toiletV2-onnx — StyleGAN2-ADA 128×128 (ONNX, opset 17)

ONNX export of the G_ema generator from the PyTorch repo RaspizdAI/toiletV2 (StyleGAN2-ADA, unconditional, 128×128, truncation_psi = 0.7).

🚀 Live preview

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

🎞️ Training timelapse

Smooth timelapse, kimg 0 → kimg 2252.

Higher-quality version: timelapse_toiletV2.mp4

File

network-snapshot-002252.onnx — opset 17, a single self-contained file (no external .data).

Input / output

	name	type	shape
input	z	float32	[batch, 512] (dynamic batch)
output	image	float32	[batch, 3, 128, 128], range [-1, 1]

Post-process: img = ((image * 0.5) + 0.5).clip(0, 1).

Usage

import numpy as np, onnxruntime as ort
sess = ort.InferenceSession("network-snapshot-002252.onnx",
                            providers=["CUDAExecutionProvider", "CPUExecutionProvider"])
z = np.random.randn(1, 512).astype(np.float32)
img = sess.run(["image"], {"z": z})[0]        # [1, 3, 128, 128] in [-1, 1]

Export details

• truncation_psi = 0.7 is baked into the graph; noise_mode = "const", fused_modconv = False, no conditioning (c is not an input).
• opset 17 (broadest compatibility: onnxruntime, web/mobile, HF). IR version 10.
• Numerically verified against PyTorch: max|Δ| ≈ 5e-3 (ONNX CPU vs PyTorch CUDA, fp32); output is bit-identical to the opset-18 export.
• Original weights (safetensors/pkl) — see RaspizdAI/toiletV2.

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.