search

6o6o / chainer-fast-neuralstyle

Chainer implementation of "Perceptual Losses for Real-Time Style Transfer and Super-Resolution".
28 stars 5 forks source link

Larger style features #8

Open CJHFUTURE opened 7 years ago

CJHFUTURE commented 7 years ago

Any advice on training setting to achieve larger or resulting in more abstract features? For example seeing more detail in large brush strokes etc?

6o6o commented 7 years ago

Well, to make features bigger in size, the script doesn't have an explicit option for that, but you can try:

a) train with bigger --image_size like 728 or 1024 which will require 6 or 12 GB of GPU memory and take 2x or 4x amount of time respectively;

b) process a smaller version of an input image and upscale it back to initial resolution afterwards with any super-resolution technique, like waifu2x. This is what I do to process full HD 1080p frames, which wouldn't just fit in my 4GB of memory otherwise, so my input images are 960x540.

Regarding more abstract and complex features transferred, this is a well-known limitation of this algorithm. Fast implementations like this with a single feed-forward pass sacrifice high level features for speed, as opposed to optimization-based techniques. I discuss this issue in more detail here. One thing you may try is increase --lambda_style to smth like 20.0 before training, but it might not give the desired result.