Art2Mus: Artwork-to-Music Generation via Visual Conditioning and Large-Scale Cross-Modal Alignment

Music generation has advanced markedly through multimodal deep learning, enabling models to synthesize audio from text and, more recently, from images. However, existing image-conditioned systems suffer from two fundamental limitations: (i) they are typically trained on natural photographs, limiting their ability to capture the richer semantic, stylistic, and cultural content of artworks; and (ii) most rely on an image-to-text conversion stage, using language as a semantic shortcut that simplifies conditioning but prevents direct visual-to-audio learning. Motivated by these gaps, we introduce ArtSound, a large-scale multimodal dataset of 105,884 artwork-music pairs enriched with dual-modality captions, obtained by extending ArtGraph and the Free Music Archive. We further propose ArtToMus, the first framework explicitly designed for direct artwork-to-music generation, which maps digitized artworks to music without image-to-text translation or language-based semantic supervision. The framework projects visual embeddings into the conditioning space of a latent diffusion model, enabling music synthesis guided solely by visual information. Experimental results show that ArtToMus generates musically coherent and stylistically consistent outputs that reflect salient visual cues of the source artworks. While absolute alignment scores remain lower than those of text-conditioned systems-as expected given the substantially increased difficulty of removing linguistic supervision-ArtToMus achieves competitive perceptual quality and meaningful cross-modal correspondence. This work establishes direct visual-to-music generation as a distinct and challenging research direction, and provides resources that support applications in multimedia art, cultural heritage, and AI-assisted creative practice. Code and dataset will be publicly released upon acceptance.

Evaluating Disentangled Representations for Controllable Music Generation

Recent approaches in music generation rely on disentangled representations, often labeled as structure and timbre or local and global, to enable controllable synthesis. Yet the underlying properties of these embeddings remain underexplored. In this work, we evaluate such disentangled representations in a set of music audio models for controllable generation using a probing-based framework that goes beyond standard downstream tasks. The selected models reflect diverse unsupervised disentanglement strategies, including inductive biases, data augmentations, adversarial objectives, and staged training procedures. We further isolate specific strategies to analyze their effect. Our analysis spans four key axes: informativeness, equivariance, invariance, and disentanglement, which are assessed across datasets, tasks, and controlled transformations. Our findings reveal inconsistencies between intended and actual semantics of the embeddings, suggesting that current strategies fall short of producing truly disentangled representations, and prompting a re-examination of how controllability is approached in music generation.

Musical Metamerism with Time--Frequency Scattering

The concept of metamerism originates from colorimetry, where it describes a sensation of visual similarity between two colored lights despite significant differences in spectral content. Likewise, we propose to call ``musical metamerism'' the sensation of auditory similarity which is elicited by two music fragments which differ in terms of underlying waveforms. In this technical report, we describe a method to generate musical metamers from any audio recording. Our method is based on joint time--frequency scattering in Kymatio, an open-source software in Python which enables GPU computing and automatic differentiation. The advantage of our method is that it does not require any manual preprocessing, such as transcription, beat tracking, or source separation. We provide a mathematical description of JTFS as well as some excerpts from the Kymatio source code. Lastly, we review the prior work on JTFS and draw connections with closely related algorithms, such as spectrotemporal receptive fields (STRF), modulation power spectra (MPS), and Gabor filterbank (GBFB).

Stemphonic: All-at-once Flexible Multi-stem Music Generation

Music stem generation, the task of producing musically-synchronized and isolated instrument audio clips, offers the potential of greater user control and better alignment with musician workflows compared to conventional text-to-music models. Existing stem generation approaches, however, either rely on fixed architectures that output a predefined set of stems in parallel, or generate only one stem at a time, resulting in slow inference despite flexibility in stem combination. We propose Stemphonic, a diffusion-/flow-based framework that overcomes this trade-off and generates a variable set of synchronized stems in one inference pass. During training, we treat each stem as a batch element, group synchronized stems in a batch, and apply a shared noise latent to each group. At inference-time, we use a shared initial noise latent and stem-specific text inputs to generate synchronized multi-stem outputs in one pass. We further expand our approach to enable one-pass conditional multi-stem generation and stem-wise activity controls to empower users to iteratively generate and orchestrate the temporal layering of a mix. We benchmark our results on multiple open-source stem evaluation sets and show that Stemphonic produces higher-quality outputs while accelerating the full mix generation process by 25 to 50%. Demos at: https://stemphonic-demo.vercel.app.

