Better than the real thing

[bmaltais]

trafficstars

Based on some of my testing I have to say that your implementation is actually producing better results that the original neural-doodle: https://twitter.com/netputing/status/710119813847207937

[DmitryUlyanov]

Thanks!

[alexjc]

The differences come down to the two core algorithms: gram-based (as this version) works well for more creative or painted styles, and patch-based (as original) works quite well for higher fidelity or photos—as indicated by your other experiment: https://twitter.com/netputing/status/713879168866512900

I've just recently added a new --variety flag that allows the patch-based approach to perform better in creative painted styles too, or texture generation. This means for the first time you can blend between fidelity and creativity: https://github.com/alexjc/neural-doodle

Here's an example rendering of your lily-pond with --variety=0.5 and --semantic-weight=1.0:

The version with --variety=0.0 (as from the paper pre-print) is a bit more accurate but also more boring:

It's great to have such a wide variety of projects to be able to compare results! Thanks for doing this too, it drives progress forward very quickly indeed ;-)