While diffusion models produce high-quality images, they often lack consistency in sequences, especially for non-linear visual sequences. To solve this, we introduce an inference-time scaling beam search strategy for latent space exploration, enabling the conditional generation of coherent image sequences through beam search diffusion. In contrast to earlier methods that rely on fixed latent priors, our method dynamically samples past latents to search for an optimal sequence of latent representations, ensuring coherent visual transitions. As the latent denoising space is explored, the beam search graph is pruned with a cross-attention mechanism that efficiently scores search paths, prioritizing alignment with both textual prompts and visual context. Human and automatic evaluations confirm that BeamDiffusion outperforms other baseline methods, producing full sequences with superior coherence, visual continuity, and textual alignment.