Diff2Mix: AI music mixing with style control via diffusion + audio effects

Automatic music mixing has long struggled to balance objective quality with creative flexibility. Most existing systems treat standard mixing and style control as separate tasks, forcing engineers to choose between a polished but fixed result or a manually editable but lower-quality pipeline. Diff2Mix, a new paper accepted to ISMIR 2026, tackles this by building a single generative system that does both. Developed by Yisu Zong, Jinjie Shi, and Joshua Reiss, Diff2Mix leverages diffusion models—the same class of generative AI behind modern image and audio synthesis—to combine multitrack recordings into a balanced musical piece. Crucially, it couples the diffusion backbone with a differentiable mixing console, meaning every audio effect parameter (gain, EQ, compression, etc.) is exposed as a continuous, optimizable value.

This architecture enables two distinct levels of user control. First, an engineer can provide a reference audio track, and Diff2Mix will match its overall production style—taking care of genre-specific loudness, tonal balance, and spatial characteristics automatically. Second, the differentiable console allows direct manipulation of individual effect parameters, giving users interpretability and fine-grained control over the final mix. The authors demonstrate competitive performance through both objective metrics and subjective listening tests, showing that Diff2Mix produces mixes on par with or better than existing automatic systems while remaining far more editable. Because the console is differentiable, it also opens the door to gradient-based optimization for custom mixing targets. The project page includes code and audio samples, making it easy for researchers and practitioners to experiment with this new approach.