Our Google Flow Workflow: One Clear Action per 8-Second Scene

Early scene plans often read well as a sentence: a character enters, notices an object, turns, reacts and moves toward a new destination. Inside an eight-second generation, however, that sentence becomes several anatomy, timing and camera problems. FourFeetz compared those overloaded plans with scenes built around one observable action and a clear final pose.

The simpler scene did not have to be motionless. HARU could take a few careful steps while breathing and blinking, or RURU could peek out and make one small retreat. The difference was that every supporting movement belonged to the same story beat. The final second no longer had to rush through an unfinished reaction merely to satisfy the written list of events.

When a scene combined a large posture change, travel across the frame and a camera move, legs and paws were more likely to lose structure. Faces could also drift during a fast turn or after a brief exit from the composition. Even when the middle looked energetic, the last frame often landed on an unbalanced pose that was difficult for Extend to continue.

Unplanned camera motion made the same problem harder to diagnose. A push-in or orbit changed character scale, viewing angle, background geometry and light at once. Stabilizing the finished clip could reduce shake, but it could not restore the original body proportions or reconstruct the environment that had changed during the move.

FourFeetz now defines the first eight seconds with one primary action, one end state and a restrained camera. A character may perform the action once or repeat a small motion twice when that repetition is natural, but the shot does not introduce a second major event. The fixed or gently controlled camera leaves more of the generation’s effort available for the character and makes the result easier to compare with the First Shot.

The action is written as a continuation of visible evidence. The approved environment, light and character appearance are not reinvented in a long block of new description. The instruction identifies what happens next, what remains still and where the performance should settle. This protects continuity without exposing the studio’s private master prompts or internal settings.

Extend works best when the last frame contains usable evidence: a balanced body, a readable face, visible limbs and a camera that has completed its move. The next instruction continues the same action or begins one modest follow-up beat. It does not ask for a new location, a dramatic angle change and a different body pose at the same time.

The join is checked at normal speed, without sound and frame by frame. FourFeetz compares face, age, ears, paws, accessories, background lines, shadow direction and camera speed across the handoff. If Extend returns a smooth motion with a different character, the take is rejected. Visual continuity is part of the performance, not a finishing detail.

This workflow did not make every generation usable. Some clips began correctly and developed a different face near the end. Others kept the character but added unwanted camera drift or left a paw in an unstable position. FourFeetz recorded the first failure, then decided whether the usable portion could be trimmed, whether the action needed to be simplified, or whether the entire shot required replacement.

CapCut supports the final assembly, timing, short transitions and loop checks, but it is not treated as a repair tool for broken anatomy. A clean edit can connect two valid clips; it cannot make an invalid body correct. The practical lesson is to give each eight-second scene a job it can complete and let editing refine successful material rather than conceal failed generation.