shift across lighting conditions - **Wardrobe drift**: Clothing details, texture, and fit change subtly - **Proportion drift**: Body proportions shift between wide and close-up shots The human visual system is extraordinarily sensitive to face recognition. Your audience will notice inconsistencies you might miss. The uncanny valley effect is compounded when a face keeps changing.

The Multi-Anchor Workflow

The most reliable consistency technique uses multiple anchors to constrain each generation: **Anchor 1 - Character Reference Sheet**: The canonical images of your character from Lesson 3. Upload as reference for every generation. **Anchor 2 - Previous Shot**: Use the last frame of the previous shot as a reference for the next shot. This creates a chain of visual continuity. **Anchor 3 - Text Description**: Include a frozen character description in every prompt. Never paraphrase -- copy and paste the exact same description every time. ``` FROZEN CHARACTER BLOCK (copy-paste into every prompt): "A 40-year-old Japanese woman with short black hair streaked with grey, angular face with prominent cheekbones, dark circles under deep-set brown eyes, slim build. Wearing a heather grey wool crew-neck sweater under a navy blue double-breasted trench coat with brass buttons. Silver stud earrings." ``` **Anchor 4 - Negative Prompts**: Explicitly exclude inconsistency: ``` Negative: "different face, different person, changing appearance, morphing features, inconsistent clothing, wrong hair color" ``` Using all four anchors simultaneously gives you the best chance at consistency. Drop any one and drift increases.