previs-anchor-keyframes

Previs Anchor Keyframes

Reference-guided video models (Seedance 2.0 and similar) stay dramatically closer to a CG previs when you feed them photoreal ANCHOR KEYFRAMES: still images generated from previs frames at key story beats, supplied as image references alongside the driving video. This skill is the production recipe for building anchors that actually match the CG.

The pipeline: previs render (driving video) -> image-model edits of selected previs frames (anchors) -> video model in reference mode (driving video + anchors + prompt).

---

1. Shade the previs for AI readability

Image and video models follow the CG far better when they can read its form:

• Wireframe-over-clay shading on the CG subjects beats flat silhouettes. A black or unlit subject only constrains outline - the model invents form, orientation and lighting. Wireframe lines communicate topology and pose.
• Tint the subjects a flat saturated color (e.g. red) when they risk blending into the environment (grey subjects on concrete). Then name that color in the prompt: "replace the red CG horse with..." - it gives the model an unambiguous edit target.
• Keep the environment photoreal if you have it (projected plate or stills); the contrast between "obviously CG subject" and "finished environment" focuses the edit.

2. Pick anchor frames at event beats

• Anchor at each subject's first CLEAN appearance (fully in frame), plus the moments the composition changes (second subject enters, action peaks).
• The shot's first frame works as a no-cost anchor if it is empty of subjects: use the raw plate/render directly - it pixel-locks the environment.
• Never anchor on a frame where a subject is partially out of frame. Image models refuse half-objects at frame edges and "complete" them into full subjects in the wrong place - repeatedly, regardless of prompt language. Slide to the nearest fully-entered frame instead.

3. Build an action-pose reference set

The single biggest cause of generations ignoring the CG pose: reference images whose pose or orientation contradicts the shot action. One standing reference will pull every generation toward standing, even over an obviously mid-stride CG input.