AI-Sketcher : A Deep Generative Model for Producing High-Quality Sketches
Authors: Nan Cao, Xin Yan, Yang Shi, Chaoran Chen2564-2571
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed technique was evaluated based on two large-scale sketch datasets, and results demonstrated its power in generating high-quality sketches. Our evaluation showed that AI-Sketcher outperformed the baseline models in generating coherent sketches on both datasets. We performed three experiments based on the above datasets to validate AI-Sketcher s drawing quality, its capability to generate multi-class sketches, and generation diversity. |
| Researcher Affiliation | Academia | Nan Cao, Xin Yan, Yang Shi, Chaoran Chen Intelligent Big Data Visualization Lab, Tongji University, Shanghai, China {nancao, xinyan, yangshi, crchen}.idvx@gmail.com |
| Pseudocode | No | The paper describes algorithms and models using mathematical equations and textual descriptions, but it does not include any explicit pseudocode blocks or algorithm listings. |
| Open Source Code | No | The paper does not contain any explicit statement or link indicating that the source code for the AI-Sketcher model is publicly available. |
| Open Datasets | Yes | We verified the performance of the proposed technique by comparing it with Sketch-RNN based on the Quick Draw dataset 1 and the Face X dataset 2, a collection of high-quality sketches of cartoon facial expressions. 1https://quickdraw.withgoogle.com/ 2https://facex.idvxlab.com |
| Dataset Splits | No | The paper states the datasets used for training and evaluation but does not explicitly specify the training, validation, and test split percentages or sample counts for reproduction. |
| Hardware Specification | Yes | AI-Sketcher was trained based on a Nvidia Tesla K80 graphic card. It takes approximately 0.013 seconds on average on an i Mac machine (3.3 GHz Intel Core i5, 8 GB RAM) to produce each stroke. |
| Software Dependencies | No | The paper mentions software components such as LSTM, Adam optimizer, ReLU, and tanh, but it does not provide specific version numbers for these software dependencies or libraries. |
| Experiment Setup | Yes | LSTM (Hochreiter and Schmidhuber 1997) with layer normalization (Ba, Kiros, and Hinton 2016) is used as both the encoder and decoder, which respectively consist of 512 and 2048 hidden nodes. The amount of GMM, which is denoted as m, is equal to 20. Our model is trained by the Adam optimizer (Kingma and Ba 2014). The learning rate of the optimizer is 0.001 and the gradient clipping is 1.0, which is used to avoid the exploding gradient problem. The batch size of the input data for each training step is set as 100. In our implementation, we set nz = 256, na = 512, and β = 0.1. In our implementation, we set ϵ = 0.20 and the upper bound of α as 1.00. |