INITIALIZING BUNKROS IDENTITY LAB
LOC UNDERGROUND
SYS --:--:--
Learning / Video Generation

The hardest medium.

AI video is a shot machine, not a film machine. Every frame has to agree with the last about gravity, faces and physics — and the longer the clip, the more that agreement breaks down into morphing and warping. This class teaches why video is exponentially harder than images, and how to direct it like a shoot.

You'll be able to

Explain the limits

Say why long, complex shots break — temporal memory.

You'll be able to

Direct a shot

Prompt for camera, motion and physics, not just a subject.

Watch for

Artifacts & rights

Warping, the uncanny valley, and unclear usage rights.

01 — Warm-up

True or false?

"To make a 30-second AI video, you should ask for the whole 30 seconds in one long generation." True or false?

Self-Check · True / False
02 — Landscape

The state of AI video

The field is fast and specialised — no single best model, just a best one per shot. Clips are usually a few seconds, increasingly with native audio. Your real job is generating many short shots and editing them into a sequence — exactly like a film shoot. Version numbers turn over monthly, so learn what each lab is good at, not its current number.

All-round + audio

Veo-class

Strong all-rounders, some generating synchronised dialogue and sound, not just silent footage.

Control

Runway-class

Reference images, explicit camera control and character consistency — the studio-workflow favourite.

Value

Kling-class

Cheap, high-volume iteration with lip-sync and native dialogue.

Physics / social

Luma / Pika-class

Physically-accurate motion and HDR (Luma); fast effects and lip-sync for social (Pika).

03 — Foundation

Why it's so hard

A still image only has to look right once. A video has to look right across time — every frame agreeing with the last. Two limits make this brutal.

Limit 1

Temporal memory

The model has limited memory of earlier frames. The longer the clip, the more it "forgets" — so faces drift, objects morph, and backgrounds warp.

Limit 2

Physics

It learned statistics of how pixels move, not actual physics. Water, hands, hair and reflections are where it most often breaks.

Core insight

Artifacts aren't random — they're predictable. They grow with clip length and motion complexity. Short shots with simple motion stay coherent; long, busy shots fall apart. The Coherence Lab below lets you feel exactly that.

04 — Interactive

The coherence lab

Each tile is a "frame" of one clip. Push clip length and motion complexity up and watch later frames drift, warp and sprout artifacts as the artifact-risk meter climbs. Keep it short and simple to stay coherent.

Sandbox · Watch coherence break down
Artifact risk: low. Short, simple shots hold together.
05 — Craft

Prompting motion

A video prompt directs a shot, not a subject. Build it from six slots, and the most controllable path is image-to-video — lock the look in a still, then add motion.

Anatomy

Six slots of a shot prompt

1 · Subject · 2 · Action (what moves, how) · 3 · Camera move & framing · 4 · Lighting · 5 · Style/lens · 6 · Physics + duration + aspect + "stable, no warping".

Weak: "a car driving, cool." Strong: "matte-black car on a wet road at dusk, low side tracking shot, slow dolly, golden-hour rim light, cinematic 35mm teal-orange, realistic spray, ~5s, 16:9, stable motion, no warping."
Workflow — image-to-video

1. Lock a hero still (composition, character, wardrobe). 2. Upscale it. 3. Load as the start frame. 4. Describe only the motion. 5. Keep it short, generate several takes. 6. Select, upscale, edit. Shoot coverage; build the film in the cut.

06 — In the Wild

A teaser in a day

A small collective needs a 20-second teaser for a community film night.

Before

One long prompt

They asked for the whole 20 seconds at once. Faces morphed, the logo melted, hands turned to soup — unusable.

After

Short shots, then cut

They storyboarded 5 short shots, used image-to-video from locked stills with directed camera moves, generated several takes each, then graded and cut in an editor.

Result: a clean, cohesive teaser — built in the edit, not one generation.
Why it matters · rights

Coherence came from short shots and coverage, not luck. Before posting, they checked the model's commercial terms and avoided any real person's likeness without consent.

07 — Research

Scout the models

Video models move fast. Find which one fits a specific need and note its clip-length limit.

Search for

What to look up

"best AI video model for character consistency" · "image-to-video tool camera control" · "AI video clip length limit".

Judge the source: recent comparisons and the makers' docs; check the date.
Your findings
Saved locally on your device.
08 — Assignment

Direct a shot

Write one shot prompt using all six slots, then describe in one line how you'd assemble a 15-second piece from short shots.

Your shot prompt + assembly plan
Saved locally.
Success rubric — tap to expand

Needs work: a bare subject and one 20s prompt.
Getting there: a decent shot prompt but no assembly plan.
Solid: six-slot prompt + a short-shots plan.
Excellent: a directed six-slot prompt with a stability note, plus a plan to generate coverage and finish in an edit with a unifying grade.

09 — Retrieval practice

Quiz — 5 questions

Question 1 · Medium
Why do longer AI clips degrade?
A
The internet connection slows down.
B
Limited temporal memory — the model loses track of earlier frames, so things drift and warp.
C
Longer clips use a worse model automatically.
Question 2 · Medium
Which workflow gives the most control over the look?
A
One long text-to-video prompt.
B
Random generation and hope.
C
Image-to-video — lock the still, then add motion.
Question 3 · Easy
How should you make a 30-second piece?
A
Generate several short shots and cut them together in an editor.
B
One 30-second generation.
C
Slow the playback of a 5-second clip.
Question 4 · Medium
Which is most likely to produce artifacts?
A
A static wide landscape.
B
Close-up hands manipulating water with fast motion.
C
A slow push-in on a single object.
Question 5 · Easy · True / False
"Raw AI clips are final deliverables — no editing or rights check needed." True or False?
T
True
F
False
10 — Avoid these

Common mistakes

Trap

One long generation

Asking for the whole scene at once. Fix: short shots, assembled in an edit.

Trap

Over-described motion

Cramming five movements into one shot. Fix: one clear movement per shot.

Trap

Fresh text-to-video for continuity

Expecting a character to match across generations. Fix: image-to-video from a locked reference.

Trap

Shipping raw clips

No grade, no rights check. Fix: finish in post, grade to one look, clear the rights.

Debug your thinking
11 — Reflect

Shot machine, not film machine

Now that you can see artifacts grow with length and motion, how does that change what you'd attempt with AI video — and where you'd still want a real camera?

Private journal
Saved locally to your device.
Well done

You understand video's hard limits and how to direct around them. Last stop: turning all these skills into real-world business value.

12 — Reference

Glossary — precise words

Temporal memory
How well a model tracks earlier frames; limited memory causes drift over long clips.
Artifact
A visual error — morphing, warping, extra limbs — that grows with length and motion.
Text-to-video
Generating a clip from a written prompt; fastest, least controllable.
Image-to-video
Animating an approved still; the most art-directable path.
Character consistency
Keeping a subject's identity stable across shots, via references.
Camera control
Directing pans, tilts, dollies and tracking shots via text or tools.
Coverage
Generating extra shots/angles so the edit has options.
Uncanny valley
The unsettling effect when near-realistic humans look subtly wrong.
Upscaling
Raising a clip's resolution after generation.
Colour grade
Matching colour/contrast so mixed clips share one look.
Storyboard
A shot-by-shot plan made before generating.
Commercial rights
Terms governing paid use of generated video; check before shipping.