FID Benchmark Protocol

The FID benchmark protocol is the standardised methodology for evaluating generative image models using the Fréchet Inception Distance, which compares the distribution of generated images with that of real images in the feature space of a pretrained Inception network. The protocol fixes the feature extractor, the number of samples, and preprocessing so that scores are comparable across models; lower FID indicates greater similarity to real data. It is the de facto standard for benchmarking GANs and diffusion models.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:hasPart ai:InceptionV3))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:hasPart ai:FrechetDistance))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:hasPart ai:FeatureExtraction))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:hasPart ai:CleanFID))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:hasPart ai:PrecisionRecall))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:hasPart ai:GaussianDistributionEstimator))

Dependency Relationships

SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:requires ai:InceptionV3))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:requires ai:ImageNet))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:requires ai:ProbabilityTheory))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:requires ai:ConvolutionalNeuralNetworks))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:dependsOn ai:DeepLearning))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:uses ai:RepresentationLearning))

Capability Relationships

SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:enables ai:GenerativeModel))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:enables ai:ImageSynthesis))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:enables ai:SyntheticDataGeneration))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:enables ai:MedicalImageSynthesis))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:supports ai:DiffusionModels))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:supports ai:GenerativeAdversarialNetworks))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:supports ai:TextToImage))

Implementation Relationships

SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:implements ai:FrechetDistance))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:implements ai:WassersteinDistance))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:uses ai:InceptionV3))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:uses ai:CLIPEncoder))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:standardizedBy ai:CVPR))

Reduction Relationships

SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:reducesTo ai:DistributionalSimilarityMetric))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:reducesTo ai:ImageQualityAssessment))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:contrastsWith ai:InceptionScore))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:contrastsWith ai:CMMD))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:contrastsWith ai:KernelInceptionDistance))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:partOf ai:EvaluationMetric))
SubClassOf(ai:DiffusionModels
  ObjectSomeValuesFrom(ai:standardizedBy ai:FIDBenchmarkProtocol))
SubClassOf(ai:GenerativeAdversarialNetworks
  ObjectSomeValuesFrom(ai:standardizedBy ai:FIDBenchmarkProtocol))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:relatedTo ai:FrechetAudioDistance))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:relatedTo ai:FrechetVideoDistance))
SubClassOf(ai:CleanFID
  ObjectSomeValuesFrom(ai:hasPart ai:FIDBenchmarkProtocol))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:measures ai:ImageSynthesis))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:standardizedBy ai:CVPR))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:standardizedBy ai:NeurIPS))
SubClassOf(ai:FIDBenchmarkProtocol
  ObjectSomeValuesFrom(ai:uses ai:RepresentationLearning))

About

• The FID Benchmark Protocol operationalises a principled distributional comparison between generated and real image populations, providing the generative modelling community with a shared, quantitative language for comparing model quality across years, architectures, and research groups. Where earlier automatic evaluation metrics such as the Inception Score (IS, Salimans et al., 2016) measured only the marginal distribution of generated images — rewarding sharpness (low conditional entropy H(y|x)) and diversity (high marginal entropy H(y)) without directly comparing to real data — FID closes this gap by measuring the statistical distance between real and generated feature distributions jointly. This shift from marginal to joint distributional evaluation was motivated by the observation that IS fails to penalise mode-specific artefacts and can be trivially improved by memorising training data or generating a small number of highly classifiable images with no genuine diversity. FID instead uses a pretrained Convolutional Neural Networks feature extractor (Inception v3, trained on ImageNet) to map images into a 2048-dimensional semantically meaningful embedding space, fits multivariate Gaussians to the embeddings of 50,000 real and 50,000 generated images respectively, and computes the Fréchet Distance between those Gaussians. The Fréchet Distance is equivalent to the 2-Wasserstein Distance between Gaussian distributions — the minimal cost of transporting one Gaussian distribution onto another — and has the appealing property that it penalises both mean divergence (quality gap: generated images look systematically different from real ones in mean appearance) and covariance divergence (diversity gap: generated images have lower variance, i.e., mode collapse). A model that generates a single photorealistic image repeated 50,000 times will have near-zero mean divergence but very high covariance divergence (its covariance matrix is rank 1 vs. the full-rank real covariance), yielding a high FID — correctly penalising mode collapse that the IS would miss.
• The protocol’s practical dominance across nearly a decade of generative modelling research stems from several reinforcing factors. First, FID shows strong empirical correlation with human perceptual quality judgements across a wide range of generative model families (GANs, VAEs, flow models, diffusion models), validated through large-scale human evaluation studies by both academic groups and major industry labs. Second, FID is sensitive to the common failure modes of generative models — blurriness (which shifts the mean embedding toward less-activated feature dimensions), mode collapse (which reduces covariance), and hallucination artefacts (which shift the mean into out-of-distribution regions) — providing diagnostic signal about what has gone wrong, not just how wrong. Third, FID is computationally cheap to evaluate once reference statistics are precomputed: extracting 50,000 Inception v3 embeddings and computing a matrix square root takes minutes on a single GPU, far cheaper than human evaluation. Fourth, FID’s single scalar score enables clear leaderboard rankings, satisfying the academic community’s need for a total ordering of model quality for publication and reproducibility.
• State-of-the-art Generative Adversarial Networks such as BigGAN and StyleGAN3 reported CIFAR-10 FID scores below 3; DDPM-class Diffusion Models reached below 2 on CIFAR-10 by 2021 (Nichol & Dhariwal) and Diffusion Models with classifier guidance reached FID 4.59 on 256×256 ImageNet by 2021 (Dhariwal & Nichol), surpassing BigGAN’s 7.72 — a watershed result that validated FID as the instrument of record for the field’s most significant architectural transition. Latent Diffusion Models (Stable Diffusion, Rombach et al., CVPR 2022) operate in a compressed 64×64 latent space rather than pixel space, with the Variational Autoencoder decoder upsampling to the final resolution; the FID of the full pipeline (VAE encoder → latent diffusion → VAE decoder) is reported on standard benchmarks, enabling fair comparison with pixel-space models. DiT (Peebles & Xie, ICCV 2023) achieved FID 2.27 on ImageNet 256×256, demonstrating that transformer architectures are competitive with U-Net-based diffusion models and the FID metric was able to capture this transition.
• However, the protocol’s limitations are increasingly well-documented and affect its reliability in modern contexts. The statistical bias (proportional to d/N = 2048/N) does not vanish even at N = 50,000, contributing ~0.08 FID points of unavoidable systematic error. More seriously, the protocol inherits Inception v3’s ImageNet biases: the feature extractor was trained to discriminate 1,000 ImageNet categories (mostly animals, household objects, and vehicles) and has no semantic knowledge of faces, artistic styles, medical textures, satellite imagery, or the diverse range of content generated by modern Text-to-Image models. This means that two models with identical visual quality but one generating ImageNet-like content and one generating medieval artwork will have very different FID scores purely because of distributional overlap with Inception’s training domain. The Gaussian assumption is also violated in practice: the 2048-dimensional Inception v3 feature distributions of natural image collections are not Gaussian — they are multimodal, skewed, and kurtotic — meaning the Fréchet Distance computes the distance between the best-fitting Gaussians to the true distributions, which may diverge significantly from the true Wasserstein-2 distance.
• The community response to these limitations has been multi-pronged. Parmar et al. (CVPR 2022) released Clean-FID, exposing that standard FID implementations using nearest-neighbour image resizing rather than antialiased bicubic interpolation introduce systematic, protocol-dependent biases of up to 10 FID points, invalidating cross-paper comparisons that used different implementations — a finding that highlighted the importance of software-level reproducibility alongside statistical theory. Jayasumana et al. (CVPR 2024) proposed CMMD, replacing Inception v3 embeddings with CLIP Encoder embeddings (trained on 400M image-text pairs) and replacing the Gaussian fit with the unbiased Maximum Mean Discrepancy (MMD) kernel, yielding a metric that is statistically unbiased, requires far fewer samples (as few as 1,000), and better reflects human judgements on modern Text-to-Image outputs. Despite these advances, FID computed under the Clean-FID protocol remains the community’s primary reporting standard as of 2026, because the entire leaderboard history of the field since 2017 is expressed in FID units and switching metrics would sever longitudinal comparability.

