Capturing high-quality tactile signals typically requires specialized hardware and controlled laboratory conditions, limiting the scalability and diversity of haptic content. Generative models, which have transformed digital language, vision, and audio content, offer a promising alternative for haptics. We propose a two-stage latent diffusion framework for generating tactile texture signals conditioned on psychophysical descriptors. In the first stage, a diffusion model learns a compact latent representation of friction signals produced by a finger sliding over diverse surfaces and reconstructs them with high temporal fidelity. In the second stage, a diffusion-based encoder maps perceptual ratings, such as roughness, bumpiness, and slipperiness, into this latent space, enabling texture generation from perceptual input. Reconstruction results demonstrate low error and a realistic signal structure. However, conditioning on psychophysical descriptors produces limited variations, primarily affecting signal amplitude, highlighting an open challenge in perceptually conditioned generative haptics.