Music2Palette: Mapping Music to Color

• Music2Palette is a cross-modal approach that translates musical cues and affective signals into visually harmonious color palettes using signal processing and emotion analysis.
• It integrates methods such as Fourier/wavelet expansion, interval-to-wavelength mapping, and real-time visualization for applications like image recoloring and synesthetic displays.
• The framework optimizes color accuracy, palette diversity, and emotion alignment through multi-objective learning while addressing challenges like basis selection and octave discontinuities.

Searching arXiv for the cited Music2Palette-related papers to ground the article in current preprints. Using the arXiv search facility to confirm the core sources and metadata. Music2Palette denotes a class of cross-modal methods that translate music into color palettes or palette-like visual structures. In the recent literature, the term is formalized most directly as emotion-aligned color palette generation from audio via cross-modal representation learning, but its methodological lineage also includes orthogonal-basis matching between paintings and musical tracks, interval-to-wavelength mappings grounded in a wave analogy between sound and light, and real-time synesthetic visualization systems that map pitch, energy, and timbre to hue, saturation, brightness, texture, or particle color (Hu et al., 7 Jul 2025). Taken together, these strands define Music2Palette as a problem at the intersection of signal processing, color science, affective computing, human-computer interaction, and computational aesthetics (Gervasio et al., 2022).

1. Genealogy of the problem

A mathematically explicit precursor appears in "Let the paintings play" (Gervasio et al., 2022), which introduces “a mathematical method to extract similarities between paintings and musical tracks.” The method digitalizes both paintings and musical tracks by finite expansions in orthogonal basis functions, specifically “both Fourier and wavelet bases,” and identifies the best fit “via an $L^2$ projection upon a finite-dimensional subspace.” Although the paper is formulated as painting–music matching rather than direct palette generation, the coefficient vectors produced by this projection admit a downstream mapping to representative colors in RGB or HSV, making it an immediate antecedent of Music2Palette-style pipelines.

A second lineage is explicitly color-theoretic. "The wave method of building color palette and its application in computer graphics" (Sabo et al., 2017) constructs harmonious color combinations from musical consonance by identifying musical interval ratios with color-wavelength ratios. Here the central object is not a learned embedding or a spectral fit but a deterministic transformation from acoustic relations such as $2:1$, $3:2$, and $5:4$ to visible-light wavelengths, followed by conversion to device colors in sRGB.

A third lineage is interactive visualization. "musicolors: Bridging Sound and Visuals For Synesthetic Creative Musical Experience" (Lee et al., 18 Mar 2025) implements a web-based, real-time audio-to-visual mapping, while "An Artistic Visualization of Music Modeling a Synesthetic Experience" (Adiletta et al., 2020) uses FFT-derived features and a MAX 8 particle simulator to produce a palette-like field of twelve RGB entries. These systems emphasize live feature extraction, rendering latency, and user interaction rather than supervised palette prediction.

This suggests that Music2Palette is not a single algorithmic recipe. It is a family of cross-modal translation problems whose outputs range from weighted mixtures of track-associated colors, to interval-derived sRGB triplets, to dynamically updated HSV or RGB palettes, to variable-length CIELCh palettes predicted from latent music representations.

2. Projection-based and basis-expansion formulations

In the projection-based formulation, a painting is modeled as a real function

$$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$

with finite expansion

$$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$

where

$$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$

A music track is represented as

$s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$

with expansion

$$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$

After sampling, the image and each track are reduced to vectors $a\in\mathbb{R}^{N}$ and $2:1$0, respectively (Gervasio et al., 2022).

The basis can be Fourier or wavelet. The Fourier basis is given as

$2:1$1

with “global frequency content” and very fast computation via FFT. The wavelet option includes “1D or 2D Daubechies, Morlet, Haar, etc.” and is described as suitable for “time–frequency and multi–resolution analysis.” In practical implementations, the image can be handled either by “1D unrolling” of pixels followed by 1D DWT/DFT or by a full 2D DWT followed by flattening; audio is processed by direct 1D DFT or DWT of the sampled waveform. The DFT coefficients are written

$2:1$2

while DWT coefficients are obtained as $2:1$3 via a Fast Wavelet Transform using Mallat’s algorithm (Gervasio et al., 2022).

Best-fit matching is posed as least squares. If

$2:1$4

the target coefficients are

$2:1$5

The normal equations are

$2:1$6

so that, when $2:1$7 is invertible,

$2:1$8

An alternative uses the compact SVD $2:1$9, giving

$3:2$0

The vector $3:2$1 is then interpreted as the set of weights with which the chosen tracks contribute to the best-fit spectrum (Gervasio et al., 2022).

