Unverified70% confidenceCautionExact time
FID is calculated by comparing the differences in means and covariance matrices between real images and generated images in the feature space of the Inception-v3 network, but it is insensitive to local semantic quality.
2
Sources
70%
Confidence
Long-term
Relevance
9/5/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedFID measures the distributional distance between generated images and real images in feature space; a lower value indicates higher generation quality.75% similarUnverifiedFID 由 Heusel 等人于2017年提出，通过比较真实图像与生成图像在 Inception 网络特征空间中的分布距离来衡量生成质量，数值越低越接近真实图像分布72% similarUnverifiedOn the dSprites, 3D Shapes, and MPI3D datasets, the researchers find a dissociation between forward decodability and latent-aligned variance: probe accuracy is high, but the latent factors do not dominate the representation geometry.69% similarUnverifiedPerceptual Loss uses a pre-trained VGG network to extract high-level feature representations of images, computing the gap between generated images and target images in feature space rather than pixel space.67% similarUnverifiedThe FID metric is insensitive to differences in high-frequency details (such as sharp edges and fine textures), so generated images may appear visually blurry or "overly smooth" while the FID score remains good.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/860742API
curl https://kongchang.com/api/v1/knowledge/claims/860742MCP
get_claim(id=860742)