One of the structural flaws of the generate-then-compress pipeline is that the geometry and appearance encoders usually contain non-differentiable operations, so gradients cannot flow back smoothly during training, making the pipeline hard to optimize end-to-end.