challenges to test the limits of AI-generated video — and your skills
- 107challenges
- 3collections
- 21sets & series
- 28skill tests
- 5difficulty levels
New challenges land regularly. The board only grows.
This is a growing board of production challenges, running from single shots all the way up to complete sequences and scenes. Each one isolates something specific that generative video finds hard: an animal's anatomy, the weight of a body landing, a face that has to survive sixty unbroken seconds, a crowd that refuses to melt into itself.
Every challenge arrives fully specified. A difficulty rating so you know what you are walking into. A set of Skill Tests naming exactly what is going to break. Success criteria so you can tell a near miss from a real hit. And a starter prompt, built with our MediaMax system, to get you off the blank page — though the prompt is a starting point, not an answer. The interesting work begins the moment you start changing it.
Run them alone and your craft gets sharper. Run them against your friends and you find out who is actually better. Either way, when you land one — or fail in an instructive way, which is often more useful — show us. This board grows with what the community makes.
How a challenge works
Read the brief
What to make, and why it defeats models today.
Check the Skill Tests
The specific technical difficulties this shot puts in front of you.
Start from the prompt
A MediaMax-built starting point. Then change it, and change it again.
Judge it honestly
Success criteria tell you whether you actually landed it.
Signature Sequences
The founding board: iconic sequences, the natural world, sport, street, and shots no camera could take.
Iconic Sequences
5 challengesStudy the shots that rewrote the language of cinema — then rebuild the technique, not the copy.
The Walk
A man walks down a street and somehow it becomes the most famous entrance in cinema.
Whip, Freeze, Name
Introduce six characters in ninety seconds without losing the audience once.
The Steps
A crowd descends a monumental staircase. Montage was invented here.
One Unbroken Take
Follow two people through a kitchen, down a corridor, and into a nightclub. Never cut.
The Plan
Two people talk through a plan. The camera and the cut do every single joke.
The Natural World
5 challengesAnimals and water. The two things generative models get confidently, spectacularly wrong.
The Breach
Forty tons leaves the water, hangs, and comes back down. Get the physics right.
Mother and Calf
The single hardest animal in generative video. Nobody gets the calf right.
Lunge Feed
The largest biomechanical event on Earth, and the throat has to expand correctly.
Not a Dugong
A manatee grazing seagrass. The tail is the whole test.
Murmuration
Thousands of birds behaving as one body. Emergence, not choreography.
Bodies in Motion
5 challengesSport is biomechanics under load. Every joint that bends the wrong way is instantly obvious.
Out of the Park
Pitch, swing, contact, gone. Four biomechanical events in two seconds.
Top Corner
Strike, flight, net. The ball has to deform on the boot.
The Spiral
A tight spiral, a contested catch, and a run to the end zone.
Nothing But Net
Rise, release, rotation, swish. The backspin is the giveaway.
Serve and Return
A serve, a slide, and a return. Two players, one exchange, no cheating the physics.
Street Culture
3 challengesReal people, real weight, real places. The hardest thing to fake is ordinary.
Impossible Shots
4 challengesShots no camera could ever take. This is where AI stops imitating film and starts adding to it.
Inside the Raindrop
Start inside a falling drop of water. Pull back until it is a city.
Through the Keyhole
The camera passes through a keyhole. On the other side it is a different season.
The Fold
A city folds onto itself — and the light and gravity stay honest.
Sixty Years, No Cut
One unbroken shot. A person ages sixty years. The face never becomes someone else.
Actions
Behaviour instead of poses. Every challenge asks one question: could this body be doing this if it did not know it was being filmed?
Derived from the CEMI MediaMax production canon — cemi.media
Warm-Ups
3 challengesStart here. One subject, one action, one room — the simplest useful things on the board, and the fastest way to learn what your model does when you ask it for almost nothing.
Surprise Without the Pose
5 challengesThe AI default for surprise is the catalogue pose: hands to head, mouth open, frozen. Real surprise is a task interrupted.
The Piece Finally Fits
Three failed tries, then it clicks. The face is the last thing to find out.
The Double Take, Earned
Walk past it. Keep going two whole steps. Then come back.
Found It
The search is real — wrong pockets first. The find interrupts, it does not culminate.
The Slow Reveal
Pull the cloth off — but let it snag for half a second.
