![]() In Advances in neural information processing systems, pages Sherjil Ozair, Aaron Courville, and Yoshua Bengio. ![]() Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, The cramer distance as a solution to biased wasserstein gradients. Lakshminarayanan, Stephan Hoyer, and Rémi Munos. ![]() Marc G Bellemare, Ivo Danihelka, Will Dabney, Shakir Mohamed, Balaji Generalization and equilibrium in generative adversarial nets (gans). Sanjeev Arora, Rong Ge, Yingyu Liang, Tengyu Ma, and Yi Zhang. Martin Arjovsky, Soumith Chintala, and Léon Bottou. Towards principled methods for training generative adversarial The cover illustration of Doujinshi is created by Zhihao Fang, and Jiakai Zhang helps create the website. (a) blonde hair, twintails, blush, smile, ribbon, red eyes (b) silver hair, long hair, blush, smile, open mouth, blue eyes (c) aqua hair, long hair, drill hair, open mouth, glasses, aqua eyes (d) orange hair, ponytail, hat, glasses, red eyes, orange eyes 5.3 Quantitative Analysis 5.3.1 Attribute Precision blonde hair Figure 9: Generated images under fixed conditions. All characters in (a)(b) appear to be attractive, but most characters in (c)(d) are distorted. On contrast, (c)(d) are associated with “glasses”, “drill hair”, which is not well learned because of the insufficiency of corresponding training images. In Figure 9, (a)(b) are generated with well learned attributes like “blonde hair”, “blue eyes”. Table 1: Number of dataset images for each tag Figure 2: t-SNE visualization of 1500 dataset images 3.3 VisualizationĪs Table 1 states, labels are not evenly distributed in our training dataset, which results that some combinations of attributes cannot give good images. We use 0.25 as the threshold and estimate each attribute’s presence independently. We choose the one with maximum probability from the network as the estimated tag.įor orthogonal tags (e.g. We show the selected tags and the number of dataset images corresponded to each estimated tag in Table 1.įor set of tags with mutual exclusivity (e.g. Given an anime image, this network can predict probabilities of belonging toĥ12 kinds of general attributes (tags) such as “smile” and “weapon”,Īmong which we select 34 related tags suitable for our task. We use Illustration2Vec saito2015illustration2vec, a CNN-based tool for estimating tags of anime illustrations 7 7 7Pre-trained model available on for our purpose. ![]() Red line indicate the original bounding box and blue line indicate the scaled bounding box. 3.1 Image Collection Figure 1: Sample Getchu page and the detection result. We propose to use a more consisting, clean, high-quality dataset,Īnd in this section we introduce our method of building such a dataset. We hypothesize that it is due to the fact that image boards allow uploading of images highly different in style, domain, and quality,Īnd believe that it is responsible for a non-trivial portion of quality gaps between the generation of real people faces and anime character faces. Previous works mentioned above all base their approaches on images crawled from one of these image boards,īut their datasets suffer from high inter-image variance and noise. Provide access to a large number of images enough for training image generation models. Web services hosting images such as Danbooru 2 2 2 and Safebooru 3 3 3 , It is well understood that image dataset in high quality is essential, if not most important, to the success of image generation.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |