Welcome to StormRaiser's experimental anime portrait generator with independently controllable content and style. In case you are not coming from our project page, this is also an online demo of our ICCV 2021 paper DisUnknown: Distilling Unknown Factors for Disentanglement Learning
Note that this demo is based on, but not an exact implementation of, our paper: in addition to disentangling style and content as in the paper, here we also attempted to disentangle color from non-color attributes. Fine-grain color labels were not available, which means we had to disentangle one labeled factor (style) and two unknown factors (color and non-color content) from each other, so domain-specific methods were required on top of the general framework proposed in our paper. We also had to reduce the capacity of our network to keep the computation cost reasonable for a web demo. Meeting all these requirements at once turned out to be not easy and consequently we have not achieved the ideal result yet, but we still hope that you find this work interesting.
"Randomize all" will randomize all dimensions. "Randomize dimensions higher than" will randomize higher dimensions in the code while keeping lower diensions fixed. Individual dimensions can be adjusted using the sliders. Since the code is quite long they are grouped into pages. Use the page selector to switch pages. For completeness all dimensions are adjustable, but we employed Nested Dropout to concentrate important modes of variation in the earlier dimensions, so basically only about the first 6 pages in the "non-color attributes" panel and the first 2 pages in the "drawing style" panel are actually useful.
Drawing style can be set to that of one of several specific artists. Just type in their name and don't forget to click the button. Our training data was obtained from Danbooru so artist names should match their tags on Danbooru.
In color settings, hair and eye colors are set by combining a base color and deltas beetween the base color and the color for sub-regions. The effect of the deltas and the face colors are somewhat subtle, but their range in the training data is small to begin with.