desktop[ library ]communitycoursesorbits / membership
<- back to libraryworkflows_
[ notebook ]2026-08-04

Lock the still. Move one action. (MiniMax H3)

I did not win MiniMax H3 by prompting harder

I locked the still first. Then I moved one action. Then I fixed one variable.

That is the whole craft story.

Conduct: Terra ran the desk

This run was orchestrated in Codex. I call that session Terra.

Terra did not paint the final frames alone. It ran the chain:

  1. Build still references with GPT Image 2.
  2. Lock the vehicle with front and rear identity plates.
  3. Send short, controlled motion jobs to MiniMax H3.
  4. Review, kill, correct, keep the dailies.

The intelligence is the stack and the loop. Not a single lucky prompt.

The locks (conduct + five craft locks)

LockWhat it does
ConductOne orchestrator (Codex / Terra) owns the chain
Product platesFront + rear identity so the black performance SUV survives every angle
World keyframesGPT Image 2 locks composition, light, and world
One-action motionH3 gets one verb and one camera or subject move
One-variable fixWhen it breaks, change only the broken axis
DeliveryKeep native 21:9, 2944 by 1248, 24 fps, H.264

What MiniMax H3 felt like

H3 is shot-scoped, not film-scoped.

  • Strong image refs beat a giant sequence prompt.
  • Narrow shot language beats a multi-beat screenplay in one box.
  • Pretty mid-frames lie. Check first, mid, and end of every take.
  • Identity can hold for a Defender-class silhouette and still lose a rear spare or paint lock. That is a QC problem, not a vibe problem.

I am not selling H3 as magic. I am reporting what held under an automotive identity test in ultrawide.

Proof strip (raw H3 selects)

These are raw selects, not a finished ad with approved music.

  1. Ice wall: product clarity under cold light.
  2. Glacial river: kinetic water and body solidity.
  3. Controlled data rupture: front half solid, rear dissolves on purpose.
  4. Basalt arches: world scale at 21:9.
  5. Reality exit: rear geometry + spare held.

Fail to fix (one variable)

First splat attempt tried too many state changes at once. Camera jumped. Body lost.

Fix: lock the camera. One controlled dissolve from the rear. Front stays solid metal.

That single constraint is the difference between a morph accident and a readable thesis beat.

Delivery truth

I did not invent an upscale story for this pack.

What shipped from the motion model:

  • 21:9
  • 2944 by 1248
  • 24 fps
  • H.264

Keep the ultrawide. Do not treat a phone crop as the creative act.

There is an assembled 30 second study in the archive. It is reference only. Music and voiceover experiments are not selected.

What I am not claiming

  • Not manufacturer work. Independent AI production proof.
  • Not "one prompt, whole commercial."
  • Not a finished sound mix.
  • Not a mystery finish pipeline. This post is stills to H3 to QC to native delivery.

Closer

Lock the still. Move one action. Fix one variable.

Lock the still. Move one action. Fix one variable.

Deep dive: the desk, the locks, the QC

The skool cut is the craft thesis. Below is the same run with desk views, identity plates, fail pairs, and native delivery proof.

G1 · Conductor

One orchestrator owns the chain. Still build, plates, motion, and QC do not negotiate ownership mid-run.

Conduct · Terra. One desk owns every stage.

G2 · Still stack

Lock composition and product identity before any motion job leaves the desk.

Still stack. Keyframe plus front and rear identity.

G3 · One-action card

H3 gets one verb and one camera or subject move. Multi-beat language is the failure mode.

One-action card. One verb. One move. Native 21:9.

G4 · Correction dial

When a take breaks, change only the broken axis. Do not re-roll the whole brief.

Correction dial. Only the broken axis moves.

G5 · Fail to fix · splat

First splat tried multi-state change. Camera jumped. Body lost. The fix is one controlled dissolve with a locked camera.

Fail to fix · splat. One variable: locked camera, one rear dissolve.

G6 · 21:9 delivery

Keep the native ultrawide. Phone crop is not the creative act.

21:9 delivery. Native H.264. Crop is damage.

Lock 1 · Product plates

Front and rear identity plates keep the black performance SUV stable across angles. Without them, H3 invents rear geometry and loses the spare.

Front identity plate of the black performance SUV
Plate front. Identity lock before motion.
Rear identity plate of the black performance SUV with spare tyre
Plate rear. Spare and rear geometry must survive every take.

Lock 2 · World keyframes

GPT Image 2 locks composition, light, and world before H3 sees a motion job. The ice wall keyframe is the still the first select answers to.

Ice wall world keyframe still used as motion reference
World keyframe · ice wall. Composition and light locked first.

Lock 3 · What H3 held under test

Shot-scoped models reward narrow language and strong refs. Under an automotive identity test in ultrawide:

  • Product silhouette held when plates were present.
  • Pretty mid-frames still failed end-frame QC on spare and paint.
  • Multi-beat prompts collapsed body or camera. One-action cards held.

That is the operating envelope I am reporting. Not a brand claim for the model.

Lock 4 · Fail pairs (splat + exit)

Splat: multi-state fail → one dissolve fix

First splat tried too many state changes at once. The fix is one variable: camera locked, rear dissolves, front stays metal.

FAIL · multi-state splat
FIX · controlled one dissolve

Exit: no spare fail → rear geometry held

Identity can look solid in a mid-frame and still drop the rear spare. QC first, mid, and end.

FAIL · exit without spare lock
FIX · reality exit, rear + spare held

Thumbs for the same pairs:

Thumb of multi-state splat fail
fail splat multi
Thumb of controlled splat fix
splat controlled
Thumb of exit fail without spare
fail exit no spare
Thumb of reality exit with spare held
reality exit

Lock 5 · Delivery restatement

What shipped from the motion model, without an invented upscale story:

  • 21:9
  • 2944 by 1248
  • 24 fps
  • H.264

Keep the ultrawide. Crop to 16:9 is a damage zone on this pack, not a refinement.


These are raw selects, not a finished ad.

01 · Ice wall

Product clarity under cold light.

Thumbnail of ice wall select

02 · Glacial river

Kinetic water and body solidity.

03 · Controlled data rupture

Front half solid. Rear dissolves on purpose.

04 · Basalt arches

World scale at 21:9.

05 · Reality exit

Rear geometry + spare held.


Assembled study

There is a 30 second corrected conform in the private archive. It is reference only. Music and voiceover are not selected. That file is over the public site asset size cap, so it is not embedded here. The raw H3 selects above are the public proof.

What I am not claiming

  • Not manufacturer work. Independent AI production proof.
  • Not "one prompt, whole commercial."
  • Not a finished sound mix.
  • Not a mystery finish pipeline. This post is stills to H3 to QC to native delivery.

Lock the still. Move one action. Fix one variable.

[ comments ]

guests welcome. members show first in the list

guests

  • Lights, Camera, ACTION

    guest · Wed, 05 Aug 2026 02:50:12 GMT