Mi Ripple

Restore digital-ripple artifacts

Use the MIYANG mi_ripple package to diagnose first, choose only a compatible treatment, inspect the visual evidence, and return the actual output files to the user.

Do not treat this as a generic denoiser. The workflow distinguishes:

1. isolated periodic lattice artifacts, which can be selectively notched;
2. granular artifacts in unstructured areas, which can receive masked reduction;
3. artifacts entangled with hair, foliage, fabric, stone, or other content, which require human review or optional cleaned-reference regeneration.

Non-negotiable rules

• Work on the original image at native resolution. Do not resize, recompress, or screenshot it before diagnosis.
• Run the deterministic local pipeline before proposing regeneration.
• Never run paid or content-changing regeneration without the user's explicit approval in the current conversation.
• Never ask the user to paste an API key into chat. Ask them to set MIYANG_API_KEY in their environment.
• Files containing _refclean are model references, not deliverable images.
• Do not claim an artifact is removed only because a numeric score decreased. Inspect the heat map and comparison boards.
• Preserve generated JSON sidecars and the XML trace with the output.
• A pipeline passed result means measured filtering damage stayed within the configured limits. It does not mean a person has accepted the image.

Locate or install the tool

First check whether the command is available:

mi-ripple --help

If this skill is being used from a checkout of the repository, install that checkout into an isolated environment:

python3 -m venv .venv
.venv/bin/python -m pip install -e .

Otherwise install from the canonical repository into an isolated environment:

python3 -m venv .mi-ripple-venv
.mi-ripple-venv/bin/python -m pip install \
  "git+https://github.com/miyang-ai/Mi-Ripple.git"

Use the corresponding environment's mi-ripple executable for subsequent steps. Do not modify the user's global Python environment.

Inputs

Obtain:

• the exact local path of one source image;
• an output directory that will not overwrite the source;
• optional user priorities such as preserving hair, paper texture, foliage, or color.

If the image path is missing or ambiguous, ask for it. Do not choose a recently used image on the user's behalf.

Supported image decoding is provided by Pillow. Prefer PNG for intermediate and final files.

Step 1: run the local pipeline

mi-ripple "/absolute/path/input.png" "/absolute/path/output"

This performs diagnosis, deterministic routing, safe local treatment when available, aligned verification, and provenance recording.

Read:

• <stem>_restored.json for the outcome and action sequence;
• <stem>_input_diag.json for measurements and selected windows;
• <stem>_input_diag_board.png for native-pixel diagnostic crops;
• <stem>_input_scaleheat.png for whole-frame flagged tiles;
• <stem>_verify_board.png when filtering was applied;
• <stem>_restored.png when a final local result was delivered.

Open the generated boards with the agent's image-reading capability. Do not interpret JSON without checking the images.

Step 2: handle the outcome

delivered

Inspect the source, final image, heat map, and verification board.

Check specifically:

• hair remains continuous rather than becoming smooth blocks;
• foliage, flowers, gravel, fabric, stone, and paper texture were not erased;
• flat gradients did not gain cloudy stains or ringing;
• no new repeated scales, honeycomb cells, woven meshes, or wave bands appeared;
• framing, dimensions, and color remain unchanged on local filtering routes.

If the visual check passes, give the user:

1. the final image path or attachment;
2. the comparison/verification board;
3. a short statement of the route taken;
4. any remaining suspected artifact.