DiffH₂O is a novel diffusion-based framework designed to synthesize realistic hand-object interactions from textual descriptions. The method addresses the challenges of generating plausible and semantically meaningful hand-object motions, especially when dealing with unseen objects. Key contributions include: 1. **Decoupling Grasping and Interaction**: The task is split into two stages: grasping and interaction. The grasping stage generates hand motions, while the interaction stage synthesizes both hand and object poses. 2. **Compact Pose Representation**: A compact representation that tightly couples hand and object poses, reducing physical artifacts such as interpenetration. 3. **Controllability**: Two guidance schemes are introduced to enhance control over the generated motions: grasp guidance and detailed textual guidance. Grasp guidance uses a single target grasp pose to guide the diffusion model, while detailed textual guidance provides comprehensive text descriptions to the GRAB dataset, enabling more fine-grained control. The evaluation demonstrates that DiffH₂O outperforms baseline methods in terms of physics-based metrics, motion diversity, and action recognition accuracy. The method also shows practicality by utilizing a hand pose estimate from an off-the-shelf estimator to guide the diffusion process and sample multiple actions in the interaction stage. The project page and videos are available at https://diffh2o.github.io/.DiffH₂O is a novel diffusion-based framework designed to synthesize realistic hand-object interactions from textual descriptions. The method addresses the challenges of generating plausible and semantically meaningful hand-object motions, especially when dealing with unseen objects. Key contributions include: 1. **Decoupling Grasping and Interaction**: The task is split into two stages: grasping and interaction. The grasping stage generates hand motions, while the interaction stage synthesizes both hand and object poses. 2. **Compact Pose Representation**: A compact representation that tightly couples hand and object poses, reducing physical artifacts such as interpenetration. 3. **Controllability**: Two guidance schemes are introduced to enhance control over the generated motions: grasp guidance and detailed textual guidance. Grasp guidance uses a single target grasp pose to guide the diffusion model, while detailed textual guidance provides comprehensive text descriptions to the GRAB dataset, enabling more fine-grained control. The evaluation demonstrates that DiffH₂O outperforms baseline methods in terms of physics-based metrics, motion diversity, and action recognition accuracy. The method also shows practicality by utilizing a hand pose estimate from an off-the-shelf estimator to guide the diffusion process and sample multiple actions in the interaction stage. The project page and videos are available at https://diffh2o.github.io/.