Climber-Pilot: A Non-Myopic Generative Recommendation Model Towards Better Instruction-Following

Generative retrieval has emerged as a promising paradigm in recommender systems, offering superior sequence modeling capabilities over traditional dual-tower architectures. However, in large-scale industrial scenarios, such models often suffer from inherent myopia: due to single-step inference and strict latency constraints, they tend to collapse diverse user intents into locally optimal predictions, failing to capture long-horizon and multi-item consumption patterns. Moreover, real-world retrieval systems must follow explicit retrieval instructions, such as category-level control and policy constraints. Incorporating such instruction-following behavior into generative retrieval remains challenging, as existing conditioning or post-hoc filtering approaches often compromise relevance or efficiency. In this work, we present Climber-Pilot, a unified generative retrieval framework to address both limitations. First, we introduce Time-Aware Multi-Item Prediction (TAMIP), a novel training paradigm designed to mitigate inherent myopia in generative retrieval. By distilling long-horizon, multi-item foresight into model parameters through time-aware masking, TAMIP alleviates locally optimal predictions while preserving efficient single-step inference. Second, to support flexible instruction-following retrieval, we propose Condition-Guided Sparse Attention (CGSA), which incorporates business constraints directly into the generative process via sparse attention, without introducing additional inference steps. Extensive offline experiments and online A/B testing at NetEase Cloud Music, one of the largest music streaming platforms, demonstrate that Climber-Pilot significantly outperforms state-of-the-art baselines, achieving a 4.24\% lift of the core business metric.

TADA! Tuning Audio Diffusion Models through Activation Steering

Audio diffusion models can synthesize high-fidelity music from text, yet their internal mechanisms for representing high-level concepts remain poorly understood. In this work, we use activation patching to demonstrate that distinct semantic musical concepts, such as the presence of specific instruments, vocals, or genre characteristics, are controlled by a small, shared subset of attention layers in state-of-the-art audio diffusion architectures. Next, we demonstrate that applying Contrastive Activation Addition and Sparse Autoencoders in these layers enables more precise control over the generated audio, indicating a direct benefit of the specialization phenomenon. By steering activations of the identified layers, we can alter specific musical elements with high precision, such as modulating tempo or changing a track's mood.

IncompeBench: A Permissively Licensed, Fine-Grained Benchmark for Music Information Retrieval

Multimodal Information Retrieval has made significant progress in recent years, leveraging the increasingly strong multimodal abilities of deep pre-trained models to represent information across modalities. Music Information Retrieval (MIR), in particular, has considerably increased in quality, with neural representations of music even making its way into everyday life products. However, there is a lack of high-quality benchmarks for evaluating music retrieval performance. To address this issue, we introduce \textbf{IncompeBench}, a carefully annotated benchmark comprising $1,574$ permissively licensed, high-quality music snippets, $500$ diverse queries, and over $125,000$ individual relevance judgements. These annotations were created through the use of a multi-stage pipeline, resulting in high agreement between human annotators and the generated data. The resulting datasets are publicly available at https://huggingface.co/datasets/mixedbread-ai/incompebench-strict and https://huggingface.co/datasets/mixedbread-ai/incompebench-lenient with the prompts available at https://github.com/mixedbread-ai/incompebench-programs.

