Princeton study: one-step image generators hide the denoising trajectory in network depth

cindy_x_wu · x · 2026-10-13

Princeton researchers present "Depth as Time in One-Step Generative Models," asking where the denoising trajectory goes when multi-step diffusion is compressed into a single forward pass.

Key findings:

Depth as time: the denoising computation that multi-step diffusion spreads across sampling steps unfolds across the depth of a one-step generator, and can be recovered by decoding intermediate layers with the model's own output head.
Depends on the transport task: the depthwise behavior hinges on the transport task a flow map is trained on. MeanFlow is the surprise — probing shorter transport intervals reveals both denoising and renoising within a single network evaluation, while drifting models without time-indexed transport tasks lack the effect.
Compression payoff: models exhibiting depthwise denoising are more compressible. Treating layerwise computation explicitly as a flow, a single time-conditioned block can be trained to denoise across layers, shrinking a MeanFlow SiT-L/2 model by 16.6× in parameters.

The results suggest the "temporal computation" of diffusion can be restructured into depth or a single block, opening new angles for one-step generative model compression.

Original post →

Princeton study: one-step image generators hide the denoising trajectory in network depth

More from Multimodal