PasteGAN: A Semi-Parametric Method to Generate Image from Scene Graph

Authors: Yikang LI, Tao Ma, Yeqi Bai, Nan Duan, Sining Wei, Xiaogang Wang

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Evaluated on Visual Genome and COCO-Stuff dataset, our proposed method significantly outperforms the SOTA methods on Inception Score, Diversity Score and Fréchet Inception Distance. Extensive experiments also demonstrate our method s ability to generate complex and diverse images with given objects.
Researcher Affiliation Collaboration Yikang Li1 , Tao Ma2 , Yeqi Bai3, Nan Duan4, Sining Wei4, Xiaogang Wang1 1The Chinese University of Hong Kong, 2Northwestern Polytechnical University 3Nanyang Technological University, 4Microsoft
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks. Diagrams and mathematical formulas are present, but not pseudocode.
Open Source Code Yes The code is available at https://github.com/yikang-li/Paste GAN.
Open Datasets Yes COCO-Stuff [28] and Visual Genome [25] are two datasets used by previous scene image generation models [4, 17].
Dataset Splits Yes Table 1 displays the attributes of the datasets. (Train 74121 / Val. 1024 / Test 2048 for COCO; Train 62565 / Val. 5506 / Test 5088 for VG)
Hardware Specification Yes training takes about 3 4 days on a single Tesla Titan X.
Software Dependencies No The paper mentions using "Adam [29]" as an optimizer but does not specify version numbers for other key software components or libraries (e.g., Python, PyTorch, TensorFlow, CUDA).
Experiment Setup Yes We train all models using Adam [29] with learning rate 5e-4 and batch size of 32 for 200,000 iterations; training takes about 3 4 days on a single Tesla Titan X. The λ1 λ8 are set to 1, 10, 1, 1, 1, 1, 0.5 and 10 respectively.