Creative Teams

July 21, 2026

We Put Gemini Omni Flash To The Test

# Creative generation

# Video generation

See The Difference It's Made For Our Creative Process

Chung Puo Chong

If you’re generating videos as often as we do, you would inevitably encounter the same obstacles we have: hands that morph mid-motion, or a liquid that moves like it's forgotten its own weight. Far from being edge cases, these are the shots that show up constantly in real production briefs, and are difficult to get right.

So when Omni Flash landed on Pencil, the team couldn't wait to put it through its paces. We ran it across ten shots that, until now, seemed to resist every attempt we made. Here’s how well Omni Flash performed each time.

What we didn't expect was how much of the troubleshooting disappeared. Far less time was spent re-generating, and in some cases, our shot came out right on the first try. Gone are the days where we had to pick apart individual frames trying to explain to a client why a hand looked wrong. We can now spend our time iterating on ideas instead of fighting the tool.

A quick note: every clip here started from a pre-generated still image, with the prompt below applied to animate it into a video. As the image prompt has no bearing on how Omni Flash handles the motion, we're only showing the video prompt, as it is the part actually being tested.

Unwrapping Candy

Wrapper physics have to get one specific thing right: twisted cellophane needs to visibly untwist, not tear. The material has to stay translucent and catch light as it deforms, and whatever's doing the unwrapping needs to stay anchored to the part of the wrapper actually being manipulated.

In our past attempts at this shot, wrappers would detach or tear rather than untwist. Sometimes it would be left hanging lopsided and seemed to behave more like tissue than cellophane — stretching and thinning before snapping, but never unwinding cleanly. In some cases, the wrapper would just stay twisted and sealed no matter how much the fingers moved.

Prompt: Animate the hands unwrapping the candy. The twisted ends untwist, the wrapper crinkles and folds naturally, and the candy is revealed inside without changing shape. Keep the wrapper texture and candy position consistent.

Google Gemini Omni Flash Result:

The twist ends untwist rather than tear. The wrapper stays intact and translucent the whole way through, both catching and reflecting light as it creases, while the candy stays centered and undistorted. It's also the only one where the grip stays plausibly anchored to the twist points for the full motion, rather than drifting away or dislodging itself from the wrapper.

Toothpaste Squeeze

Viscous fluid extrusion is a deceptively layered test. There's the physics — the paste has to hold its shape under gravity while still flowing, without collapsing into a drip or freezing into a rigid worm. There's the behavioral layer too: notice how when you apply toothpaste, you do so in a dragging motion across the brush head, and not simply dispense it at a single point.

Toothpaste squeezing would trip us up in one of two ways. First, the paste would behave wrongly, emerging as a thin, gravity-pulled bead rather than holding its viscous shape. It might also maintain an unnaturally consistent ribbon width with no natural taper. Second, the application logic would not match real life behavior, with the paste being piled onto a single static point on the brush with no spreading or dragging motion.

Prompt: Extreme close-up of toothpaste being squeezed from a tube onto toothbrush bristles, smooth controlled extrusion, product holding shape naturally, clean bathroom lighting, macro commercial realism.

Google Gemini Omni Flash Result:

The toothpaste curls forward with a believable thick, heavy consistency, tapering naturally as the flow loses momentum. Toward the end of the clip it visibly separates cleanly from the nozzle under its own weight. This small but crucial detail that sells the whole motion as one continuous, physically consistent event.

Chocolate Pour

Thick pours have a lot to get right. The flow has to respond to whatever it's landing on. Where food is involved, crumbs and chip pieces should redirect it, not get ignored. Also, viscous liquids don't behave like thin ones. They extend in a continuous ribbon rather than snapping apart, and the pour has to mind its own volume, visibly depleting as it continues.

We've managed to get the chocolate flowing previously, but something else would always break the visual harmony. Some streams thinned to a hairline and snapped before reforming, or stayed confined to one narrow channel instead of spreading across the surface it's landing on. In other cases, the pour seemed infinite, never depleting its source no matter how long the clip ran. Even the color would drift unnaturally from brown to a reddish or pink tone.

Prompt: Extreme macro shot of glossy melted chocolate being slowly poured onto a dessert, rich flowing texture, realistic viscosity and reflections, cinematic food commercial lighting, shallow depth of field.

Google Gemini Omni Flash Result:

The pour visibly reacts to what it's landing on, with crumbs and chip pieces holding pockets of chocolate back while the rest flows through open space. The covered area widens naturally as the pour continues, speeding up or slowing depending on where the spoon is. The color and consistency stay stable and correct for the entire clip.

Finger Snapping

A snap is a small, specific mechanical action — the thumb and middle finger have to make contact and release in one motion. It's easy for a model to substitute a generic hand gesture, but getting it to perform the action is something that’s never seemed possible.

This is a shot that consistently failed at depicting contact. If it wasn’t the thumb and finger hovering close without closing the gap, then it’d be the fingers posing or drifting between unrelated shapes with no coherent motion holding it together. Elsewhere, hands would alternate between a loose fist and an open palm instead of snapping.

Prompt: Medium shot of a person's hand snapping their fingers naturally.

Google Gemini Omni Flash Result:

Thumb and middle finger actually meet and release in a repeating motion consistent with a real snap, with audio in sync with the contact points. Getting the mechanics and the sound to agree with each other is what makes it appear as a real gesture rather than an approximation of one.

Deodorant Application

Product application shots need three things to agree with each other: the contact point, the motion, and the target area. If the applicator touches the wrong surface, drifts off-target, or moves in a way that doesn't match how the product is actually used, the whole shot falls apart.

