We propose a framework for creating articulated human avatars, editing their styles, and animating the human avatars from three different types of text instructions. The three types of instructions, identity, edit, and action, are fed into three models that generate, edit, and animate human avatars. Specifically, the proposed framework takes identity instruction and multi-view pose condition images to generate the images of a human using the avatar generation model. Then, the avatar can be edited with text instructions by changing the style of the images generated. We apply the Neural Radiance Field (NeRF) and Poisson reconstruction to extract a human mesh model from images and assign linear blend skinning (LBS) weights to the vertices. Finally, the action instructions can animate human avatars, where we use the off-the-shelf method to generate the motions from text instructions. Notably, our proposed method adapts the appearance of hundreds of different individuals to construct a conditionally editable avatar-generated model, allowing easy creation of 3D avatars using text instructions. We demonstrate high-fidelity 3D animatable avatar creation with text instructions on various datasets and highlight a superior performance of the proposed method compared to the previous studies.

Recent advances in deep learning have significantly improved the performance of video prediction, however, top-performing algorithms start to generate blurry predictions as they attempt to predict farther future frames. In this paper, we propose a unified generative adversarial network for predicting accurate and temporally consistent future frames over time, even in a challenging environment. The key idea is to train a single generator that can predict both future and past frames while enforcing the consistency of bi-directional prediction using the retrospective cycle constraints. Moreover, we employ two discriminators not only to identify fake frames but also to distinguish fake contained image sequences from the real sequence. The latter discriminator, the sequence discriminator, plays a crucial role in predicting temporally consistent future frames. We experimentally verify the proposed framework using various real-world videos captured by car-mounted cameras, surveillance cameras, and arbitrary devices with state-of-the-art methods.