In the trenches, hands-on as it happens.

cat .verdict TOOL Google Whisk SCORE ███████░░░ 7/10 FREE 3 generations · experimental · shared with Flow PAID Google AI Plus / Pro / Ultra OUTPUT8s clips · GIF 270p + MP4 720p BEST concept iteration · storyboard-style clips · guided prompts with LLM help AVOID high-seed creative mode · complex anatomy · precise-mode without good prompts TESTED2026-02-23 · 3 generations
cat .tools-manifest MODELS drwxr-xr-x Kimi K2.5 Moonshot Free # Prompt engineering drwxr-xr-x ChatGPT OpenAI Free/paid # Image generation + prompts PLATFORM drwxr-xr-x Google Whisk Google Experimental (3 generations) # Subject-scene-style video drwxr-xr-x Google Flow Google Free (50 credits/day) # Text-to-video comparison ─────────────────────────────────────────

Google Labs has another experimental project: Whisk. Available in G1 countries, you get three generations to try it out.

nanni-complaint-clip.mp4
Whisk · Nanni of Ur complaint · 8s · subject-scene-style approach

What is Whisk?

Google Whisk sits under the Google Labs FX umbrella alongside Flow. Where Flow uses text prompts and reference images as “ingredients,” Whisk tries a different approach: three structured inputs.

From their FAQ:

Behind the scenes, the Gemini model automatically writes a detailed caption of your images. It then feeds those descriptions into Google’s latest image generation model.

This is a convenience layer — previously you might ask an LLM to describe an image and re-feed it. Whisk automates that step.


The three-ingredient model

Whisk asks you to provide:

  1. Subject — What the image is about. Characters, objects, or combinations. A mysterious renaissance vampire. A cardboard movie display. Your own face as a directional reference.
  2. Scene — Where the subjects show up. A fashion runway. A pop-up holiday card. You can bring characters in or swap them.
  3. Style — The aesthetic, material, or technique. Clay-documentary. Cinematic noir. Hand-drawn animation.

You can provide text or images for each. Whisk generates intermediate images from your inputs, then animates.


Example 1: DJ in the Arctic

First attempt: a DJ putting on a show near a glacier.

The prompt

Rather than using the three-ingredient structure, I started with a single detailed description from ChatGPT:

ai@hackers:~$ user
describe the movements of a DJ on stand mixing the beats and rhythmic movements, then the look of their local area and stage, and the feeling of the audience perfect description for an image generation LLM to create a great scene, arctic near a glacier, but a female DJ with headphones on scratching a beat
ChatGPT assistant

A cinematic wide shot on an arctic shoreline at blue hour: towering glacier walls in the background, fractured ice glowing cyan from within, slow mist drifting across a frozen fjord. In the foreground, a female DJ stands behind a minimalist, frost-rimmed DJ booth carved from translucent ice, subtle LED light embedded inside the ice casting soft neon blues and violets through the crystal structure. She wears over-ear headphones tilted to one side, wind lifting a few strands of her hair, breath faintly visible in the cold air.

Her pose is dynamic and rhythmic: left hand on the vinyl, fingers splayed as she scratches the record, right hand riding the crossfader with precise motion. Shoulders angled forward, knees slightly bent in a groove, head nodding to an unheard beat. A slim laptop sits beside the turntable, waveform glow reflecting across the ice surface. Cables run beneath the booth, partially frozen into the frost.

Lighting is dramatic and atmospheric: low arctic sun just below the horizon creating a magenta-to-deep-blue sky gradient, aurora beginning to form faint green ribbons above. A soft key light from the booth illuminates her face with cool tones, while a subtle rim light from the glacier edge outlines her silhouette. Fine snow particles drift through the air, catching light like glitter.

Style: ultra-realistic, high dynamic range, cinematic composition, shallow depth of field on the DJ with glacier softly blurred, crisp ice textures, volumetric fog, dramatic color contrast between neon booth lighting and natural arctic blues, 50mm lens look, high detail, motion energy frozen in a single frame.

