UIST 2026

SpeLLA Generative Framework for Synthesizing 3D Visual Effects from Text Prompts with Large Language Models

Jeongseok OhSeungJun Kim

HCIS Lab, GIST

From a text prompt to an editable 3D VFX draft in Unity.

Generated results

Meshes, textures, and particles composed into a single effect.

“A purple horizontal beam”
“Blue transparent spherical shield”
“A fire tornado effect”
“Falling raindrop effect”
“Green magic aura effect”
“Round golden portal effect”
“Red explosion and smoke”
“A blue sword slash effect”

Examples from the paper’s supplementary video. Each effect is rendered in Unity; the camera views are synchronized for this gallery.

Abstract

Creating visual effects for 3D content requires specialized design knowledge and technical expertise. SpeLL produces editable 3D VFX drafts from text prompts using large language models and retrieval-based asset assembly. It combines prompt enhancement, spatial design for component retrieval and layout arrangement, and temporal design for activation sequencing.

Informed by a formative study with professional VFX designers, SpeLL generates parameterized meshes, textures, and particle systems that can be refined in Unity. A user study with twenty professional designers evaluated the framework against ablated systems, 3D generation baselines, and manual workflows. Results showed higher intention alignment, spatial-temporal layout quality, and editability. The goal is to support early exploration and drafting while leaving final expressive control to designers.

Framework

From a prompt to a coordinated effect

Component selection, spatial arrangement, and temporal sequencing work together.

The SpeLL pipeline: prompt enhancement provides descriptions and layout guidance; LLMs propose spatial and temporal layouts; retrieval and refinement inform component matching; meshes, textures and particles are integrated into the generated VFX.
Original framework figure from the paper. View full-size figure ↗

Prompt enhancement

Expand a high-level request into component descriptions and spatial-temporal guidance.

Spatial and temporal design

Retrieve suitable components, arrange their positions and sizes, and refine their activation sequence.

Editable Unity output

Assemble a parameterized VFX draft that supports further adjustment by the designer.

Three simple meshes paired with noise textures form sub-components; these are combined with particles into one complete VFX asset.

Built from VFX components

Mesh
Provides the basic shape and spatial structure.
Texture
Tileable noise patterns and shaders create visual detail and flow.
Particle
Adds the emission and motion of small finishing details.

Demonstration

See SpeLL in use

The supplementary video shows generation, results, and parameter-based editing.

Download video · A web-sized copy of the original supplementary video.

Code release

Code coming soon.

We are preparing the implementation and documentation for public release. The repository currently contains the project page and a placeholder structure for the Unity client, backend, and data preparation tools.

View repository

Citation

@inproceedings{oh2026spell,
  author    = {Oh, Jeongseok and Kim, SeungJun},
  title     = {SpeLL: A Generative Framework for Synthesizing 3D Visual Effects
               from Text Prompts with Large Language Models},
  booktitle = {Proceedings of the 39th Annual ACM Symposium on
               User Interface Software and Technology},
  series    = {UIST '26},
  year      = {2026},
  publisher = {Association for Computing Machinery},
  doi       = {10.1145/3830398.3830697},
  url       = {https://doi.org/10.1145/3830398.3830697}
}