top of page
Creative Process
Generative identity studies exploring visual stability, constraint systems, and motion translation.
These short-form works explore rhythm, emotional tone, and visual continuity through AI-assisted motion design and curated narrative constraints.
01 — Exploration & Identity Definition
Objective: Define a stable generative character identity before introducing controlled variation.
Before developing a modular character system, I explored multiple interpretations within a constrained thematic direction: eco-technical companion.
This phase focused on identifying which visual traits remained stable across generations and which introduced instability or drift.
Exploration Focus Areas
• Head geometry consistency
• Eye rendering stability
• Body proportion retention
• Material behavior under diffused lighting
• Botanical integration placement
• Silhouette clarity locking and controlled variation testing.
Early Exploration Outputs
Exploration 01 | Exploration 02 | Exploration 03
Stability Observations
• Rounded helmet geometries retained proportion more reliably than angular designs.
• High-detail metallic textures increased surface noise and reduced repeatability.
• LED-style eye panels produced stronger identity cohesion than reflective glass eyes.
• Minimal studio backgrounds improved proportion stability compared to dense environmental scenes.
• Crown-based botanical integration maintained visual balance better than side-mounted elements.
Conclusion
This exploration phase isolated the core traits capable of forming a stable character blueprint. Only after identifying repeatable geometry and material behaviors did I proceed to formal character
02 — Character Blueprint & Constraint Lock
Objective: Formalize a stable character blueprint by defining locked variables before introducing controlled variation.
After the exploration phase, I selected a single character iteration and translated its most stable traits into a structured constraint system. This ensured identity consistency across further testing and reduced generative drift.
Character Blueprint (Locked Variables)
Head Geometry — Rounded capsule helmet with smooth edge continuity
Eye System — Cyan LED panel display (non-reflective, low specularity)
Body Proportion — Compact torso with large head ratio (approx. 1.4:1 scale bias)
Surface Finish — Gloss white polymer with controlled micro-wear
Accent Lighting — Soft green central core indicator
Botanical Integration — Crown-mounted sprout growth
Framing — Center-weighted composition
Lighting — Soft diffused overhead key, low contrast
Background — Minimal neutral studio
Why Locking Matters
In generative workflows, identity drift often occurs when geometry, surface treatment, or lighting descriptors are loosely defined.
By explicitly locking structural and material traits, the character becomes a stable asset rather than a stylistic prompt outcome.
This distinction is critical when moving from exploratory generation into production-level asset development.
Base Prompt Architecture
“A high-resolution cinematic render of a compact eco-robot with a rounded capsule helmet, cyan LED eye panels, gloss white polymer body, crown-mounted sprout, soft diffused overhead lighting, centered composition, minimal neutral background, shallow depth of field.”
Structural descriptors remain fixed.
Environmental and secondary variables are introduced separately.
Swappable Variables (For Controlled Testing)
• Environment
• Pose
• Camera distance
• Lighting temperature
• Minor prop interaction
All variations are applied without altering locked structural descriptors.
03 — Controlled Variation & Stability Testing
Objective: Stabilize character identity through visual referencing and iterative aesthetic calibration rather than rigid prompt scripting.
After selecting the strongest character direction, I did not rebuild the prompt from scratch. Instead, I used Omni-reference and stylistic reinforcement to guide consistency.
My process is image-driven first, language-driven second.
Method
Rather than engineering complex prompt trees, I:
• Reused the strongest output as a visual anchor
• Applied Omni-reference to preserve geometry and proportion
• Adjusted stylization strength incrementally
• Refined lighting and surface tone visually
• Repeated short iteration cycles
The goal was not to “describe the character better” in text — it was to visually lock the identity through reference continuity.
What I Was Watching For
• Eye shape drift
• Helmet curvature distortion
• Material gloss inconsistency
• Proportion shifts between head and torso
• Botanical element reinterpretation
Each iteration was evaluated visually rather than linguistically.
If something drifted, I corrected by:
– Increasing reference weight
– Reducing stylization
– Simplifying environmental inputs
Controlled Variation Outputs
Variation 01 — City background | Variation 02 — Park background | Variation 03 — High-detail environmental scene
Each output was evaluated for:
• Proportion drift
• Material instability
• Eye-shape consistency
• Botanical integration accuracy
• Lighting contamination across surfaces
Insight
Generative systems respond differently to textual precision versus visual anchoring.
In this case, identity stability improved more through controlled visual referencing than through expanded prompt specificity.
That distinction matters.
Result
The character stabilized through:
• Consistent silhouette
• Repeated material reinforcement
• Controlled stylization
• Reduced descriptor noise
This prepared the asset for motion testing without reintroducing structural drift.
04 — Image-to-Video Translation & Motion Stability
Objective: Evaluate how the locked character identity performs under motion generation and temporal interpolation.
Static consistency does not guarantee temporal stability.
This phase tested whether the character could maintain structural and material integrity once translated into video using Runway and Pika.
Tools Used
• Runway (image-to-video)
• Pika (motion stylization testing)
• Reference frame anchoring from final stabilized still
Motion Testing Approach
Rather than generating new scenes from scratch, I:
• Imported the stabilized hero image
• Applied minimal motion prompts (micro head turn, breathing motion, subtle plant movement)
• Avoided large environmental changes in early tests
• Observed deformation across frame transitions
The goal was controlled motion, not spectacle.
What I Was Evaluating
Geometry Stability
– Helmet curvature retention
– Eye panel distortion during head movement
– Limb proportion consistency
Material Continuity
– Surface gloss consistency under shifting light
– Reflection behavior across frames
– Specular highlight stability
Botanical Integrity
– Sprout movement realism
– Leaf shape drift
– Edge artifacting
Temporal Artifacts
– Warping between frames
– Limb jitter
– Background hallucination
Observed Instabilities
• Larger head rotations introduced subtle helmet deformation
• Fast camera movement caused eye panel flicker
• Detailed foliage backgrounds increased frame hallucination
• High stylization amplified edge shimmer
Adjustments
• Reduced motion amplitude
• Locked camera position before introducing subject movement
• Lowered stylization intensity
• Simplified background planes
• Shortened motion loop duration
Small motion + stable anchor = higher coherence.
Result
The character maintained identity under:
• Subtle head movement
• Soft environmental parallax
• Minimal lighting variation
• Close crop framing
This confirmed that the asset could translate from still identity study to short-form motion without structural collapse.
bottom of page