Mi-Ripple: Restoring Images Degraded by Iterative AI Editing
Jiayin Chen, Yicheng Xu, Muting Wang
Miyang Technology (Shanghai) Co., Ltd. · Key Laboratory of System Software, Chinese Academy of Sciences · Institute of Software, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Shanghai Jiao Tong University · Tianjin University
Published September 10, 2026 · on the feed September 11, 2026 · ▲ 52 · arXiv 2609.11317
The abstract
Iterative reference-conditioned image editing can introduce grid-like and granular textures, commonly described as digital ripple. We present Mi-Ripple, a diagnosis-guided restoration workflow that suppresses this digital ripple while protecting image structure. Mi-Ripple separates periodic lattice artifacts from content-entangled granular texture, then combines selective spectral notching, structure-aware smoothing, and cleaned-reference regeneration. This separation enables low-distortion filtering when artifacts are spectrally isolated and visual reconstruction when filtering would erase legitimate detail. Across fourteen notch-only executions, whole-image residual standard deviation is 0.08--0.44 in CIELAB lightness units. In a paired regeneration example, reference cleaning reduces output debris density by 45\%. Mi-Ripple links measurable artifact reduction to visibly cleaner generated images, rather than optimizing a spectral score alone.
The Brief
1:54A short podcast summary of the paper. Written and voiced by AI.
Transcript
Mi-Ripple. Restoring Images Degraded by Iterative AI Editing. By Jiayin Chen and two others, from Miyang Technology, the Key Laboratory of System Software of the Chinese Academy of Sciences, and four other institutions. Published on the tenth of September. Its code repository is public and real: thirteen source files.
When an image is edited again and again with an AI editor, each pass conditioned on a reference, the result can pick up textures that were never there: grid-like patterns, and granular grain. This damage is commonly called digital ripple. The problem is that the ripple and the real image are mixed together, and a clean-up that is too aggressive removes both.
Mi-Ripple is a restoration workflow guided by diagnosis. It suppresses digital ripple while protecting the structure of the image. Its first step is to tell two kinds of damage apart: periodic lattice artifacts, and granular texture that is entangled with the content. The distinction matters because the two call for different treatment.
It then combines three tools. Selective spectral notching, structure-aware smoothing, and regeneration from a cleaned reference. The separation decides which tool fits. Where artifacts are isolated in the spectrum, filtering can remove them with low distortion. Where filtering would erase legitimate detail, the workflow reconstructs the image visually instead.
Across fourteen executions that used notching alone, the residual standard deviation over the whole image ranged from zero point zero eight to zero point four four, in CIELAB lightness units. In one paired example of regeneration, cleaning the reference reduced the density of debris in the output by forty-five percent.
The authors frame the goal in terms of what a person sees. Mi-Ripple ties measurable artifact reduction to generated images that look visibly cleaner, instead of optimizing a spectral score for its own sake. Diagnose first, and only then choose between filtering and rebuilding.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is damage you cannot remove without risking the picture itself.
You edit an image again and again with an AI editor, each time feeding the last output back in as the reference. At full size, the picture now shows grid patterns and a grainy, honeycomb-like surface.
The road ahead
- 1 · The problem ~7 min Repeated AI edits leave grids and grain in a picture, and one blanket clean-up cannot remove them without also destroying real detail.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Diagnose the damage first, then route it: notch out isolated grid peaks, and regenerate or carefully smooth grain that is tangled with content.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Separate the damage types, choose a treatment for each, and for entangled grain clean the reference and regenerate from it.
- 6 · The experiment ~6 min You’ll guess: In a paired regeneration example, how much did cleaning the reference first cut the density of debris in the output?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.