Crate blazediff_milo 

MILO: a learned perceptual image quality metric, run on the CPU.

MILO (Çoğalan, Bemana, Myszkowski, Seidel and Groth, ACM TOG 2025) is a full-reference metric that models visual masking: an error hidden in texture counts for less than the same error on a flat surface. Where SSIM is a formula, MILO is a small convolutional network, 44,930 parameters, whose trained weights ship inside this crate. Given a reference and a distorted image it predicts a per-pixel visibility mask, weights the absolute error by it, and pools that into one number.

This is a port of the authors’ PyTorch implementation, held to its outputs by tests/parity.rs. Same pyramid, same network, same pooling; only the execution differs: a line-buffer pipeline so memory stays flat whatever the image size, lane-generic SIMD kernels picked at compile time, and row bands spread over threads with results that don’t depend on the thread count.

use blazediff_milo::{milo, MiloOptions, Rgba8};

let outcome = milo(
    Rgba8::new(&rgba1, width, height),
    Rgba8::new(&rgba2, width, height),
    &MiloOptions::default(),
)?;
println!("raw error {:.6}, MOS {:.3}", outcome.raw_error, outcome.mos);

Two numbers come out. raw_error is the metric proper: the masked mean absolute error, exactly zero for identical images and growing with visible damage (the fixtures in this repo land between 0.0002 and 0.01). mos maps it onto KADID-10k’s five-point mean opinion score through a learned calibration; note the calibration tops out around 4.35 rather than 5 for identical input, so threshold on raw_error, not mos.

Structs

MiloOptions

Knobs for a comparison.

MiloOutcome

A score and the maps it was pooled from.

Rgba8

A borrowed RGBA8 image, 4 bytes per pixel, row-major. Alpha is ignored, as the reference converts to RGB.

Enums

MiloError

Why MILO could not produce a score.

Constants

MIN_SIDE

The smallest side the reference implementation accepts: three halvings must leave at least two pixels for the seed mask.

Functions

milo

Compare distorted against reference.

render_map

Paint map, one value per pixel of a width x height image, over an RGBA8 output buffer as an opaque grayscale image, bright where the perceived error is high. Values are clamped to 0..=1, the range the reference’s own colormap render assumes.