Same face: lock one clean reference still before adding a co-star

Where this lesson fits

Shot Daily 01 covered what makes a shot inspectable. Lesson 02 focuses on minimizing facial drift. All three language editions teach the same studio discipline.

Our Consistent Character manual documents API limits (up to 7 reference images, optional voice, 720p cap). Here we focus purely on the practical shooting order.

Core rule: the still locks identity, not pose

A reference portrait constrains bone structure, hair silhouette, skin tone, and build. It is not a contract for blocking, camera angle, or room layout — and wardrobe holds much better when restated in text. Treating a portrait as a full scene layout causes stiff posture and portrait backdrops to bleed into dramatic shots.

Separate who from what: the image defines identity; your text prompt defines room, lighting, gesture, and lens. Crop out watermarks, text overlays, or background clutter before uploading.

• Lock: facial geometry, hair length/parting, wardrobe fabric, headcount
• Write separately: blocking, room, key light, single gesture, camera move
• Strip out: watermarks, captions, sunglasses, and cluttered backgrounds

Start with one master still, not seven

While the endpoint accepts up to seven images, first-pass testing works best with one canonical portrait. Mixing seven lighting setups and outfits averages features into a stranger. Validate 4 seconds from a clean, front-or-three-quarter still in matching wardrobe before adding secondary angles.

Voice references never replace visual references — audio without a portrait still generates an unanchored face.

• Unobstructed face within 30° of front-facing, even lighting
• Hair and outfit match tonight's scene
• Original fictional character sheet or explicitly licensed talent still
• Name <IMAGE_0> in text and repeat hair, dress, and solo headcount

Why 4 seconds beats 15 seconds for identity checks

Identity drift compounds with heavy motion, occlusion, and long duration. Four seconds is plenty to verify one slow turn, one blink, and one hand resting on a chair. Fifteen-second requests are far more likely to invent cuts or extra characters.

Our public benchmark uses 1 still, 4 seconds, 3:4, 480p — holding the gold-dress identity while documenting slight pose and facial drift on the verified examples page.

Post-generation checklist

Scrub the downloaded MP4 from start to finish. When drift appears, subtract variables — swap in a cleaner portrait or reduce motion beats rather than writing longer paragraphs.

• Compare frame 1 and final frame jawline, eyes, and nose bridge
• Check hair silhouette and dress neckline through the turn
• Ensure no unrequested second person or text overlay appeared
• Confirm the gesture stayed strictly turn, blink, and hand on chair
• If features soften, match prompt key-light color temperature closer to the still

Clean character assets first

Consistent Character workflows are designed for original fictional character sheets and authorized talent assets. Minor-related imagery and unauthorized real-person likenesses are prohibited.

Building a clean, reusable fictional Master Reference portrait gives multi-episode short dramas a reliable foundation.

Next lesson

Once solo identity holds, you are ready for two-handers. Shot Daily 03 covers anchoring two distinct leads, preventing unrequested third characters, and keeping hands/limbs clean across overlaps.

Lock this solo 4-second test first.