Reinforcement Learning Camera Preset Generator
- 6 Devlogs
- 22 Total hours
A reinforcement learning agent that can determine the settings needed to make a raw image look like it’s been through a certain filter.
A reinforcement learning agent that can determine the settings needed to make a raw image look like it’s been through a certain filter.
I fixed the video demo to show in a reasonable amount of time and built a demo cli with the sunglasses filter as an exe.
I built a program that makes a cool animation to show before and after the filter
A good model run testing with 28 photos with my sunglasses as the filter: it’s finally scoring well ~88% and actually looks like it should.
I’ve also worked on a cli tool to work with the ml_model output to apply filters, view histograms, list settings, and score models
first full model training and it went TERRIBLE. My scoring was based on the histogram for the image, so color is not preserved
. These photos have very similar average light values. I’m going to rerun the model with a new scoring algorithm that scores each color channel individually.
Test model done on basic settings (Averaging around 92% accuracy, which is good, but it did it in a completely different way)
Next I’m going to train it on a more complex filter
I’m working on a reinforcement learning photo model that can determine what post-processing settings can be used to make an image look like certain filters (infrared filters, sunglasses, etc)
Right now my biggest bottleneck is waiting for training. This test on a simple setting took over 16 hours to train. I am also not very familiar with RL, so I’m learning as I go.