neural-style
neural-style copied to clipboard
jcjohnson
About people face not good
When I use the original VGG-19 model,I found when face in the picture,the face in the finished picture is not good.How could I change the options to get good effect.Who can give me advise? Thanks!
My experiments (not vary familiar with ML) showed some options for this problem
- Try to play with style_scale and style_layers. I have python scripts that try almost all valuable combinations of these params (keep num_iterations to greatly speed up processing).
- I have better results with Texture nets in term of less distortion. Also you can try additional params that have some forks of this repository.
- There are some implementation of neural styling that provides masks for processing. You can use OpenCV to detect a face and apply a mask to it.
I've tried style_scale,style_layers,content_weight,style_weight and other options,but the face in the finished picture is still not good. Can you tell your options? Thanks
My last time I've used these options, but the result depends on the styling image anyway:
python run_neural_style.py -style_scale 1.2 -init image -style_image .... -content_image ~... -image_size 750 -output_path ... -content_weight 5 -style_weight 50 -save_iter 100 -num_iterations 1000 -gpu 0 -optimizer adam -tv_weight 0.001 -learning_rate 6 -seed 123 -style_layers relu1_2,relu2_2,relu3_2,relu4_2 -original_colors True -content_layers relu5_2
In my opinion, you may need to find out what kind of exact style you want. You may not have to use all the initial style layers( relu1_2,relu2_2,relu3_2,relu4_2 ). I had a good result with relu2_1,relu3_1,relu4_1. So you can start with -content_weight 0 to find the style you expected. What's more, L-BFGS tends to give better results.
How could I choose style_layers?I found when I userelu1_2,relu2_2,relu3_2,relu4_2 or relu1_2,relu2_2,relu3_2,relu4_2,the face is worse... What's the mean of style_layers?I know this is neural network layers,but how I know what layer fit me?
I may misleaded you, but I actually got a nice result. This page may help you #363
@devpuck You can check out my experiments in fine tuning a model to improve facial feature preservation here: https://github.com/jcjohnson/neural-style/issues/292#issuecomment-300685209
