The recent developments in the field of computer vision and graphics have been significantly influenced by the integration of diffusion models and disentangled representations, aiming to enhance realism, control, and generalization in various applications such as person image synthesis, object manipulation, dance generation, human-object interaction synthesis, 3D human generative modeling, and garment animation. A common theme across these advancements is the focus on overcoming the limitations of existing methods by introducing novel frameworks that leverage the strengths of diffusion models for capturing complex distributions and disentangled representations for improved interpretability and control. These approaches not only address specific challenges such as detail preservation, compositional generalization, coherence, synchronization, and physical plausibility but also set new benchmarks for performance and quality in their respective domains.

Noteworthy Papers

• DRDM: A Disentangled Representations Diffusion Model for Synthesizing Realistic Person Images: Introduces a novel approach for person image synthesis with enhanced control over poses and appearances, achieving superior detail preservation and realism.
• EC-Diffuser: Multi-Object Manipulation via Entity-Centric Behavior Generation: Presents a behavioral cloning approach that enables efficient learning and compositional generalization in multi-object environments.
• CoheDancers: Enhancing Interactive Group Dance Generation through Music-Driven Coherence Decomposition: Develops a framework for generating coherent group dances by decomposing the problem into synchronization, naturalness, and fluidity.
• SyncDiff: Synchronized Motion Diffusion for Multi-Body Human-Object Interaction Synthesis: Introduces a synchronized motion diffusion strategy for synthesizing realistic multi-body interactions.
• JADE: Joint-aware Latent Diffusion for 3D Human Generative Modeling: Proposes a generative framework for 3D human bodies with fine-grained control over skeleton structures and local surface geometries.
• Diffgrasp: Whole-Body Grasping Synthesis Guided by Object Motion Using a Diffusion Model: Offers a novel approach for generating realistic whole-body grasping motions guided by object motion.
• Learning 3D Garment Animation from Trajectories of A Piece of Cloth: Introduces a disentangled scheme for animating garments based on the constitutive behaviors learned from a piece of cloth, reducing the need for extensive garment data.