Toward Real-World Adoption of Portrait Relighting via Hybrid Domain Knowledge Fusion

arXiv cs.CV / 4/28/2026

📰 NewsIdeas & Deep AnalysisModels & Research

共有:

Key Points

Real-world portrait relighting adoption is slowed by domain gaps between datasets, differences in camera sensitivity, and high computational costs.
The paper proposes “Hybrid Domain Knowledge Fusion,” which combines synthetic OLAT (One-Light-at-a-Time) data with real-world datasets to train a compact model.
It uses domain-aware adaptation with specialized prior models, then applies augmented knowledge distillation to transfer multi-domain expertise into a lightweight student network.
Experiments report a 6x to 240x inference speedup while retaining state-of-the-art visual quality, and the training pipeline is supported by a large, high-fidelity synthetic dataset with varied ground-truth intrinsics.
Overall, the work targets practical deployment by explicitly addressing both data mismatch issues and runtime efficiency through hybrid data and distillation.

Abstract

The real-world adoption of portrait relighting is hindered by dataset domain gaps, camera sensitivity, and computational costs. We address these challenges with Hybrid Domain Knowledge Fusion, a paradigm that fuses the specialized strengths of synthetic, One-Light-at-A-Time (OLAT), and real-world datasets into a compact model. Our approach features specialized prior models hardened by domain-aware adaptation, followed by augmented knowledge distillation into a lightweight student model with multi-domain expertise. Our method demonstrates a 6x to 240x inference speedup while maintaining state-of-the-art (SOTA) visual quality in the experiments. Additionally, we construct a massive, high-fidelity synthetic dataset with diverse ground-truth intrinsics to support our training pipeline.