DOI: 10.35377/saucis...1871911 ISSN: 2636-8129

High-Fidelity Robotic Portraiture via Patch-Based Generative Adversarial Networks

Prawit Boonmee, Jirayus Arbking, Prajaks Jitngernmadan, Pinphong Ruangraweenukit, Natthawat Boonkhirassamee, Kittipong Nitsaisook, Kawin Yosmao, Prawee Jarujit, Kritsanapon Worrasuwatthanakul, Suriyen Kongtip, Ponlawat Chopuk
Show PDF Cite
he rapid advancement of humanoid robotics has intensified the demand for “embodied AI” systems capable of translating abstract perception into precise physical manipulation. While robotic art serves as an excellent benchmark for such dexterity, existing systems often struggle to preserve high-frequency details, particularly in complex facial regions like the eyes, or rely on prohibitively expensive industrial hardware. To address these limitations, this research presents a novel algorithmic pipeline for high-fidelity robotic portrait drawing. We propose a “split-transform-merge” methodology utilizing a patch-partitioned Generative Adversarial Network (P2LDGAN). Unlike traditional global inference methods, which lose fine detail, our approach partitions input images into 256 × 256 patches, processes them independently to maximize local feature retention, and spatially reconstructs them for execution by a low-cost Dobot Magician robotic arm. Qualitative results demonstrate that this patch-based strategy significantly outperforms current state-of-the-art competitors in rendering smooth arcs and fine facial features. By successfully bridging modern generative AI with precise physical execution, this work provides a robust, low-cost solution for automated artistic creation and fine-motor robotic control.

More from our Archive