The corresponding implementation pipeline is explicit: choose $3:2$2 music tracks, load their sampled waveforms, choose a painting image and convert it to grayscale, transform the image into $3:2$3, compute $3:2$4 for each track, assemble $3:2$5, solve for $3:2$6, optionally synthesize a best-fit audio signal, and then generate a color palette from the resulting coefficients. In the reference app hosted at github.com/pgerva/playing-paintings, the GUI uses PySide2, DWT is computed with PyWavelets “(e.g. db5, 8 scales),” and display is limited to $3:2$7 tracks per run (Gervasio et al., 2022).

The Morandini case studies illustrate the method’s behavior. For the painting 340C/1989 projected onto $3:2$8 Tchaikovsky tracks, the reported weights are $3:2$9, normalized to percentages $5:4$0. Across comparisons of 15 Morandini paintings against “9 composers×4 tracks=36 audio samples” under four transforms, “Gershwin consistently ranks highest similarity,” whereas “Bach and Tchaikovsky” rank lowest in most configurations. Here the coefficient vector $5:4$1 becomes the bridge from spectral matching to palette construction: either each track is assigned a base color and mixed according to normalized weights, or the spectral signature is grouped into low-, mid-, and high-frequency bands and mapped to RGB channels (Gervasio et al., 2022).

3. Wave-ratio methods and the construction of harmonious palettes

The wave method starts from the claim that “sound and color are both propagating waves” and that consonance in music arises from “simple ratios of sound-wave periods,” with an analogous construction for colors based on “simple ratios of light-wave wavelengths” (Sabo et al., 2017). Consonant intervals are classified by frequency ratios

$5:4$2

and the key assumption is to identify this interval ratio with a color-wavelength ratio: $5:4$3 If the resulting wavelength falls outside the visible interval $5:4$4, it is returned to the visible band by multiplying or dividing by $5:4$5, described as octave shifting in the color domain. The basic formula is

$5:4$6

with $5:4$7 chosen so that $5:4$8 (Sabo et al., 2017).

The operational pipeline proceeds as follows. First, a reference note $5:4$9 is chosen, for example $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$0 Hz, and paired with a reference wavelength $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$1, for example $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$2 nm. For any target note $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$3, the ratio $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$4 is computed, the raw wavelength $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$5 is formed, repeated octave shifts bring it into the visible range, and the resulting $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$6 is converted from wavelength to CIE XYZ and then to linear sRGB, followed by standard sRGB gamma correction and clamping to $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$7 (Sabo et al., 2017).

The paper gives three concrete interval examples. For the octave $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$8, $$f:\Omega\subset\mathbb{R}^2\to\mathbb{R}, \qquad f\in L^2(\Omega), \qquad \Omega=[0,X]\times[0,Y],$$9 nm is shifted down to $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$0 nm, yielding the reference green, with gamma-corrected output $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$1, $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$2, $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$3. For the perfect fifth $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$4, $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$5 nm is shifted down to $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$6 nm, producing a violet–blue, with interpolated XYZ values $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$7, $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$8, $$f(x,y)\simeq \sum_{m,n=1}^{N} a_{mn}\,\phi_{mn}(x,y),$$9 and gamma-corrected output $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$0, $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$1, $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$2. For the major third $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$3, $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$4 nm remains in range and maps to a deep red, with $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$5, $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$6, $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$7 and output $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$8, $$a_{mn}=\langle f,\phi_{mn}\rangle_{L^2(\Omega)}=\int_\Omega f(x,y)\phi_{mn}(x,y)\,dx\,dy.$$9, $s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$0. The resulting green, violet–blue, and deep red are described as a consonant triad analogous to $s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$1–$s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$2–$s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$3 (Sabo et al., 2017).

This framework is deterministic and physically parameterized, but it is also explicitly contingent on reference choices. The same source notes that the mapping depends on $s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$4 and $s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$5, that single wavelengths produce only spectral colors, that some wavelengths map outside the sRGB gamut and must be clamped, and that visible-range octave shifting introduces “step-function discontinuities.” A plausible implication is that the wave method is best understood as a structured harmony model rather than a unique perceptual truth about music–color correspondences (Sabo et al., 2017).

4. Real-time synesthetic systems and palette-like visual outputs

Real-time systems operationalize Music2Palette as interactive audiovisual rendering. In musicolors, the architecture is “getUserMedia → Web Audio API → AudioStream” for input and a “Three.js WebGL canvas in browser” for output. Feature extraction uses the Pitchy library for pitch at “~100 Hz update,” Meyda’s “energy” feature for short-time loudness, and Meyda features including spectralFlatness, spectralKurtosis, perceptualSpread, and spectralCentroid for timbre. The render loop runs at requestAnimationFrame (~60 fps), all processing is client-side in JavaScript, and the reported audio-to-feature latency is “≈ 20–50 ms,” with render latency reduced by WebGL GPU acceleration; “no external hardware” is required (Lee et al., 18 Mar 2025).

