AuraTracer智迹闻
中文

EVENT DOSSIER

LensStyle: Learning the Optical Aesthetics for Controllable Stylized Lens Effect Rendering

2026-09-07 12:00 Models 🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
1mentions
SummaryAI generated

On September 7, 2026, arXiv published the new paper LensStyle, which proposes a unified framework for achieving controllable stylized lens effect rendering. The model uses a dual-path controller design, decoupling continuous optical parameters (such as focal length and blur intensity) from discrete lens style conditions (such as circular, polygonal, star-shaped, etc.), enabling physically-based fine-grained control. The research team constructed the MultiLens dataset, which includes multiple lens image pairs synthesized under real-world optical constraints for training. Experimental comparisons show that LensStyle outperforms existing methods in terms of realism, controllability, and aesthetic quality, driving the development of computational photography in the field of multi-lens style simulation.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
LensStyle

SignalsSIGNALS

Keyword heat
  • LensStyle1

All reports (1)SOURCES

A arXiv cs.CV en 2026-09-07 12:00

LensStyle: Learning the Optical Aesthetics for Controllable Stylized Lens Effect Rendering

arXiv:2609.04939v1 发布新论文 LensStyle,提出一种通过联合连续 - 离散控制显式建模镜头美学的可控风格化镜头效果渲染统一框架。该模型引入双路径控制器,将连续光学参数调制(如焦距和模糊强度)与离散镜头风格条件(如圆形、多边形、甜甜圈、猫眼及星芒效果)解耦,实现了单一框架内的细粒度、可解释且基于物理的镜头操控。研究团队构建了包含在真实光学约束下合成多镜头图像对的 MultiLens 数据集以支持训练。实验表明,与现有镜头效果渲染方法及扩散模型相比,LensStyle 在真实性、可控性和美学质量上均表现更优,推动了计算摄影向多镜头风格模拟的发展。