Base Composition Confounds Classifier-Free Guidance in DNA Diffusion Models

In diffusion models that design cell-type-specific enhancer sequences, stronger classifier-free guidance raises predicted chromatin accessibility while shifting base composition away from real sequences and distorting motif frequencies. Composition-neutral guidance removes sequence-wide base preferences from the guidance update without retraining and improves the accessibility-quality trade-off of a compact convolutional discrete diffusion model.