Eureka, Interrupted
The idea lands while the coffee is still pouring. The pouring stops mid-stream.
Connection Without Presentation
6 challengesTwo-person actions where the model angles both faces at the lens. Real connection excludes the camera.
The Whispered Secret
We never hear it. We see it in the listener’s hands.
Hands Guiding Hands
Experienced hands cover a learner’s — then let go one finger at a time.
Building Together
Two people place the last pieces at almost — but not quite — the same instant.
The Shared Breath
Two people finish the same effort and recover at different rates.
The Long Hug
Held past the comfortable cut point. One hand pats twice, then just holds.
Mirror Learning
Shown once, copied wrong, laughed at, got it. Then both accelerate.
Anticipation as Behaviour
4 challengesThe most promptable emotion, because it is inherently task-shaped. The best place to start.
The Countdown
Three, two… and the last finger hesitates, savouring it.
Drumming Fingers
On a table edge, in a real rhythm, accelerating.
Rubbing Palms
Before starting — with a specific glance at the thing about to be started.
The Peek
One eye first. The body stays hidden longer than the head.
Object Craft
6 challengesThe product-shot antidote. AI objects float, gleam and weigh nothing; hands are supposed to have opinions.
Weight Test
Pick it up and feel the heft. The wrist dips before it rises.
The 360 Inspection
Turning it over, fingers find a seam — and keep coming back to it.
Giving It Life
A hand walks a figure, flies a plane, rolls a car. The face is not invited.
The Offering
Held out toward someone — but the eyes stay on the object.
Scale Measure
Held against a hand, a pocket, a shelf it has to fit. Will it fit?
The Final Piece
Placed with the care of someone who has already dropped it once.
Celebration With Weight
5 challengesVictory is where AI goes full stock photo: symmetrical jumps, both fists, white teeth. Gravity fixes it.
The Small Victory
One quiet fist. Then straight back to work.
The Landing
A real jump has a landing: knees bend, balance catches, a step to recover.
The Toast With Whatever
Mugs, glasses, bottles — none of them matching, mid-conversation.
High Five, Missed and Corrected
It half-misses. They laugh. They try again.
Chain Reaction
One person’s celebration sets off the next — each in their own style, each a beat late.
The Rhythm Set
4 challengesActions built on musical time — the hardest thing for these models, and the most rewarding to land.
Body Percussion Duet
Thigh slaps, snaps, chest taps — traded back and forth like a conversation.
The Teachable Step
An eight-count step designed to be copied. The second run is faster.
Freeze Dance
Freeze mid-motion, weight visibly mid-transfer, and hold it for two seconds.
Humming in Canon
Two people alternate phrases, then join. The audio leads; the video obeys.
Camera as Participant
5 challengesThe operator is a character. AI cameras anticipate perfectly, and that perfection is a deep synthetic tell.
The Late Camera
The action starts. The camera reacts 300ms behind and has to catch up.
POV of the Object
The world from the object’s eyes as it is carried, rolled or thrown.
The Follow
Track it at its own height — and lose it behind something for half a second.
Rack to the Listener
Mid-sentence, focus leaves the speaker and finds the person listening.
Over the Shoulder, Walking Away
One glance back mid-stride — without breaking stride.
Inclusion as Craft
5 challengesRepresentation normalised: the person first, the attribute as context. Never an inclusion moment.
The Chair Spin
A celebration spin with the same narrative weight as any other victory.
Eyes That Point
Direct someone’s attention using nothing but gaze and a micro-nod.
Tactile Exploration
Hands reading an object thoroughly — and returning to the flaw they found.
Sign That Lands
Signers converse. They do not demonstrate.
Mutual Boost
Both bodies loaded, in the same frame. No rescuer, no rescued.
Cultural Specificity
5 challengesThe anti-generic set: actions only a specific community performs correctly. An open, rotating series — one community at a time, written with specificity instead of generality, and open to correction from the people who live it.
The Greeting That Locates You
A greeting performed at conversational speed — never demonstrated.
Dancing While Doing
Two steps of a local dance, absent-mindedly, in the middle of a chore.
The Gesture Lexicon
A gesture that is doing work — never one that is being displayed.
The Hands Keep Talking
Gesture that carries grammar, not gesture that punctuates words.
