Where does AI video-to-audio fall short of a Foley pass on surface and timbre control?

Problem. A Foley artist changes shoes, surfaces and props per scene; V2A gives what the model infers. Conditions. Current open V2A models (MMAudio-class) with text or video conditioning. Method. AC-Foley (ICLR 2026) states existing systems cannot synthesise creator-specified variations such as footsteps varying by surface material, attributing it to coarse training labels and text that cannot describe micro-acoustic features; its remedy is conditioning on reference audio. Its own limitations section adds that with several concurrent sources (dialogue, ambience, object interactions) the model may fail to align specific sounds with their visual triggers. Use: for hero footsteps and cloth, treat V2A as a draft layer; supply reference audio of the surface or record the pass. Failure modes: same footstep timbre on all surfaces; ambiguous mixed scenes. Limitations. Authors' own framing of related work; results are their benchmarks, not a human-Foley comparison; none of this was run by me. Version/date. arXiv 2603.15597 (ICLR 2026). Curator qualifications: All cited documentation excerpts located on the live pages by curator GET; no video sources. Limitations: Statements are the authors' own about existing systems; no direct comparison to a human Foley pass is reported. Recommendation to treat V2A as draft is my inference. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://arxiv.org/html/2603.15597",
      "note": "Introduction; Introduction; Appendix E"
    }
  ]
}