NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI · #5772 most downloaded on PyPI
High-fidelity performance metrics for generative models in PyTorch
Last release 7 months ago
17 Feb 2026
Ships fairly regularly
a new release about every 2.9 years
Most releases are documented
notes for 3 of 4 stable releases
Nothing withdrawn
no release was ever pulled
6 years old
4 releases · first in 2020
One column per quarter.
Refer to the CHANGELOG.md for details
Refer to the CHANGELOG.md for details
cifar100-train, cifar100-valclip-vit-b-32, vgg16, dinov2-vit-s-14, dinov2-vit-b-14, dinov2-vit-l-14, dinov2-vit-g-14calculate_metrics
prc: Calculate PRC (Precision and Recall)prc_neighborhood: Number of nearest neighbours to consider in PRCprc_batch_size: Batch size in PRCfeature_layer_prc: Name of the feature layer to use with PRC metricsamples_resize_and_crop: Transform all images found in the directory to a given size and square shapefeature_extractor: Accepts a new feature extractors clip-vit-b-32, vgg16, dinov2-vit-s-14, dinov2-vit-b-14, dinov2-vit-l-14, dinov2-vit-g-14feature_extractor_internal_dtype: Allows to change the internal dtype used in the feature extractor's weights and activations; might be useful to counter numerical issues arising in fp32 implementations, e.g. those seen with the growth of the batch sizefeature_extractor_compile: Compile feature extractor (experimental: may have negative effect on the metrics numerical precision)kid_kernel: Allows choosing between the default poly (polynomial, default) and rbf (RBF) kernelskid_kernel_rbf_sigma: Specifies RBF kernel sigma in KIDprc: Calculate PRC (Precision and Recall)prc-neighborhood: Number of nearest neighbours to consider in PRCprc-batch-size: Batch size in PRCfeature-layer-prc: Name of the feature layer to use with PRC metric--samples-resize-and-crop: Transform all images found in the directory to a given size and square shape--feature-extractor: Accepts a new feature extractors clip-vit-b-32, vgg16, dinov2-vit-s-14, dinov2-vit-b-14, dinov2-vit-l-14, dinov2-vit-g-14--feature-extractor-internal-dtype: Allows to change the internal dtype used in the feature extractor's weights and activations; might be useful to counter numerical issues arising in fp32 implementations, e.g. those seen with the growth of the batch size--feature-extractor-compile: Compile feature extractor (experimental: may have negative effect on the metrics numerical precision)--kid-kernel: Allows choosing between the default poly (polynomial, default) and rbf (RBF) kernels--kid-kernel-rbf-sigma: Specifies RBF kernel sigma in KIDInputs can be models Perceptual Path Length Documentation
Inputs can be models Perceptual Path Length Documentation
calculate_metrics
ppl: Calculate PPL (Perceptual Path Length)ppl_epsilon: Interpolation step size in PPLppl_reduction: Reduction type to apply to the per-sample output valuesppl_sample_similarity: Name of the sample similarity to use in PPL metric computationppl_sample_similarity_resize: Force samples to this size when computing similarity, unless set to Noneppl_sample_similarity_dtype: Check samples are of compatible dtype when computing similarity, unless set to None.ppl_discard_percentile_lower: Removes the lower percentile of samples before reductionppl_discard_percentile_higher: Removes the higher percentile of samples before reductionppl_z_interp_mode: Noise interpolation mode in PPLinput1_model_z_type: Type of noise accepted by the input1 generator modelinput1_model_z_size: Dimensionality of noise accepted by the input1 generator modelinput1_model_num_classes: Number of classes for conditional generation (0 for unconditional) accepted by the input1 generator modelinput1_model_num_samples: Number of samples to draw from input1 generator model, when it is provided as a path to ONNX model. This option affects the following metrics: ISC, FID, KIDinput2_model_z_type: Type of noise accepted by the input2 generator modelinput2_model_z_size: Dimensionality of noise accepted by the input2 generator modelinput2_model_num_classes: Number of classes for conditional generation (0 for unconditional) accepted by the input2 generator modelinput2_model_num_samples: Number of samples to draw from input2 generator model, when it is provided as a path to ONNX model. This option affects the following metrics: ISC, FID, KID--ppl: Calculate PPL (Perceptual Path Length)--ppl-epsilon: Interpolation step size in PPL--ppl-reduction: Reduction type to apply to the per-sample output values--ppl-sample-similarity: Name of the sample similarity to use in PPL metric computation--ppl-sample-similarity-resize: Force samples to this size when computing similarity, unless set to None--ppl-sample-similarity-dtype: Check samples are of compatible dtype when computing similarity, unless set to None.--ppl-discard-percentile-lower: Removes the lower percentile of samples before reduction--ppl-discard-percentile-higher: Removes the higher percentile of samples before reduction--ppl-z-interp-mode: Noise interpolation mode in PPL--input1-model-z-type: Type of noise accepted by the input1 generator model--input1-model-z-size: Dimensionality of noise accepted by the input1 generator model--input1-model-num-classes: Number of classes for conditional generation (0 for unconditional) accepted by the input1 generator model--input1-model-num-samples: Number of samples to draw from input1 generator model, when it is provided as a path to ONNX model. This option affects the following metrics: ISC, FID, KID--input2-model-z-type: Type of noise accepted by the input2 generator model--input2-model-z-size: Dimensionality of noise accepted by the input2 generator model--input2-model-num-classes: Number of classes for conditional generation (0 for unconditional) accepted by the input2 generator model--input2-model-num-samples: Number of samples to draw from input2 generator model, when it is provided as a path to ONNX model. This option affects the following metrics: ISC, FID, KIDstl10-train, stl10-test, stl10-unlabelednormal, uniform_0_1, unitlerp, slerp_any, slerpexamples/sngan_cifar10.py)Fidelity and determinism tests passing, relative errors updated in README.md. Inception weights reuploaded from https://github.com/mseitzer/pytorch-fi
Fidelity and determinism tests passing, relative errors updated in README.md. Inception weights reuploaded from https://github.com/mseitzer/pytorch-fid/releases in order to assign DOI to this release.
Nothing published for this version
Your coding agent can read these notes before it upgrades. Set up the MCP server →