DRAGON: Distributional Rewards Optimize Diffusion Generative Models

We present Distributional RewArds for Generative OptimizatioN (DRAGON), a versatile framework for fine-tuning media generation models towards a desired outcome. Compared with traditional reinforcement learning with human feedback (RLHF) or pairwise preference approaches such as direct preference optimization (DPO), DRAGON is more flexible. It can optimize reward functions that evaluate either individual examples or distributions of them, making it compatible with a broad spectrum of instance-wise, instance-to-distribution, and distribution-to-distribution rewards. Leveraging this versatility, we construct novel reward functions by selecting an encoder and a set of reference examples to create an exemplar distribution. When cross-modality encoders such as CLAP are used, the reference examples may be of a different modality (e.g., text versus audio). Then, DRAGON gathers online and on-policy generations, scores them to construct a positive demonstration set and a negative set, and leverages the contrast between the two sets to maximize the reward. For evaluation, we fine-tune an audio-domain text-to-music diffusion model with 20 different reward functions, including a custom music aesthetics model, CLAP score, Vendi diversity, and Frechet audio distance (FAD). We further compare instance-wise (per-song) and full-dataset FAD settings while ablating multiple FAD encoders and reference sets. Over all 20 target rewards, DRAGON achieves an 81.45% average win rate. Moreover, reward functions based on exemplar sets indeed enhance generations and are comparable to model-based rewards. With an appropriate exemplar set, DRAGON achieves a 60.95% human-voted music quality win rate without training on human preference annotations. As such, DRAGON exhibits a new approach to designing and optimizing reward functions for improving human-perceived quality. Sound examples at https://ml-dragon.github.io/web.

DiffRhythm: Blazingly Fast and Embarrassingly Simple End-to-End Full-Length Song Generation with Latent Diffusion

Recent advancements in music generation have garnered significant attention, yet existing approaches face critical limitations. Some current generative models can only synthesize either the vocal track or the accompaniment track. While some models can generate combined vocal and accompaniment, they typically rely on meticulously designed multi-stage cascading architectures and intricate data pipelines, hindering scalability. Additionally, most systems are restricted to generating short musical segments rather than full-length songs. Furthermore, widely used language model-based methods suffer from slow inference speeds. To address these challenges, we propose DiffRhythm, the first latent diffusion-based song generation model capable of synthesizing complete songs with both vocal and accompaniment for durations of up to 4m45s in only ten seconds, maintaining high musicality and intelligibility. Despite its remarkable capabilities, DiffRhythm is designed to be simple and elegant: it eliminates the need for complex data preparation, employs a straightforward model structure, and requires only lyrics and a style prompt during inference. Additionally, its non-autoregressive structure ensures fast inference speeds. This simplicity guarantees the scalability of DiffRhythm. Moreover, we release the complete training code along with the pre-trained model on large-scale data to promote reproducibility and further research.

MusicInfuser: Making Video Diffusion Listen and Dance

We introduce MusicInfuser, an approach for generating high-quality dance videos that are synchronized to a specified music track. Rather than attempting to design and train a new multimodal audio-video model, we show how existing video diffusion models can be adapted to align with musical inputs by introducing lightweight music-video cross-attention and a low-rank adapter. Unlike prior work requiring motion capture data, our approach fine-tunes only on dance videos. MusicInfuser achieves high-quality music-driven video generation while preserving the flexibility and generative capabilities of the underlying models. We introduce an evaluation framework using Video-LLMs to assess multiple dimensions of dance generation quality. The project page and code are available at https://susunghong.github.io/MusicInfuser.

DGFM: Full Body Dance Generation Driven by Music Foundation Models

In music-driven dance motion generation, most existing methods use hand-crafted features and neglect that music foundation models have profoundly impacted cross-modal content generation. To bridge this gap, we propose a diffusion-based method that generates dance movements conditioned on text and music. Our approach extracts music features by combining high-level features obtained by music foundation model with hand-crafted features, thereby enhancing the quality of generated dance sequences. This method effectively leverages the advantages of high-level semantic information and low-level temporal details to improve the model's capability in music feature understanding. To show the merits of the proposed method, we compare it with four music foundation models and two sets of hand-crafted music features. The results demonstrate that our method obtains the most realistic dance sequences and achieves the best match with the input music.

DanceMosaic: High-Fidelity Dance Generation with Multimodal Editability

Recent advances in dance generation have enabled automatic synthesis of 3D dance motions. However, existing methods still struggle to produce high-fidelity dance sequences that simultaneously deliver exceptional realism, precise dance-music synchronization, high motion diversity, and physical plausibility. Moreover, existing methods lack the flexibility to edit dance sequences according to diverse guidance signals, such as musical prompts, pose constraints, action labels, and genre descriptions, significantly restricting their creative utility and adaptability. Unlike the existing approaches, DanceMosaic enables fast and high-fidelity dance generation, while allowing multimodal motion editing. Specifically, we propose a multimodal masked motion model that fuses the text-to-motion model with music and pose adapters to learn probabilistic mapping from diverse guidance signals to high-quality dance motion sequences via progressive generative masking training. To further enhance the motion generation quality, we propose multimodal classifier-free guidance and inference-time optimization mechanism that further enforce the alignment between the generated motions and the multimodal guidance. Extensive experiments demonstrate that our method establishes a new state-of-the-art performance in dance generation, significantly advancing the quality and editability achieved by existing approaches.

SoK: How Robust is Audio Watermarking in Generative AI models?

