toiletV2 — StyleGAN2-ADA 128×128 (RaspizdAI)

Unconditional StyleGAN2-ADA (NVIDIA PyTorch) generator of 128×128 toilet images, trained from scratch on the dataset_128 dataset — 9,966 images.

🚀 Live preview

Run the model directly in your browser — no install required: ai.amogus.bond

🎞️ Training timelapse

Smooth timelapse, kimg 0 → kimg 2252 (last 4 snapshots locally, every snapshot also uploaded here).

Higher-quality version: timelapse_toiletV2.mp4

Model specs

Parameter	Value
Architecture	StyleGAN2-ADA (PyTorch)
Resolution	128×128
z_dim / w_dim	512 / 512
Conditioning	none (c_dim = 0)
Trained for	~2,252 kimg out of 3,000 (run stopped manually)
Best FID50k	27.5 (kimg 2211)
truncation_psi	0.7

Files

File	Contents
network-snapshot-002252.safetensors	generator G_ema weights (113 tensors) — inference only
network-snapshot-002252.pkl	full training snapshot (G, D, G_ema, augment_pipe) — for resume/finetune

Usage (PyTorch + safetensors)

import sys, torch
from safetensors.torch import load_file
sys.path.insert(0, "stylegan2-ada-pytorch")          # original StyleGAN2-ADA repo
from training.networks import Generator

G = Generator(z_dim=512, c_dim=0, w_dim=512, img_resolution=128, img_channels=3).cuda().eval()
G.load_state_dict(load_file("network-snapshot-002252.safetensors"))

z = torch.randn(1, 512, device="cuda")
c = torch.zeros([1, 0], device="cuda")
img = G(z, c, truncation_psi=0.7, noise_mode="const", fused_modconv=False)  # [1,3,128,128], range [-1,1]

Note: on recent PyTorch versions the custom CUDA kernels (bias_act/upfirdn2d) from the original repo may fail to build — the pure-PyTorch reference implementations are then selected automatically.

Training

• Dataset: 9,966 images, 128×128.
• --cfg=auto --mirror=1 --batch=128; FID50k computed during training.
• On every snapshot a log + a grid of samples were pushed to Telegram, and the weights to this repo.

License & Attribution

This model is licensed under CC-BY-SA-4.0.

• You are free to share and adapt the material.
• You must give appropriate credit, provide a link to the license, and indicate if changes were made.
• You cannot remove the RaspizdAI copyright and attribution, nor can you close the source of derivative models.