Persistence Forcing: Exploiting Feature Specialization in Pixel-Space Diffusion
PerF splits diffusion features into persistent global carriers and active detail refiners, then uses that split to guide image generation.
The paper says heterogeneous refinement in pixel-space DiTs makes some features encode global structure while others focus on local, high-frequency detail. Persistence Forcing uses the persistent features to condition the active ones during sampling. The authors report PerF-L reaches 1.91 FID on ImageNet 256x256 with about half the parameters of JiT-H, while PerF-H reports 1.63 at 256x256 and 1.76 at 512x512. HF Daily Papers' note
The paper says heterogeneous refinement in pixel-space DiTs makes some features encode global structure while others focus on local, high-frequency detail. Persistence Forcing uses the persistent features to condition the active ones during sampling. The authors report PerF-L reaches 1.91 FID on ImageNet 256x256 with about half the parameters of JiT-H, while PerF-H reports 1.63 at 256x256 and 1.76 at 512x512. HF Daily Papers' note
score 4