RelightFormer: Feed-forward Generative Transformer for Multiview Object Relighting
RelightFormer relights objects from one or more views without first reconstructing intrinsic scene properties.
The paper describes a feed-forward generative Transformer adapted from a video foundation model. It injects target environment maps through a latent illumination module and uses permutation-invariant positional encodings for unordered multi-view inputs. The authors trained it on Laval Objaverse Dataset, with 90K objects and 39K illuminations, and report state-of-the-art results across single-view, multi-view, and novel-view relighting. HF Daily Papers' note
The paper describes a feed-forward generative Transformer adapted from a video foundation model. It injects target environment maps through a latent illumination module and uses permutation-invariant positional encodings for unordered multi-view inputs. The authors trained it on Laval Objaverse Dataset, with 90K objects and 39K illuminations, and report state-of-the-art results across single-view, multi-view, and novel-view relighting. HF Daily Papers' note
score 4