Diffusion models have gained significant attention for their generative capabilities in computer vision and natural language processing. Their application in recommender systems is promising, given their ability to recover meaningful user interactions from noisy historical data. However, existing diffusion-based recommenders face critical challenges, including exacerbation of popularity bias and the susceptibility of conditional signals to noise, particularly in modeling long-tail item preferences. To address these issues, we propose Tail-aware Conditional Diffusion Recommender Model (TCDRec), which introduces dual conditional signals to balance popularity bias and diversity. For long-tail items, TCDRec constructs interaction graphs using co-occurrence relationships to reveal latent preferences, while employing separate cross-attention modules to decouple global (popular) and local (long-tail) user interests. To enhance guidance signals’ robustness, conditional embeddings are parameterized within learnable Gaussian distributions, reducing noise interference during denoising. Experiments on three real-world datasets demonstrate TCDRec's superior performance in recommendation accuracy and diversity compared to state-of-the-art baselines.