Basic codes for inversion in StyleGANXL
Check out the colab notebook. Inversion for Imagenet works best at 512x512 resolution.
General Observations :
Works reasonably well for OOD Natural Images and even shapes but requires a number of iterations.
Hard for facial images using Imagenet weights
Hard for cases where main object occupies proprotionally less space to the fore/back ground
Adding codes for more experiments
1 Super Resolution : We follow PULSE/BRGM's method. Harder than naive inversion requires more iterations till convergence
2 Masking : We follow PULSE/BRGM's method. Harder than naive inversion requires more iterations till convergence
3 Colorization : We follow PULSE/BRGM's method. Harder than naive inversion requires more iterations till convergence