Joint audio--video generators are trained on data in which what an event looks like and what it sounds like are strongly, often spuriously, correlated: a particular material, texture, or object appearance co-occurs with a particular sound. This paper is a controlled causal study of the resulting failure mode. Building an AV structural causal model in which the audio is, by construction, independent of the video's nuisance appearance, we show that models which let audio read video directly-through cross-attention or a shared latent-learn a visual shortcut: they predict sound from appearance rather than from the causal event, and collapse when the appearance-event correlation is broken at test time, literally synthesizing the wrong event's sound. Crucially, the popular remedy of routing both modalities through a shared common-cause latent does not fix this: a bottleneck, an unsupervised shared/private factorization, and a faithful shared-prior model all grab the appearance proxy and fail like the direct model. Blocking the shortcut instead requires an intervention on the nuisance. Under the stated SCM and intervention assumptions we prove that counterfactual invariance is necessary and sufficient to identify the causal predictor, and we verify the mechanism across a feature-vector SCM, procedural pixel video, real images with spectrogram audio and a pretrained backbone, moving real digits, and a conditional generator. On a \emph{real, pretrained} video-to-audio generator, an input-intervention test shows the model is far from invariant to sound-irrelevant edits (recolouring or graying a video substantially changes the sound it generates).

翻译：暂无翻译