Teaching stage lighting has a physical bottleneck. To show a student what a warm backlight at 30 degrees does to a body on stage, you need the stage, the fixtures, the rig time and someone standing in the light. Most programmes have a fraction of what they need, and the student who isn’t at the console is watching.
VR removes the physical constraint — you can hang a hundred virtual fixtures — but, as LumiNote’s authors point out, existing VR lighting systems offer very little support for live instruction. They’re practice environments, not teaching ones.
LumiNote (arXiv, Sept 15), from Danxuan Liang, Chun Yin Li, Zheng Wei, Xian Xu, Meng Xia, Huamin Qu and Wai Tong, is a system built for the teaching itself.
What it does
An instructor speaks. The system converts that spoken pedagogical intent into three things:
- Spatial annotations the instructor reviews before students see them — arrows, highlights, marks in the 3D space
- Executable demonstrations — the lighting state described, actually applied to the virtual rig
- Linguistic support for explaining the result
The review step is the important design decision: the LLM proposes, the instructor approves. Nothing reaches students unvetted.
What the study found
Three instructors, twenty-four students, working through familiar lighting topics. The results are refreshingly specific about where the model helps:
- LLM assistance was most valuable for expressive, under-specified goals — “make this feel like late afternoon,” “isolate her without making it cold.” Vague intent is exactly what language models are good at turning into a concrete starting point.
- It required greater expert intervention for fixture-specific or spatial configuration requests — which channel, which position, which angle. Precise technical instructions are where it needed correcting.
- Instructors treated generated suggestions as a controllable refinement process, shifting their effort from manual setup toward pedagogical expression. That’s the actual win: less time patching a rig, more time teaching.
- Representations that externalized expert reasoning did not always align with novice comprehension. Making an expert’s thinking visible doesn’t automatically make it legible to a beginner.
Why it generalizes
That third finding applies well beyond lighting. Every creative discipline with a complex tool — DAWs, node graphs, modular synths, game engines — has the same teaching problem: the expert’s intent is invisible, and the student sees only the parameter changes. Systems that capture and show intent rather than actions are a genuinely useful idea.
The fourth finding is the caution. Externalising expert reasoning is not the same as explaining it. An annotation that makes perfect sense to the instructor who generated it can be meaningless to someone who doesn’t yet know what to look at. Anyone building AI-assisted teaching tools for creative software should expect that gap and design for it, rather than assuming visibility equals understanding.
For the performance-technology audience specifically, this is also a reminder that VR’s most credible near-term use in theatre isn’t performance — it’s training on expensive rigs that students otherwise queue for.