[AI튜토리얼] AI 영화, 이 순서대로 만들면 됩니다 | 메카닉 영화 단편 '제로' 제작기

Summary of the “Zero” AI Mecha Short Film Production

The video outlines an end-to-end workflow for creating an AI-generated mecha short film titled “Zero.” It emphasizes that the biggest determinant of quality is scenario structure, supported by high-detail “character sheets” covering design, costumes, props, and space consistency.

The creator describes iterating through dozens of revisions, using an agent/workflow with near-semi-automation, and repeatedly testing outputs at multiple resolutions to catch errors early.

A major theme is copyright and guideline compliance. The creator notes that copyright issues can influence design choices even when visible differences (e.g., weapon/color changes) exist.

---

Artistic Techniques / Creative Processes Shown

Storyboarding & Breakdown

• Break a short segment (e.g., a ~40-second clip) into time slices (noted: 0.3-second intervals).
• Plan the scenario from the resulting frames.
• Use storyboards as reference, but avoid producing the final AI narrative directly from storyboard frames alone.

Scenario Writing Using “Drama / Screenwriting Formulas”

• Build the script with scene-by-scene setups, payoffs, and creephangers.
• Keep the writing concise:
  • Overly long/detailed drafts can blur priorities and lead to strange scenes.
• Design the scenario to be prompt-friendly, aiming for large portions to be directly visualizable from prompts.

Character / Mecha “Sheet” System

Create extensive character sheets for every scenario element, including:

• Characters (and their animation needs)
• Props
• Spaces / locations for continuity
• Multiple versions for mechanisms (robot mechanics) and creatures

These sheets are described as essential for a semi-automated pipeline and for reducing inconsistencies later.

Mechanical design scope is mentioned as producing ~60 mechanical designs (with an additional reference to 85 robot mechanics at another point), then selecting final candidates.

Multi-Model Iterative Art Direction

• Generate images using multiple tools/models, then refine until the look is correct.
• Tools mentioned include:
  • GPT
  • C-Dream (also referenced with variants like C-Dance / C-Dream)
  • Nanobar / Nanobara
• If results feel clumsy, mottled, or lack charm, the method is to keep refining via round-robin generation across tools.

Reference-Driven Style Development

• Use explicit visual references, including:
  • Berserk
  • Akira
• Iteratively refine designs until proportions, textures, and faces feel cinematic and consistent.

Cinematography & Motion Through Tools

• Add camera movement using “movement camera motion” via Sitance (as described).
• Apply AI texture processing.
• Editing order matters: the creator stresses that results change significantly depending on sequence and structure.

Resolution & QA Workflow

• Generate/test at 480p first, then move to 720p with stronger prompts, and finally upscale to 4K using Topaz.
• Review multiple variations (noted: roughly five versions in 480p) to catch:
  • detail breakdowns
  • contextual errors
  • dialogue blurring

Workflow Automation (“Semi-Automation”)

• The pipeline is described as automating roughly 70–80%, with manual effort covering the remaining 20–30% revisions.
• Automation requires significant upfront “know-how,” trial-and-error, and careful preparation of assets and processes.

Time-Limited Production Discipline

• The lecture/video was produced under a 3–4 day deadline to upload, otherwise the creator claims it could take weeks or months.
• They frame quality as an “equivalent exchange”: more effort/detail refinement yields better output.

Guideline / Copyright Compliance Considerations

• Even when designs are “different,” similarity in training-influenced style can be disappointing.
• Mentions avoiding/mitigating copyright filter issues (described as “bypassing copyright filters” in lectures).
• Notes handling NSFW guideline / aspect ratio issues depending on the platform.
• Uses naming/notation rules to reduce tool errors (including references to Claude-related problems).

---

Steps / Advice Presented (Consolidated)

1. Scenario Planning

   • Break the clip into frames at 0.3-second intervals.
   • Use a structure system with:
     • scene-by-scene setups
     • payoffs
     • creephangers
   • Keep the script not overly long, focusing on necessary elements to avoid priority confusion.

2. Design & Revision Loop

   • Generate many concepts (mechanical/creature ideas in the dozens).
   • Select candidates and refine repeatedly.
   • If faces/mechs look wrong, switch among GPT ↔ C-Dream ↔ Nanobar until the “feel” is optimal.

3. Character Sheets & Consistency

   • Build detailed sheets for:
     • every character
     • every prop
     • every space/location
   • Treat sheets as continuity infrastructure: they must connect properly.

4. Video Generation QA

   • Validate outputs in this order:
     • 480p
     • 720p (improved prompts)
     • upscale to 4K using Topaz
   • Test multiple versions in 480p early to catch errors quickly.

5. Editing & Styling

   • Don’t rely on storyboards blindly—restructure/edit flexibly.
   • Create typography layout (English fonts via a “Fonts” site; Hangul via “Nun(nu)”).
   • Apply texture processing and add camera motion using Sitance.

6. Compliance

   • Check copyright carefully and repeatedly.
   • Follow guideline/notation rules to reduce tool errors (including Claude workflow considerations).
   • Consider platform-specific NSFW/aspect ratio constraints, and handle bypass cases as needed.

---

Creators / Contributors Featured (As Stated)

• AI Spa (the channel/creator introducing the tutorial: “hello. This is AI Spa.”)
• Claude (referenced as a collaborator/tool during scenario development)
• Tools referenced:
  • GPT
  • C-Dream / C-Dance
  • Nanobar / Nanobara
  • Sitance
  • Topaz

No additional human co-creators by name are clearly identified beyond AI Spa.

---