Wepresentanewcompactbutdenserepresentationofscenegeometrywhichisconditionedontheintensitydatafromasingleimageandgeneratedfromacodeconsistingofasmallnumberofparameters.Weareinspiredbyworkbothonlearneddepthfromimages,andauto-encoders.Ourapproachissuitableforuseinakeyframe-basedmo