3 ms·

Very interesting! This seems somewhat similar to the recently published GIFnets[1]. However, I believe GIFnets is training a reusable network to a predict pale

by dbaupp 6y ago

Very interesting! This seems somewhat similar to the recently published GIFnets[1]. However, I believe GIFnets is training a reusable network to a predict palettes, and pixel assignments, while this post is focusing on optimising the "weights" (i.e. pixel values) for a single image. I wonder if the loss functions from GIFnets could be applied to this single-image approach to potentially solve the banding problem via something a little more "perceptual" than the variance term mentioned. [1]: "GIFnets: Differentiable GIF Encoding Framework" https://arxiv.org/abs/2006.13434 https://arxiv.org/abs/2006.13434

underanalyzer 6y ago
That's interesting! One thing I was surprised about is that they don't address optimizing the palette and dither pattern across time (b/c most gifs are animated). This feels to me like it would be really interesting and a hard problem for traditional algorithms. They do mention it as a possibility for future work at the end tho. They also seem to have separate losses for the palette net and the dither net instead of just adjusting both to optimize a general image quality metric (although it does look like they have some kind of perceptual loss, it's just not the only objective)