Image pipeline spec template

Reference template for the image pipeline. Every encode strategy and every generate option is documented inline. Copy this into ~/.config/mm/pipelines/image/{mode}.yaml (replacing {mode} with fast or accurate) and edit only the fields you want to override — omitted keys fall back to the built-in defaults.

Source: python/mm/pipelines/image/spec.yaml.template

# Image pipeline spec — all available strategies and their options.
#
# Copy this file to ~/.config/mm/pipelines/image/{mode}.yaml
# and customise the values you need.  Omitted strategy_opts fall back
# to their defaults.

kind: image
mode: fast        # fast | accurate

# ── encode ──────────────────────────────────────────────────────────

encode:

  # ── resize (default) ────────────────────────────────────────────
  # Fits the image into a bounding box while preserving aspect ratio
  # and EXIF orientation.  Uses the Rust fast-path when available.
  strategy: resize
  strategy_opts:
    max_width: 1024           # max dimension in pixels (width or height)

  # ── tile ────────────────────────────────────────────────────────
  # Resized overview + grid of tile crops in a single Message.
  # Good for high-resolution images where fine detail matters.
  # strategy: tile
  # strategy_opts:
  #   max_width: 1024         # tile dimension and overview bounding box

  # Custom post-processing transform (optional).
  # pyfunc: null

# ── generate ────────────────────────────────────────────────────────

generate:
  prompt: >-
    Describe this image in 10 words or less.
    Then list exactly 5 keyword tags.

    ## FORMAT
    --description--

    Tags: Up to 5 keyword tags

    Example:
    A cat sitting on a windowsill.

    Tags: cat, windowsill, sunlight, pet, animal
  max_tokens: 256
  # model: null              # pin a specific LLM model (overrides profile)
  # temperature: null        # sampling temperature (null = model default)
  # json_mode: false         # request JSON-formatted response
  # think: false             # enable extended thinking
  # reasoning_effort: none   # none | low | medium | high
  # extra_body: {}           # provider-specific pass-through kwargs