pix2pixHD_ACM published on github

How our system works

We have open sourced our modifications to the pix2pixHD project.  Our modifications allow users to produce high definition images combined with text descriptive data to control the attributes of the generated image.

pix2pixHD_ACM enhancements

• Affine Combine Module (ACM) layer

• Optional Self-attention layer

• Spectral Normalization

• Adaptive Instance Normalization

• Data manipulation routines including color jitter, crop+stretch, rotate

• Both network types supported with ACM: global and local enhancer

• Tested up to 1024x1024 image output

• Updated to use new torch.cuda.amp API (pytorch 1.7+)

• Support non-cuda devices for training/inference.

New generate.py script

• Supports linux V4L video input and output

• Live generating from directory (inotify-based)

• Generate from template yaml

https://github.com/deepglugs/pix2pixHD_ACM