metadata
dataset_info:
features:
- name: latent
sequence:
sequence:
sequence: float16
splits:
- name: train
num_bytes: 5878040000
num_examples: 70000
download_size: 3475551903
dataset_size: 5878040000
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
task_categories:
- unconditional-image-generation
size_categories:
- 10K<n<100K
FFHQ Dataset (pravsels/FFHQ_1024) encoded using the dc-ae-f128c512-mix-1.0 auto encoder.
Example usage
import sys
sys.path.append('../dcae') # https://github.com/vladmandic/dcae
from dcae import DCAE
from datasets import load_dataset
import torch
import torchvision
dataset = load_dataset("SwayStar123/FFHQ_1024_DC-AE_f128", split="train")
dc_ae = DCAE("dc-ae-f128c512-mix-1.0", device="cuda", dtype=torch.bfloat16).eval() # Must be bfloat. with float16 it produces terrible outputs.
def denorm(x):
return (x * 0.5 + 0.5).clamp(0, 1)
latent = next(iter(dataset))["latent"]
latent = torch.tensor(latent, device=DEVICE, dtype=DTYPE)
with torch.no_grad():
recon = dc_ae.decode(latent.unsqueeze(0)).squeeze(0)
recon = denorm(recon).to(torch.float32)
torchvision.utils.save_image(
recon,
"recon.png",
normalize=False,
)