I am happy to share the results of my first serious attempt at training a model based on the insane Marigold V2 by @AntonObukhov1 and his team! AI generated depth of field has been historically unusable, lacking detail and temporal coherence. MarigoldV2 solved the detail part and I set out to solve the coherence. Through a series of targeted training passes focusing on large scale morphing and flicker first, I then moved to boundary and edge detection before finally squeezing back in the detail using Marigold V2s Sink Loss technique The end result speaks for itself and while still not perfect, I have learned tons and have set out to expand my personal library of high quality CGI, photoreal scenes so I can do a much larger training run in future. Training was completed over 400 H200 Hours.