Jinho Chang

I'm Jinho Chang, currently a Ph.D. student at KAIST and a member of BioImaging-Signal Processing & Learning Lab(BISPL).

I'm working on the applications on continuous/discrete diffusion and flow-matching models with their various inference-time guidance and alignment methods, including reward-guided editing, negative sampling, score distillation, etc. Also, leveraging my chemistry background, I aim to solve diverse molecular tasks that are hard to be solved in principle, including conditional molecule design and editing, synthesizability and retro-synthesis prediction, intrinsic property prediction, etc.

Email  /  CV  /  Scholar  /  LinkedIn  /  Github

Research

I'm currently interested in Diffusion and flow-based models with their applications, Multimodal learning, and AI for chemistry.

Generative models for chemistry.

We present a new latent diffusion model LDMol for molecule generation with complex conditions like natural texts. Leveraging a chemically informative latent space obtained with contrastive learning, LDMol not only outperforms previous text-to-molecule generative models, but can also be applied to downstream tasks like retrieval and molecule editing.

We introduce a multimodal molecular pre-trained model that integrates molecular structures and their biochemical properties, which enables various multimodal and unimodal tasks like conditional molecule generation, property prediction, molecule classification, and reaction prediction.

We proposed a novel unsupervised learning model that could quantify the drug-likeness of a given molecule, by adopting a language model trained via unsupervised learning. Our model showed more consistent performance across different datasets and gave more reasonable scores.

Image & text generative models.

Prior reward-guided sampling methods in diffusion/flow-matching largely focused on the generation task and remain underexplored for image editing. We propose a training-free editing framework that casts the reward-guided editing as a trajectory optimal control problem, and iteratively updates adjoint states to steer the trajectory endpoint into a reward-optimized output. Our method surpasses inversion-based training-free guidance baselines across diverse editing tasks, achieving stronger rewards with higher source image fidelity without reward hacking.

We point out the drawback of naively negating the Classifier-Free Guidance (CFG) term, and propose a novel approach to enhance negative CFG in conditional diffusion models using contrastive loss. By aligning or repelling the denoising direction based on the given condition, our method overcomes the distortions caused by traditional negative CFG, achieving superior sample quality and effective removal of unwanted features across various scenarios.

Taking inspiration from previous work on the transformer and VQ-GAN combination for bidirectional image and text generation, we developed a method for instruction-tuning a text-only LLM to gain vision-language capabilities for medical images. We let LLM understand tokenized visual inputs by instructing it to answer questions about image inputs and generate image outputs.

We present a score distillation sampling for video editing to circumvent the standard reverse diffusion process. We propose to match the space-time self-similarities during the score distillation, demonstrating its superiority in altering appearances while accurately preserving the original structure and motion.

We propose a simple yet powerful model-agnostic image editing method that tackles complex image editing queries in a manner of divide-and conquer. Moreover, the selective application with a new penalty coefficient helps to precisely target editing areas while preserving the integrity of the objects in the source image.

Experiences

Development of a reinforcement learning model to generate molecules with specific chemical or pharmaceutical properties guided by the given RL reward.

Reviewer participation

Education

Last update: Oct 15, 2025