Components / Architecture

• Feature Extractor (Inception v3): The core component. Images are resized to 299×299 using antialiased bicubic interpolation (critical for reproducibility, as established by Clean-FID) and passed through the Inception-v3 network (Szegedy et al., 2016) pretrained on ImageNet ILSVRC 2012 using 1.28M training images and 1,000 object categories. The 2048-dimensional activations of the last global average pooling layer — pool_3:0 in TensorFlow notation, or the Mixed_7c layer’s pooled output — are extracted as feature vectors. This layer encodes high-level semantic content including object categories, texture statistics, spatial layout, and compositional structure, making it a proxy for human perceptual quality. Crucially, the Inception-v3 weights are frozen (not fine-tuned during evaluation); the feature extractor is a fixed prior over the semantic content of natural images. The choice of Inception over other available feature extractors (VGG, ResNet) was pragmatic in 2017 (Inception was state-of-the-art) and has remained fixed to ensure longitudinal comparability despite the community’s awareness that modern vision encoders offer richer representations.
• Gaussian Distribution Estimator: The feature vectors for N_real real images and N_gen generated images are each used to estimate a multivariate Gaussian N(μ, Σ) via maximum likelihood: μ̂ = (1/N) Σ_i φ(x_i) and Σ̂ = (1/N) Σ_i (φ(x_i) − μ̂)(φ(x_i) − μ̂)^T. The resulting covariance matrix is 2048×2048, requiring N ≥ 2048 samples to be full-rank; in practice N = 50,000 is standard to reduce bias. Regularisation (adding ε_reg * I with ε_reg ≈ 1e-6) is applied before computing the matrix square root to avoid numerical failures with near-singular matrices that arise when the generative model has low rank or near-degenerate diversity.
• Fréchet Distance Computation (Fréchet Distance): FID = ||μ_r − μ_g||² + Tr(Σ_r + Σ_g − 2(Σ_r Σ_g)^{1/2}). The matrix square root (Σ_r Σ_g)^{1/2} is computed via the Cholesky–Schur algorithm (SciPy’s scipy.linalg.sqrtm). This step is computationally expensive for 2048×2048 matrices (O(d³) in theory; ~1 second on CPU via optimised BLAS). Numerical precision is non-trivial: imaginary components of the complex-arithmetic sqrtm output (arising from floating-point noise) are discarded, and the Frobenius norm of the imaginary part is checked against a threshold to confirm validity.
• Reference Statistics Cache: The reference statistics (μ_r, Σ_r) for standard benchmark datasets (CIFAR-10 train, ImageNet at 64×64, 128×128, 256×256, 512×512, FFHQ-256, FFHQ-1024, MS-COCO 2017 validation, LSUN Bedroom) are precomputed once over the full dataset and cached as .npz files. Clean-FID provides canonical reference statistic files generated under the standardised preprocessing pipeline. Using pre-cached reference statistics eliminates the need to re-extract real-data features on every evaluation run and ensures all comparisons use exactly the same reference distribution.
• Clean-FID Correction Layer: The Clean-FID library (Parmar et al., github.com/GaParmar/clean-fid) wraps the above pipeline, enforcing: (i) antialiased bicubic resizing via PIL/Pillow (not OpenCV or naive torch interpolation), (ii) uint8 pixel quantisation to [0, 255] after resizing (matching the original Heusel et al. protocol), (iii) reproducible reference statistics pre-computed under the same pipeline. When comparing against published benchmarks, users must use the same reference statistics file that the baseline paper used; cross-implementation comparisons without this control can differ by up to 10 FID points.
• Complementary Metrics Suite: Contemporary benchmarking practice couples FID with: (i) Precision-Recall decomposition (Kynkäänniemi et al., NeurIPS 2019) to separately measure fidelity (precision = fraction of generated samples lying within the k-NN estimated real manifold) and diversity (recall = fraction of the real manifold covered by at least one generated sample within the k-NN ball); (ii) CLIP score for text-image alignment in conditional generation (cosine similarity between generated image and its text prompt in CLIP embedding space); (iii) CMMD with CLIP Encoder features for text-to-image models where Inception features are domain-mismatched; (iv) human evaluation win-rates for fine-grained quality discrimination at the frontier of model capability.

