ImageBind links six data types, with limits in the training mix

Meta introduced ImageBind on 9 May, linking images and video, text, audio, depth, thermal information and motion-sensor readings in a shared representation. The release includes code and pretrained weights.

Photo: LPS.1, “Meta Platforms Headquarters Menlo Park California” (2022). CC0 1.0. Resized; cropped; responsive crops.

The launch documentation describes image-paired training and a shared embedding space.

The launch model card limits thermal training to outdoor streets and depth to indoor scenes. The text encoder likely works only in English. The model is for research, not deployment.

The README names CC BY-NC 4.0; the licence and model card specify CC BY-NC-SA 4.0. Both restrict commercial use; only the latter names ShareAlike.