Some combination of the contact point, the motion, or the target area would land wrongly in our past generations of deodorant application. We often noticed a visible product residue left behind that a solid stick shouldn't leave. These were often paired with diagonal dragging motions instead of a smooth rub, or a contact point drifting entirely off the applicator's own surface, sliding down past the armpit toward the collarbone or side of the torso.

Worth noting more broadly: across all of our video generation testing prior to Omni Flash, models defaulted to a dabbing motion rather than a wiping one, despite clear prompting.

Prompt: Close-up of a person applying stick deodorant to their underarm, natural arm raise and smooth product roll on glide, realistic skin movement and compression, clean commercial lighting, grounded realism.

Google Gemini Omni Flash Result:

Application stays confined to the armpit, controlled and gentle, without straying into surrounding skin. For once, the application motion actually comes off as a wipe rather than a dab.

Keyboard Typing

Typing is a rhythm problem as much as a motion one. Each finger has to travel a short distance, land on the correct key, and release at a tempo that reads as purposeful rather than searching. Realism is compromised with just a small mismatch between where a finger is and which key visibly depresses.

Previous models would project cadences that felt paused and not fluid, with fingers hovering mid-air between presses instead of holding a rhythm. Non-typing fingers would arc high above the keys in a motion that resembled playing a piano; and keys would be visibly depressing with no finger anywhere near them.

Prompt: Close-up of fingers typing naturally on a laptop keyboard, realistic key depression and finger rhythm, soft monitor light reflections, shallow depth of field, grounded realism.

Google Gemini Omni Flash Result:

Finger motion looks natural and varied, matching the tempo of relaxed, real typing rather than a looping animation. Each finger travels a short, purposeful distance and lands convincingly on its key. There is just one small exception: a key appears to depress without a clearly corresponding finger contact at the start of the clip.

UGC Vlog

Talking-head UGC footage lives or dies on whether the camera feels human. Handheld footage naturally drifts and jerks in step with the speaker's own energy. Too smooth and it comes across as locked-off. Too uniform and it suggests a shake effect applied after the fact, rather than an actual hand holding a phone. Format compliance matters here too: UGC content has to be usable on the platforms it's built for.

We've seen this lean too far in one direction or the other before. Unconvincingly smooth camera motion would break the illusion of the handheld nature of the filming. This would also lead to framing being so locked and centered that it looked more like a stabilized shot rather than something being held by a human hand. Some models were even unable to support the basic 9:16 portrait format that defines UGC content, or failed to produce any audio at all.

Prompt: Influencer woman speaks directly to camera in a casual handheld-style shot, natural blinking and head movement, subtle handheld phone shake, authentic TikTok-style realism, casual indoor lighting. She is talking about her day.

Google Gemini Omni Flash Result:

Camera motion syncs to the beats of her speech. The motion and jerks land more pronounced on the moments spoken with more force, settling elsewhere. It shows the difference between a camera that's shaking and one that's reacting.

Pastry Kneading

Rolling dough is a continuous cause-and-effect problem. Pressure from the roller should visibly widen and flatten the dough with each pass, the stroke should alternate back and forth, and the hands need to maintain a believable grip throughout.

A different link in the chain breaks each time we've attempted this generation. The hand might fully release the pin mid-motion with fingers splayed with no contact while the pin keeps moving on its own; or the pin would visibly gather more and more flour with continued rolling while the dough beneath narrows instead of widening.

Prompt: Animate the hands rolling the pin forward across the dough. The dough should flatten gradually, flour should shift subtly, and the rolling pin should rotate naturally as pressure is applied.

Google Gemini Omni Flash Result:

From frame to frame, the dough visibly widens and flattens with each pass, and the stroke alternates convincingly in both directions. Pressure and shape change stay causally connected the whole way through, the way real dough behaves under a roller, with the hands maintaining a continuous, believable grip throughout.

Lip Balm Application

Application shots are only as good as their contact accuracy. The right surface of the product needs to make contact with the right area, with the right pressure, and the right coverage. In this case, the lip balm needs contact across both lips, not just one, and it's the product's rounded tip that should be touching skin, not the packaging around it.

We've only gotten pieces of this right before. Contact with the lips would come in with too much force for a soft balm application. We also noticed application motions would move in small halting increments rather than one continuous sweep, and the touch point would move from the balm itself to the flat edge of the tube's cap as the clip progressed. Application always seemed to be confined to the lower lip, never making it to the upper one.

Prompt: Animate the hand applying lip balm across the lips in one smooth motion. The balm glides naturally, the lips move subtly, and the product remains aligned with the hand. Avoid mouth distortion or unnatural hand movement.

Google Gemini Omni Flash Result:

Tracked across the full clip, the balm makes contact across both the upper and lower lip as it moves with the actual product (not the packaging) doing the work throughout. The final frames show a believable post-application lip press with a habitual motion that comes right after applying balm.

So, It’s Pretty Good Right?

We think so, and not just because the outputs looked better. Across all nine shots, cause and effect was never broken, Pressure would shape what was touched, contact actually happens, and motions would hold up from the first frame to the last.

This consistency allowed us to close the gap between an idea and producing something client-ready. Landing a usable shot is no longer a long-drawn and hours-long affair.

But it's not flawless. The recurring abrupt zoom/framing jump is worth highlighting, and there's still real value in testing your specific use case before committing to it for production. Although ten scenarios isn't enough to make a definitive verdict, they're common failure points, so we think it's a fair snapshot.

We'd like to hear from you. Did it hit the mark, miss it, or land somewhere in between? Drop your results in the community page and let us know what else it's good for.