A paper from Disney Research Studios presented at EGSR 2026 (the Eurographics Symposium on Rendering) tackles one of computer graphics’ most stubborn tradeoffs: the gulf between the gorgeous, physically-accurate images of path tracing and the interactive speeds artists actually need to work. Titled “Neural Render Proxies for Interactive and Differentiable Lighting,” the method lets you relight a static scene at roughly 30–60 Hz — interactive frame rates — while keeping the result close to full path-traced quality. For anyone who has waited minutes per frame to see how a lighting change looks, that’s a meaningful shift.
The core trick: split the problem
The key idea is to separate rendering into two parts: path sampling (tracing how light could travel through the scene) and emission computation (how bright each of those paths ends up, which depends on the lights). Disney’s approach runs a light-agnostic pass that collects the transport data — the geometry of how light moves — once, in a way that doesn’t depend on the specific lighting. A lightweight, scene-specific neural network then reuses that collected data to compute the final image under new lighting, cheaply and repeatedly. Because the expensive, lighting-independent work is done up front, changing the lights afterward becomes fast enough to feel live.
Why “differentiable” is in the title
The method is also differentiable, which is a bigger deal than it sounds. Differentiable rendering means you can run the process backward: given a target look, the system can compute how to adjust the lights to get there, rather than making an artist tweak by hand and re-render. That opens the door to inverse problems — “find the lighting that makes the scene look like this” — and to optimization and AI-driven workflows layered on top. Interactive relighting is the immediately useful part; differentiability is what makes it a research building block others can extend.
What it means for creators
The honest framing: this is a research result, aimed (for now) at static scenes, and it’s one paper among many at a rendering conference rather than a tool you can download tonight. But the direction is exactly the one that reshapes creative pipelines. Fast, accurate relighting collapses the feedback loop that dominates lighting and look-development work in film, animation, and increasingly real-time art — letting artists dial lighting by eye at interactive speed instead of rendering, waiting, and guessing. It fits a broader neural-rendering wave (the same current running through world models and real-time graphics research) that is steadily dragging path-traced quality toward interactive speed.