chicagoypark/ImageNet_FID50K

ImageNet-256 Inception pool3 features (50k) — reference and VAR samples

Precomputed Inception-v3 pool3 features (2048-d) for a 50,000-image ImageNet-256 reference set and for 50,000 generated samples from each of the three VAR checkpoints (Tian et al., NeurIPS 2024).

These are the exact features behind the energy-distance / FID comparison in our work. They let you reproduce distributional metrics without re-running Inception over 200,000 images (~1 GPU-hour), and without downloading ~40 GB of raw pixels.

Files

File	Shape	dtype	Size	Contents
imagenet256_train_ref50k_inception_pool3.npz	(50000, 2048)	float32	391 MB	Reference — ImageNet-2012 train, 50 images/class × 1000 classes
var_d16_imagenet256_50k_inception_pool3.npz	(50000, 2048)	float32	391 MB	VAR-d16 samples
var_d20_imagenet256_50k_inception_pool3.npz	(50000, 2048)	float32	391 MB	VAR-d20 samples
var_d24_imagenet256_50k_inception_pool3.npz	(50000, 2048)	float32	391 MB	VAR-d24 samples

Every array is stored under the key arr_0 (the ADM/guided-diffusion convention).

import numpy as np
from huggingface_hub import hf_hub_download

p = hf_hub_download("chicagoypark/ImageNet_FID50K",
                    "var_d24_imagenet256_50k_inception_pool3.npz",
                    repo_type="dataset")
feats = np.load(p)["arr_0"]        # (50000, 2048) float32

How these were produced

Reference. ImageNet ILSVRC-2012 train, stratified 50 images per class over all 1000 classes, seed 0. Preprocessing is the standard ADM evaluation pipeline: resize short side to 256, then center crop to 256×256.

python pack_imagenet_subsample.py \
  --imagenet_train_root /path/to/ILSVRC2012_img_train \
  --output_npz imagenet_ref50k.npz \
  --per_class 50 --image_size 256 --seed 0

Generated samples. VAR-d{16,20,24} at ImageNet 256×256, 50 images per class × 1000 classes, cfg=1.5, top_k=900, top_p=0.96.

Features. Extracted with the frozen Inception-v3 graph classify_image_graph_def.pb — the same graph OpenAI's guided-diffusion evaluator uses for FID — taking the pool_3 layer.

python extract_inception_pool.py --in_npz <images>.npz --out_npz <features>.npz

Verification

Exact energy distance V-statistic over all 50,000 × 50,000 pairs (2.5 billion), reference vs each model. These are deterministic: recomputing on the same data returns bit-identical values (we verified reproduction four months apart on different GPUs).

model	E|X-Y|	E|X-X'|	E|Y-Y'|	D²_V	D²_U
VAR-d16	1.84103684e+01	1.88215388e+01	1.79656066e+01	3.35914e-02	3.28556e-02
VAR-d20	1.85036478e+01	1.88215388e+01	1.81584873e+01	2.72695e-02	2.65299e-02
VAR-d24	1.85712376e+01	1.88215388e+01	1.83037014e+01	1.72350e-02	1.64925e-02

where D² = 2E‖X−Y‖ − E‖X−X′‖ − E‖Y−Y′‖.

D²_V vs D²_U. The two differ only in whether the within-set averages include the i = j diagonal. D²_V divides by n² (diagonal included, contributing zeros); D²_U divides by n(n−1). Since the diagonal is exactly zero, this is an exact algebraic relation:

D²_V = D²_U + E_XX/n + E_YY/m          (here: +4.3% of D²_V, and it decays as 1/n)
D²_U = 2·E_XY − (n/(n−1))·E_XX_V − (m/(m−1))·E_YY_V

D²_V is biased upward but is guaranteed non-negative (it is the energy distance between the two empirical distributions). D²_U is unbiased but can go negative when the true distance is near zero. At n = 50,000 either is safe; below ~5,000 the D²_V bias becomes a large fraction of the signal.

Caveats — please read before using

1. The arrays are class-sorted, exactly 50 images per class. Row i belongs to class i // 50. Consequences:

• A contiguous slice is not a random subsample. feats[:5000] is 100% of the first 100 classes and 0% of the other 900. Always draw explicit random indices, ideally stratified per class.

• Inception Score requires shuffling first. The ADM evaluator splits activations contiguously at split_size=5000, so on a class-sorted array each split spans only 100 classes, collapsing the marginal entropy that IS depends on. Measured on these exact files:

  	as-stored (sorted)	shuffled	published
  VAR-d16	62.25	280.20	274.4
  VAR-d20	65.07	304.42	302.6
  VAR-d24	66.26	312.26	312.9
  reference (real ImageNet)	67.63	335.47	—

  The last row is the tell: real photographs show the same depressed value, so this is an ordering artifact, not a property of the samples. FID, sFID, precision and recall are permutation-invariant and unaffected.

2. FID computed against this reference will not match published VAR numbers. Published FID for these checkpoints (3.60 / 2.95 / 2.33) is computed against VIRTUAL_imagenet256_labeled.npz, whose mu/sigma were precomputed from the full ~1.28M-image ImageNet train set — not from a 50k subsample. Precision/recall in that file use only its 10,000 stored images. Comparing to published numbers requires that reference; this 50k reference is the one the energy distances above are computed against.

3. Pixel-space metrics are not reproducible from these files. These are 2048-d features. Anything in raw pixel space needs the original (50000, 256, 256, 3) uint8 arrays, which are not part of this upload.

Citation

Features derive from ImageNet (Deng et al., 2009), the Inception-v3 graph distributed with guided-diffusion (Dhariwal & Nichol, 2021), and VAR checkpoints (Tian et al., 2024). Please cite those works alongside this dataset.