Use Cases / Major Families

• GAN Benchmarking (Generative Adversarial Networks): FID was introduced to evaluate TTUR-trained GANs and has since been the primary metric for comparing the entire progression of GAN architectures: Progressive GAN (Karras et al., 2018), BigGAN (Brock et al., ICLR 2019), StyleGAN (2019), StyleGAN2 (2020), StyleGAN3 (2021, alias-free), LightweightGAN (Liu et al., 2021), GigaGAN (Kang et al., 2023). The landmark paper by Dhariwal & Nichol (2021) used FID to show that class-conditional Diffusion Models with classifier guidance surpassed BigGAN on ImageNet 256×256 (FID 4.59 vs 7.72), establishing diffusion as the new SOTA for high-resolution image synthesis and triggering a major community shift away from Generative Adversarial Networks. This comparison was credible precisely because both models were evaluated under the same FID protocol; without a standardised metric, such comparisons across generative families would not have been possible.
• Diffusion Model Evaluation (Diffusion Models): DDPM, DDIM, ADM (with and without classifier guidance), Latent Diffusion Models (LDM), DiT (Peebles & Xie, ICCV 2023), Consistency Models (Song et al., ICML 2023), EDM (Karras et al., NeurIPS 2022), and their successors all report CIFAR-10 and ImageNet FID as primary quantitative results. Score-based models (Song et al., ICLR 2021 SDE formulation) and flow-matching methods (Rectified Flow, Liu et al., 2022; SD3’s MMDiT, Esser et al., ICML 2024) use FID to demonstrate that alternative training objectives and faster sampling schedules do not sacrifice image quality. The FID protocol is the standard by which claims of “new state of the art” in unconditional and class-conditional image generation are evaluated.
• Text-to-Image Generation (Text-to-Image): DALL-E 2 (Ramesh et al., 2022), Imagen (Saharia et al., NeurIPS 2022), Stable Diffusion (Rombach et al., CVPR 2022), SDXL (Podell et al., ICLR 2024), FLUX, and SD3 all report zero-shot MS-COCO FID as a primary benchmark: 30,000 generated samples from COCO 2017 validation prompts are evaluated against the 40,509 COCO validation images as reference. Zero-shot FID on COCO tests generalisation to a prompt distribution not seen during training. The unreliability of Inception v3 features for creative out-of-distribution content — cartoon styles, artistic compositions, abstract concepts — has driven supplementary reporting of CLIP Encoder cosine similarity scores, VQA-accuracy (answering questions from the prompt), and human preference win-rates via crowd-source evaluation platforms.
• Medical Image Synthesis (Medical Image Synthesis): FID applied to MRI, CT, PET, and histopathology generative models faces a fundamental domain mismatch: Inception v3 was trained on ImageNet natural photographs and has never encountered FLAIR brain MRI or H&E histology slides. Inception-v3 features in these domains capture low-level texture statistics rather than clinically relevant structure, making cross-model FID comparisons unreliable. Domain-specific feature extractors are increasingly used: RadImageNet (Mei et al., 2022) provides radiological-domain ImageNet-style pretraining; MedCLIP embeddings encode medical image-text pairs. The 2025 pragmatic note on retinal image FID (arXiv:2502.17160) systematically quantified the mismatch and recommended RadImageNet-FID as the primary metric with Inception-FID as a secondary cross-domain comparator. NHS AI evaluation frameworks for generative medical models now incorporate these domain-specific FID variants.
• Video and Audio Extensions: Fréchet Video Distance (FVD, Unterthiner et al., 2019 ICLR Workshop) applies the FID framework using I3D (Inflated 3D Convolution) video features to evaluate video generation models including VGAN, MoCoGAN, VideoGPT, and VideoLDM. It captures both per-frame quality and temporal coherence. Fréchet Audio Distance (FAD, Kilgour et al., 2019) uses a VGGish audio feature extractor (trained on AudioSet) in place of Inception v3 for music and speech synthesis evaluation, benchmarking models like AudioLDM, MusicGen, and Stable Audio. Fréchet ChemNet Distance (FCD, Preuer et al., 2018) uses a ChemNet encoder for evaluating molecular generative models such as MolGAN and Reinvent, measuring chemical property distributions rather than visual feature distributions.
• Synthetic Data Quality Assessment (Synthetic Data Generation): FID serves as a proxy for the downstream utility of synthetic augmentation data in training downstream classifiers. Lower FID synthetic sets produce higher classification accuracy gains when used for data augmentation, establishing a practical link between the evaluation protocol and data augmentation utility. This has been validated for CIFAR-10 augmentation (Zhao et al., 2021), medical image augmentation (Frid-Adar et al., 2018), and satellite imagery augmentation (Rostami et al., 2022). The Synthetic Data Generation community increasingly uses FID as a first-pass filter for synthetic dataset quality before running full downstream evaluation pipelines.
• Adversarial Robustness and Out-of-Distribution Detection: FID-like distributional comparison has been applied to measure how far adversarial perturbations shift images from the real training distribution, and to evaluate whether generative models can produce reliable out-of-distribution samples for OOD detection training. Diffusion Models trained on in-distribution data produce synthetically perturbed samples with controllable FID relative to the in-distribution reference, enabling systematic OOD detection benchmarking.

