QC-StyleGAN - Quality Controllable Image Generation and Manipulation

Authors: Dat Viet Thanh Nguyen, Phong Tran The, Tan M. Dinh, Cuong Pham, Anh Tran

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments. We conduct experiments on the common datasets used by Style GAN, including FFHQ, AFHQ-Cat, and LSUN-Church. We compare the quality of our QC-Style GAN models with their Style GAN2-Ada counterparts in Table 1, using the FID metric.
Researcher Affiliation Collaboration 1Vin AI Research 2MBZUAI 3Posts & Telecommunications Institute of Technology
Pseudocode No No explicit pseudocode or algorithm blocks were found.
Open Source Code Yes We generated some quantitative videos and provided them in this link. For each video, we show the same manipulation in 6 runs.
Open Datasets Yes Datasets. We conduct experiments on the common datasets used by Style GAN, including FFHQ, AFHQ-Cat, and LSUN-Church. FFHQ [1] is a large dataset of 70k high-quality facial images collected from Flickr, introduced since the first Style GAN paper. ... AFHQ [50] is a HQ dataset for animal faces... LSUN-Church is a subset of the LSUN [51] collection.
Dataset Splits No The paper uses "test set" (e.g., Celeb A-HQ test set, AFHQ-Cat test set) but does not provide specific details about training, validation, or test dataset splits (e.g., percentages, counts, or explicit mention of validation set).
Hardware Specification No The paper states in the checklist that total amount of compute and type of resources are included, but the main body of the paper does not specify particular hardware details such as GPU/CPU models or specific compute resources used for the experiments.
Software Dependencies No The paper mentions 'Style GAN2-Ada' as a reference implementation but does not provide specific version numbers for any software dependencies used in the experiments.
Experiment Setup Yes Synthesis network. We use Style GAN2-Ada as reference to implement our QC-Style GAN. The quality code has size Dq = 16. In Degrad Block, we use L = 32 and P = 3. The weight for the distillation loss λKD = 3. Our networks were trained using the same settings as in the original work until converged.