- Author
- Zeyue Tian, Zhaoyang Liu, Yizhu Jin and others, HKUST
- Published
- Undated (accepted to ICLR 2026)
- Kind
- Research project page with abstract and audio and video samples
- Original
- zeyuet.github.io/AudioX/
- Subject
- Audio models and analysis
- Themes
- Art through code
What it is
A team at HKUST and one independent researcher present AudioX. The model generates audio and music from text, video, image and audio conditions. The page says the work was accepted to ICLR 2026.
What it says
- AudioX is one framework making audio from text, video, image or audio inputs.
- Its core piece is a Multimodal Adaptive Fusion module that merges the different inputs.
- The team built IF-caps, a training set of over 7 million samples.
- The authors claim superior results to prior methods, especially in text-to-audio and text-to-music.
- The page offers samples for six tasks, including inpainting and music completion.
- Comparison sections place AudioX beside models such as MMAudio, Tango 2 and AudioGen.
What evidence it gives
The page offers an abstract, a figure caption and many sample comparisons. The caption mentions Inception Score and instruction-following benchmarks. Named test sets include AudioCaps and MusicCaps. No result numbers appear in the page text.
Interests and limits
The authors present their own model and chose the samples. The linked paper, code and dataset were not opened. Only the page text was read; audio, video and figures were not viewed or heard. Claims of superior performance are the authors' own.