Datasets:
latent_mean array 3D | latent_mean_flip array 3D | label int64 0 150 |
|---|---|---|
[
[
[
13.770496368408203,
11.50731086730957,
10.199234962463379,
9.990827560424805,
2.8142683506011963,
-0.3535628914833069,
3.510664463043213,
6.67972469329834,
-1.177479863166809,
9.455290794372559,
3.1241936683654785,
5.325014114379883,
... | [
[
[
13.526000022888184,
11.420271873474121,
11.14678955078125,
10.942617416381836,
-2.342175245285034,
-3.1613192558288574,
1.379578948020935,
-0.6782270669937134,
-1.6897149085998535,
3.346569299697876,
3.79329514503479,
3.2261769771575928... | 0 |
[[[2.7381083965301514,-10.102185249328613,1.2500375509262085,2.3481552600860596,0.5691646933555603,-(...TRUNCATED) | [[[14.297260284423828,16.872407913208008,12.039960861206055,13.573067665100098,10.523987770080566,-3(...TRUNCATED) | 0 |
[[[-6.093410968780518,-12.264487266540527,1.3106073141098022,2.0087454319000244,0.7022301554679871,-(...TRUNCATED) | [[[-4.086976528167725,-3.8903286457061768,-2.185616970062256,-2.9218592643737793,1.650349497795105,2(...TRUNCATED) | 0 |
[[[-3.3233046531677246,5.40574312210083,3.147181510925293,2.680140495300293,4.706805229187012,2.0946(...TRUNCATED) | [[[9.847652435302734,5.955145835876465,6.343406677246094,8.050045013427734,5.093027591705322,5.27234(...TRUNCATED) | 0 |
[[[-5.807473182678223,-3.9747936725616455,-4.09126615524292,-0.8862267732620239,1.3231884241104126,-(...TRUNCATED) | [[[0.6593751311302185,-4.620868682861328,-5.6849493980407715,-0.9953986406326294,-1.3549933433532715(...TRUNCATED) | 0 |
[[[11.955228805541992,11.479147911071777,10.639253616333008,11.267666816711426,11.209165573120117,11(...TRUNCATED) | [[[9.731059074401855,-1.2644717693328857,1.0425288677215576,6.542402744293213,7.332821369171143,7.16(...TRUNCATED) | 0 |
[[[-4.385206699371338,4.352762699127197,-3.817986249923706,-0.6614940762519836,6.1144819259643555,-0(...TRUNCATED) | [[[11.918990135192871,9.579286575317383,0.9102331399917603,0.9657902121543884,8.909244537353516,13.9(...TRUNCATED) | 0 |
[[[-7.787033557891846,-3.53220534324646,-0.8615802526473999,-2.4414920806884766,4.325491905212402,-6(...TRUNCATED) | [[[-1.0689473152160645,1.8585505485534668,-6.175002098083496,4.516266345977783,9.943082809448242,-2.(...TRUNCATED) | 0 |
[[[-9.005204200744629,0.34543466567993164,-7.671469211578369,0.9892041683197021,13.598747253417969,7(...TRUNCATED) | [[[0.27867668867111206,-5.350423336029053,-6.565067291259766,-8.305253982543945,-1.180970549583435,-(...TRUNCATED) | 0 |
[[[-0.2541891932487488,-2.874506950378418,-1.2002344131469727,-3.8453099727630615,3.6042640209198,-2(...TRUNCATED) | [[[4.9600725173950195,18.721717834472656,8.209339141845703,-7.915995121002197,-1.5551550388336182,-3(...TRUNCATED) | 0 |
ImageNet-256 SD-VAE-ft-MSE Latents
Pre-computed posterior means (no variance/std) from the Stable Diffusion VAE (stabilityai/sd-vae-ft-mse) for the full ImageNet-1K training set at 256×256 resolution, stored as Parquet shards. Each example includes latents for both the original and horizontally flipped image, enabling flip augmentation without re-encoding at training time.
Dataset Description
Each example contains:
| Column | Shape | Type | Description |
|---|---|---|---|
latent_mean |
(4, 32, 32) |
float32 |
Posterior mean of the original image |
latent_mean_flip |
(4, 32, 32) |
float32 |
Posterior mean of the horizontally flipped image |
label |
scalar | int64 |
ImageNet class label (0–999) |
- Number of examples: 1,281,167 (full ImageNet-1K train split)
- Latent spatial size: 32×32 (8× downsampled from 256×256 pixels)
- Latent channels: 4
Creation
Images were center-cropped and resized to 256×256 using the Dhariwal (ADM) cropping method, then encoded with stabilityai/sd-vae-ft-mse. Only the posterior mean is stored (not variance), for both the original and horizontally flipped image. Pixels are normalised to [-1, 1] before encoding, consistent with DiT / SiT conventions.
Usage
from datasets import load_dataset
ds = load_dataset("yuanchenyang/imagenet-256-sd-vae-ft-mse-latesnts", split="train")
example = ds[0]
latent = example["latent_mean"] # numpy array (4, 32, 32)
latent_flip = example["latent_mean_flip"] # numpy array (4, 32, 32)
label = example["label"] # int
Intended Use
Training latent diffusion models (e.g., DiT, SiT, SR-DiT) on ImageNet-256 without needing to run the VAE encoder during training.
Source Code
Preprocessing scripts are based on: https://github.com/Martinser/REG/tree/main/preprocessing
License
The latent representations inherit the ImageNet license terms. The VAE weights are from Stability AI's sd-vae-ft-mse (CreativeML Open RAIL-M license). The preprocessing code is licensed under the MIT license.
- Downloads last month
- 376