WeditGAN: Few-Shot Image Generation via Latent Space Relocation

Authors: Yuxuan Duan, Li Niu, Yan Hong, Liqing Zhang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on a collection of widely used source/target datasets manifest the capability of Wedit GAN in generating realistic and diverse images, which is simple yet highly effective in the research area of few-shot image generation.
Researcher Affiliation Collaboration 1Mo E Key Lab of Artificial Intelligence, Shanghai Jiao Tong University 2Tiansuan Lab, Ant Group
Pseudocode No The paper does not include any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Codes are available at https://github.com/Ldhlwh/Wedit GAN.
Open Datasets Yes Dataset Following previous works, we mainly focus on model transfer from face photos to artistic portraits, including FFHQ Sketches (Wang and Tang 2009), Babies, Sunglasses, paintings by Amedeo Modigliani, Raphael, and Otto Dix (Yaniv, Newman, and Shamir 2019). We also test Wedit GAN on LSUN Church Haunted houses, and LSUN Car Wrecked cars.
Dataset Splits No All the target datasets contain only ten training images, with resolution of 256^2. The paper specifies the size of the few-shot training data but does not provide explicit train/validation/test splits for the datasets to enable reproducibility of data partitioning.
Hardware Specification No The paper does not provide any specific hardware details (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes We conduct the experiments with Wedit GAN and its three variants. Wedit GAN perp impose perpendicular regularization, with λperp = 10^-4. Wedit GAN alpha finetunes the editing intensity after learning w, with λα reg = 1/0.1/0.01 for different cases. Wedit GAN CL appends contrastive losses on feature changes for multiple layers in both the generator and the discriminator, with λCL = 0.5.