MagicBokeh, developed by vivo Blueprint Lab, is a single-step unified generative framework designed for realistic telephoto bokeh rendering. The framework tackles common problems in high-zoom photography, such as noise amplification, boundary blur, and error accumulation, which are exacerbated by low-quality inputs. It employs an alternating training strategy, focus-aware mask attention, and degradation-aware depth estimation to overcome these challenges. The method achieves state-of-the-art results on both synthetic and real-world datasets, demonstrating its effectiveness and robustness. This innovation is particularly significant for mobile imaging, where hardware constraints limit optical capabilities. By enabling efficient and realistic bokeh simulation, MagicBokeh enhances the photographic experience on smartphones, offering users DSLR-like depth-of-field effects without the need for specialized lenses. The research was accepted as an Oral presentation at CVPR 2026, underscoring its impact and novelty in the field of computational photography.
vivo's MagicBokeh, a CVPR 2026 Oral, presents a unified generative framework for realistic telephoto bokeh rendering, solving noise and blur issues in high-zoom inputs with novel attention and depth estimation techniques.