AI-Generated Music Detection in Broadcast Monitoring

AI music generators have advanced to the point where their outputs are often indistinguishable from human compositions. While detection methods have emerged, they are typically designed and validated in music streaming contexts with clean, full-length tracks. Broadcast audio, however, poses a different challenge: music appears as short excerpts, often masked by dominant speech, conditions under which existing detectors fail. In this work, we introduce AI-OpenBMAT, the first dataset tailored to broadcast-style AI-music detection. It contains 3,294 one-minute audio excerpts (54.9 hours) that follow the duration patterns and loudness relations of real television audio, combining human-made production music with stylistically matched continuations generated with Suno v3.5. We benchmark a CNN baseline and state-of-the-art SpectTTTra models to assess SNR and duration robustness, and evaluate on a full broadcast scenario. Across all settings, models that excel in streaming scenarios suffer substantial degradation, with F1-scores dropping below 60% when music is in the background or has a short duration. These results highlight speech masking and short music length as critical open challenges for AI music detection, and position AI-OpenBMAT as a benchmark for developing detectors capable of meeting industrial broadcast requirements.

Video-based Music Generation

As the volume of video content on the internet grows rapidly, finding a suitable soundtrack remains a significant challenge. This thesis presents EMSYNC (EMotion and SYNChronization), a fast, free, and automatic solution that generates music tailored to the input video, enabling content creators to enhance their productions without composing or licensing music. Our model creates music that is emotionally and rhythmically synchronized with the video. A core component of EMSYNC is a novel video emotion classifier. By leveraging pretrained deep neural networks for feature extraction and keeping them frozen while training only fusion layers, we reduce computational complexity while improving accuracy. We show the generalization abilities of our method by obtaining state-of-the-art results on Ekman-6 and MovieNet. Another key contribution is a large-scale, emotion-labeled MIDI dataset for affective music generation. We then present an emotion-based MIDI generator, the first to condition on continuous emotional values rather than discrete categories, enabling nuanced music generation aligned with complex emotional content. To enhance temporal synchronization, we introduce a novel temporal boundary conditioning method, called "boundary offset encodings," aligning musical chords with scene changes. Combining video emotion classification, emotion-based music generation, and temporal boundary conditioning, EMSYNC emerges as a fully automatic video-based music generator. User studies show that it consistently outperforms existing methods in terms of music richness, emotional alignment, temporal synchronization, and overall preference, setting a new state-of-the-art in video-based music generation.

MOSS-Audio-Tokenizer: Scaling Audio Tokenizers for Future Audio Foundation Models

Discrete audio tokenizers are fundamental to empowering large language models with native audio processing and generation capabilities. Despite recent progress, existing approaches often rely on pretrained encoders, semantic distillation, or heterogeneous CNN-based architectures. These designs introduce fixed inductive biases that limit reconstruction fidelity and hinder effective scaling. In this paper, we argue that discrete audio tokenization should be learned fully end-to-end using a homogeneous and scalable architecture. To this end, we first propose CAT (Causal Audio Tokenizer with Transformer), a purely Transformer-based architecture that jointly optimizes the encoder, quantizer, and decoder from scratch for high-fidelity reconstruction. Building on the CAT architecture, we develop MOSS-Audio-Tokenizer, a large-scale audio tokenizer featuring 1.6 billion parameters, pre-trained on 3 million hours of diverse, general audio data. We show that this simple, fully end-to-end approach built from homogeneous, causal Transformer blocks scales gracefully and supports high-fidelity reconstruction across diverse audio domains. Across speech, sound, and music, MOSS-Audio-Tokenizer consistently outperforms prior codecs over a wide range of bitrates, while exhibiting predictable improvements with increased scale. Notably, leveraging the discrete tokens from our model, we develop the first purely autoregressive TTS model that surpasses prior non-autoregressive and cascaded systems. Furthermore, MOSS-Audio-Tokenizer enables competitive ASR performance without auxiliary encoders. Our findings position the CAT architecture as a unified, scalable interface for the next generation of native audio foundation models.