Blog Art et Visuels IA : La créativité à l'ère numérique
WAR COMES Visual Creator Lab: Leonardo AI Teaser Dry Run #1
Creator Lab case study: Three Leonardo AI opening tests, one controlled start-frame variable, two songs to serve, and a documented decision process for building the WAR COMES teaser without spending the story reveal too early.
WAR COMES Visual Creator Lab: Leonardo AI Teaser Dry Run #1
This is not a tutorial built from a perfect final result. It is a controlled Creator Lab documenting how one opening shot is tested, compared and revised before it earns a place in the WAR COMES teaser.
The project begins with a clear story constraint: Jack Righteous receives a vision from God and sees War approaching through the Four Horsemen. The visual must create dread before spectacle, and the first rider reveal must still matter when it arrives.
The output matters. The decision process matters more.
Across the JR Creator Labs, the strongest work follows the same discipline: define the job, protect what must not change, isolate a meaningful variable, compare against a baseline, record the evidence, and make a decision. This WAR COMES dry run uses that same method for AI video.
That puts this case study beside the logic already used in the Suno Studio 2 Creator Lab, Controlled Parameter Testing Lab, AI Music Creative Direction Lab, and AI Song Revision Lab.
Five disciplines are being applied to one five-second test.
For this lab, the work is being evaluated through five expert lenses rather than one generic “does it look cool?” reaction.
Does the shot protect the reveal?
The image must support the prophetic vision without spending the red-horse moment too early.
Did the prompt control the right variable?
The model should not be asked to reinvent composition, character, motion and lore at the same time.
Does the camera create dread?
Negative space, slow movement and restrained light should carry the opening before action appears.
Can the world remain consistent?
Start frames and later character references reduce drift across a sequence built from short clips.
Can another creator learn from the test?
The article records the input, variable, result and decision instead of only presenting the winner.
Does it fit the timeline?
A beautiful clip is still wrong if it arrives before the audio and story have earned it.
One opening has to serve two different pieces of music.
1:44 teaser
The teaser begins near silence and moves through suspicion, warning, bass impact and the final message: “Dis nuh di war. Is di warnin’.” The opening therefore needs mystery before identification.
Main song · first 0:44
The main song moves through “Mi nah run.” → “One rider.” → “Red horse.” → “War.” Early imagery must leave room for those reveals to become progressively stronger.
Lab success condition: Clip 1 should work under the teaser first, while remaining reusable in the main song without forcing an early story rewrite.
What cannot be lost while testing
- Prophetic, not generic warfare. The scene is a vision from God, not random battle footage.
- Jack is the witness. He should not become a conventional action hero.
- War arrives on the red horse. That reveal must remain available for later clips.
- The other Horsemen stay back. Showing all four immediately destroys the progression.
- Black, restrained crimson, ash, storm and negative space. The visual world should feel serious and cinematic, not bright fantasy or sci-fi.
Dry Run #1 changes one major thing at a time.
The baseline asked Leonardo Wan 3.0 to invent the world from text. The next test added a start frame. The third kept the same start frame and changed the prompt toward restraint. That gives the comparison actual value.
| Test | Start frame | Prompt strategy | Question |
|---|---|---|---|
| 01-A | No | Text-to-video baseline | Can the model invent the right opening world from text alone? |
| 01-B | Yes | Original broad prompt | What changes when composition is anchored? |
| 01-C | Same frame | Rewritten for restraint | Can motion and mystery improve without changing the visual anchor? |
The dry-run configuration
| Model | Wan 3.0 |
|---|---|
| Clip length | 5 seconds |
| Speed | Standard |
| Format | 9:16 vertical |
| Resolution | Full HD 1080 × 1920 for start-frame tests |
| Generated audio | Off — the existing music is the timing source |
| Private Mode | On |
| Fixed seed | Off during exploration |
| Negative prompt | No separate field was visible in this Wan 3.0 interface, so exclusions were written into the main prompt. |
“JR Universe” is meaningful to us, not to the model.
A private world name is not a visual instruction. The usable prompt had to translate the concept into things the model can render: cracked black earth, ruined iron, huge storm clouds, drifting ash, restrained crimson light, one distant horse and a supernatural warning atmosphere.
