Tumor synthesis plays a crucial role in enhancing AI models for tumor detection and segmentation, particularly in challenging cases where real tumor data is scarce. Existing synthesis methods are typically unconditional, generating tumors without precise control, or mask-based, which only regulates size and shape. While text-driven approaches enable finer control over tumor characteristics (e.g., texture, heterogeneity, boundaries, and pathology), they require textual annotations, which require medical experts to annotate and are thus non-trivial to obtain. To address these challenges, we propose an image-guided tumor synthesis framework that uses tumor masks and reference tumor images to guide synthesis, eliminating the need for textual annotations while ensuring fine-grained control over key tumor properties. Built upon a Latent Diffusion Model, our approach integrates a tumor feature extraction module to incorporate reference tumor characteristics and introduces an inference-time intensity control mechanism to align HU intensity distributions, reducing artifacts and enhancing realism. We validate our method through liver tumor segmentation experiments, demonstrating that synthetic tumors generated by our approach effectively enhance model training and improve segmentation performance. Our results highlight the potential of image-guided tumor synthesis in expanding training datasets and improving the performance of medical models. Code and trained models are available at: https://github.com/DahangYe/Image-Guided-Liver-Tumor-Synthesis .