WorldSonus

What it does

WorldSonus is a research video-to-audio model from the Hong Kong University of Science and Technology and Noiz AI, with co-authors at MetaX and Shanghai Jiao Tong University. It writes 48 kHz stereo in 100 millisecond chunks from frames that have already arrived, so a world model can keep drawing while sound follows. A text prompt can change at the next chunk without recomputing earlier audio. The 6 October 2026 release includes the paper arXiv:2610.08760, inference code, and weights. Noiz says it is not in Noiz Studio. The project code and these weights are CC BY-NC 4.0: attribution, and non-commercial use only. The paper's 41.2 millisecond timing is a measurement on one H100; the public script does not promise that latency. No price is listed.

Where it fits

Read the project page, then run the published inference script on a video and a sound prompt. Use streaming mode if you want each chunk as three frames arrive. Treat the audio as generated research output, and check the non-commercial license before any other use.

Before you commit

Non-commercial license. Not a control in Noiz Studio, and the public code does not guarantee the paper's latency. Generated sound can miss events in the picture.

Release & update notes

Dates below refer to the linked announcement or Product Hunt launch. They are separate from the date this profile was added.

• Official announcement / update · 2026-10-06 ↗

  Research release: paper, code, and non-commercial weights.