Adversarial Masked Diffusion for Path-Generation Decoupling

Diffusion models usually learn generation and reverse-path construction through the same regression objective. This coupling is stable, but it also makes generation a conditional point-estimation problem: when a corrupted state admits many plausible clean completions, full-sample regression tends to favor averaged solutions. We propose Adversarial Masked Diffusion (AMD), a masked diffusion framework for path-generation decoupling. In AMD, generation and path formation are assigned to different learning signals: clean-sample generation is learned through adversarial distribution matching, while reverse-path preservation is enforced only through regression on the coordinates that remain visible in the corrupted input. Because generation is handled adversarially rather than by a timestep-specific full regression target, the generator does not require an explicit timestep input; it is used as a shared clean predictor across all reverse steps. The AMD discriminator also receives no timestep embedding, corrupted state, mask, or conditioning image. It receives only a clean-space sample, either a real clean image or a generated clean prediction, and outputs multiple timestep heads. The sampled reverse step selects which discriminator head supplies the adversarial feedback. The masked diffusion process uses nested Bernoulli masks in data space: as the corruption level increases, the set of replaced coordinates grows monotonically, while replacement values are sampled independently at each level. A visible-coordinate path loss anchors the prediction to the unmasked part of the input, preventing the adversarial generator from ignoring the reverse state. Thus, AMD decouples generation from path preservation: adversarial learning matches the clean data distribution, and regression constructs the reverse path.