pluralistic inpainting

[tumble-weed]

Hello, thanks for the great work. I have a use case to generate multiple samples for inpainting an area on imagenet and pascal. I had some questions surrounding this. I am only tangentially familiar with flow based models, so had some questions

• do you have any opinion on the feasibility of using your training script for inpainting on these datasets. i.e. any comment on the time or any instability that may be likely on larger datasets like imagenet.
• how do i generate multiple samples for the same mask during inference
• as far as i am aware one of the properties of flow based models is to also be able to give the likelihood of the samples. is this possible? thanks again.

[p0p4k]

Hi, for your third question you might refer to this https://github.com/cantabile-kwok/VoiceFlow-TTS#easter-eggs--citation and FFJORD (https://github.com/rtqichen/ffjord). For the second question, I am thinking it will be like diffusion, condition the cfm estimator (unet usually) with mask of inpainting target. However, since I am exploring that area, I will let you know once I get more details.