Skip to main content

Style LoRAs — Vision Document

Research / vision document

This page describes the design vision for our Style LoRAs system, which will integrate with Amanda Co-Pilot's generative motion model post-launch. The Casual / Combat / Dance / Romance modules described below are the planned style adapters; the LoRA pipeline is in active training. See Amanda Co-Pilot for delivery timeline.

Low-Rank Adaptation (LoRA) enables fine-tuning of large pretrained models by training only a small number of additional parameters. Applied to motion synthesis, this allows domain-specific control without full model retraining.

What are Style LoRAs?

Style LoRAs are lightweight adapter modules that modify the base Kinetiq model to produce motion in specific styles. Instead of training a new model for each domain, we:

  1. Train a base model on diverse motion data
  2. Create small LoRA adapters for specific styles
  3. Inject the adapter at inference time
  4. Generate styled motion without base model changes

Available Style Modules

Casual

Everyday movement patterns:

  • Walking variations (confident, relaxed, tired)
  • Standing poses and idle animations
  • Sitting and getting up
  • Casual gestures and reactions

Best for: NPC idle animations, background characters, slice-of-life content

Combat

Martial arts and action sequences:

  • Punches, kicks, and strikes
  • Blocking and dodging
  • Weapon handling (sword, staff, gun)
  • Combat stances and transitions

Best for: Action games, fighting sequences, RPG combat

Dance

Choreographed and rhythmic movement:

  • Club/party dancing
  • Formal ballroom styles
  • Contemporary and modern dance
  • Synchronized group motion

Best for: Music videos, party scenes, performance content

Romance

Couples interaction and intimate gestures:

  • Embracing and holding hands
  • Dance partners
  • Emotional reactions
  • Close proximity interactions

Best for: Relationship content, drama scenes, social simulation

Gender Expression

Unlike models that encode gender as a binary switch, Kinetiq supports continuous gender expression through dedicated LoRA modules.

Why This Matters

Motion style varies independently of skeletal structure. A "masculine walk" or "feminine gesture" is a learned cultural pattern, not an anatomical constraint.

Available Expression Modules

  • Neutral - Balanced expression, suitable for most content
  • Masculine-leaning - Broader gestures, wider stance
  • Feminine-leaning - Fluid movement, expressive hands
note

Gender expression LoRAs can be combined with style LoRAs (e.g., "feminine combat" or "masculine dance").

Using Style LoRAs

In the Editor

  1. Open the Kinetiq Editor
  2. Generate or import a base animation
  3. Select a Style LoRA from the dropdown
  4. Adjust the style weight (0.0 - 1.0)
  5. Preview and export

Style Weight

The style weight controls how much the LoRA affects output:

WeightEffect
0.0No style influence (base model only)
0.5Balanced blend
1.0Full style influence
>1.0Exaggerated style (may cause artifacts)

Combining Styles

Multiple LoRAs can be combined with weighted blending:

Final = Base + (0.7 × Combat) + (0.3 × Masculine)

This produces combat motion with a masculine expression undertone.

Technical Details

LoRA Architecture

  • Rank: 16 (balance of quality vs. file size)
  • Target Layers: Attention and feedforward modules
  • Parameters: ~2M per adapter
  • File Size: ~8 MB per style

Training Data

Each style LoRA is trained on curated subsets of our Ecological Interaction Data:

StyleTraining HoursSources
Casual500+ hoursSocial interactions
Combat200+ hoursMartial arts, action
Dance300+ hoursPerformances, clubs
Romance150+ hoursCouples, intimacy

Requesting New Styles

We're continuously expanding our style library. To request a new style:

  1. Join our Discord
  2. Post in #feature-requests
  3. Describe the style with examples
  4. Vote on community suggestions

Popular requests may be prioritized for training.


Style LoRAs are available on all subscription tiers.