KI-Visuals für Kreative: Erstellen, Abgeben & Skalieren

WAR COMES Creator Lab: What Production Run #2 Taught Us About AI Video

Published August 28, 2026Last updated August 28, 2026By Gary Whittaker
What this guide will help you do

A JR Creator Lab case study on start-frame design, prompt translation, model comparison, motion control, reveal timing, and when a high-cost 30-second AI video generation is justified.

JR Creator Labs · WAR COMES · Production Run #2 Lessons

When AI Video Stops Being a Prompt Test and Starts Becoming Production

Production Run #2 taught us that better AI video does not come from writing longer prompts. It comes from controlling what the model is allowed to invent.

WAR COMES began with a simple problem: introduce Jack Righteous into the approved visual world for “Mi nah run” without letting the model spend later story beats too early. That one shot forced us to work through start-frame design, prompt translation, motion control, model comparison, reveal timing, cost, and eventually a full 30-second production decision.

The production problem

The model was creating visually good footage that told the wrong story.

Our first moving Jack candidate looked cinematic, but Jack began walking toward the threat. That changed the meaning of “Mi nah run” from refusal to retreat into early confrontation. The distant threat also became too ambiguous and briefly read more like a giant humanoid than War mounted on horseback.

Creator Lab lesson: a generation can look good and still be wrong. Evaluate what story information the clip spends, not only whether the image is attractive.
Playable production test

First Production Run #2 Wan attempt

Training note: preserve this clip as evidence of a strong visual result that accidentally advanced Jack's confrontation beat too early.

Lesson 1 · Start-frame design

Fix ambiguity in the image before asking motion to fix it.

The first Jack reference gave the video model too much freedom. Jack's stance could be interpreted as ready to walk, while the distant threat was not structurally clear enough to guarantee “rider on horseback.” We rebuilt the start frame instead of endlessly adding corrections to the motion prompt.

Before

Too much interpretation

  • Jack could plausibly step forward.
  • The distant threat was vague.
  • The model had to infer horse-versus-humanoid structure.
After

Production information baked into frame 1

  • Both of Jack's feet visibly planted.
  • Horse-and-rider silhouette structurally readable.
  • Threat smaller and farther away.
  • Reveal budget protected by composition itself.
Actual start frames

Reference still comparison

WAR COMES Production Run 2 initial Jack Righteous reference start frame
Initial Jack reference start frame.
WAR COMES Production Run 2 optimized start frame with Jack Righteous and distant rider on horseback
Optimized frame: planted Jack, smaller distant mounted threat, clearer horse-and-rider structure.
General rule: when the first frame already contains the information the model needs, the prompt can spend more of its attention on motion.
Lesson 2 · Prompt translation

Translate creative meaning into things a video model can see and move.

Human creative language such as “peace-first freedom fighter,” “War remains protected,” or “spiritual and digital tension” is useful for the production team, but not precise enough for a generation model. We translated those ideas into posture, scale, anatomy, motion, lighting, distance, and exclusions.

Creative direction Model-readable instruction
Jack stands his ground Both feet stay fixed on the cracked ground for the entire shot.
War remains protected One small human-sized rider remains seated on one horse, far away, approximately the same scale and visibility.
Do not create a giant Readable horse head, neck, body and four-legged base with a human-sized rider mounted above.
Peace-first resolve No attack pose, no raised weapon, no dramatic turn; calm upright posture and subtle breathing only.
Digital/spiritual overlap Restrained distant interference in haze/light plus ominous cloud and light behavior—not interfaces or neon effects.
Prompt coaching: do not ask the model to understand your lore. Ask it to render the physical evidence of your lore.
Lesson 3 · Reveal control

AI video models naturally want something to happen.

Across multiple generations, the mounted threat repeatedly moved closer even when the creative job was simply to establish Jack. That is useful behavior when the story needs escalation, but destructive when the shot needs restraint.

For WAR COMES the reveal order matters:

Beat What is earned
Mi nah run Jack exists in the warning world and refuses to retreat.
One rider The mounted threat becomes readable.
Red horse The horse's identity is visually confirmed.
War The confrontation is unmistakable.
Production principle: movement is story information. A horse moving closer, Jack taking one step, a face becoming visible, or fire appearing are not merely animation choices—they spend narrative beats.
Lesson 4 · Same brief, different models

The “best” video model depends on the job.

We kept the optimized start frame and core shot assignment substantially consistent and compared several models. The purpose was not to crown a universal winner. It was to see which model best handled stillness, world continuity, human motion, mounted-threat structure, cinematic quality, and reveal restraint.

Model What we observed in this shot Production use
Wan 3.0 Accelerated Corrected horse/rider structure but compressed the reveal quickly; threat became dominant. Useful evidence, not the Clip 2 benchmark.
Wan 3.0 Standard Better control and strong continuity, but the rider still advanced more than the shot required. Strong visual candidate; reveal rate too aggressive.
Seedance 2.5 Best balance of planted Jack, restrained camera pressure, distant threat, negative space and story control. Selected leader for “Mi nah run.”
Kling 3.0 Turbo Strong cinematic polish and continuity, with more natural escalation toward the mounted threat. Potentially more useful for a later “One rider” reveal.
Hailuo Very strong atmosphere and mounted-threat readability, but the rider increasingly became the subject. Strong challenger and possible later-shot model.
Controlled model comparison

Wan 3.0 Accelerated

Controlled model comparison

Wan 3.0 Standard

Five-second leader

Seedance 2.5

Controlled model comparison

Kling 3.0 Turbo

Controlled model comparison

Hailuo comparison

Creator Lab lesson: do not ask “which model is best?” Ask “which model's natural behavior best fits the creative job I need right now?”
Lesson 5 · Use model bias instead of fighting it

A failure for one shot can reveal the right model for another.

