KinemaFX: Kinematic-Driven Effects

• KinemaFX is a kinematic-driven system that separates semantic descriptors from kinematic behaviors to enable nuanced control over particle effects.
• It combines natural-language input with simple graphical manipulation, allowing intuitive exploration and customization of dynamic visual effects.
• The system employs an iterative, preference-guided search to efficiently retrieve and compose complex effect artworks from extensive particle effect libraries.

KinemaFX is a kinematic-driven interactive system for exploring and customizing particle effect artworks in games and animation. Introduced as an LLM-powered workflow for non-expert users, it targets a specific gap in existing particle authoring environments: conventional tools such as Unity’s Shuriken and VFX Graph and Unreal’s Cascade and Niagara provide low-level parameter control, but weak support for high-level intent expression and for searching large effect libraries by motion behavior rather than by visual style alone. KinemaFX addresses this by explicitly separating semantic features from kinematic behaviors, combining natural-language input with simple graphical specification, and using iterative, preference-guided exploration to support the construction of customized effect artworks from existing particle effects (Zhang et al., 26 Jul 2025).

1. Problem Setting and Design Rationale

The system is motivated by the observation that particle effects are widely used to simulate natural phenomena and stylized visual effects, yet creating effect artworks remains difficult for non-expert users. The stated target users include indie game developers, level designers, and educators. The emphasis is not merely on single effects, but on effect artworks: compositions built by combining multiple individual particle effects such as fire bursts, magic circles, and portals (Zhang et al., 26 Jul 2025).

A formative study comprising 6 semi-structured interviews and analysis of 839 effects from 147 artworks identifies two central difficulties. First, kinematic behavior is crucial but hard to express in words. Semantic descriptors such as “blue magical aura” are comparatively easy to articulate, whereas motion descriptions such as “a ring that shrinks, then spirals upward and breaks into sparks” are longer, more ambiguous, and poorly handled by text-only pipelines. Second, exploration via text search is inefficient. The study reports that users often prefer example-based exploration—effectively, “I like this effect, show me more like this”—over repeated prompt reformulation (Zhang et al., 26 Jul 2025).

These observations motivate three system-level requirements. The first is a conceptual and computational model that distinguishes what is naturally expressed semantically from what is better represented kinematically. The second is an interactive workflow in which users can mix natural language with simple graphical input. The third is an exploration loop driven by implicit preferences, where user selections among intermediate results steer subsequent search. A plausible implication is that KinemaFX is designed less as a generative model of particle systems than as a retrieval-and-composition environment over a structured asset space.

2. Conceptual Model of Particle Effects

The core formalization represents each particle effect by a structured pair

$R = (S, K),$

where $S$ is the semantic component and $K$ is the kinematic descriptor (Zhang et al., 26 Jul 2025).

The semantic features cover aspects readily conveyed in natural language: category, stylistic descriptors, emotional tone, color, and overall style. Formally,

$S \in \mathbb{R}^d,$

with similarity measured via cosine similarity. The semantic embedding is produced from normalized textual descriptions using sentence-transformers.

The kinematic behavior is defined as

$K = (\text{shape}, \text{trail}, \text{duration}).$

Here, duration is the time from first particle generation to last disappearance. Emission shape is abstracted as a small set of primitives, based on the dataset analysis: point ($\text{radius} < 0.1$), circle, cylinder, and sphere. The structured shape representation is

$\text{shape} = (s, r, h),$

where $$s \in \{\text{circle}, \text{cylinder}, \text{sphere}\}$$, $r$ is outer radius, and $h$ is height for cylinders.

The trail representation models the motion of the outer boundary of the emission shape in a spherical coordinate system aligned with the effect’s axis of symmetry. For a boundary particle position

$S$0

translation along the axis corresponds to changes in $S$1, radial expansion or contraction to changes in $S$2, and rotation around the axis to changes in $S$3. The trail is discretized over $S$4 time intervals, with default $S$5:

$S$6

The dataset analysis further reports that 97.4% of effects have symmetry along at least one axis, and this regularity is explicitly exploited to simplify representation (Zhang et al., 26 Jul 2025).

This abstraction is central to the system’s notion of “kinematic-driven.” Rather than encoding low-level particle-system parameters, KinemaFX encodes complex multi-particle motion as a compact shape-transformation curve. This suggests a deliberate trade-off: the representation sacrifices some physical and procedural detail in exchange for searchability, alignment, and user-facing interpretability.

3. Search, Alignment, and Preference-Guided Exploration

KinemaFX supports interactive search by converting diverse constraints into structured representations, defining distance metrics over those representations, and searching a preprocessed database of 839 effects (Zhang et al., 26 Jul 2025).

Three kinds of constraints are converted into $S$7. For user input, semantic text is normalized by an LLM and embedded as $S$8, while graphical input is directly mapped to $S$9. For an individual particle effect instance, semantic information is built by concatenating the effect’s textual description with its artwork description, while kinematic information is extracted by simulation: particles are sampled over time, positions and related attributes are collected, and spherical-coordinate averages estimate shape and trail. For extrapolation between two instances, used in directional exploration, semantics are generated by LLM analysis of commonalities and differences, while kinematics are extrapolated dimension-wise and logarithmically to keep values in reasonable ranges.

