ParticleGen: A Multi-Agent System for Particle Effects Generation

ParticleGen is a multi-agent system aimed at generating particle effects from scratch from plain-language prompts, with a focus on structured, editable output rather than one-off visuals. The workflow splits the job into planning and parameterization, then loops in rendered feedback to iteratively improve the effect. That matters because particle authoring still tends to live in a high-friction space of procedural logic, dense parameter sets, and a lot of trial and error.

The system is built and evaluated in Unreal Engine 5’s Niagara, which makes it immediately relevant to teams already shipping with UE5 VFX pipelines. It was tested across a range of scenarios including elemental spells, natural phenomena, and fireworks, showing that the approach can cover both stylized gameplay effects and more general-purpose spectacle. The emphasis on editable simulation logic is important: teams need outputs they can tune, not just pretty renders.

A notable piece of the pipeline is a diagnostic mechanism that maps visible artifacts back to likely procedural causes. In practice, that could make AI-assisted VFX generation more usable for technical artists and graphics programmers, since the system is not just generating an effect but also helping explain what went wrong when the result drifts off target. The work positions LLM-driven authoring as a production aid rather than a replacement for hands-on control.

For developers, the practical upside is faster iteration on effects that normally require specialized Niagara knowledge. If these kinds of systems hold up outside curated demos,...