Using the linear-color-transfer.py Python script from my Neural-Tools project for Histogram Matching style transfer, greatly improves Neural-Style outputs as per the examples here, and countless experiments conducted by others.
Using a modified version of my code and GIMP, I took the masked regions of each style image and then applied them to the target content image regions. Essentially this was a partially manual version of Histogram Matching style transfer. The result from one of my experiments with masked linear color transfer can be seen below, along with the control test:
Regular Masked Style Transfer:
Masked Histogram Matching:
The above images were produces with the deep-photo-styletransfer, which is based on the Neural-Style code. This modified script adds the Spatial Control features from deep-photo-styletransfer into Neural-Style: neural_style_seg.lua
Notice the grey/faded areas on the non-histogram matching image, compared to the masked histogram matching image.
I think that an improved version of the linear-color-transfer.py Python script with support for masked regions (using input masks) would create better outputs with deep-photo-styletransfer, by eliminating the issues caused by extremely dark and light colors, and non-matching histograms. Though I am not sure how to go about adding support for masked regions in the script.
Using the linear-color-transfer.py Python script from my Neural-Tools project for Histogram Matching style transfer, greatly improves Neural-Style outputs as per the examples here, and countless experiments conducted by others.
Using a modified version of my code and GIMP, I took the masked regions of each style image and then applied them to the target content image regions. Essentially this was a partially manual version of Histogram Matching style transfer. The result from one of my experiments with masked linear color transfer can be seen below, along with the control test:
Regular Masked Style Transfer:
Masked Histogram Matching:
The above images were produces with the deep-photo-styletransfer, which is based on the Neural-Style code. This modified script adds the Spatial Control features from deep-photo-styletransfer into Neural-Style:
neural_style_seg.luaNotice the grey/faded areas on the non-histogram matching image, compared to the masked histogram matching image.
I think that an improved version of the linear-color-transfer.py Python script with support for masked regions (using input masks) would create better outputs with deep-photo-styletransfer, by eliminating the issues caused by extremely dark and light colors, and non-matching histograms. Though I am not sure how to go about adding support for masked regions in the script.