Can we use VL-BERT as image or text generator?

[xealml]

use VL-Bert as a generator for text or image or both?

[jackroos]

Yeah, this is an interesting topic. Theoretically, VL-BERT can be used as image/text generator since it models joint distribution of image RoIs and text tokens (e.g., this paper is a method to use BERT for text generation). But pre-training tasks of VL-BERT is not designed for generation, so other pre-training tasks may be needed to obtain good results. Thanks!

[xealml]

Thanks for your reply.