Happy to share our work on discrete denoising diffusion models (D3PMs) @NeurIPSConf 2021: arxiv.org/pdf/2107.03006…. D3PMs are diffusion models for discrete data like text or (quantized) images, and they’re flexible! A thread (with code!) 1/n
3
29
176
0
43
Download Image
Diffusion models generate data by denoising corrupted inputs. D3PMs do the same for discrete data with a much more flexible family of corruption processes! They can randomly replace words, mask pixels, or even replace tokens with synonyms! 2/n