You're right that torchvision.datasets.ImageFolder doesn’t natively support loading images directly from S3. The 2019 limitation still stands — it expects a local file system path. However, AWS released the S3 plugin for PyTorch in 2021, which allows you to access S3 datasets as if they were local, using torch.utils.data.DataLoader. Alternatively, you can mount the S3 bucket using s3fs or fsspec, copy data to a temporary local directory, or create a custom Dataset class that streams images directly from S3 using boto3. For large datasets and training at scale, the S3 plugin is the cleanest and most efficient path.