Search is defined as

$K$0

where $K$1 is the database, $K$2 is the constraint representation, and $K$3 denotes allowable transformations including translation, rotation, scaling, and duration adjustment (Zhang et al., 26 Jul 2025).

Semantic similarity uses cosine similarity:

$K$4

The paper states that semantic distance $K$5 is derived from this similarity, though the exact formula is not explicitly given.

Kinematic distance is defined over the evolved shapes induced by trails. At each time step $K$6,

$K$7

where $K$8 is Hausdorff distance and $K$9 is rotational difference. Trail distance is then

$S \in \mathbb{R}^d,$0

Duration mismatch is penalized by a non-linear factor

$S \in \mathbb{R}^d,$1

yielding overall kinematic distance

$S \in \mathbb{R}^d,$2

The total distance is described conceptually as a weighted sum,

$S \in \mathbb{R}^d,$3

with $S \in \mathbb{R}^d,$4 and $S \in \mathbb{R}^d,$5 controlled by a user-facing semantic–kinematic weighting slider.

On top of this retrieval layer, the exploration loop alternates between local exploration and directional exploration. Local exploration retrieves the top-$S \in \mathbb{R}^d,$6 most similar effects; the default is $S \in \mathbb{R}^d,$7. Directional exploration uses two previously liked effects to infer a preference direction, semantically through the LLM and kinematically through extrapolation. User selections are treated as implicit preferences and become new anchors for search. The paper explicitly notes that this is not explicit Bayesian preference learning or classic relevance-feedback modeling with formal equations; the preference model is embodied in the interaction design itself (Zhang et al., 26 Jul 2025).

4. Intent Expression, LLM Integration, and Interface Structure

KinemaFX accepts three forms of user control: natural-language semantic input, graphical kinematic input, and a weighting slider that controls the relative importance of semantics and kinematics in search (Zhang et al., 26 Jul 2025).

Semantic input is free-form text, such as “A dark purple vortex that charges then explodes into sparks” or “Gentle frost aura spreading on the ground.” The pipeline sends this text to GPT-4o-mini for normalization and compression into concise effect descriptions, then embeds the result using sentence-transformers (all-MiniLM-L6-v2).

Graphical input is intentionally restricted to three key properties. Duration is set by a slider. Emission shape is specified by manipulating 3D primitives—circle, sphere, and cylinder—through translation, rotation, and scaling. Emission trail is specified by dragging a mouse to draw an axisymmetric vector, interpreted as an endpoint projection in the plane containing the axis and the vector, together with a control for spiraling. These inputs map directly into the $S \in \mathbb{R}^d,$8, $S \in \mathbb{R}^d,$9, and $K = (\text{shape}, \text{trail}, \text{duration}).$0 components of $K = (\text{shape}, \text{trail}, \text{duration}).$1.

The LLM has two stated functions. First, it normalizes both user text and dataset descriptions to produce more consistent semantic descriptors. Second, during directional exploration, it analyzes two reference effects, identifies shared features and differences, and generates candidate semantic descriptions that preserve common properties while varying along differing dimensions, constrained by the original user intent (Zhang et al., 26 Jul 2025).

The interface is organized into three principal views. The Intent Input Panel provides the text box, 3D graphical editor, duration control, trail input mechanism, and semantic–kinematic weighting slider. The Exploration View shows the current effect and top-$K = (\text{shape}, \text{trail}, \text{duration}).$2 candidates with animated previews and supports click-based selection for the next iteration. The Composition View provides a canvas for arranging multiple effects in space and time, with controls for transforms and temporal parameters and a preview of the combined animation.

Two usage scenarios in the paper illustrate how this interaction design supports partial and evolving intent. In the “Charge–Explosion” scenario, the user decomposes the target effect temporally into a charging phase and an explosion phase, emphasizing semantics for the charging stage and kinematics for the explosion stage. In the “Frost Range” scenario, the user decomposes the target effect spatially into atmosphere above ground and frost spreading on the ground, using different mixtures of semantic and kinematic constraints for each component. In both cases, the system supports starting from incomplete ideas and refining them through retrieval and selection rather than through exhaustive prompt engineering (Zhang et al., 26 Jul 2025).

5. Composition Workflow and Empirical Evaluation

After exploration, KinemaFX supports the creation of effect artworks by allowing users to select components from exploration history and combine them through spatial and temporal editing. For each component, the system exposes translation, rotation, scaling, and temporal properties including duration, playback speed, and start delay. The paper emphasizes that effects are pre-aligned kinematically during search, reducing manual adjustment at composition time (Zhang et al., 26 Jul 2025).

The evaluation is a within-subjects ablation study with 16 participants aged 19–29 years old, evenly split by gender, all non-experts in VFX. The dataset remains the same 839 particle effects and 147 artworks used elsewhere in the paper. Each participant created four effects, one under each of four conditions: Baseline, Preference-Guided Exploration, Kinematic-Driven Search, and KinemaFX (Full System). Themes were Buff, Spell, Attack, Teleport, with counterbalancing of condition order and theme assignment (Zhang et al., 26 Jul 2025).

