DiffusionPack: Bin Packing with Custom Human Preferences

The three-dimensional bin packing problem (3D-BPP) is a challenging NP-hard task with real-world applications in logistics and manufacturing. Traditional approaches rely on manually designed heuristics and lack flexibility in handling complex, customizable preferences. We introduce a diffusion-based offline framework that operates in a continuous solution space and supports test-time integration of human preferences. We utilize pretrained large language models (LLMs) as preference interpreters and interaction mediators. LLMs can help understand and interpret natural language preferences into guidance functions and inpainting rules. Our method enables zero-shot adaptation to user-defined requirements, achieving efficient, high-quality packing without retraining. We view this as a step toward agentic systems that unite multimodal planning and human-AI collaboration for real-world decision-making tasks.