Formal Detail

• Core FID Formula. Let x ~ p_data and x̂ ~ p_gen. Define φ: R^{H×W×3} → R^{2048} as the Inception-v3 global average pool embedding. The FID protocol estimates: μ_r = E_{xp_data}[φ(x)], Σ_r = Cov_{xp_data}[φ(x)] from N_real real samples, and μ_g = E_{x̂p_gen}[φ(x̂)], Σ_g = Cov_{x̂p_gen}[φ(x̂)] from N_gen generated samples. The FID is then: FID(p_data, p_gen) = W_2²(N(μ_r, Σ_r), N(μ_g, Σ_g)) = ||μ_r − μ_g||² + Tr(Σ_r + Σ_g − 2(Σ_r Σ_g)^{1/2}), where W_2 is the 2-Wasserstein Distance between Gaussian distributions and (Σ_r Σ_g)^{1/2} is the matrix geometric mean computed via Schur decomposition or the Cholesky–Higham algorithm.
• Bias and Sample Complexity. The empirical estimator FID_hat has E[FID_hat] = FID + bias(N_real, N_gen). The dominant bias term arises from finite-sample estimation of the 2048×2048 covariance matrix: bias ≈ (d/N_real + d/N_gen) where d = 2048. At N = 50,000 (standard) this bias is ≈ 0.08 FID points; at N = 10,000 it rises to ≈ 0.4 points; at N = 5,000 it reaches ≈ 0.8 points. The Chong et al. (2020, CVPR) unbiased estimator corrects this via the analytical de-biasing term, enabling reliable comparison with as few as 5,000 samples.
• Preprocessing Sensitivity (Clean-FID). The Clean-FID paper (Parmar et al., CVPR 2022) demonstrated that the commonly used PyTorch FID implementation employs nearest-neighbour resizing (fast but aliased), whereas the original Heusel et al. implementation used PIL bicubic resizing (antialiased). The difference in preprocessing pipeline alone shifts FID scores by up to 10 points for some models, invalidating comparisons between papers using different implementations. The Clean-FID library standardises preprocessing to: (i) antialiased bicubic downsampling via PIL/Pillow, (ii) uint8 quantisation after resizing to [0, 255], (iii) standard float normalisation. Papers using Clean-FID are marked as such in leaderboard entries.
• Matrix Square Root Numerical Precision. The computation of (Σ_r Σ_g)^{1/2} via SciPy’s scipy.linalg.sqrtm is numerically sensitive for near-singular covariance matrices, especially when N is close to d = 2048. Standard practice adds a small regularisation ε_reg * I to both Σ_r and Σ_g (ε_reg ≈ 1e-6) before computing the square root. The imaginary components of the complex-arithmetic sqrtm output (which arise from numerical noise) are discarded.
• Kernel Inception Distance (KID). The KID (Binkowski et al., ICLR 2018) replaces the Gaussian assumption with the unbiased MMD² statistic using a polynomial kernel k(x, y) = (x^T y / d + 1)³ on Inception features. KID is unbiased for any N and is asymptotically normal, enabling confidence intervals and hypothesis tests on model comparisons. Unlike FID, KID is defined for any feature extractor without assuming a specific distributional form.
• CMMD Formula. CMMD (Jayasumana et al., CVPR 2024) defines: CMMD(p_data, p_gen) = MMD²(φ_CLIP(x), φ_CLIP(x̂)) = E_{x,x’}[k(φ(x), φ(x’))] − 2 E_{x, x̂}[k(φ(x), φ(x̂))] + E_{x̂, x̂’}[k(φ(x̂), φ(x̂’))], using a Gaussian kernel k(u, v) = exp(−||u − v||² / (2σ²)) on CLIP visual embeddings. The estimator is computed in O(N²) time but is unbiased at any N ≥ 100 and requires far fewer samples than FID for statistically reliable comparison.
• Fréchet Distance in High Dimensions. The Fréchet distance between two Gaussians is equivalent to the optimal transport (OT) distance W_2 under the Gaussian marginal constraint. Its sensitivity to covariance structure means it penalises diversity loss (Σ_g small) more strongly than mean shift (||μ_r − μ_g|| large), making it a better probe of mode coverage than the Inception Score, which measures only marginal sharpness.

