Google Flow Complete Guide: Directing HARU as a Multi-Shot Production

This page asks one production question: can reference-led multi-shot production in Google Flow support a result where separate HARU shots feel like one directed scene after they enter the edit? The hard part is keeping identity, screen direction, camera height and afternoon light coherent across independently generated shots. It does not claim a universal winner, score, success rate or fixed production time.

The review boundary is beat-sheet intent, shot cards, motion candidates, rough-edit continuity and final approval were reviewed as separate stages. Anything beyond the files, frames and public project links shown here is a question for a future project, not a completed FourFeetz test.

For Google Flow Complete Guide, the repository contains the featured HARU image and the character reference shown below. They support a discussion of composition, identity and review criteria; they do not by themselves prove hidden settings, generation counts or a tool-wide benchmark.

The comparison anchor is the approved HARU reference, beat sheet, shot cards and neighboring frames in the rough edit. This keeps the page tied to observable material instead of borrowing evidence from an unrelated character article.

The material is most vulnerable when a beautiful clip changed camera height, reversed screen direction and raised the scene energy before the story was ready. A polished frame is not enough if it cannot preserve the intended character, action or edit relationship.

The appropriate response is returning to the shot card, removing simultaneous camera and subject changes, and rebuilding the sequence around one clean HARU action. This is a review decision, not a claim that every model or project will behave the same way.

Flow's Extend workflow is most useful when an approved clip reaches the right moment but needs more screen time or a new story beat. Instead of regenerating the opening, continue from a stable end frame so the next clip inherits the current subject, environment, light and camera relationship. Treat the extension as a new shot decision, not as free extra seconds.

For the HUGO and HARU meadow sequence, the safest extension point was a frame with readable silhouettes, planted feet and no fast camera move. We described only the next action, kept the camera behavior modest and generated short alternatives. Each result was placed immediately after the source clip to check motion cadence, screen direction and whether the cut seam became visible.

Extend can preserve useful context, but it does not lock every detail. Character age, anatomy, collar placement, background geography and light can drift as the generated duration grows. A clean first second can also hide a failure near the end, so review the full continuation frame by frame before approving another extension.

The main production risk is compounding drift. A slightly different face or body proportion may look harmless in one continuation, then become the new starting point and grow more obvious in the next. Return to the last approved source whenever identity or anatomy changes; do not ask a later extension to repair an unstable earlier one.

Motion continuity needs the same attention as appearance. If the source ends during a step, head turn or camera push, the extension must respect direction and speed. Generating from a calm hold is usually more predictable than continuing from motion blur, crossed legs, occlusion or a whip pan. Preserve extra handles around the seam so the editor can trim through the cleanest transition.

Budget by accepted seconds, not generated seconds. Extensions that pass identity review can still fail because they offer no cut point, repeat an action, change the landscape or move the light source. Save the source clip, extension prompt, model setting, date and rejection reason together so another attempt changes one known variable instead of restarting the search.

The public method is build the sequence outside the generator, approve ingredients, request one shot function at a time and judge every result in the edit. The useful controls are approved ingredient references, one shot function per generation, locked screen direction, rough-edit test before finishing. They are planning prompts rather than private presets or guaranteed settings.

The Google Flow Complete Guide example below demonstrates instruction order only. It deliberately omits master prompts, character-bible details, weighting logic and private production configuration.

This Google Flow Complete Guide article does not establish a numerical rating, a winning tool, a repeatable success percentage or a fixed budget. Tool access, model behavior, credits and terms can change, so those facts must be checked in the live service before production.

A future test should keep the approved HARU reference, beat sheet, shot cards and neighboring frames in the rough edit fixed, change one relevant variable and record whether the result can enter the intended edit. The page remains useful only while its claims stay inside that evidence boundary.