Generators balance quality (how realistic each image looks) with coverage (how much variety they produce). Sharpening individual samples often shrinks variety: a model forced to output one average answer for a noisy input that could lead to many different clean images.
This approach adds a random code so the model can propose several candidate next steps for the same noisy state. Candidates compete during training; only the best match per example is updated. Different image types shape different predictions instead of being blended.
On an image benchmark it improves realism and variety together, and combining it with existing steering methods expands the frontier further. A single model can cover many modes without losing fidelity, and the extra training cost is small.