How would you use multiple style images with their corresponding masks? about neural-style-tf HOT 4 CLOSED

--style_imgs s1.jpg s2.jpg --style_mask --style_mask_imgs m1.jpg m2.jpg

You need a mask for each style image though.

from neural-style-tf.

To recreate the last segmentation example as given in the Readme, I executed the following, which follows the pattern you mentioned above.

python neural_style.py --content_img 02270.jpg --content_img_dir examples/segmentation --style_imgs okeffe_red_canna.jpg okeffe_iris.jpg --style_imgs_dir examples/segmentation --style_mask_imgs 02270_mask_face.jpg 02270_mask_face_inv.jpg --verbose --model_weights "/xx/imagenet-vgg-verydeep-19.mat" --optimizer adam --max_iterations 1000 --print_iterations 10

(The png images were converted to jpg prior to use)

This is the associated metadata :

image_name: result
content: 02270.jpg
styles[0]: 1.0 * okeffe_red_canna.jpg
style_masks[0]: 02270_mask_face.jpg
style_masks[1]: 02270_mask_face_inv.jpg
init_type: content
content_weight: 5.0
style_weight: 1000.0
tv_weight: 0.001
content_layers: ['conv4_2']
style_layers: ['relu1_1', 'relu2_1', 'relu3_1', 'relu4_1', 'relu5_1']
optimizer_type: adam
max_iterations: 1000
max_image_size: 512

I am unable to upload result here. Only one style image is being used for stylization.

from neural-style-tf.

Hi njordsir,

I think you need to add the flag --style_mask. I should remove that and just check if the length of the style mask images list is >0.

Cam

from neural-style-tf.

Adding the --style_mask flag and setting the blending weights as --style_imgs_weights 0.5 0.5 generated the desired result.

Maybe the blending weights could be equally distributed among all style images instead of the current default of 1.0, unless otherwise specified.

Thanks for the very well-written code and your help.

from neural-style-tf.