The same problem, compressed into four papers
Audio–visual condensation faces the vision–language problem in a sharper form. Each sample pairs a visual frame with an audio clip, the pairing is what a joint model learns from, and matching each modality’s distribution separately can keep both marginals while losing the correspondence between them. The literature is small, but it has already followed the arc the larger settings took over several years.
- Explicit cross-modal terms on random networks. AVDD extends distribution matching to audio and vision with two added losses: one matching the joint distribution, one matching the gap between modalities. Evaluation networks are trained from scratch. On VGGSound-10K at 10 samples per class it reports 55.0%, against 43.9% for plain distribution matching and 68.2% with the full training set.
- Pretrained encoders and a shared/private split. DAVDD replaces random networks with a bank of frozen pretrained audio–visual encoders and separates shared from modality-private features. Its ablation credits roughly equal gains to the pretrained bank and to the decoupling.
- Condense inside an aligned space. ImageBindDC skips bespoke matching networks and condenses directly in ImageBind’s pretrained joint embedding space, matching characteristic functions. It is the clearest case in multimodal condensation of a later method beating its predecessor on accuracy and speed at once, running about 5.7× faster than AVDD under a shared protocol.
- More than two modalities. HoPA carries the LoRS-style recipe (trajectory matching plus a similarity structure) to video, audio and text together. It uses a rank-1 spectral proxy to avoid modeling every pair of modalities, and beats the pairwise variants the authors construct.
What the arc shows
The movement is the same one seen in image classification and image–text condensation: from random networks with hand-designed alignment terms, to pretrained encoders, to condensing inside a representation that is already aligned. Each step improves results and adds a dependency on a large pretrained model.
Open questions
- Scale of modalities. No paper yet distills more than three modalities jointly, so the claim that these methods scale gracefully with the number of modalities is argued, not demonstrated.
- Baselines. Comparisons are mostly against unimodal methods applied per modality, which is a low bar.
- Evaluation. Benchmarks (VGGSound subsets, AVE, MUSIC-21) are small, and no study yet controls for the label-regime and recipe effects that proved decisive for images. See Evaluation.