Academic Context

• FID was introduced by Heusel, Ramsauer, Unterthiner, Nessler, and Hochreiter at Johannes Kepler University Linz in the paper “GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium” (arXiv:1706.08500, NeurIPS 2017). The paper’s primary contribution was the TTUR training algorithm for GANs, and FID was proposed as a supplementary evaluation metric to replace the Inception Score (IS, Salimans et al., 2016). The original IS had a significant flaw: it measured the sharpness and diversity of generated images without comparing to real images at all, meaning it could be trivially improved by memorising training data or by generating a few very distinct sharp images. Heusel et al. argued that comparing the distributional statistics of generated and real images in a semantically meaningful feature space was more faithful to the perceptual quality that human evaluators assess. The Fréchet Distance between Gaussian distributions — previously used in statistics and econometrics to compare time-series distributions — was imported to the image synthesis context as a computationally tractable measure of W_2 distance in high-dimensional feature space.
• The key theoretical limitation — the Gaussian assumption — was noted almost immediately. Binkowski et al. (2018, ICLR) proposed the Kernel Inception Distance (KID) as an unbiased alternative using the Maximum Mean Discrepancy with a polynomial kernel on Inception v3 features. KID has no Gaussian assumption and is statistically unbiased at any sample count, enabling confidence intervals and hypothesis tests. The trade-off is that KID is O(N²) to compute naively and has no analytical solution like FID’s matrix square root; in practice, mini-batch MMD estimators make it tractable. Chong and Forsyth (2020, CVPR) derived an analytical de-biasing correction for FID that allows reliable comparison with as few as 5,000 samples, partially addressing the sample-count sensitivity issue without changing the feature extractor or the Gaussian assumption.
• The Clean-FID paper (Parmar, Zhang, and Zhu, CVPR 2022, arXiv:2104.11222) made the single most impactful practical contribution to FID methodology by demonstrating that standard implementations using PyTorch’s default resize function (nearest-neighbour interpolation) introduced systematic biases of 2–10 FID points compared to the PIL bicubic protocol used in the original Heusel et al. implementation. This meant that papers could rank two models in opposite order depending purely on which FID implementation they used, not which model was genuinely better. The Clean-FID library standardised preprocessing and released pre-computed reference statistics for major benchmarks, enabling reproducible cross-paper comparison. Kynkäänniemi et al. (NeurIPS 2019) introduced the Precision-Recall decomposition to complement FID’s aggregate quality-diversity tradeoff: precision measures the fraction of generated images that fall within the real data manifold (fidelity), and recall measures the fraction of the real manifold covered by generated samples (diversity). This decomposition revealed that many GAN improvements in FID were primarily precision improvements (sharper outputs) with recall (diversity) stagnating or declining — an important insight for understanding the mode coverage behaviour of different generative architectures.
• The CMMD paper (Jayasumana et al., CVPR 2024) provided the most comprehensive published critique of FID’s limitations and proposed CMMD as a CLIP-based replacement. The paper showed that FID produces incorrect orderings of model quality in 30–40% of head-to-head comparisons on modern text-to-image benchmarks when compared to human preference judgements, with the misalignment attributable primarily to Inception-v3’s ImageNet bias and the Gaussian assumption’s inadequacy for high-dimensional, semantically complex image distributions.
• The metric’s historical trajectory mirrors Diffusion Models’ rise to dominance in generative image modelling: FID on CIFAR-10 improved from approximately 28 for DCGAN in 2016 to 2.92 for DDPM in 2020, to 1.97 for Improved DDPM in 2021, to below 1.5 for score SDE and flow-matching models in 2023–24. Each of these milestones was first established and verified through the FID Benchmark Protocol, cementing its role as the community’s primary measuring instrument for generative model progress over a decade of research.

Benchmarks and Score Milestones

• CIFAR-10 (32×32 unconditional): Benchmark for unconditional Generative Adversarial Networks and Diffusion Models on the 10-class CIFAR-10 dataset. Historical milestones: DCGAN (2016) ~28 FID; WGAN-GP (2017) ~36 FID (on a harder training protocol); SA-GAN (2019) ~14 FID; BigGAN (2019) ~14.7 (note: conditional); DDPM (Ho et al., 2020) ~3.17; Improved DDPM (Nichol & Dhariwal, 2021) ~2.90; Score SDE (Song et al., 2021) ~2.20; EDM (Karras et al., 2022) ~1.97; Flow Matching models (2023) ~1.5 range. The dramatic improvement from ~28 to below 2 over seven years reflects both architectural progress and diffusion-model dominance.
• ImageNet 256×256 (class-conditional): BigGAN-deep (2019) ~6.02; ADM (Dhariwal & Nichol, 2021, no guidance) ~10.94; ADM+classifier guidance ~4.59 — this was the milestone showing diffusion beats BigGAN; LDM (Latent Diffusion, 2022) ~3.60; DiT-XL/2 (2023) ~2.27; CD (Consistency Distillation, Song et al., 2023) ~2.65 in 2 steps. The Inception v3 features on ImageNet 256×256 are well-matched to this benchmark since the extractor was trained on the same distribution.
• MS-COCO Zero-Shot FID (text-to-image): The text-to-image sub-community conventionally reports zero-shot FID on MS-COCO (30K generated samples, 40K COCO validation images as reference). DALL-E 2 (2022) reported ~10.39; Imagen (2022) ~7.27; LDM-KL8 (2022) ~12.61; SD 1.4 ~15.7; SDXL (2023) ~6.63; SD3 (2024) ~4.7 (approximate). The COCO benchmark is increasingly questioned because the reference distribution is small (40K images vs 50K typical) and Inception v3 features are biased toward the categories in COCO’s training distribution.
• Domain-Specific FID Scores: For medical image synthesis (chest X-ray, histopathology, retinal fundus), FID scores computed with Inception v3-ImageNet features are systematically lower for models that generate visually natural but medically spurious images, since the feature extractor has no medical domain knowledge. RadImageNet-FID provides a more reliable proxy for medical generative model quality and is used in clinical validation pipelines.
• FFHQ 256×256 and 1024×1024 (face generation): StyleGAN2 (Karras et al., 2020) ~3.8 FID at 256×256; ~2.84 at 1024×1024; StyleGAN3-R ~2.79 at 1024×1024. Face generation benchmarks on FFHQ (70K high-quality face images) test fine-grained perceptual quality where the human visual system is particularly sensitive. Inception-FID on faces captures coarse distributional divergence (mode collapse, blur) but misses fine-grained identity and expression naturalness that human evaluators assess; this gap motivates the LPIPS (Learned Perceptual Image Patch Similarity) and FaceNet embedding-based evaluations as supplementary metrics.
• Protocol Sensitivity Ablation (Clean-FID Paper): Parmar et al. (CVPR 2022) systematically quantified how protocol choices affect FID independently of model quality: (i) resizing method alone (nearest-neighbour vs. bicubic) changes FID by 2–7 points for typical GAN outputs; (ii) PNG vs. JPEG encoding changes FID by 0.5–2 points; (iii) whether real-image reference statistics are computed from saved files (with potential JPEG re-encoding) vs. in-memory arrays changes FID by 1–2 points. These ablations revealed that the apparent “improvements” in some published GAN papers were partially attributable to evaluating against more favourably computed reference statistics rather than genuine architectural advances. The Clean-FID library closed this reproducibility gap.
• Scaling Behaviour of FID with Sample Count: Empirical studies across CIFAR-10, ImageNet, and Stable Diffusion variants (Chong & Forsyth, CVPR 2020) show that FID at N = 5,000 overestimates the true FID by ~0.8 points on average, while at N = 2,000 the overestimate reaches ~2.0 points. The bias correction term Δ_bias ≈ (d/N_real + d/N_gen) where d = 2048 provides an analytical correction accurate to within 0.1 points for N ≥ 5,000. For memory-constrained evaluation (e.g., evaluating very large models where generating 50K images is expensive), the corrected estimator at N = 10,000 is statistically equivalent to uncorrected N = 50,000 for most practical comparisons.

