Changing Color and Material with Omni Flash in Google Flow: Three Tiers, One Prompt Pattern

The episode makes one claim up front: recoloring and resurfacing sound trivial, yet reflection physics splits the job into three distinct difficulty tiers. The host demonstrates the split hands-on with the Omni Flash model inside Google Flow, in the fourth video of the series. The frame is what makes this installment valuable in my eyes: not a demo reel, but a map of what is safe and what is risky.

The recolor pattern is a single sentence: name the object, name the new shade, done. The first trial turns a basketball purple, and the result reads clean at first glance. My note of caution: a lone subject and barely any motion is the model's comfort zone, so this test sets the ceiling, not the average.

The convincing trial is the red car: several cuts from different angles all shift into a deep matte black finish under one command. Landing it on the first attempt, with no extra credits spent, explains why color sits in the safe column. Holding one shade across edited angles is the detail that makes this daily-usable.

Before surfacing comes a detour into light: kill the studio lamps and the image collapses, and the host argues generated video obeys the same law. Out of that come three tiers: non-reflective surfaces such as wool, paper, rust, and marble are safe; light-passing ones such as glass and ice carry medium risk; mirrors such as chrome, gold, and mercury are the riskiest. That ladder steers every test that follows.

In the safe tier the red car becomes rusted metal first, then wool, each in a single command with strong results. The wool pass leaves a slightly toy-like feel, though shape and motion stay intact. My reading: the texture swap convinces, the physical weight does not quite, yet for social output it is more than enough.

The light-passing tier starts with the same car turning into thick clear glass, consistent if a step below the rust and wool passes. The hard trial is a watermelon: its flesh is asked to become blown glass with water visible inside, and the water never fully persuades. The lesson generalizes beyond this clip: pile surface, content, and physics changes onto one request and something gives; fewer reflections, more consistency.

The mirror tier is the weakest link. The basketball turned into polished chrome looks fine frozen, but once it moves the reflections read fake and the court markings warp. Trials pushing the watermelon toward polished gold and then flowing molten gold, both with water inside, repeat the pattern: passable water, reflections that lag the action. The model repaints the surface but cannot refresh the surrounding light frame by frame.

The closing verdict comes with a number: recoloring lands nearly every time, the first two tiers hold, mirrors have limits. The host points to a document with every command used and to earlier volumes linked from the cards. My takeaway is operational: ship color jobs straight to production, and set expectations for surfacing jobs by tier.