MusicMorpher

Music Morphing via Style Tokens

New for the rebuttal: in addition to our original cherry-picked examples, the Examples section now includes morphs on real music, randomly drawn (non-cherry-picked) pairs, swapped endpoints, ablations, and baselines (FreeMorph, naive crossfade, MorphFader).

Abstract

Continuous-strength instruction editing lets users specify what to change and by how much. Such methods provide this control for images, but remain underdeveloped for music because they require source-target examples or morph paths for supervision, which are hard to obtain. Image-based Kontinuous Kontext derives such trajectories from FreeMorph outputs, but requires post-filtering since interpolating zero-noise latents is non-smooth. We find image-based methods transfer poorly to music: they act like a volume slider and ignore musical traits like tempo and beat alignment. We therefore identify music morphing, not editor fine-tuning, as the upstream bottleneck. We introduce MusicMorpher, a music morphing method that generates instruction-conditioned morph paths between a source and target. We do so by learning a per-song style token, applying it via modulation rather than attention, and interpolating it continuously. At inference, we tempo- and beat-align songs before style-token computation to prevent rhythmic discontinuities. We compare MusicMorpher against SDEdit, RF-Inversion, FreeMorph, and FlowMorph on morph quality and smoothness, introducing trajectory-level measures absent from prior single-sample metrics. Results show our method produces higher-quality, smoother morphs, and, without supervised instruct-edit training, extends to continuous-strength semantic edits reaching quality-alignment trade-offs beyond baseline sweeps; our analyses explain prior methods’ inconsistent results.

Method

MusicMorpher separates a song’s style — its performance, instrumentation, and mix character — from its structure, and interpolates each independently. A single per-song style token is injected via AdaLN modulation and blended between two songs with spherical interpolation (slerp); each song’s structural content is separately captured by inverting its latent to noise and slerp-interpolating the inverted noise trajectories. The two endpoints are also tempo- and beat-aligned before style-token computation, so the resulting morph path does not fight two competing rhythmic pulses. Full details, baselines, and quantitative comparisons are in the paper.

Examples

Notation. Throughout this page, λ is the interpolation weight toward Right: λ = 0 is Left and λ = 1 is Right. Equivalently, λ = 1 − w, where w is the weight on Left.
1Choose an example suite — our original cherry-picked examples are shown by default