The feature-to-visual mapping is explicit. Pitch drives “discrete note hue + octave→saturation (predefined 12-note palette),” energy controls sphere scale, and timbre modulates texture parameters such as “surface roughness, sharpness, hue shifts.” All color work is performed in HSV and then converted to RGB by Three.js. The mapping for pitch is summarized by

$s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$6

with value fixed, after which HSVtoRGB produces the display color. The system further notes that hue changes are discrete on pitch onset, while continuous interpolation by

$s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$7

is recommended for smooth transitions, although such smoothing equations are not part of the official formulation (Lee et al., 18 Mar 2025).

The user-study findings in musicolors are methodological as well as experiential. Composers reported timbre-to-texture mappings as particularly inspiring and requested representations that preserve a “memory” of past chords or pitches. Developers emphasized modularity and ease of integration with other music APIs. Listeners described vivid synesthetic experiences and requested presets expressing moods such as a “melancholy” palette. The paper’s recommendations are correspondingly practical: provide intuitive default mappings but expose them for customization, integrate visualization into streaming and interface contexts, and support export or sharing features (Lee et al., 18 Mar 2025).

A related but more explicitly signal-driven implementation appears in MAX 8. "An Artistic Visualization of Music Modeling a Synesthetic Experience" (Adiletta et al., 2020) uses adc~ → buffer~ 1024, jit.fft (N=1024), cartopol, and a fold to the first $s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$8 FFT points. The frequency domain is partitioned into twelve contiguous bins $s:[0,T]\to\mathbb{R}, \qquad s\in L^2(0,T),$9, each summarized by

$$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$0

with optional temporal averaging. These normalized magnitudes and average phases are mapped into HSV parameters, converted to RGB, and used to drive twelve particle groups. The paper then defines “12 discrete $$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$1 entries, one per frequency bin,” together with a continuous palette interpolation

$$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$2

when $$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$3 lies within bin $$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$4. Although framed as a synesthetic particle simulator, this is also a Music2Palette mechanism: it derives a structured, frequency-conditioned palette directly from audio (Adiletta et al., 2020).

5. Emotion-aligned cross-modal representation learning

The most direct formalization of Music2Palette is "Music2Palette: Emotion-aligned Color Palette Generation via Cross-Modal Representation Learning" (Hu et al., 7 Jul 2025). The system begins with MuCED, described as “the first large-scale music–palette dataset enriched with Russell vectors and expert refinement.” MuCED is assembled from music sources Emotify (400 clips), DEAM (1 802 clips), and PMEmo (794 clips), together with 5 992 palettes containing 3–5 colors each from Color Hunter, ColourLovers, and Kobayashi’s Color Image Scale. Automated matching computes an 8-dimensional Russell emotion vector for every music clip and every palette, then ranks candidate pairs by cosine similarity

$$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$5

For each music item, the top five palettes are retained for expert refinement (Hu et al., 7 Jul 2025).

Expert curation is two-stage. In Stage 1, “20 design experts” independently adjust each candidate set through hue shifts, saturation tweaks, and re-ordering according to the criteria of “emotion consistency,” “relationship to musical features,” and “color complexity/diversity.” In Stage 2, “25 experts (with at least 3 from different cultures per sample)” rate each refined pair on a 5-point Likert scale, and pairs with mean score below $$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$6 are removed. The final MuCED dataset contains “2 634 high-quality music–palette pairs,” with “average automatic emotion similarity 0.76” and “average expert rating 4.36/5” (Hu et al., 7 Jul 2025).

Emotion is represented through Russell’s circumplex model using eight prototypical emotions: excited, happy, content, calm, depressed, sad, afraid, and angry. The corresponding vectors are

$$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$7

$$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$8

Music emotion scores are obtained from a pretrained music–text contrastive model, whereas palette emotion scores are obtained from a text-embedding model applied to a short color description. On the input side, audio is resampled at $$g(t)\simeq \sum_{k=1}^{M} b_k\,\psi_k(t), \qquad b_k=\langle s,\psi_k\rangle_{L^2(0,T)}=\int_0^T s(t)\psi_k(t)\,dt.$$9 Hz, most clips are $a\in\mathbb{R}^{N}$0 s long, and each time frame contains a 539-dimensional feature vector: 128-dimensional Mel-spectrogram, 12-dimensional chroma, 7-dimensional spectral contrast, 6-dimensional tonal centroid (“tonnetz”), 384-dimensional rhythm features, 1-dimensional RMS energy, and 1-dimensional pitch. On the output side, palettes are represented in CIELCh and deduplicated by sorting in LCh space (Hu et al., 7 Jul 2025).

