Jiho Park

Representation for Generation: Toward End-to-End Latent Diffusion Modeling
Feb 2026

• Research Question: Prior work searches for better representations to aid generation; instead, can we directly optimize the representation through the generative objective?
• Designed the framework from a self-supervised learning perspective
  (e.g., stop-grad targets, regularization).
• Effective on toy experiments, but scaling to images consistently led to noisy representations and training instability, — leaving practical E2E-LDM an open question.

[Slides]

Video-to-3D: Recurrent Feed-Forward Semantic Gaussian Splatting
Oct 2024 - Dec 2024

• Fine-tuned recurrent 3D geometry models to predict semantic Gaussian Splatting.
• Implemented an open-vocab 3D segmentation pipeline.

Coarse-to-Fine Imitation Learning with Hierarchical Action Sequence Quantization
Apr 2024 - Jun 2024

• Implemented VAR (hierarchical VQ-VAE & autoregressive transformer) with
  FSQ (Finite Scalar Quantization) on state-action sequence datasets.