Working with neural-style

Justin Johnson provides detailed information on GitHub concerning how to use the various neural-style parameters. The following will only discuss the settings specific to Painting Movies.

Style and target imageStyle image (Kandinskys Black and Violett from 1923) and target (Source: PD/Rada)

The GPU-based rendering time for a film frame with a resolution of 1280 x 720 pixels and 1000 iterations using the VGG-19 model by Leon Gatys totals around seven and a half minutes. The processing time for a 800 x 450 pixel image is three minutes.

Computed imagesComputed image with VGG-19 and NIN-Imagenet model

By using the NIN-Imagenet model, the processing time for a film frame with a resolution of 1280 x 720 pixels and 1000 iterations can be reduced to around 45 seconds. However, the quality of the results can fluctuate strongly depending on the source and target images selected.

Test image with original coloursTest image with original colours and 2000 iterations

The maximum output on the test workstation is 1600 x 900 pixels. The computing time for 2000 iterations is 23 minutes. The test image output in the example features the original colours.

Installation of neural-style
Working with neural-style
Style Transfer for Video Clips
Post-production
Alternative Techniques