This is the same principle used in the Build Your Own Universe: Story Intake Lab: story identity becomes more usable when it is converted into documented constraints and repeatable creative decisions.
No start frame: strong world, weak control.
The first generation proved the broad direction could work. Leonardo found cracked ground, red fissures, storm clouds and a dark horse presence. But it also revealed the weakness of asking text-to-video to solve too many jobs at once.
The horse became too clear too quickly, and the model introduced a hooded person before Jack was supposed to enter.
Partial fit
Atmosphere works, but the image gives away too much for the “Hear war? / Nah yet.” opening.
Weak without recutting
The early person and horse reduce the impact available for “One rider / Red horse / War.”
Decision: do not lock. Preserve the world language, change the control method.
Anchoring the composition changes the experiment.
The second test uses the approved start frame as the visual anchor. That removes one major source of randomness: Leonardo no longer has to invent the entire world before it can animate it.
The improvement is not simply “better art.” The test becomes more controlled because composition is held constant while Leonardo is asked to solve motion.
Decision: keep the start-frame method. Revise the prompt once before choosing the clip.
Same frame. Tighter prompt. Better story discipline.
The third generation keeps the horse small and distant, preserves the empty road and storm-cloud negative space, avoids introducing Jack, and lets the red horizon intensify without turning the scene into a battle.
The revised motion prompt
Strongest so far
It supports the “not yet” idea: threat is present, identity is withheld.
Strong candidate
Jack, the rider and the red horse still have room to arrive as separate escalating beats.
Current decision: 01-C is the leading Clip 1 candidate. Do not call it final until it is tested on the actual timeline.
What changed — and what we learned
| Generation | Variable | What improved | What failed | Decision |
|---|---|---|---|---|
| 01-A | Text only | World and mood were viable | Premature reveal; invented character | REVISE METHOD |
| 01-B | Add start frame | Composition became intentional | Prompt still encouraged too much storytelling | KEEP FRAME / REVISE PROMPT |
| 01-C | Same frame, restraint prompt | Mystery, negative space, delayed reveal | Still needs timeline test | CURRENT LEADER |
Do not optimize an AI generation in isolation.
- Start with the job. What must this clip accomplish in the larger piece?
- Protect what matters. Identify the story, sound or brand value you refuse to lose.
- Change one meaningful variable. Otherwise you cannot explain why the result improved.
- Compare against a baseline. A new result should beat something specific, not merely feel novel.
- Judge in context. Put the clip against the song, scene or sequence before calling it successful.
- Record the decision. KEEP, REVISE, RESTORE or ROUTE — then know why.
This is the practical connection between AI visual creation and the wider JR Creator Labs: creation becomes more useful when experimentation produces evidence, not just outputs.
Clip 2 changes the problem: Jack enters the vision.
The environment is now established. The next job is character continuity. Jack Righteous can enter for the “Mi nah run” reuse beat: hooded, face hidden, standing as witness rather than action hero.
That means the next lab variable should not be another free text-to-video attempt. It should be a locked Jack reference still followed by image-to-video, while the Clip 1 environment language remains protected.
Use the layer that matches the blocker you actually have.
Free Creator Labs: this case study is already useful as public learning. For the transferable video workflow, continue to the AI Video Generation Guide or return to the Free Creator Labs front door.
Optional DEVELOP / specialist depth: use the Story Intake Lab when world boundaries and story decisions are the blocker, or the Controlled Parameter Testing Lab when experimental control is the blocker. These are deeper support layers, not required graduation steps for the free curriculum.
Already have JR access: go through Member Home so your existing Creator Library, focused road, Creator Pro, Complete Access or legacy entitlement can route you correctly. If your access explicitly includes project review and one judgment call is still blocking progress, use Member Project Guidance.
The article will evolve with the production.
Each future WAR COMES test should be added using the same record: brief → protected value → variable → prompt/reference → result → comparison → decision → next action.
That keeps the work useful beyond this one teaser. The finished video is one outcome. The repeatable creator process is the larger asset.
Create What You Love | Love What You Create.
Dry Run #1 · August 28, 2026 · Leonardo AI · Wan 3.0 · JR Universe / WAR COMES · Platform interfaces and model behavior can change.
Build visual recognition
Make the project easier to recognize and understand.
Connect covers, visual content and identity to the platform that gives the work a permanent home.
Discussion