"Input Perturbation Reduces Exposure Bias in Diffusion Models. (arXiv:2301.11706v1 [cs.LG])" — An exploration of the fact that the the long sampling chain in Denoising Diffusion Probabilistic Models (DDPM) leads to an error accumulation phenomenon, which is similar to the exposure bias problem in autoregressive text generation.

Paper: http://arxiv.org/abs/2301.11706

#AI #CV #NewPaper #DeepLearning #MachineLearning

<<Find this useful? Please boost so that others can benefit too 🙂>>
The inputs and the prediction t…