Trained in One Leap

At the end of the chain the state keeps essentially nothing of the image it came from, so near-identical inputs were paired with wildly different targets. Squared error is smallest when the answer given is the average of the targets an input is compatible with, and the average of a pile of unrelated images is a smeared wash.

Squared error can never produce a sharp image; pushing outputs towards realism is precisely what the discriminator in a GAN was for. With no adversarial term in the objective, a network trained this way is bound to hedge, and no amount of data or capacity will change that.

The forward process destroys the image irreversibly, so nothing can be recovered from its final state. Diffusion only appears to invert it because it always restarts from a partly-noised real image rather than from noise, and true generation from scratch is not actually happening.

The plateau is a capacity limit: no single network can hold the entire mapping from noise to images in its weights, so it settles for something generic. The multi-step method wins because spreading the same work across hundreds of applications gives hundreds of times the effective capacity.