PrismAudio: Decomposed Chain-of-Thoughts and Multi-dimensional Rewards for Video-to-Audio Generation

TL;DR

PrismAudio tackles video-to-audio generation by decomposing reasoning into four specialized CoT modules aligned with semantic, temporal, aesthetic, and spatial goals, optimizing them via multi-dimensional reinforcement learning. It introduces Fast-GRPO, a hybrid ODE-SDE sampling approach that enables efficient training for diffusion-based audio generation, and AudioCanvas, a rigorous benchmark with rich multi-event scenarios and CoT annotations. Empirical results show state-of-the-art performance across all perceptual axes on both in-domain VGGSound and out-of-domain AudioCanvas, with extensive ablations confirming the benefits of multi-dimensional CoT-RL and the efficiency of the Fast-GRPO framework. The work advances controllable, interpretable V2A generation and provides practical insights into balancing competing perceptual objectives in multimodal synthesis.

Abstract

Video-to-Audio (V2A) generation requires balancing four critical perceptual dimensions: semantic consistency, audio-visual temporal synchrony, aesthetic quality, and spatial accuracy; yet existing methods suffer from objective entanglement that conflates competing goals in single loss functions and lack human preference alignment. We introduce PrismAudio, the first framework to integrate Reinforcement Learning into V2A generation with specialized Chain-of-Thought (CoT) planning. Our approach decomposes monolithic reasoning into four specialized CoT modules (Semantic, Temporal, Aesthetic, and Spatial CoT), each paired with targeted reward functions. This CoT-reward correspondence enables multidimensional RL optimization that guides the model to jointly generate better reasoning across all perspectives, solving the objective entanglement problem while preserving interpretability. To make this optimization computationally practical, we propose Fast-GRPO, which employs hybrid ODE-SDE sampling that dramatically reduces the training overhead compared to existing GRPO implementations. We also introduce AudioCanvas, a rigorous benchmark that is more distributionally balanced and covers more realistically diverse and challenging scenarios than existing datasets, with 300 single-event classes and 501 multi-event samples. Experimental results demonstrate that PrismAudio achieves state-of-the-art performance across all four perceptual dimensions on both the in-domain VGGSound test set and out-of-domain AudioCanvas benchmark. The project page is available at https://PrismAudio-Project.github.io.

Figures (5)

• Figure 1: Overview of PrismAudio. Left panel: the progress of CoT training data construction using Gemini 2.5 Pro and then fine-tuning VideoLLaMA2 for decomposed CoT generation (Section \ref{['subsec:multidimensional-cot']}). Right panel: the Fast-GRPO multi-dimensional CoT-RL framework (Section \ref{['subsec:fast-grpo']}) for post-training the Audio Foundation Model (Section \ref{['subsec:audio_foundation_model']}).
• Figure 2: Training convergence on Semantic reward measured by the CLAP score.
• Figure 3: Qualitative comparison of PrismAudio against baseline models.
• Figure 4: The bar chart illustrates the distribution of audio event classes within the AudioCanvas benchmark.
• Figure 5: Empirical validation of ODE--SDE distribution equivalence during training.