Post by National AI Research Lab (NAIRL)

3D Gaussian Splatting (3DGS) is incredibly fast, but highly sensitive to photometric inconsistencies like varying scene illumination and camera exposure. What if we could solve this artifact issue in just seconds? We are thrilled to share that "CHROMA," a groundbreaking project led by Jisu Shin (GIST(Gwangju Institute of Science & Technology)) from Prof. Hae-Gon Jeon’s group at Yonsei University, has been accepted to the premier #ICLR2026! 🚀 Developed in global collaboration with Huawei Noah’s Ark Lab (UK), this research elegantly breaks through the computational bottleneck of 3D reconstruction. 💡 How CHROMA changes the game: • The Bottleneck: Addressing appearance inconsistencies typically requires optimizing per-scene embeddings, which becomes the main compute bottleneck (taking up to 40 minutes for competing methods). • A Single-Step Solution: CHROMA approaches inconsistent appearances by estimating per-frame bilateral grids, allowing for efficient, resolution-agnostic processing in a single step. • Unmatched Efficiency: The model achieves State-of-the-Art (SOTA) performance in appearance harmonization with an overhead of just a few seconds! A massive congratulations to Prof. Hae-Gon Jeon's team and Jisu Shin for this outstanding global collaboration, pushing the practical boundaries of 3D vision technology! 🔗 Read the ICLR 2026 Paper: https://lnkd.in/eJMm7wtf #NAIRL #YonseiUniversity #HaeGonJeon #ICLR2026 #3DVision