MMDisCo: Multi-Modal Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation

This study aims to construct an audio-video generative model with minimal computational cost by leveraging pre-trained single-modal generative models for audio and video. To achieve this, we propose a novel method that guides singlemodal models to cooperatively generate well-aligned samples across modalities. Specifically, given two pre-trained base diffusion models, we train a lightweight joint guidance module to adjust scores separately estimated by the base models to match the score of joint distribution over audio and video. We show that this guidance can be computed using the gradient of the optimal discriminator, which distinguishes real audio-video pairs from fake ones independently generated by the base models. Based on this analysis, we construct a joint guidance module by training this discriminator. Additionally, we adopt a loss function to stabilize the discriminator's gradient and make it work as a noise estimator, as in standard diffusion models. Empirical evaluations on several benchmark datasets demonstrate that our method improves both single-modal fidelity and multimodal alignment with relatively few parameters. The code is available at: https://github.com/SonyResearch/MMDisCo . Published as a conference paper at ICLR 2025 a trade-off between the quality of generated samples and model dependency, which increases the computational cost. In this paper, we propose a novel method that is training-based but model-agnostic. Our method does not require backpropagation through the base models for the optimization. Specifically, we introduce a lightweight joint guidance module on top of audio and video base models that adjust their outputs for audio-video joint generation. We assume that pre-trained base models are black box diffusion models (i.e., we can access only their outputs and do not depend on a specific architecture design like a cross-attention module to construct a joint generation model). We formulate the joint generation process as an extension of the classifier guidance (C-guide) for single-modal data (Song et al., 2021; Dhariwal and Nichol, 2021) . We show that this joint guidance can be computed through the gradient of the optimal discriminator that distinguishes real audio-video pairs from the fake ones independently generated by base models. We only train the discriminator with proper regularization inspired by Denoising Likelihood Score Matching (DLSM) (Chao et al., 2022) . Extensive experiments on several benchmark datasets demonstrate that our proposed method can efficiently integrate single-modal base models for audio and video into a joint generation model, maintaining the performance of each single-modal generation without incurring a significant computational cost (see Appendix A.9). RELATED WORK AUDIO-VIDEO JOINT GENERATION BY DIFFUSION MODELS Since an audio-video pair is one of the most popular types of multimodal data, several works train diffusion models with such pairs to achieve a conditional single-modal generation: video-conditional audio generation (