The model architecture is an encoder–decoder Transformer. The music encoder is an Audio Spectrogram Transformer modified so that the input is the concatenation along time of Mel, chroma, contrast, tonnetz, rhythm, energy, and pitch features. Temporal positional encoding is purely sinusoidal: $a\in\mathbb{R}^{N}$1 The encoder maps the feature sequence $a\in\mathbb{R}^{N}$2 to

$a\in\mathbb{R}^{N}$3

To encourage multiple plausible palettes for the same music, Gaussian noise is injected: $a\in\mathbb{R}^{N}$4 A Transformer decoder with “4 layers, 8 heads each,” cross-attention to $a\in\mathbb{R}^{N}$5, and a learnable stop token $a\in\mathbb{R}^{N}$6 outputs a variable-length palette

$a\in\mathbb{R}^{N}$7

through a feed-forward layer, a linear layer, and a Sigmoid to keep each channel in $a\in\mathbb{R}^{N}$8 (Hu et al., 7 Jul 2025).

6. Optimization, evaluation, applications, and open issues

The learning objective is explicitly multi-objective: $a\in\mathbb{R}^{N}$9 with “typical weights” $2:1$00, $2:1$01, and $2:1$02. The color term uses CIEDE2000 in LCh together with bipartite matching via the Hungarian algorithm: $2:1$03 The diversity term encourages hue spread,

$2:1$04

and the emotion-consistency term aligns predicted palette emotion both to music emotion and to the ground-truth palette emotion: $2:1$05 Training uses an “80 % train, 10 % val, 10 % test” split, approximately $2:1$06 samples, AdamW with initial learning rate $2:1$07, CosineAnnealingLR, batch size $2:1$08, $2:1$09 epochs, and an NVIDIA A6000 GPU with $2:1$10 GB (Hu et al., 7 Jul 2025).

Evaluation is correspondingly multi-faceted. Palette diversity and distribution are measured by Div., Multi., Convex Hull Overlap, and Bhattacharyya Coefficient; emotion alignment is measured by Emotion Similarity and Jensen–Shannon Divergence; application-level performance is tested in image recoloring through CLIP-score, Emotion Accuracy, and Emotion Incremental Score; and a user study includes four tasks concerning palette preference, quality of music–palette matching, and the degree to which palettes help understand or enhance felt emotion (Hu et al., 7 Jul 2025). In the reported Music→Palette experiment, the proposed method achieves Div. $2:1$11, Multi. $2:1$12, CHO $2:1$13, BC $2:1$14, ES $2:1$15, and JS $2:1$16, compared with LP+T2C at $2:1$17, $2:1$18, $2:1$19, $2:1$20, $2:1$21, $2:1$22; LP+GPT at $2:1$23, $2:1$24, $2:1$25, $2:1$26, $2:1$27, $2:1$28; Img+CT at $2:1$29, $2:1$30, $2:1$31, $2:1$32, $2:1$33, $2:1$34; and ground truth at Div. $2:1$35, ES $2:1$36, JS $2:1$37. The same source states that the method leads in Table 2 for image recoloring and in Table 3 for all four subjective tasks (Hu et al., 7 Jul 2025).

The demonstrated applications are “music-driven image recoloring,” “video generating,” and “data visualization” (Hu et al., 7 Jul 2025). These should be understood alongside earlier implementations: projection-based systems can synthesize a best-fit audio reconstruction and derive palettes from spectral weights (Gervasio et al., 2022); interval-based systems can generate consonance-driven sRGB schemes from musical ratios (Sabo et al., 2017); browser systems can render low-latency, customizable music visualizations with note-conditioned hue and timbre-conditioned texture (Lee et al., 18 Mar 2025); and FFT-bin systems can expose a twelve-color continuous palette function for particle-based display (Adiletta et al., 2020).

The open issues are correspondingly heterogeneous. Projection methods note “dimensional alignment” problems between 2D image transforms and 1D audio transforms, dependence on basis choice, and ill-conditioning of $2:1$38, with Tikhonov regularization and non-negative least squares suggested as remedies (Gervasio et al., 2022). The wave method explicitly depends on reference note and wavelength choices, experiences gamut clipping, and introduces octave-shift discontinuities (Sabo et al., 2017). musicolors states that normalization and scaling are “not explicitly specified,” and its smoothing equations are recommended rather than official (Lee et al., 18 Mar 2025). The learning-based framework itself proposes future work in “real-time interaction,” “user customization,” “cross-cultural tuning,” and “multi-modal fusion” (Hu et al., 7 Jul 2025). A plausible implication is that Music2Palette remains an active research area in which the central unresolved question is not whether music can be mapped to color, but which target notion of correspondence—spectral similarity, wave-theoretic harmony, synesthetic immediacy, or emotion alignment—should govern the mapping in a given application.