AI Transfer Style - Reinforcement Learning

Hi D5 and Community,

I’ve been diving into AI Transfer Style to create a watercolor-style image. While the AI TS works really great and creates a nice style, it has a hard time detecting two different roof types.

This is the original image that I would like to create a watercolor image from:

The first image is using the smallest weight transfer style and max shape matching:

As you can see, it created all the roofs with the same copper texture.

Using the same settings and about 6 or 7 tries later, the AI made all the roofs dark to match the asphalt:

There was one time that the cartoon style recognized it. So I tried to use that as the base, but that was unsuccessful:


My workaround is to take different images and piece them together in Photoshop.

I have two suggestions that that could be implemented in the future and would help in creating a better image and better workflow:

  1. Lock certain areas I like and regenerate other areas that need work (Similar to Photoshop Generate, where you can generate in a specific area). In this case, it would be nice to save the copper roof on the right and regenerate the asphalt roof in the middle until it was dark.

  2. Another idea that could help is to create a reinforcement learning system that would help push the AI in a certain direction that meets your needs for a specific render. Rather than generating image after image, hoping to find the result you are looking for.

I hope this information is useful and helpful. Thank you for taking the time to consider this.

Best,

KB