xunhuang1995/AdaIN-style

Introduces an adaptive instance normalization layer that dynamically computes affine parameters from style features, enabling arbitrary style transfer at 15 FPS on 512×512 images—720× faster than prior methods. Implements fine-grained control through content-style weighting (alpha), color preservation, multi-style interpolation, spatial masking for region-specific styles, and video processing. Built in Torch with VGG-19 feature extraction and trainable decoder, supporting both inference and training on COCO and WikiArt datasets.

1,557 stars. No commits in the last 6 months.