WorldCompass teaches video world models to follow actions better and keep pictures pretty by using reinforcement learning after pretraining.
ThinkRL-Edit teaches an image editor to think first and draw second, which makes tricky, reasoning-heavy edits much more accurate.