GMM
K1/K4 Gaussian reference assets and offline feature-evaluation assets. K4: seed3407, absolute fitting jitter1e-5; NPZ stores raw covariance (do not add jitter twice).
Files: SigLIP_K1.npz, MAE_K1.npz, Inception_K1.npz, and corresponding _K4.npz;
Stats.tar contains frozen reference moments; Encoders.tar contains frozen representation weights.
No generator base checkpoints or training checkpoints are included.
Upstream representation weights retain their respective upstream licenses; this repository does not relicense them. Sources:
- https://huggingface.co/timm/vit_so400m_patch16_siglip_256.v2_webli
- https://huggingface.co/timm/vit_large_patch16_224.mae
- https://huggingface.co/timm/vit_large_patch14_dinov2.lvd142m
- https://huggingface.co/timm/vit_large_patch14_clip_224.openai
- https://huggingface.co/timm/convnextv2_base.fcmae_ft_in22k_in1k
- https://github.com/toshas/torch-fidelity
- Reference statistics and model implementation: https://github.com/Jiawei-Yang/FD-Loss
Generator bases are available directly upstream: https://huggingface.co/jjiaweiyang/FD-Loss/tree/main/checkpoints/base
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support