Fixing AI editing artifacts starts with admitting they aren't all the same problem

When you feed a generated image back into an AI editor repeatedly, the outputs accumulate structured noise: grid patterns, honeycomb textures, granular surfaces that look fine in thumbnails but are conspicuous at native resolution. Mi-Ripple is a restoration workflow built around one diagnostic insight — that these artifacts are not all the same kind of problem, and treating them as if they were causes more damage than it fixes.

The core distinction is between lattice artifacts and granular texture. Lattice artifacts produce isolated peaks in the frequency spectrum and can be removed with selective notch filtering that touches a tiny fraction of frequency bins — in one example, just 0.11% — leaving the rest of the image essentially intact. Across fourteen notch-only runs, whole-image residual standard deviation stays between 0.08 and 0.44 CIELAB lightness units, with high-frequency retention above 98% in the portrait cases. Granular texture, by contrast, has no isolated spectral peak to target. It is entangled with legitimate scene content — dense foliage, hair, material surfaces — and cannot be filtered without erasing real detail. For those cases, the workflow either applies structure-aware masked suppression or cleans the reference image before regeneration.

The reference-cleaning route is the more consequential finding. Rather than trying to fix a degraded output directly, Mi-Ripple prepares a cleaner version of the image to feed into the next generation call. In a paired example, this reduced output debris density from roughly 1,842 to 1,020 components per megapixel — a 45% drop. Scale-index coverage in a moss scene fell from 35.3% to 15.8% after cleaning the reference. The tradeoff is explicit: regeneration can alter scene details and does not recover a known pixel-level ground truth. The workflow flags this and requires human review rather than presenting regenerated outputs as aligned corrections.

The characterization results are honest about configuration dependence. Channel B showed lattice artifacts in all 43 tested outputs; Channel A was negative in 20 outputs at one resolution and positive in 6 at another. A vendor name alone predicts nothing. The Banana100 external analysis of seven model families found that some maintain stable characteristic periods across ten editing steps while others show rising artifact strength without a stable period — persistence and accumulation are separate phenomena worth tracking independently.

Prompt-based mitigation doesn't hold up. Paired tests of foliage texture constraints, plant noun additions, and negation phrases produced mean differences near zero with sign-test p-values around 0.73. The workflow's practical recommendation is structural: use a star-shaped editing pattern where each generation draws from an approved reference rather than the previous output, breaking the accumulation chain at its source.

The thresholds throughout are study-specific and the authors say so plainly. Single-draw regeneration comparisons demonstrate routes, not population-average gains. That candor is the paper's most useful quality.