Taehong Moon

I am a deep learning researcher working on robot foundation models and agentic systems.

My research focuses on developing efficient multimodal generative models for real-world environments. My previous work spans vision, language, and speech, including real-time interactive agents, speech generation, image generation, and efficient inference. I am currently exploring how these ideas can be extended to embodied systems, with a particular focus on perception, reasoning, and action.

My broader goal is to build efficient and generalizable multimodal systems that connect vision, language, speech, and action for meaningful interaction with the physical world.

Previously, I was a research engineer at KRAFTON. I received my M.S. in Artificial Intelligence from KAIST, where I was advised by Prof. Juho Lee, and my B.S. in Industrial Engineering from Seoul National University.

Email/ Scholar/ Twitter/ Github

$Taehong Moon$

News

May 2026Joined General Robotics as a Robotics Engineer.
Released the Raon-OpenTTS dataset.
May 2025ResGen and How to Move Your Dragon accepted to ICML 2025.
May 2024A Simple Early Exiting Framework accepted to ICML 2024.
Nov 2023Joined KRAFTON as a Research Engineer.

Research

(* denotes equal contribution)

Publications

For a complete list of publications, please see my Google Scholar profile.

Efficient Generative Modeling with Residual Vector Quantization-Based Tokens

Jaehyeon Kim*, Taehong Moon*, Keon Lee, Jaewoong Cho

International Conference on Machine Learning (ICML), 2025

ResGen — a masked generative model over residual-vector-quantization tokens that predicts cumulative token embeddings rather than individual tokens, making the number of sampling steps independent of both sequence length and depth.

project

How to Move Your Dragon: Text-to-Motion Synthesis for Large-Vocabulary Objects

Wonkwang Lee, Jongwon Jeong*, Taehong Moon*, Hyeon-Jong Kim, Jaehyeon Kim, Gunhee Kim, Byeong-Uk Lee

International Conference on Machine Learning (ICML), 2025

A text-to-motion framework that generalizes to a large vocabulary of objects, generating plausible motions for diverse and even unseen objects.

project/dataset

Rare-to-Frequent: Unlocking Compositional Generation Power of Diffusion Models on Rare Concepts with LLM Guidance

Dongmin Park, Sebin Kim, Taehong Moon, Minkyu Kim, Kangwook Lee, Jaewoong Cho

International Conference on Learning Representations (ICLR), 2025

A training-free framework that uses LLM reasoning to identify rare concepts in a prompt, unlocking compositional generation where SoTA T2I models fail.

code

A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models

Taehong Moon, Moonseok Choi, Eunggu Yun, Jongmin Yoon, Gayoung Lee, Jaewoong Cho, Juho Lee

International Conference on Machine Learning (ICML), 2024

Adaptive compute for diffusion sampling — drop more blocks where score estimation is relatively easy (near noise), keep the full network where it is hard (near data).

code

Fine-tuning Diffusion Models with Limited Data

Taehong Moon, Moonseok Choi, Gayoung Lee, Jung-Woo Ha, Juho Lee

NeurIPS Workshop on Score-Based Methods, 2022

Attention-only fine-tuning with a time-aware adapter (A³FT) adapts pretrained diffusion models to small datasets without the overfitting of full fine-tuning.

Technical Reports

Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech

Semin Kim*, Seungjun Chung*, Taehong Moon, Sangheon Lee, Minyoung Ahn, Keon Lee, Nam Soo Kim, Jaewoong Cho, Ludwig Schmidt, Kangwook Lee, Dongmin Park

arXiv:2605.20830, 2026

Open data (615K hours), pipeline, code, and models for robust zero-shot TTS — competitive with closed-data SoTA, with best-in-class robustness across clean, noisy, wild, and expressive speech.

code/dataset

Blog

Coming soon

Notes on world models, generative modeling, and Physical AI will show up here.