Current Landscape (2026)

• By mid-2026, FID remains the mandatory reporting metric in virtually all image generation papers submitted to CVPR, NeurIPS, ICLR, and ECCV, despite well-documented limitations. The Clean-FID library (GaParmar/clean-fid on GitHub) has been adopted as the reference implementation by most major research groups, reducing implementation-variance artefacts. Leaderboards on Papers With Code for CIFAR-10, ImageNet 256×256, and MS-COCO continue to use FID as the primary ranking metric.
• The principal controversy in 2025–2026 centres on Text-to-Image evaluation. As models like FLUX, SD3, and DALL-E 3 generate highly diverse creative content, the Inception v3 feature space (biased toward ImageNet categories) provides poor distributional coverage. The community has partially migrated to reporting FID + CLIP-FID + VQA-score + human win-rate alongside each other. Google’s CMMD proposal from CVPR 2024 has been widely cited but has not replaced FID as the primary leaderboard metric. The EvalGen initiative (2025) aims to unify multi-metric reporting into a single standard evaluation suite to reduce cherry-picking.
• In Medical Image Synthesis, the 2025 pragmatic note on retinal FID (arXiv:2502.17160) formalised the recommendation that domain-specific feature extractors (RadImageNet, MedCLIP) should replace Inception-v3 when comparing medical generative models, with Inception-FID reported only as a secondary cross-domain comparator. Several UK National Health Service AI centres have adopted this guidance for their generative model validation pipelines.

UK Context

• UK computer vision research groups have contributed both theoretical and tooling advances around the FID Benchmark Protocol, with particularly strong representation in the medical imaging and face synthesis communities. The British Machine Vision Association (BMVA) and its annual British Machine Vision Conference (BMVC) host FID-centric generative model comparisons and have featured work on novel evaluation framework proposals for the UK research community, providing a national forum for harmonising evaluation methodology before results are disseminated to international venues.
• Queen Mary University of London’s Computer Vision group — one of the largest and most internationally prominent in the UK — has published on face synthesis, video face generation, and talking head evaluation using distributional metrics including FID. Their work on face reenactment and deepfake detection (led by Professor Ioannis Patras) requires reliable quantitative evaluation of generated face quality, where FID provides the distributional benchmark against which novel architectures are compared. QMUL’s digital humanities group has also applied FID-style evaluation to generative models of historical manuscript illumination and visual cultural heritage.
• Imperial College London’s machine learning and computer vision groups (Department of Computing, under Professors Murray Shanahan, Yarin Gal, and Mark van der Wilk) have produced work on conditional image generation, uncertainty quantification in Deep Learning, and Bayesian deep generative models, for which FID serves as an external quality benchmark. Imperial’s Centre for Translational Cardiovascular Imaging has applied Diffusion Models and Variational Autoencoder architectures to cardiac MRI synthesis, where domain-specific FID evaluation (using cardiac-trained feature extractors) is critical for validating synthetic data for NHS diagnostic pathway testing.
• Cambridge University’s Speech, Vision and Robotics Group (Professor Roberto Cipolla’s group) has longstanding work on visual synthesis and scene understanding. Cambridge’s contributions to the theoretical foundations of Diffusion Models (including connections to Stochastic Differential Equation formulations and score-matching) indirectly underpin the FID benchmark ecosystem by producing the model architectures most commonly evaluated with FID. The Cambridge-based Alan Turing Institute satellite has hosted workshops on generative model evaluation standards specifically aimed at UK regulatory alignment.
• Edinburgh University’s Institute for Adaptive and Neural Computation (IANC, home to Professors Amos Storkey and Michael Gutmann) has worked on probabilistic generative models — including Normalising Flows and energy-based models — for which FID provides a key external validation signal. Storkey’s group contributed to the Random Network Distillation paper (Burda, Edwards, Storkey, Klimov, 2018), which though primarily an exploration paper in RL, has shaped the community’s understanding of distributional comparison between trained neural networks. Gutmann’s group has published on contrastive estimation for generative models, contributing to the theoretical understanding of distribution comparison in feature spaces that underpins FID.
• In the Northern English context, the University of Manchester’s Computer Science and Imaging Sciences departments have active groups in medical image analysis using generative models, where the retinal and histopathology FID debates are immediately relevant. The Manchester Centre for Imaging Sciences works on synthetic MRI data generation for rare diseases and paediatric populations, where small dataset sizes make FID sensitivity at low sample counts a practical concern. The University of Leeds Institute of Medical Research employs FID-based metrics to validate synthetic brain MRI and chest X-ray data for augmenting small clinical datasets — a critically important use case given NHS data access constraints around patient privacy under the UK GDPR framework. Newcastle University’s Digital Institute has applied generative model evaluation in cultural heritage digitisation projects, using FID to benchmark image super-resolution and inpainting models for archival document restoration where domain-specific feature extractors are needed. The University of Sheffield’s Computer Vision group (Professor Moi Hoon Yap’s group) contributes to generative model evaluation for skin lesion synthesis and wound imaging applications within the NHS digital pathology pipeline.

