Idea of this thread is to share our failed ML neural style transfer attempts, with some insights in what work, what doesn't. Also to share some learning resources that would help us understand better how to successfully train our own ML models.
I successfully trained ML style transfer example with noise generator given from SideFX.
After that I thought to train it on our own terrain example. Here is first result of that, posting them in the order they were trained:
First two were trained with very big range of the lowest and highest point in the terrain, while 90% of terrain was plateau. Moment I isolated only that, thing started working better. I trained with different data set size, resolution, training rate...
This kind of visualizing onnx generations alognside with graph seems to me like a good way to make a sense of the process.
Did anyone try to use this? If so, what would be best learning resources for us VFX artists to take, so we are not left behind
