Procedural Fish Modeling.

Abstract

It is time- and man-power intensive to craft various fish species for underwater animations and games. Even professionals spend hours to days for one. Therefore, we propose Procedural Fish Generation, which presents an innovative and automatic approach to generate 3D fish models with one lateral image. The core lies in parameterizing the ray-finned fish with curves and optimizing them with textures to fit the input using differentiable rendering, greatly reducing the manual modeling work. It presents advantages over multi-image reconstruction in requiring single image, while state-of-the-art methods suffer from such a scenario to achieve informative reconstruction. Also, our method outputs a polygon mesh, widely compatible with modern graphics hardware and software, thus facilitating further editing. Furthermore, we fine tune the prompts for Stable Diffusion while users can type a name to find high-quality lateral images. Extensive ablation studies and comparisons have proved its effectiveness and efficiency for experts and non-experts.