Condition	Search basis	Preference guidance
Baseline	Semantic similarity only	No
Preference-Guided Exploration	Semantics only	Yes
Kinematic-Driven Search	Semantics + kinematics	No
KinemaFX (Full System)	Multimodal, semantics + kinematics	Yes

The study uses 15 Likert-scale items (1–7) grouped into 7 dimensions: Intent Expression, Convergent Exploration, Divergent Exploration, Kinematic Harmony, Effect Artworks, Effort, and Experience. Statistical analysis uses repeated-measures ANOVA and Friedman tests, with Bonferroni-corrected pairwise comparisons. Significance indicators are reported as $K = (\text{shape}, \text{trail}, \text{duration}).$3 (Zhang et al., 26 Jul 2025).

The reported findings are primarily qualitative, but several patterns are explicit. The combination of semantic and graphical input improves intent expression, especially for motion trajectories that participants found difficult to describe in text. Preference-guided exploration improves both convergent and divergent exploration, allowing users to lock onto promising results and also discover surprising variations. Kinematic-driven search improves kinematic harmony, reduces the need to tweak durations and transforms manually, and yields more coherent compositions. The full system is reported to best support artworks that capture users’ core intended features while remaining diverse beyond that core. Participants also described the interaction as “more relaxed,” and the paper attributes lower mental effort to the reduction in repeated prompt reformulation (Zhang et al., 26 Jul 2025).

A common misconception in this problem space is that stronger LLMs alone can close the gap between user intent and motion specification. The evaluation, together with the system design, argues against that position: the paper’s central claim is not that semantics are unimportant, but that semantics alone are insufficient for kinematics.

6. Relation to Prior Work, Claimed Contributions, and Limitations

KinemaFX is positioned against three lines of prior work. In particle effect and VFX creation, traditional platforms such as Unity, Unreal, Houdini, and After Effects provide rich parametric control but assume expertise. Gesture-based systems such as Energy-Brushes and MagicalHands are described as supporting limited control dimensions and often 2D or constrained 3D motions. Evolutionary and search-based approaches, including Chueca et al. 2024 and Hastings et al. 2008, are characterized as optimizing parameter vectors without explicitly modeling kinematic distance or exploration direction. KinemaFX departs from these by representing motion at a shape-transform level and coupling it to preference-guided exploration (Zhang et al., 26 Jul 2025).

In human–AI collaborative visual design, systems such as Spellburst, PromptPaint, DesignPrompt, AutoSpark, GenQuery, WorldSmith, DirectGPT, and 3DALL-E are described as focusing mainly on static images or simple animations, typically driven by text prompts with modest visual constraints. KinemaFX instead targets 3D, multi-particle, dynamic effects, and formalizes a semantic–kinematic split because semantics alone inadequately capture motion. In user-guided exploration and preference integration, systems such as Luminate, Prompting for Discovery, RoomDreaming, and BO as Assistant explore design spaces via structured prompt generation, Bayesian optimization, or reuse of selected final outputs as prompts. KinemaFX’s distinctive move is to keep preferences in the visual effect space itself: intermediate effect selections seed the next search directly, without retranslating visual preferences back into text (Zhang et al., 26 Jul 2025).

The authors identify three primary contributions. First, a conceptual model and an LLM-based interactive system that jointly capture semantic features and kinematic behaviors while supporting combined semantic and graphical input, implicit preference-guided exploration, and effect artwork customization. Second, a structured representation and distance metric based on $K = (\text{shape}, \text{trail}, \text{duration}).$4, with $K = (\text{shape}, \text{trail}, \text{duration}).$5, discretized spherical-coordinate trail representation, cosine-based semantic similarity, and kinematic similarity derived from Hausdorff distance, rotational penalty, and duration adjustment. Third, empirical validation through the 16-participant ablation study, which the paper presents as evidence that the system improves intent expression, exploration efficiency and diversity, mental effort, and the quality of resulting effect artworks (Zhang et al., 26 Jul 2025).

The paper also states several limitations. KinemaFX does not suggest high-level effect structures such as “a charging phase then an explosion phase,” nor does it recommend logical or hierarchical grouping of component effects. Composition tools are limited to basic spatial and temporal transforms; richer mechanisms such as hierarchical organization, dependency graphs, or node-based editing are left for future work. The graphical kinematic interface is intentionally simple and therefore less expressive for advanced users. Proposed future directions include sketch-based motion input, VR/AR embodied interaction, and kinematic templates or recommendations. The work also does not systematically compare different LLMs or evaluate LLM performance in isolation. Finally, although the authors suggest that the conceptual model and kinematic abstraction may generalize to domains such as character animation, gesture or interaction design, and scientific visualization of multi-particle dynamics, that transfer is not empirically tested in the reported study (Zhang et al., 26 Jul 2025).

Taken together, KinemaFX defines particle-effect authoring as a hybrid problem of semantic interpretation, kinematic representation, retrieval, and composition. Its technical distinctiveness lies in making kinematics first-class in both representation and interaction, rather than treating motion as a residual attribute of text prompts or low-level engine parameters.