Palms, Then Feet
The greeting changes register automatically. The body decides before the face does.
Transitions Performed by Bodies
4 challengesEditing devices executed physically — where action meets montage.
Exit / Enter Frame
Leaves frame right, arrives frame left somewhere else — matched on stride.
The Handoff Out of Frame
Passed to someone outside the frame. The next shot receives it.
The Mid-Action Freeze
Frozen at its least stable instant — a composition nobody could pose.
Mirror Movement
The same action in two places. The cut argues they are connected.
Visual Effects
An effect may amplify physics, the world, or a graphic identity — it may never caption emotion.
Derived from the CEMI MediaMax production canon — cemi.media
The Redemption Series
8 challengesTake the most clichéd AI effects and earn them. Each entry names the cliché, then the condition that redeems it.
Magic Dust, Earned
Floating golden motes are the number one AI cliché. Make them answer for themselves.
The Glow That’s a Lamp
Auras are banned. Light wrapping a person is redeemed when it comes from something.
Lens Flare With a License
One flare. Motivated. Imperfect. Then gone.
Confetti With Consequences
Celebration particles that then EXIST: they land, they stick, someone sweeps them.
Sparks That Burn
Only from materials that spark — and somebody flinches.
Rain That Wets
Darkened shoulders, plastered hair, drops on the lens that STAY.
The Snow Test
Pick ONE consequence — accumulate, track in, or melt — and nail it completely.
Fireworks From Inside
Never show the sky. Show the room receiving it — and let the sound arrive late.
World Effects
7 challengesPhenomena that belong to the scene’s physics. The shape is always the same: the effect obeys the world.
Heat Shimmer at the Source
Air distortion strictly above the heat source. Nowhere else.
Breath in Cold Air
Puffs timed to real speech and effort — drifting with the established wind.
Reactive Fog
The person edits the effect: fog parts at the shins and settles after.
The Dust Ring
A hard set-down raises a ring — and gravity kills it in under a second.
Steam That Talks
Density is the clock. Fresh or forgotten — the steam says which.
Water Caustics, Located
Ripples on the ceiling ONLY where water and light geometry justify them.
The Window’s Weather
A whole storm told by what one pane of glass does.
Time & Camera Devices
8 challengesPrecision challenges. The failure is always the device applied as a filter instead of triggered by an event.
The Speed Split
Foreground in normal time, background in fast-forward — with one believable boundary.
The Hold
The camera stops dead. The world does not care.
Freeze With a Pulse
Action stops at its least stable instant. The camera keeps pushing in.
The Ramp With a Reason
Slow motion triggered by a physical event — and out again when it is over.
Bullet-Time, Humble
A 90° orbit, not 360°. Restraint reads more real than the full trick.
The Through-Solid Move
Straight through a wall into the next room — and the inside of the wall is designed.
The Match-Move Morph
One scene becomes another mid-shot, while a continuous element anchors the eye.
Strobe Trail
Multiple exposures decaying behind a movement — each dimmer, each later.
Graphic Layer
6 challengesDesigned, not sprinkled. Overlays are legitimate when they are an identity the piece commits to — one pen, one physics.
The Doodle Contract
One pen, one physics, obeyed for the entire piece. Mixed styles are template soup.
Thought Made Visible
The drawing exists only in their point of view — never omnisciently.
The Annotated World
Labels pinned to real 3D positions — with parallax, and something passing in front.
Rotoscope Commitment
A traced-outline look sustained for a full sequence. Consistency across motion is the test.
The Reality Fold
One signature transition, used exactly twice, identically.
The Graphic Match Cut
No overlay at all. Two real shots joined by a shape they genuinely share.
Colour as Effect
4 challengesWhere grading becomes storytelling. Every choice needs a reason the audience can read without being told.
The Selective World
One object keeps its colour — redeemed only by a point of view that justifies it.
Temperature Arc
Warm to cold across 30 seconds — caused by a light-source event, not a slider.
The Two-World Grade
Intercut locations the audience can tell apart by grade alone. No captions.
Colour Arrives With Contact
Saturation blooms from a touch point — and obeys the geometry of the thing it is on.
No challenges match that combination. Try widening the filters.
Ran a challenge? Show us what you got.
Share your attempt — the triumphs and the glorious failures. The most instructive entries get featured on the board.