LightBridge: relighting a Gaussian splat in a single pass
A feed-forward method that relights an entire 3D Gaussian Splatting scene in one pass, without per-scene optimisation, trained on a large multi-illumination dataset. It targets the baked-in lighting that limits how splat captures can be reused.
Splats are the fastest way to capture a real location, and baked lighting is what keeps them from matching anything shot on another day. Relighting without per-scene training is the step that turns a scan into a usable set extension.
Gaussian splatting has quietly become the fastest way to turn a real place into something a camera can move through. The catch has always been that the light comes with it. A splat captured at four in the afternoon stays at four in the afternoon, and matching it to a plate shot on an overcast morning means either living with the mismatch or rebuilding the scene. LightBridge goes after that constraint directly, and the interesting part is not that it relights a splat but how little work it asks for to do so.
Previous approaches fell into two camps, as the authors describe them. Inverse rendering fits simplified reflectance and lighting models scene by scene, which is slow and limits quality. Generative methods built on large diffusion models produce more convincing light, but then need a separate per-scene optimisation to bake the edited look back into the 3D representation. LightBridge replaces both with a feed-forward pipeline: a latent bridge diffusion stage that treats relighting as transport from a source to a target lighting condition and extracts visual tokens in one step, and a Gaussian propagation transformer that spreads those cues across the whole splat without full attention over every token. Training relies on a purpose-built multi-illumination dataset of paired observations of the same scenes under different light.
In pipeline terms, this touches the handoff between capture and set extension. Today a virtual production or VFX team scanning a location usually has to schedule the scan to match the intended lighting, or treat the splat as reference only. A single-pass relight would let a supervisor capture once and adjust the light to match each shooting day, or to match a plate from a different unit. It also matters for previs: a scanned location that can be relit is a set a director can block at the time of day the scene is written for.
The open questions are significant, and they are the ones a supervisor would ask first. The abstract reports “competitive” quality without naming baselines or figures, and does not say what the control signal is: an environment map, a reference image, or a text description. Those are very different workflows. Speed is described only as a single forward pass, with no timings. The dataset and code are promised on acceptance, so for now nobody outside the team can test it against their own captures. Relit splats also have to hold up next to photographed plates in a grade, which is a stricter test than a side-by-side in a paper.
Read alongside the other 3D work in the feed, the direction is consistent. Mise-en-Video stages multi-shot generation in 3D boxes, AMD’s planned acquisition of World Labs puts a price on 3D world models, and LightBridge addresses one of the reasons real captures do not slot neatly into those pipelines. Our view is that lighting control, more than resolution, decides whether a scan is an asset or a reference. When the code is public, the test is simple: capture a location, relight it to match a plate from another day, and see whether a compositor accepts it.