Jongju Kim

Team Leader at BLUEDOT Inc.

Led the transition from GAN-based production pipeline to a large-scale controllable diffusion architecture. Defined R&D roadmap for transitioning from GAN-based super-resolution to diffusion-based methods. Coordinated cross-functional collaboration among research, engineering, sales, and marketing teams. Deployed enhancement system to 6+ enterprise clients including broadcasting and film production companies.

Applied AI Research Scientist of Computer Vision at BLUEDOT Inc.

Designed and upgraded Walrus v1 (GAN-based SR) and Walrus v2 (controllable one-step text-to-image diffusion SR for old footage). Core methodology published as IFCSR and accepted to CVPR 2026. Upgraded pipeline to a 1.3B-parameter controllable diffusion model with inference-free fidelity-realism control. Implemented image-space blending via dual LoRAs enabling controllable fidelity–realism trade-off. Designed 2-stage training for specialized LoRAs enabling controllable generation without extra inference steps. Modified LPIPS to generate more natural results. Trained the 1.3B model with mixed precision (FP32, FP16) on NVIDIA H100. Optimized inference using ONNX/TensorRT for deployment. Deployed enhancement system to 6+ enterprise clients.