This model is the VQGAN+ tokenizer with a vocabulary size of 12 bits. It uses a downsampling factor of 16 and is trained on ImageNet for images of resolution 256.
You can find more details on the project page and in the paper.
Dataset used to train markweber/vqgan_plus_12bit
Evaluation results
- rFID on ILSVRC/imagenet-1kself-reported1.390
- InceptionScore on ILSVRC/imagenet-1kself-reported193.900
- LPIPS on ILSVRC/imagenet-1kself-reported0.315
- PSNR on ILSVRC/imagenet-1kself-reported21.000
- SSIM on ILSVRC/imagenet-1kself-reported0.550
- CodebookUsage on ILSVRC/imagenet-1kself-reported1.000