A pipeline that introduces explicit 3D intermediate representations—first recovering scene geometry using depth estimation or NeRF/3D Gaussian Splatting, then rendering a new view from the target character's position, and finally using an image model to fill in details—may be more controllable than an end-to-end approach.