Raw waveform diffusion can now deliver the same—or even higher—audio fidelity that autoencoder‑based pipelines have long claimed as their exclusive domain. By discarding any latent compression step, WavFlow produces samples that listeners cannot distinguish from those generated by established latent diffusion models.
For years the community has built audio generators on top of semantic‑acoustic autoencoders, a strategy epitomized by Stable Audio 3, which first compresses waveforms into a compact latent space before applying diffusion. This two‑stage design has been justified as necessary to tame the high dimensionality of raw audio and to keep training tractable. WavFlow’s VGGSound results prove that a pure‑waveform approach is competitive: “Experimental results show that WavFlow achieves competitive results on the video‑to‑audio benchmark VGGSound (FD 59.98, IS 17.40, DeSync 0.44) … matching or exceeding the performance of established latent‑based methods” [1]. The FD score of 59.98 sits squarely within the range of top latent models, while the IS and DeSync numbers confirm comparable perceptual quality and temporal alignment.
On the text‑to‑audio front, WavFlow even sets new records: “Our model attains the best FD (10.63) and IS (12.62) reported to date, rivaling dedicated T2A systems” [1]. Those figures surpass the best published latent‑based scores on AudioCaps, demonstrating that raw‑space diffusion does not sacrifice semantic relevance for fidelity.
When the architecture is scaled to the 16 kHz “L” variant, the gap widens: “Scaling to WavFlow‑L‑16kHz yields consistent improvements, surpassing MMAudio‑L‑44.1kHz in distributional fidelity (FD: 59.98 vs. 60.60) while matching its performance in perceptual and alignment metrics (IS 17.40, DeSync 0.44)” [1]. This head‑to‑head comparison shows that raw‑waveform diffusion can outpace a leading latent system on the most demanding distributional metric while staying on par elsewhere.
The study’s scope still leaves open several practical concerns. Training required five million video‑text‑audio triplets and a custom amplitude‑lifting scheme to keep optimization stable, implying a higher data and compute budget than many latent pipelines. Moreover, the best results are reported at 16 kHz, whereas many production scenarios demand 44.1 kHz or higher fidelity, raising the question of whether the same gains will hold at those rates.
If these results generalize, the default assumption that audio synthesis must pass through an encoder‑decoder bottleneck should be revisited. Future benchmark suites ought to include a raw‑waveform diffusion baseline, and engineers can consider dropping the autoencoder stage altogether when building multimodal generation systems.