tomg-group-umd/sphere2

Sphere Encoder 2

Paper | Code

Sphere Encoder 2 is a standalone autoencoder that generates images by decoding random points from a high-dimensional latent sphere. This repository hosts the pretrained checkpoints, the FDr⁶ reference statistics and the dataset json files used by the official code.

Models

Model	flowers-256px	flowers-512px	imagenet-256px	imagenet-512px
Sphere2-B	ckpt	ckpt	ckpt	ckpt
Sphere2-L	ckpt	ckpt	ckpt	ckpt

Models trained with the FD-lite loss (Sec. B.1 of the paper):

Model	imagenet-256px	imagenet-512px
Sphere2-B†	ckpt	ckpt
Sphere2-L†	ckpt	ckpt

Files

Folder	Content
experiments	one folder per model, with cfg.json and the checkpoint
fdr6_stats	FDr6 reference statistics for Oxford Flowers and ImageNet, at 256px and 512px
imagenet.json	train.json, val.json and folder_to_id_to_label.json for ImageNet
flowers.json	train.json and val.json for Oxford Flowers

Usage

Download a model folder into workspace/experiments of the code repository, then follow its README for sampling, evaluation and training:

hf download tomg-group-umd/sphere2 --include "experiments/sphere2-large-imagenet-512px/*" --local-dir workspace
bash scripts/sample_imagenet.sh