fussim 0.1.0

Optimized CUDA implementation of differentiable SSIM for PyTorch

fussim

Fast CUDA SSIM for PyTorch. Based on MrNeRF/optimized-fused-ssim.

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

Installation

pip install fussim

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/

Build from source

Requires CUDA Toolkit and C++ compiler.

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

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()

API

fused_ssim

from fussim import fused_ssim

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

Parameter	Type	Default	Description
img1	Tensor	required	First image (B, C, H, W). Receives gradients.
img2	Tensor	required	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)

from fussim import ssim

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	Tensor	required	First image (B, C, H, W). Receives gradients.
Y	Tensor	required	Second image (B, C, H, W).
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	Gaussian window: 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, size_average=True, win_size=11, K=(0.01, 0.03))
loss = 1 - module(pred, target)
loss.backward()

FP16 / AMP

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

Performance

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

Implementation	Forward	Backward	Total	Speedup
pytorch_msssim	28.7 ms	28.9 ms	57.5 ms	1.0x
fussim	4.38 ms	4.66 ms	9.04 ms	6.4x

Limitations

Parameter	Constraint	Reason
win_size	7, 9, or 11	CUDA kernel templates
win_sigma	1.5 (fixed)	Hardcoded in kernel
win	Not supported	Uses built-in Gaussian

Attribution

Project	Author
optimized-fused-ssim	Janusch Patas
fused-ssim	Rahul Goel

Citation

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