Comparing outputs

I ran the same prompt through multiple tools to compare:

ChatGPT image generation:

chatgpt-dj-result.webp
ChatGPT generated DJ in arctic scene
ChatGPT image generation·2026-02-22

Google Nano Banana (Gemini’s image model):

nano-banana-result.webp
Nano Banana generated DJ scene
Nano Banana · Google's image model·2026-02-22

Whisk generated images:

whisk-dj-v1.webp
Whisk DJ variant 1
Whisk variant 1·2026-02-22
whisk-dj-v2.webp
Whisk DJ variant 2
Whisk variant 2·2026-02-22

Whisk generates two versions. I would have preferred more variation — temperature or other settings might be available, but the interface doesn’t surface them clearly.

Animate and Refine

After generating images, “Animate” and “Refine” are prominently visible. I animated with the same prompt, removing the “frozen in a single frame” part:

whisk-dj-clip.mp4
Whisk · DJ in the Arctic · 8s · best result of the comparison

Nice beat. Best result overall — decent adherence to the theme, stuck to the concept nicely.

For comparison, here’s Flow’s text-to-video with the same prompt:

flow-text-to-video.mp4
Flow · DJ in the Arctic · text-to-video comparison

The library approach

whisk-library.webp
Whisk library interface
Whisk library · everything in one place·2026-02-22

The library has everything you need — so it seems optional to skip the previous steps. Or include any of them. Let’s see what it does from text-to-video to compare against Flow.

Actually, it has what we were originally looking for:

whisk-surprise.webp
Whisk surprise discovery
Found it — the subject/scene/style inputs·2026-02-22

It was missing our video, telling us to animate a picture to get started — which resolved on a refresh. So let’s try again with the three-ingredient approach.


Example 2: Nanni of Ur — the 3,750-year-old refund request

This time, let Kimi K2.5 set up the three ingredients properly.

The brief

ai@hackers:~$ user
Ea Nasir the infamous copper swindler. 'What do you take me for, that you treat somebody like me with such contempt?' — Nanni of Ur • ~1750 BCE • World's Oldest Customer Complaint -- complaining of ea nasir bad copper. We need a specific subject, scene, and style for Whisk. It should be funny, brief, and reflect bad copper. Ancient mesopotamia, inscription on tablet of his shoddy ingots. Consider an aggrieved man hammering away the infamous tablet with some real bad copper around him as he angrily complains.
Kimi K2.5 assistant

Narrative Direction: “The 3,750-Year Refund Request”

We’re visualizing the moment Nanni of Ur realizes he’s been scammed—not just as a static historical tablet, but as the active creation of the world’s first angry Yelp review. The comedy emerges from the anachronistic intensity of his consumer rage: a Mesopotamian merchant losing his mind over quality control while physically surrounded by the evidence—copper ingots so oxidized they’re literally crumbling into green dust as he gestures at them.


Subject

Nanni of Ur – A middle-aged, slightly sweating Mesopotamian merchant in authentic woolen kaunakes (sheepskin skirt), sporting an exasperated expression that transcends millennia. He’s actively engaged in hammering/inscribing cuneiform into a damp clay tablet with a reed stylus, occasionally stopping to gesture violently at the offending merchandise.

The Evidence – The physical manifestation of “bad copper”: prominently displayed ingots that are visibly oxidized to a crusty turquoise-green, misshapen with rough casting flaws, and actively shedding flaky corrosion.

The Infamous Tablet – The clay complaint tablet itself, positioned where we can see the wedge-shaped impressions being pressed into the damp surface, permanently capturing phrases like “such contempt” and “substandard copper.”


Scene

The Merchant’s Courthouse in Ur, c. 1750 BCE – An interior courtyard of a mud-brick residence with reed-mat ceilings, where Nanni has set up his “dispute station.”

  • A low wooden counting table covered in merchant tools: balance scales (visibly tipped, as if failing to measure up), additional clay envelope-tablets, and accounting tokens scattered in frustration
  • The “Wall of Shame” corner: the bad copper ingots piled haphazardly on a woven reed mat, catching shafts of dusty sunlight that illuminate their embarrassing green oxidation
  • Architectural details: dried mud-brick walls with bitumen waterproofing, a sliver of the Euphrates visible through a doorway, clay storage vessels in the background
  • Atmospheric dust motes dancing in warm, golden-hour light filtering through woven screens—emphasizing the clay-heavy environment and the “ancient” texture of the moment

Style

Cinematic Clay-Documentary – A hybrid of historical accuracy and mockumentary comedy that treats this ancient complaint with the gravitas of a modern true-crime investigation.

