StyleGAN-T releases training code, but keeps the checkpoints

What it does

StyleGAN-T is a text-conditional image generator from the original StyleGAN authors. This repository contains the full distributed training pipeline, data preparation tools, and evaluation scripts. It handles everything from small zip-based datasets to multi-million-image webdataset collections, and can train both unconditional and text-conditional models.

The interesting bit

The project treats the generator as conditional on CLIP text embeddings while keeping the text encoder frozen by default, letting the GAN architecture do the heavy lifting. It also supports a progressive-growing-style workflow where you freeze low-resolution blocks and train only higher-resolution layers when scaling up.

Key highlights

• Supports both zip archives and WebDataset formats for large-scale training.
• Includes built-in FID and CLIP score evaluation against COCO.
• Offers a lite config for ablations and a custom config for manual capacity tuning.
• Distributed training with gradient accumulation to hit target batch sizes across multiple GPUs.
• Can resume training from checkpoints and supports progressive resolution scaling.

Caveats

• No pretrained checkpoints are provided. You must train from scratch or supply your own.
• Metric computation only works on zip datasets, not webdatasets.
• The authors explicitly treat this as a one-time research code drop and do not accept pull requests.

Verdict

Researchers replicating the ICML 2023 paper or training custom text-to-image GANs from scratch will find the complete machinery here. If you were hoping to download a .pkl and start generating fox paintings immediately, this is not that repository.