fussim 0.2.0

Optimized CUDA implementation of differentiable SSIM for PyTorch

fussim

Fast CUDA SSIM for PyTorch — pip install fussim with pre-built wheels.

~7x faster than pytorch-msssim | FP16/AMP support | Drop-in replacement

Based on MrNeRF/optimized-fused-ssim.

---

Installation

pip install fussim

Works out of the box for Linux (PyTorch 2.9 + CUDA 12.8) and Windows (PyTorch 2.8 + CUDA 12.8).

Other PyTorch/CUDA versions

Pre-built wheels for PyTorch 2.5–2.9 and CUDA 11.8–12.8:

pip install fussim --extra-index-url https://opsiclear.github.io/fussim/whl/

This auto-selects the correct wheel for your installed PyTorch version.

PyTorch	Python	CUDA 11.8	CUDA 12.1	CUDA 12.4	CUDA 12.6	CUDA 12.8
2.5.1	3.10–3.12	✓	✓	✓	—	—
2.6.0	3.10–3.12	✓	—	✓	✓	—
2.7.1	3.10–3.12	✓	—	—	✓	✓
2.8.0	3.10–3.13	—	—	—	✓	✓
2.9.0	3.10–3.13	—	—	—	✓*	✓*

*Linux only. Windows has a known PyTorch bug.

→ Open Installation Configurator

Build from source

Requires CUDA Toolkit and C++ compiler.

git clone https://github.com/OpsiClear/fussim.git && cd fussim
pip install .

# For specific GPU architecture:
TORCH_CUDA_ARCH_LIST="8.9" pip install .  # RTX 4090

Other platforms

For macOS, older CUDA versions, or unsupported configurations, use the original fused-ssim which builds from source.

---

Quick Start

import torch
from fussim import fused_ssim

img1 = torch.rand(1, 3, 256, 256, device="cuda", requires_grad=True)
img2 = torch.rand(1, 3, 256, 256, device="cuda")

# Compute SSIM
ssim_value = fused_ssim(img1, img2)

# Use as loss
loss = 1.0 - fused_ssim(img1, img2)
loss.backward()

FP16 / AMP:

with torch.autocast(device_type="cuda"):
    ssim_value = fused_ssim(img1, img2)  # Uses FP16 kernel automatically

---

API

fused_ssim

fused_ssim(img1, img2, padding="same", train=True, window_size=11) -> Tensor

Parameter	Type	Default	Description
img1	Tensor	—	First image (B, C, H, W). Receives gradients.
img2	Tensor	—	Second image (B, C, H, W)
padding	str	"same"	"same" or "valid"
train	bool	True	Enable gradient computation
window_size	int	11	Gaussian window: 7, 9, or 11

Returns: Scalar mean SSIM value.

ssim (pytorch-msssim compatible)

ssim(X, Y, data_range=255, size_average=True, win_size=11, K=(0.01, 0.03), nonnegative_ssim=False) -> Tensor

Parameter	Type	Default	Description
X, Y	Tensor	—	Images (B, C, H, W). Gradients computed for X.
data_range	float	255	Value range (255 for uint8, 1.0 for normalized)
size_average	bool	True	Return scalar mean or per-batch values
win_size	int	11	7, 9, or 11
K	tuple	(0.01, 0.03)	SSIM constants (K1, K2)
nonnegative_ssim	bool	False	Clamp negative values to 0

SSIM Module

from fussim import SSIM

module = SSIM(data_range=1.0)
loss = 1 - module(pred, target)
loss.backward()

---

Performance

RTX 4090, batch 5×5×1080×1920, 100 iterations:

Implementation	Forward	Backward	Total	Speedup
pytorch-msssim	28.7 ms	28.9 ms	57.5 ms	1.0×
fussim	4.38 ms	4.66 ms	9.04 ms	6.4×

---

Limitations

Constraint	Reason
window_size: 7, 9, or 11	CUDA kernel templates
win_sigma: 1.5 (fixed)	Hardcoded in kernel
Custom win not supported	Uses built-in Gaussian

---

Attribution

• optimized-fused-ssim by Janusch Patas
• fused-ssim by Rahul Goel

Citation

@software{optimized-fused-ssim,
    author = {Janusch Patas},
    title = {Optimized Fused-SSIM},
    year = {2025},
    url = {https://github.com/MrNeRF/optimized-fused-ssim},
}