Future Directions (2026–2030)

• CMMD and CLIP-Based Metrics: The shift from Inception v3 to CLIP Encoder features as the universal feature extractor for FID-like metrics is widely expected to complete within the next two to three years, driven by the limitations of ImageNet-trained features for modern text-to-image and multi-modal generation evaluation. CMMD’s unbiasedness (no Gaussian assumption), lower sample-count requirement (reliable at ~1,000 vs ~50,000 for FID), and better alignment with human preference judgements make it a compelling default. The primary barrier to adoption is leaderboard inertia — the academic community’s need to maintain longitudinal comparability with historical baselines keeps FID entrenched even as its limitations are well understood. Transition protocols that report both FID and CMMD in parallel (similar to the transition from IS to FID in 2017–2019) are likely in the 2026–2028 timeframe.
• Perceptual Quality Beyond Gaussians: Replacing the Gaussian distributional assumption with normalising flows, kernel density estimators, or energy-based density models over the 2048-dimensional Inception v3 feature space — or higher-dimensional CLIP Encoder feature spaces — would yield a more faithful measure of distributional distance without the bias introduced by the Gaussian approximation. Computational cost has been the barrier: fitting a high-dimensional normalising flow to 50,000 feature vectors is expensive. Learned estimators trained once on large reference sets and then evaluated per-model are beginning to make this tractable. The FLD (Fréchet Distribution Distance with normalising flows) proposal extends this direction.
• Automated Multi-Metric Evaluation Suites: Initiatives like EvalGen (2025) and GenBench aim to standardise composite evaluation protocols that report FID, Precision-Recall, CLIP score, VQA accuracy (for conditional generation), DINO-FID (using DINOv2 features), and human evaluation simultaneously, packaging these into a reproducible automated evaluation pipeline. This reduces the ability of researchers to cherry-pick metrics that happen to favour a given architecture and brings generative model evaluation closer to the multi-metric leaderboard paradigm used in NLP benchmarking (GLUE, SuperGLUE, BIG-Bench).
• Video and 3D Generation Benchmarks: As text-to-video models (Sora, VideoLDM, Stable Video Diffusion) and text-to-3D models (DreamFusion, Magic3D, Zero-1-to-3) mature, the standardisation of Fréchet Video Distance and Fréchet Point Cloud Distance (FPCD) will mirror the role FID plays for 2D image generation. Temporal consistency metrics that extend FID to capture motion coherence, optical flow naturalness, and physics plausibility alongside per-frame quality are being formalised alongside FVD. The absence of a standardised video evaluation protocol analogous to the Clean-FID correction for images is a current gap that the 2026–2028 research community is beginning to address.
• Regulation-Aware Evaluation: As the EU AI Act (effective August 2026) and UK AI regulation (DSIT’s pro-innovation framework) begin to govern systems that generate realistic synthetic media of people, FID-like protocols may be embedded into mandatory conformity testing frameworks requiring standardised, auditable evaluation against real-image distributional baselines. Regulators in both jurisdictions are developing technical standards for photorealism thresholds below which a generated image of a person must be clearly labelled as synthetic. The FID Benchmark Protocol is the most natural technical operationalisation of such thresholds, positioning it for a regulatory role beyond academic benchmarking.
• Foundation Model Feature Extractors: The feature extractor backbone will migrate from Inception v3 (2016 ImageNet model) to modern foundation vision encoders trained on billions of image-text pairs: DINOv2 (Meta, 2023, 1.1B image training), SigLIP (Google, 2023, trained on WebLI), OpenCLIP ViT-bigG (LAION, trained on LAION-2B), and EVA-CLIP (BAAI, 2023). These models produce richer semantic representations with better coverage of out-of-distribution image content, yielding FID-like metrics that better align with human quality judgements across diverse domains including art, scientific imagery, and medical data. The transition to foundation model extractors will require establishing new canonical reference statistics for each benchmark dataset under the new feature space.
• Generalisation to Multi-Modal Generative Models: The FID protocol is inherently image-centric. Extending its principles to multi-modal generative models — systems that generate paired (image, audio), (image, text), or (video, action) tuples — requires distributional comparison in joint embedding spaces. CLIP provides a natural joint image-text feature space; extending to video+audio+text joint embeddings requires foundation models trained on multi-modal data (ImageBind, Meta 2023; AnyMAL, 2024). The Fréchet Distance in joint embedding spaces would provide an integrated quality-coherence metric for multi-modal generation.

Key Terminology

• Fréchet Distance (Fréchet Distance): Also known as the Wasserstein-2 distance in the Gaussian case. Measures the similarity between two probability distributions by computing the minimum-cost way of transforming one distribution into the other (optimal transport), penalising both mean displacement and covariance mismatch.
• Inception v3: A deep Convolutional Neural Networks architecture (Szegedy et al., 2016) trained on ImageNet ILSVRC 2012 (1.28M images, 1,000 classes). The 2048-dimensional global average pool layer embedding encodes high-level semantic content and serves as the fixed feature extractor in the FID protocol.
• Inception Score (IS): The predecessor metric (Salimans et al., 2016) measuring the KL divergence between the conditional label distribution p(y|x) and the marginal p(y) over generated images, averaged over all generated images. IS does not compare to real images and rewards models that generate highly classifiable, diverse outputs; it does not penalise mode collapse directly.
• Clean-FID: The corrected FID implementation (Parmar et al., CVPR 2022, github.com/GaParmar/clean-fid) that enforces antialiased bicubic resizing and uint8 quantisation, eliminating preprocessing-pipeline differences as a confound in cross-paper comparisons.
• KID (Kernel Inception Distance): An unbiased FID alternative (Binkowski et al., ICLR 2018) using MMD with a polynomial kernel on Inception v3 features. Statistically unbiased at any sample count; enables hypothesis testing on model comparisons.
• CMMD (CMMD): CLIP Maximum Mean Discrepancy (Jayasumana et al., CVPR 2024). Replaces Inception v3 with CLIP Encoder embeddings and the Gaussian assumption with a kernel MMD, yielding an unbiased metric requiring ~1,000 samples that better aligns with human judgement on modern Text-to-Image outputs.
• Precision-Recall Decomposition: Separates FID’s quality-diversity aggregate into precision (fraction of generated samples inside the real data manifold — fidelity) and recall (fraction of the real manifold covered by generated samples — diversity). Kynkäänniemi et al. (2019, NeurIPS) proposed using k-nearest-neighbour manifold estimates for both values.
• Fréchet Audio Distance (FAD): Audio analogue of FID using a VGGish feature extractor (trained on AudioSet) to compare distributions of real and generated audio clips. Used to benchmark audio synthesis models and music generation systems.
• Fréchet Video Distance (FVD): Video analogue of FID using I3D features (trained on Kinetics) to compare distributions of real and generated video clips across the temporal dimension. Introduced by Unterthiner et al. (2019).
• Reference Statistics: Precomputed (μ_r, Σ_r) over a canonical reference dataset using a fixed preprocessing pipeline. Clean-FID provides reference statistics files for CIFAR-10, ImageNet at multiple resolutions, and FFHQ, enabling reproducible FID evaluation without re-extracting real-data features.
• Mode Collapse: A failure mode of Generative Adversarial Networks where the generator produces only a small subset of the real data distribution. FID detects mode collapse via the covariance term: collapsed generation has a much smaller Σ_g, inflating Tr(Σ_r + Σ_g − 2(Σ_r Σ_g)^{1/2}).
• Zero-Shot FID: FID computed on a test dataset unseen during training, typically used in Text-to-Image benchmarking (zero-shot MS-COCO FID). Measures generalisation quality beyond the training distribution.
• Sample Efficiency: The minimum number of generated samples required to obtain a statistically reliable FID estimate. Standard FID requires ~50,000 samples; Clean-FID recommends the same; KID requires ~10,000; CMMD requires ~1,000. Sample efficiency matters for expensive generative models where generation cost is high.

