Weiyi You · Jinhua Zhang · Xingyu Zhou · Wei Long · Junyu Lou · Shuhang Gu†
University of Electronic Science and Technology of China
† Corresponding author
PixSGR is an efficient pixel-space diffusion framework for high-fidelity image generation. It starts from a supervised low-channel bottleneck that captures the low-dimensional manifold of natural images, then progressively expands the channel dimensionality and spatial resolution to recover fine-grained structures. During spatial refinement, coarse-scale attention maps preselect globally relevant interactions for fine-scale sparse attention, enabling non-local refinement without the quadratic cost of dense attention.
On ImageNet, PixSGR achieves FID scores of 1.51 at 256 × 256 and 1.60 at 512 × 512.
PixSGR couples cross-scale token scoring with guided sparse attention and convolutional upsampling.
For questions or collaborations, contact Weiyi You.
If you find this work useful, please consider starring the repository.