shreenithi20/fmnist-8x8-latents

Fashion MNIST 8×8 Latents Dataset

This dataset contains Fashion MNIST images converted to 8×8×4 VAE latent representations, optimized for training transformer-based diffusion models.

Dataset Information

• Source: Fashion MNIST (60,000 training images)
• Latent Resolution: 8×8×4 (compressed from 64×64×3)
• Total Samples: 60,000
• Data Type: float32
• Latent Range: [-3.65, 3.66]
• VAE Model: stabilityai/sd-vae-ft-mse

Files

• image_latents.npy: Latent representations (60000, 4, 8, 8)
• label_encodings.npy: One-hot encoded class labels (60000, 10)

Class Labels (0-9)

0. T-shirt/top
1. Trouser
2. Pullover
3. Dress
4. Coat
5. Sandal
6. Shirt
7. Sneaker
8. Bag
9. Ankle boot

Usage

import numpy as np

# Load latents
latents = np.load("image_latents.npy")  # Shape: (60000, 4, 8, 8)
labels = np.load("label_encodings.npy")  # Shape: (60000, 10)

# Use with diffusion models
model_input = torch.from_numpy(latents).float()
class_conditioning = torch.from_numpy(labels).float()

VAE Configuration

• Input Resolution: 64×64×3 (FMNIST images resized)
• Output Resolution: 8×8×4 (compressed latents)
• Compression Factor: 64× (64×64×3 → 8×8×4)
• Latent Scale: 0.18215

Important Note on Scaling

The latents in this dataset are scaled by a factor of 0.18215. This scaling is applied by the VAE encoder to normalize the latent representations. When using these latents:

• For training: Use the scaled latents as-is (the model learns to work with this scaling)
• For decoding back to images: Manually divide the latents by 0.18215 before passing to the VAE decoder
• For custom processing: If you need the original unscaled latents, divide by 0.18215

Example of correct decoding:

# Load scaled latents
scaled_latents = np.load("image_latents.npy")

# Unscale before decoding
unscaled_latents = scaled_latents / 0.18215

# Now pass to VAE decoder
decoded_images = vae_decoder(unscaled_latents)

This scaling factor is standard for the stabilityai/sd-vae-ft-mse model and ensures optimal training stability.