Research & Literature

• 1. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., & Hochreiter, S. (2017). GANs trained by a two time-scale update rule converge to a local Nash equilibrium. NeurIPS, 30. arXiv:1706.08500.
• 2. Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., & Chen, X. (2016). Improved techniques for training GANs (Inception Score). NeurIPS, 29.
• 3. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the Inception architecture for computer vision. CVPR.
• 4. Binkowski, M., Sutherland, D. J., Arbel, M., & Gretton, A. (2018). Demystifying MMD GANs (KID). ICLR 2018. arXiv:1801.01401.
• 5. Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models (DDPM). NeurIPS, 33.
• 6. Nichol, A., & Dhariwal, P. (2021). Improved denoising diffusion probabilistic models. ICML. arXiv:2102.09672.
• 7. Dhariwal, P., & Nichol, A. (2021). Diffusion models beat GANs on image synthesis (ADM). NeurIPS, 34. arXiv:2105.05233.
• 8. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). High-resolution image synthesis with latent diffusion models (Stable Diffusion). CVPR. arXiv:2112.10752.
• 9. Kynkäänniemi, T., Karras, T., Laine, S., Lehtinen, J., & Aila, T. (2019). Improved precision and recall metric for assessing generative models. NeurIPS, 32.
• 10. Parmar, G., Zhang, R., & Zhu, J.-Y. (2022). On aliased resizing and surprising subtleties in GAN evaluation (Clean-FID). CVPR. arXiv:2104.11222.
• 11. Chong, M. J., & Forsyth, D. (2020). Effectively unbiased FID and Inception Score and where to find them. CVPR.
• 12. Jayasumana, S., Ramalingam, S., Veit, A., Glasner, D., Chakrabarti, A., & Kumar, S. (2024). Rethinking FID: towards a better evaluation metric for image generation (CMMD). CVPR. arXiv:2401.09603.
• 13. Brock, A., Donahue, J., & Simonyan, K. (2019). Large scale GAN training for high fidelity natural image synthesis (BigGAN). ICLR.
• 14. Karras, T., Laine, S., & Aila, T. (2019). A style-based generator architecture for GANs (StyleGAN). CVPR.
• 15. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., & Aila, T. (2021). Alias-free GAN (StyleGAN3). NeurIPS, 34.
• 16. Song, J., Meng, C., & Ermon, S. (2021). Denoising diffusion implicit models (DDIM). ICLR 2021. arXiv:2010.02502.
• 17. Peebles, W., & Xie, S. (2023). Scalable diffusion models with Transformers (DiT). ICCV 2023. arXiv:2212.09748.
• 18. Esser, P., et al. (2024). Scaling rectified flow transformers for high-resolution image synthesis (SD3 MMDiT). ICML 2024. arXiv:2403.03206.
• 19. Unterthiner, T., van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., & Gelly, S. (2019). FVD: A new metric for video generation. ICLR 2019 Workshop.
• 20. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S. K. S., Gontijo-Lopes, R., Karagol-Ayan, B., Salimans, T., Ho, J., Fleet, D. J., & Norouzi, M. (2022). Photorealistic text-to-image diffusion models with deep language understanding (Imagen). NeurIPS, 35.
• 21. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., & Chen, M. (2022). Hierarchical text-conditional image generation with CLIP latents (DALL-E 2). arXiv:2204.06125.
• 22. Podell, D., et al. (2023). SDXL: improving latent diffusion models for high-resolution image synthesis. ICLR 2024. arXiv:2307.01952.
• 23. Aziz, A., et al. (2024). GLIPS: Global-Local Image Perceptual Score for generative model evaluation. arXiv:2407.XXXXX.
• 24. Arora, S., & Zhang, Y. (2025). A pragmatic note on evaluating generative models with FID for retinal image synthesis. arXiv:2502.17160.
• 25. Henny, M., et al. (2025). EvalGen: Towards unified multi-metric evaluation suites for generative image models. arXiv preprint.
• 26. StudioGAN: Kang, M., et al. (2022). StudioGAN: a taxonomy and benchmark of GANs for image synthesis. IEEE TPAMI. arXiv:2206.09479.
• 27. Croitoru, F.-A., Hondru, V., Ionescu, R. T., & Shah, M. (2023). Diffusion models in vision: A survey. IEEE TPAMI.
• 28. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Shao, W., Zhang, S., Cui, B., & Yang, M.-H. (2023). Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys.

Provenance