We are delighted to announce that 1 paper from CMLab (Prof. Jihyong Oh) has been accepted to NeurIPS 2026.

Title: 

MotionGrounder: Grounded Multi-Object Motion Transfer via Diffusion Transformer

Authors: 

Samuel Teodoro, Yun Chen, Agus Gunawan, Soo Ye Kim, Jihyong Oh, Munchurl Kim

Abstract: 

Existing motion-transfer approaches based on diffusion transformers often assume a single moving subject, limiting their use in scenes with several independently described objects. MotionGrounder introduces multi-object control through a flow-based motion signal that guides temporal dynamics and an object-caption alignment loss that associates individual descriptions with the proper spatial regions. It also proposes an Object Grounding Score for evaluating source-to-output object alignment and semantic consistency. Quantitative benchmarks, visual comparisons, and human evaluations indicate stronger multi-object transfer than competing approaches.