This package provides teaching nodes for quantifying how similar two images are, e.g. to score a reconstruction, an upscaling method, or a compression result against a reference - shown directly on the node and output as FLOAT for further use.

Features

PSNR (Peak Signal-to-Noise Ratio, dB) plus the underlying MSE it's computed from, for free
MAE (Mean Absolute Error)
SSIM (Structural Similarity Index), fixed to the standard Wang et al. 2004 parameters (11x11 Gaussian window, sigma=1.5, K1=0.01, K2=0.03) - no parameters exposed, since these defaults are what "SSIM" conventionally means in papers
A batch is scored by computing the metric per image pair and averaging - the same convention benchmark papers use over a whole test set - rather than pooling every pixel from every image together
Assumes the standard 0..1 IMAGE range; not offered for SIGNED_IMAGE, since PSNR's reference range has to match the data's true peak-to-peak range to mean anything

Included Nodes

Metrics

Image PSNR
Image MAE
Image SSIM

Sample Workflows

Lesson 1. Image PSNR

Workflow1

Downscales an image and upscales it back with three different interpolation methods (bilinear, bicubic, lanczos), then scores each reconstruction against the original with Image PSNR - a quick way to compare resampling methods numerically.

Lesson 2. Image SSIM

Workflow2

Same comparison as Lesson 1, scored with Image SSIM instead - useful for seeing where PSNR and SSIM agree or disagree on which method looks "better".

Installation

Clone this repository into your ComfyUI custom_nodes folder.

ComfyUI/
└── custom_nodes/
    └── comfyui-image-metrics/

Restart ComfyUI after installation. No extra Python dependencies are required.

License

See the repository license file for details.