Audio watermarking is increasingly used to verify the provenance of AI-generated content, enabling applications such as detecting AI-generated speech, protecting music IP, and defending against voice cloning. To be effective, audio watermarks must resist removal attacks that distort signals to evade detection. While many schemes claim robustness, these claims are typically tested in isolation and against a limited set of attacks. A systematic evaluation against diverse removal attacks is lacking, hindering practical deployment. In this paper, we investigate whether recent watermarking schemes that claim robustness can withstand a broad range of removal attacks. First, we introduce a taxonomy covering 22 audio watermarking schemes. Next, we summarize their underlying technologies and potential vulnerabilities. We then present a large-scale empirical study to assess their robustness. To support this, we build an evaluation framework encompassing 22 types of removal attacks (109 configurations) including signal-level, physical-level, and AI-induced distortions. We reproduce 9 watermarking schemes using open-source code, identify 8 new highly effective attacks, and highlight 11 key findings that expose the fundamental limitations of these methods across 3 public datasets. Our results reveal that none of the surveyed schemes can withstand all tested distortions. This evaluation offers a comprehensive view of how current watermarking methods perform under real-world threats. Our demo and code are available at https://sokaudiowm.github.io/.

Evaluating the Impact of AI-Powered Audiovisual Personalization on Learner Emotion, Focus, and Learning Outcomes

Independent learners often struggle with sustaining focus and emotional regulation in unstructured or distracting settings. Although some rely on ambient aids such as music, ASMR, or visual backgrounds to support concentration, these tools are rarely integrated into cohesive, learner-centered systems. Moreover, existing educational technologies focus primarily on content adaptation and feedback, overlooking the emotional and sensory context in which learning takes place. Large language models have demonstrated powerful multimodal capabilities including the ability to generate and adapt text, audio, and visual content. Educational research has yet to fully explore their potential in creating personalized audiovisual learning environments. To address this gap, we introduce an AI-powered system that uses LLMs to generate personalized multisensory study environments. Users select or generate customized visual themes (e.g., abstract vs. realistic, static vs. animated) and auditory elements (e.g., white noise, ambient ASMR, familiar vs. novel sounds) to create immersive settings aimed at reducing distraction and enhancing emotional stability. Our primary research question investigates how combinations of personalized audiovisual elements affect learner cognitive load and engagement. Using a mixed-methods design that incorporates biometric measures and performance outcomes, this study evaluates the effectiveness of LLM-driven sensory personalization. The findings aim to advance emotionally responsive educational technologies and extend the application of multimodal LLMs into the sensory dimension of self-directed learning.

Aligning Text-to-Music Evaluation with Human Preferences

Despite significant recent advances in generative acoustic text-to-music (TTM) modeling, robust evaluation of these models lags behind, relying in particular on the popular Fr\'echet Audio Distance (FAD). In this work, we rigorously study the design space of reference-based divergence metrics for evaluating TTM models through (1) designing four synthetic meta-evaluations to measure sensitivity to particular musical desiderata, and (2) collecting and evaluating on MusicPrefs, the first open-source dataset of human preferences for TTM systems. We find that not only is the standard FAD setup inconsistent on both synthetic and human preference data, but that nearly all existing metrics fail to effectively capture desiderata, and are only weakly correlated with human perception. We propose a new metric, the MAUVE Audio Divergence (MAD), computed on representations from a self-supervised audio embedding model. We find that this metric effectively captures diverse musical desiderata (average rank correlation 0.84 for MAD vs. 0.49 for FAD and also correlates more strongly with MusicPrefs (0.62 vs. 0.14).

Note-Level Singing Melody Transcription for Time-Aligned Musical Score Generation

Automatic music transcription converts audio recordings into symbolic representations, facilitating music analysis, retrieval, and generation. A musical note is characterized by pitch, onset, and offset in an audio domain, whereas it is defined in terms of pitch and note value in a musical score domain. A time-aligned score, derived from timing information along with pitch and note value, allows matching a part of the score with the corresponding part of the music audio, enabling various applications. In this paper, we consider an extended version of the traditional note-level transcription task that recognizes onset, offset, and pitch, through including extraction of additional note value to generate a time-aligned score from an audio input. To address this new challenge, we propose an end-to-end framework that integrates recognition of the note value, pitch, and temporal information. This approach avoids error accumulation inherent in multi-stage methods and enhances accuracy through mutual reinforcement. Our framework employs tokenized representations specifically targeted for this task, through incorporating note value information. Furthermore, we introduce a pseudo-labeling technique to address a scarcity problem of annotated note value data. This technique produces approximate note value labels from existing datasets for the traditional note-level transcription. Experimental results demonstrate the superior performance of the proposed model in note-level transcription tasks when compared to existing state-of-the-art approaches. We also introduce new evaluation metrics that assess both temporal and note value aspects to demonstrate the robustness of the model. Moreover, qualitative assessments via visualized musical scores confirmed the effectiveness of our model in capturing the note values.

ComposeOn Academy: Transforming Melodic Ideas into Complete Compositions Integrating Music Learning

Music composition has long been recognized as a significant art form. However, existing digital audio workstations and music production software often present high entry barriers for users lacking formal musical training. To address this, we introduce ComposeOn, a music theory-based tool designed for users with limited musical knowledge. ComposeOn enables users to easily extend their melodic ideas into complete compositions and offers simple editing features. By integrating music theory, it explains music creation at beginner, intermediate, and advanced levels. Our user study (N=10) compared ComposeOn with the baseline method, Suno AI, demonstrating that ComposeOn provides a more accessible and enjoyable composing and learning experience for individuals with limited musical skills. ComposeOn bridges the gap between theory and practice, offering an innovative solution as both a composition aid and music education platform. The study also explores the differences between theory-based music creation and generative music, highlighting the former's advantages in personal expression and learning.