Seedance's restraint made it the leader for Jack's stillness. Kling and Hailuo naturally increased the rider's presence. That was a weakness for “Mi nah run,” but may become a strength for “One rider.” The goal is not to force every model into identical behavior. It is to understand what each model wants to do and assign it work accordingly.

Workflow improvement: save strong rejects. Label why they failed the current shot. They may already contain the behavior required by the next shot.
Lesson 6 · Cost changes the testing strategy

A 30-second hero generation is not the place to discover the brief.

Once Seedance became the five-second leader, we moved to a full 30-second, 9:16, Full HD 1080×1920 generation. Leonardo showed a cost of 21,540 credits for that single run.

At that cost, the long generation should happen only after the short tests have already answered the expensive questions: character posture, world continuity, rider structure, reveal order, and model selection.

Long-run gate

  1. The short-form model test has a clear leader.
  2. The start frame is production-ready.
  3. The story progression is written in timed stages.
  4. Character identity and scale rules are explicit.
  5. The 30-second run has multiple potential edit uses.
  6. We know what would justify spending another 21,540 credits—and what would not.
Lesson 7 · The hero sequence changed the story again

Production should continue developing Brand and Voice—not just visuals.

The full 30-second sequence introduced a new creative decision: Jack begins to burn with supernatural fire as War approaches. That fire is not decoration. It represents pressure awakening something darker inside the peace-first Jack Righteous identity.

For now, the fire can imply the Lion/war-response aspect without rendering Lion as a separate character. That preserves unresolved canon while allowing the visual production to deepen the character.

Sound → Voice → Brand → World → Visual: when a new visual idea changes what the character means, treat it as creator development. Check it against the upstream identity—not only the VFX quality.
Production Run #2 winner

Seedance 2.5 · 30-second hero sequence

The winning sequence gives us the intended escalation: Jack holds → the city shows signs of war → the mounted rider becomes readable → the dark red horse emerges → controlled fire awakens around Jack → confrontation becomes inevitable.

Production record

What Production Run #2 actually delivered

  • An optimized Jack start frame designed for motion control.
  • A model-readable prompt strategy built from visual evidence instead of abstract lore terms.
  • A controlled comparison across multiple current video models.
  • A clear reason Seedance 2.5 became the leader for the “Mi nah run” beat.
  • Reusable evidence showing why visually strong generations can still fail story timing.
  • A 30-second hero generation created only after the expensive uncertainties were reduced.
  • A new character-development decision: fire begins to visually signal Jack's darker war-response side.
Use this on your own project

The JR Creator Lab method

  1. Define the shot job. What new information is this shot allowed to give?
  2. Build the start frame. Put identity, scale, composition and reveal limits into frame 1.
  3. Translate meaning into visuals. Describe posture, anatomy, distance, movement, lighting and exclusions.
  4. Test short first. Use inexpensive clips to expose model behavior.
  5. Compare the model, not five different briefs. Keep the assignment controlled.
  6. Record strong failures. A rejected behavior may belong later in the story.
  7. Spend credits only after uncertainty drops. Long generations should produce assets, not discover fundamentals.
  8. Return upstream when needed. If a visual changes Voice, Brand or World, treat that as creator development and make the decision deliberately.
The goal is not prompt perfection. The goal is a repeatable decision process that gives the model less room to make creative decisions you wanted to make yourself.
Where this learning goes next

Choose the next route by the blocker—not by the size of the access package.

Free Creator Labs: the public lesson is complete enough to apply without buying anything. Continue with the AI Video Generation Guide when the next problem is visual production, or return to Free Creator Labs when the project has moved back into Sound, Voice, Brand or another stage.

Optional specialist depth: if controlled comparison itself is the repeated blocker, use the Controlled Parameter Testing Lab as a deeper implementation example. If story/world boundaries are drifting, use the Story Intake Lab. Neither is required simply because you finished this article.

Already have JR access: use Member Home to resume the focused road, Creator Library, Creator Pro, Complete Access or legacy entitlement you already own. Do not repurchase a public or standalone resource just because it appears in this case study. If your current access explicitly includes consultation and a specific judgment remains unresolved, use Member Project Guidance.

Creator Labs connections

Keep the principle connected to the larger creator road.

This visual production case sits downstream of decisions made in Sound, Voice and Brand. Use those routes when the generation problem is actually an upstream creator-development problem; use the video route when the creator decisions are already clear and the blocker is visual execution.

Create What You Love | Love What You Create.

Production Run #2 media record. The stills and videos embedded here are the actual assets used in this Creator Lab comparison.

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

Leave a comment

articleall levels
On this page

    Your next move

    Turn the reading into useful work.

    Apply this now

    Complete one action before opening another guide.

    Write down the most important decision this article changes, then apply it to the project while the reasoning is still fresh.

    Continue learning

    Keep the subject connected.

    Use the public library to compare related guidance before changing the project.

    Continue with public guidance →
    Go deeper

    Use structured training for ordered work.

    Move into the member system when the project needs a sequence, templates and application—not another isolated tip.

    Explore structured training →
    Use a resource

    Support the next action.

    Use a workbook, checklist or ASK JACK route only when it reduces friction in the work.

    Open the supporting route →

    The Righteous Beat

    Get the week’s most useful creator guidance, platform changes and free resources.

    Join the free newsletter →