Reproduce “Scaling Transformer Energy-based Models For High Quality Image Synthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jungsoo Lee, Alexi Gladstone, Yilun Du
TEM lets a transformer-based energy-based model generate images by scoring samples in a pretrained autoencoder’s latent space and using Langevin sampling. Multiple energy heads and positive-only representation alignment support training, while adding energies enables the demonstrated concept composition without retraining. On class-conditional ImageNet, TEM-XL reports FID-50K 3.82 at 256×256 and 8.24 at 512×512, without classifier-free guidance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.