Depth as Time in One-Step Generative Models
One-step image generators appear to move diffusion’s denoising process from sampling steps into network depth.
The paper reports that intermediate layers in some one-step models can be decoded as stages of a denoising trajectory. The effect depends on the transport task used in training: MeanFlow shows denoising and renoising inside one forward pass, while drifting models do not show the same pattern. The authors also say this structure makes such models more compressible, including a `16.6x` parameter reduction for a MeanFlow `SiT-L/2` model into a single time-conditioned block. ArXiv · AI/CL/LG's note
The paper reports that intermediate layers in some one-step models can be decoded as stages of a denoising trajectory. The effect depends on the transport task used in training: MeanFlow shows denoising and renoising inside one forward pass, while drifting models do not show the same pattern. The authors also say this structure makes such models more compressible, including a `16.6x` parameter reduction for a MeanFlow `SiT-L/2` model into a single time-conditioned block. ArXiv · AI/CL/LG's note
score 5