Paper Explained: Mi-Ripple — Diagnose First, Then Repair the Ripple That Iterative AI Editing Leaves Behind

Feed an AI-edited image back in as the reference for the next edit, and grid-like and grainy texture — digital ripple — starts to build up. Mi-Ripple first tells apart a lattice you can cut out in the frequency domain from grain tangled up with the picture, then picks either a notch filter or regeneration from a cleaned reference.

Mi-Ripple: Restoring Images Degraded by Iterative AI Editing

The short version

This post covers "Mi-Ripple: Restoring Images Degraded by Iterative AI Editing" (arXiv:2609.11317, Jiayin Chen et al., published 2026-09-10). It proposes a restoration workflow for the damage that piles up when you edit an image with AI over and over: diagnose which kind of artifact you have, treat it with a matching tool, then check numerically that you haven't wrecked the picture.

What the abstract says: iterative reference-conditioned editing can introduce grid-like and granular textures, often called digital ripple. Mi-Ripple separates periodic lattice artifacts from granular texture that is entangled with image content, and combines selective spectral notching, structure-aware smoothing, and regeneration from a cleaned reference. When an artifact is isolated in the spectrum, low-distortion filtering suffices; when filtering would erase legitimate detail, the image is reconstructed instead. Across fourteen notch-only runs, whole-image residual standard deviation was 0.08–0.44 in CIELAB lightness units, and in one paired regeneration example, cleaning the reference cut output debris density by 45%.

An analogy: a copy of a copy

Photocopy a document, then photocopy the copy, and a few generations in you'll see a fine mesh or grain creeping in. AI editing does something similar: pass the last output in as the next reference and unwanted texture rides along with it. The paper notes the damage is easy to miss in thumbnails and only shows up at native resolution as grids, honeycomb patterns, or grainy surfaces (§1).

There are two kinds of ripple. A lattice repeats at a fixed spacing, so in the frequency domain it shows up as isolated points. Granular texture is grit mixed into fine content like moss or leaves; it has no clean period and is tangled with the scene. Nails hammered into a floor at even spacing can be pulled out; sweep the sand off a gravel path and the gravel goes with it. Hence: tell them apart first, then pick the tool (§3).

Step 1: Diagnose (§3.1)

A Fourier transform breaks an image into "how much of each stripe pattern, at each fineness" (see The FFT from Scratch). From a lightness ($L^*$) patch $P$ the paper builds:

In words: take the log amplitude of the patch after tapering its edges with a window $w$, and measure how far each component sticks out above $B(r)$, the median at the same radius $r$ (the same fineness). A lattice makes a few components spike. The whole-image detector keeps small components that rise more than 1.2 above a $21\times21$ local median, and calls a lattice when there are at least two and the largest excess is at least 2.5. Granular texture is measured in the spatial domain: the "scale index" is the share of 128-pixel tiles covered by fine, uniform, round blobs, with 6% or more graded as structured (Appendix A.2).

The widget below is about JPEG, but it builds the right intuition: regular damage comes and goes with specific frequency components. The paper itself notes that one photograph hit an anomaly of 3.12 from JPEG blocking, so spectral anomalies are not unique to generated images (§6.1).