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

TL;DR

MMDisCo introduces a training-based, model-agnostic joint guidance mechanism that leverages two pre-trained single-modal diffusion models (audio and video). A lightweight discriminator is trained to distinguish real versus independently generated pairs, and its gradient provides the cross-modal score adjustment, stabilized by a DLSM-inspired denoising regularization. The approach preserves single-modal fidelity while improving audio-video alignment across in-domain and out-of-domain benchmarks, with only a small parameter footprint and modest inference overhead. This enables scalable integration of state-of-the-art single-modal generators into joint audio-video synthesis without extensive retraining of base models.

Abstract

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 single-modal 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.

Figures (8)

• Figure 1: Overview of the training process of our proposed method. We train a joint discriminator on the top of two base diffusion models to distinguish real video-audio pairs from fake ones generated by base models. Additionally, we adopt a denoising objective, as in standard diffusion models, to match the gradient of the discriminator with regard to the inputs to the residual noise between ground truth noises and predicted noises from base models.
• Figure 2: Visualization of guidance results on toy datasets. The top row shows samples drawn from ground truth distribution (GT), and the bottom row shows generated samples (Gen.).
• Figure 3: Generated samples from AnimateDiff / AudioLDM and ours trained on the VGGSound dataset. We used the captions shown on the leftmost side and the same random seed for both settings to generate samples.
• Figure 4: Visualization of the generated samples across loss functions used to train our guidance module with the toy dataset. The top row shows the IND setting, and the bottom shows the OOD setting.
• Figure 5: Ablation study for the training epochs. We trained the discriminator with $C=128$ and $L=2$ for 300 training epochs. We generated samples with five different random seeds and computed their average and std.