Controllable and Compositional Generation with Latent-Space Energy-Based Models

Authors: Weili Nie, Arash Vahdat, Anima Anandkumar

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we show the effectiveness of our method in conditional sampling, sequential editing and compositional generation, and we also perform an ablation study on the sampling method. Experimental setting We use Style GAN-ADA [27] as the pre-trained model for experiments on CIFAR-10 [33], and Style GAN2 [29] for experiments on FFHQ [28].
Researcher Affiliation Collaboration Weili Nie NVIDIA wnie@nvidia.com Arash Vahdat NVIDIA avahdat@nvidia.com Anima Anandkumar Caltech, NVIDIA anima@caltech.edu
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes We include the code to reproduce our results in the supplementary material.
Open Datasets Yes We use Style GAN-ADA [27] as the pre-trained model for experiments on CIFAR-10 [33], and Style GAN2 [29] for experiments on FFHQ [28].
Dataset Splits No The paper does not explicitly provide specific training/test/validation dataset splits (e.g., percentages or sample counts) needed to reproduce the data partitioning for its own models. It mentions ‘10k (w, c) pairs created by [1] for training’ but no validation/test split for this data.
Hardware Specification Yes it takes less than one second to sample a batch of 64 images on a single NVIDIA V100 GPU
Software Dependencies No The paper mentions software components like ‘dopri5 solver’ and ‘Style GAN2’ but does not specify version numbers for general software dependencies (e.g., Python, PyTorch, TensorFlow, CUDA) that would be needed for replication.
Experiment Setup Yes For LACE-ODE, we use the dopri5 solver [7] with the tolerances of (1e-3, 1e-3), and we set T = 1, βmin = 0.1 and βmax = 20. For LACE-LD, the step size and the standard deviation σ of t are chosen separately for faster training [18], where the number of steps N = 100, step size = 0.01 and standard deviation σ = 0.01. Unless otherwise specified, we use truncation = 0.7 for our method.