FID
A widely used metric for evaluating the quality of image generative models. It measures generation quality by computing the distance between the distributions of generated and real images in the feature space of an Inception network, with lower scores indicating higher generation quality.
Core Facts
FID measures the distributional distance between generated images and real images in feature space; a lower value indicates higher generation quality.
FID (Fréchet Inception Distance) is a quantitative metric used to measure the distance between the distribution of generated images and real photographs.
Timeline (last 90 days)
The paper did not disclose FID scores.
FID measures the statistical distance between the generated distribution and the real distribution, with lower values being better, but it is heavily influenced by the feature extractor and sample size, and fluctuates noticeably on small datasets.
MiniDog's evaluation covers FID, HPSv2, PickScore, and various VAE reconstruction metrics.
FID is affected by the number of samples, preprocessing and the Inception version, and cross-paper comparisons should be based on the same evaluation protocol.
On the class-conditional ImageNet-1K benchmark, PixelMoE achieves FID 1.47 at 256×256 resolution and FID 1.53 at 512×512, without using representation alignment and perceptual losses.
CLIP score measures semantic consistency rather than visual quality and cannot capture perceptual issues such as image blurriness, artifacts, or color distortion; in practice it is typically used in combination with SSIM or FID.
Physics-IQ measures the correctness of causal reasoning, which differs from traditional visual quality metrics such as FID that measure the perceptual realism of images.
On the ImageNet-256 dataset, with NFE of 25 and no guidance, the time-lag-measurement-based gain selection method reduces FID from 28.0 to 12.2.
The FID value of 12.2 is estimated by linear interpolation of FID measurements from adjacent gain values.
Systematic evaluation benchmark methods in the industry include blind scoring, ELO rating systems, and automated metrics such as FVD (frame-level perceptual distance) and FID (image quality score).
16 more timeline events
All Facts (20)
FID measures the distributional distance between generated images and real images in feature space; a lower value indicates higher generation quality.
80%VerifiedFID (Fréchet Inception Distance) is a quantitative metric used to measure the distance between the distribution of generated images and real photographs.
75%UnverifiedFID 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.
70%UnverifiedFID (Fréchet Inception Distance) evaluates generation quality by comparing the statistical distribution differences between real and generated data in the feature space of a deep neural network, with lower values indicating closer resemblance to real data.
70%UnverifiedThe paper did not disclose FID scores.
50%UnverifiedFID measures the statistical distance between the generated distribution and the real distribution, with lower values being better, but it is heavily influenced by the feature extractor and sample size, and fluctuates noticeably on small datasets.
50%UnverifiedMiniDog's evaluation covers FID, HPSv2, PickScore, and various VAE reconstruction metrics.
50%UnverifiedFID is affected by the number of samples, preprocessing and the Inception version, and cross-paper comparisons should be based on the same evaluation protocol.
50%UnverifiedOn the class-conditional ImageNet-1K benchmark, PixelMoE achieves FID 1.47 at 256×256 resolution and FID 1.53 at 512×512, without using representation alignment and perceptual losses.
50%UnverifiedCLIP score measures semantic consistency rather than visual quality and cannot capture perceptual issues such as image blurriness, artifacts, or color distortion; in practice it is typically used in combination with SSIM or FID.
50%UnverifiedPhysics-IQ measures the correctness of causal reasoning, which differs from traditional visual quality metrics such as FID that measure the perceptual realism of images.
50%UnverifiedThe FID value of 12.2 is estimated by linear interpolation of FID measurements from adjacent gain values.
50%UnverifiedOn the ImageNet-256 dataset, with NFE of 25 and no guidance, the time-lag-measurement-based gain selection method reduces FID from 28.0 to 12.2.
50%UnverifiedSystematic evaluation benchmark methods in the industry include blind scoring, ELO rating systems, and automated metrics such as FVD (frame-level perceptual distance) and FID (image quality score).
50%UnverifiedCompared to ShapeLLM-Omni, OctLLM reduces the FID metric for image-to-3D generation by 17.4%.
50%UnverifiedDreamSim is a deep learning-based perceptual similarity metric trained to align with human judgments of image similarity, and it is closer to subjective visual experience than LPIPS or FID.
50%UnverifiedThe INP metric replaced the previous FID metric in 2024.
50%UnverifiedThe experimental evaluation metrics cover four dimensions: FID, number of trainable parameters, runtime, and GPU memory usage.
50%UnverifiedUnder standard DDPM settings, rank 4 achieved the best FID score of 124.1380, with rank 8 following closely at 124.2136.
50%UnverifiedHuman subjective voting can capture quality differences that are difficult to quantify, making it an important complement to automated metrics such as FID and CLIP Score.
50%