ChocolateDave/imagenet-256-sd-vae-latents

ImageNet-256 SD VAE latents

This dataset caches VAE posterior parameters for ImageNet-1k (ILSVRC 2012) images at 256 × 256 resolution. It supports latent-space generative modeling, including DiT and Flow Maps experiments, without repeatedly encoding images.

Data splits and format

Split	Images	Parquet shards	Filenames
Training	1,281,167	294	train-00000-of-00294.parquet … train-00293-of-00294.parquet
Validation	50,000	14	validation-00000-of-00014.parquet … validation-00013-of-00014.parquet

Files live under data/. Each file preserves the filename, image order, labels, and row boundaries of its corresponding original ImageNet shard. The unlabeled test split is not included. Counts above describe complete splits; completion records in metadata/ identify the published, verified splits.

Field	Type	Description
image	float32 array, (8, 32, 32)	Unscaled VAE posterior mean in channels 0–3 and standard deviation in channels 4–7, in CHW order.
label	ClassLabel	Original ImageNet integer label and 1,000-class name mapping.
source_index	int64	(source_parquet_index << 32) + row_in_source_parquet. Unique within each split; not a contiguous global index.

The image field contains posterior parameters, not decoded pixels or a sampled latent. The last four channels store standard deviation, not log variance.

Processing and provenance

Source images come from ILSVRC/imagenet-1k at revision 49e2ee26f3810fb5a7536bbf732a7b07389a47b5. The pretrained encoder is enterprise-explorers/sd-vae-ft-mse-flax at revision 7581b0a0489cc8483c876a728b830b9ce087cf03. These are Flax-converted SD VAE ft-MSE weights.

1. Convert the image to RGB.
2. Apply the ADM center crop used by MeanFlow: repeatedly halve large images with BOX filtering, resize the shorter side to 256 with BICUBIC filtering, then center crop to 256 × 256.
3. Normalize pixel values to [-1, 1].
4. Encode with the VAE and store mean and exp(0.5 * log_variance) as float32 posterior parameters.

No random horizontal flip or posterior sampling is baked into the cache. Apply training augmentation and sample the posterior when consuming the dataset. Metadata under metadata/ records source revisions, representation, row counts, and completed splits.

Loading and sampling

import numpy as np
from datasets import load_dataset

rows = load_dataset(
    "ChocolateDave/imagenet-256-sd-vae-latents",
    split="train",  # or "validation"
    streaming=True,
    token=True,  # authenticate with your Hugging Face account if access is gated
).with_format("numpy")

example = next(iter(rows))
mean, std = np.split(example["image"], 2, axis=0)
rng = np.random.default_rng(0)
noise = rng.standard_normal(mean.shape).astype(np.float32)
z = (mean + std * noise) * np.float32(0.18215)

z is a scaled training latent with shape (4, 32, 32). Before VAE decoding, divide it by 0.18215 and convert to the decoder's expected tensor layout. Remove streaming=True to download and cache a split locally.

Licensing and access

ImageNet data. This derived cache remains subject to the original ImageNet terms of access, also displayed by the source repository. Use is limited to non-commercial research and education. Recipients must agree to those terms before receiving access, and users remain responsible for rights in the source images. No additional rights to the underlying images are granted by this conversion. Public repository visibility does not replace acceptance of ImageNet's terms; downloads should retain an access gate requiring that agreement, as in timm/imagenet-1k-wds.

VAE checkpoint. The encoder model card declares the checkpoint license as MIT. See that model card and the original SD VAE ft-MSE model for attribution and model details. Model weights are not included here. The checkpoint's MIT declaration does not relicense this ImageNet-derived dataset or remove ImageNet's access restrictions.

Intended use and limitations

This cache is intended for research and evaluation on ImageNet latent-space models. Center cropping removes image content outside the crop, and VAE compression is lossy. Reconstructions can retain identifiable or copyrighted content. Source-label errors, representation biases, and encoder limitations remain relevant; this conversion does not independently audit them.

Attribution and references

ImageNet was curated by the ImageNet/ILSVRC team, including Olga Russakovsky, Jia Deng, and collaborators. The VAE checkpoint is distributed by Enterprise Explorers as a Flax conversion of the Stable Diffusion ft-MSE autoencoder. This cache was prepared by ChocolateDave for the Flow Maps project.

• Russakovsky et al., ImageNet Large Scale Visual Recognition Challenge, IJCV, 2015.
• Deng et al., ImageNet: A Large-Scale Hierarchical Image Database, CVPR, 2009.
• Original ImageNet Parquet dataset.
• SD VAE ft-MSE Flax checkpoint.
• MeanFlow dataset preparation reference.