  • Color Palette: Warm terracotta dominance (clay reds, sand browns) violently contrasted with the sickly oxidized-copper green of the bad ingots—visual shorthand for “promised bronze, got garbage”
  • Texture Emphasis: Extreme close-ups capturing the tactile quality of damp clay being impressed with cuneiform, the crumbly, flaky surface of corroded copper, and the coarse wool of Mesopotamian garments
  • Depth & Focus: Shallow depth of field keeping the tablet and Nanni’s frustrated face razor-sharp while the background bad copper blurs into bokeh-green abstraction
  • Lighting: Dust-filtered “golden hour” warmth that suggests this is a moment being preserved in amber (or clay), with dramatic side-lighting that makes the cuneiform shadows pop off the tablet surface
  • Comedic Framing: Slightly low angles on Nanni to give him underdog protagonist energy, with reaction-shot style cuts implied between his face and the offending copper

Whisk Integration Note: When prompting Whisk, emphasize the action—Nanni mid-gesture, pointing at the bad copper while simultaneously inscribing the complaint, capturing that specific moment where ancient commerce and eternal frustration collide.

Setting up the ingredients

I wasn’t sure whether to let Whisk generate images from text or provide my own. Tried both:

Subject (ChatGPT generated):

nanni-subject.webp
Nanni of Ur subject image
Subject · ChatGPT generated · Nanni the aggrieved merchant·2026-02-23

Scene (Whisk generated from text):

whisk-scene-mesopotamia.webp
Mesopotamian courtyard scene
Scene · Whisk generated · merchant's courthouse·2026-02-23

Style (from Flow — Whisk’s was awful):

flow-style-reference.webp
Cinematic clay-documentary style reference
Style · Flow generated · cinematic clay-documentary·2026-02-23
whisk-processing-spinner.webp
Whisk processing animation
Whisk processing animation·2026-02-23

The video prompt

ai@hackers:~$ user
best prompt for the video generation 6-8 seconds
Kimi K2.5 assistant
Cinematic close-up tracking shot: A sweaty Mesopotamian merchant in a woolen kaunakes angrily hammering cuneiform into a damp clay tablet with a reed stylus, periodically stopping to gesture with theatrical contempt at a pile of visibly oxidized green copper ingots actively crumbling beside him, warm golden-hour light filtering through reed mats in a mud-brick courtyard, dust motes dancing in shafts of light, shallow depth of field emphasizing the flaky texture of bad copper versus smooth clay, comedic mockumentary style, 6-8 seconds of escalating ancient consumer frustration

Result

I put that in with my ingredients. Whisk produced two images that were a mash-up of the inputs, then asked which to animate. It’s unclear if the ingredients were actually used or if it just did text-to-video from context — but the results suggest they were:

nanni-complaint-clip.mp4
Whisk · Nanni of Ur complaint · subject-scene-style · 8s

Precise vs Creative mode

whisk-google-example.webp
Whisk precise vs creative mode documentation
From Google's FAQ · use an LLM for better precise mode results·2026-02-22

Precise mode was already on for my generations. Let’s try Creative with a different seed:

whisk-seed-settings.webp
Whisk seed settings
Seed 62343 · what does this do exactly?·2026-02-23

Seed 369420 — Creative mode:

whisk-three-hands.webp
Three hands high seed creative mode
High seed + Creative = three arms and psionic anger-powers·2026-02-23

These were bad. Let’s not animate these.


What you get


Verdict

No perfect videos out of the box. With structured prompting via an LLM and the subject-scene-style approach, you get useable concept clips.

Works well:

  • Storyboard-style concept clips
  • Quick iteration when you have good reference images
  • Structured prompts from LLMs (Kimi, ChatGPT)

Plan around:

  • Creative mode + high seeds = anatomy disasters
  • Interface can be confusing — ingredients don’t always surface clearly
  • Only 3 generations to experiment with
  • Style generation can be weak — consider providing your own

For direct text-to-video with more daily capacity, see Google Flow hands-on. Both tools feel like two Git worktrees spawned from the same codebase — similar capabilities, slightly different interfaces.