eryet/.claude/skills/waifu2x

Upscale Image(s) with waifu2x

You are helping the user upscale one or more images using waifu2x-ncnn-vulkan. The input can be a single image file or a folder containing images.

Step 0: Check for default flag

Check if $ARGUMENTS contains -d or --default. If it does, enable quick mode: skip all prompts and use default settings (scale=2, noise=2, model=cunet, format=png, tile=0). Remove the -d/--default flag from the arguments before proceeding — any remaining text is treated as the image/folder path.

Step 1: Get the input path

Determine the input path using the following priority:

1. From $ARGUMENTS: If provided (after removing any flags), use that as the input path.
2. From a pasted image: If the user pasted an image in the conversation, look for the source path in the image metadata. Pasted images may include a source annotation like [Image: source: <path>] — extract the file path from that.
3. From the user's message: Check if the user's message contains a file or folder path.
4. Ask the user: If none of the above yielded a path, ask the user for the path to the image or folder they want to upscale.

Determine if path is a file or folder

After resolving the path, check whether it is a file or a directory:

• If it's a file: Verify it exists. Proceed with single-image mode (one image to process).
• If it's a directory: Use Glob to find all image files in that folder matching *.{png,jpg,jpeg,webp,bmp,tga} (non-recursive — only the top level of the folder). Skip any files that already have _waifu2x_ in their name to avoid re-processing previous outputs. If no images are found, tell the user and stop. Otherwise, proceed with batch mode (multiple images to process).

Step 2: Ask for upscaling options

If quick mode is enabled (from Step 0), skip this step entirely.

Otherwise, first ask the user whether they want to use default settings or customize. Use AskUserQuestion with a single question:

• Use defaults (Recommended): Immediately proceed with scale=2, noise=2, model=cunet, format=png, tile=0
• Customize settings: Let the user pick each option individually

If the user picks Use defaults, skip straight to Step 3 with the default values.

If the user picks Customize settings, present the following options using AskUserQuestion:

• Scale factor: 1, 2 (default), or 4
• Noise reduction level: -1 = none, 0, 1, 2 (default, medium), or 3
• Model: models-cunet (default, best quality), models-upconv_7_anime_style_art_rgb (fast, anime), or models-upconv_7_photo (fast, photos)
• Output format: png (default), jpg, or webp
• Tile size: 0 = auto (default), 32, 64, 128, 256, or 400 (lower values use less VRAM)

Step 3: Locate waifu2x executable

Check if a memory file exists at the project memory directory that stores the waifu2x executable path. Look for a memory file about the waifu2x path (e.g., reference_waifu2x_path.md).

• If a saved path is found, verify the executable still exists at that location. If it does, use it. If not, ask the user for the new path.
• If no saved path is found, try running waifu2x-ncnn-vulkan directly first. If that fails (command not found), ask the user where waifu2x-ncnn-vulkan.exe is located.
• Once the user provides a valid path, save it to a memory file (type: reference) so it can be reused in future conversations.

Step 4: Run waifu2x

Output naming

Construct each output filename as: <original-name>_waifu2x_<scale>x.<format>

For example: photo.png with scale 2 and png format becomes photo_waifu2x_2x.png

Place each output file in the same directory as its input file.

Single-image mode

Run the command once:

<waifu2x-path> -i "<input-path>" -o "<output-path>" -s <scale> -n <noise> -m "<model-path>" -f <format> -t <tile-size>

Batch mode (folder)

Process each image in the folder sequentially. For each image, run:

<waifu2x-path> -i "<input-path>" -o "<output-path>" -s <scale> -n <noise> -m "<model-path>" -f <format> -t <tile-size>

Print a progress line for each image as it completes (e.g., [1/5] Done: image1.png). If a single image fails, log the error and continue with the remaining images — do not stop the entire batch.

Use the resolved executable path from Step 3. When using a full path, also pass the model directory as a full path relative to the executable's folder.

Step 5: Report results

After all processing completes:

Single-image mode

• If it succeeded: tell the user the output file path and its file size (use ls -lh on the output)
• If it failed: show the error output and suggest possible fixes (e.g., waifu2x-ncnn-vulkan not installed, invalid model path, out of VRAM — suggest lowering tile size)

Batch mode

• Show a summary table with columns: filename, status (success/failed), output file size
• Report totals: how many succeeded, how many failed
• If any failed: list the errors and suggest fixes