The flagship architecture, V3, is much more streamlined, yet more flexible and efficient than previous models. V3 looks at the image from the UV color dimensions (which together encode chromaticity), leaving the Y dimension (luminance) unchanged. A lookup is generated from the source image, and applied to the UV dimensions. The final image is then reassembled.
Now, one may inquire as to the reason for generating a lookup for every inference image. V3 follows the 'mapping' schematic as it brings the benefit of 1) representing image transformations with significant, significant dimensionality reduction (as opposed to networks like autoencoders), and 2) image resolution-independent inference speed. Yes. An 8K image would be fixed at the same speed as it would take to fix a 256px image, all within 1-2 milliseconds average.
InnoColor utilizes a deep-learning guided image transformation to modify visual content, like images, video, and GUI, in both digital software and physical interfaces.
