mikeparcewski/alter

Image Alteration

AI-powered modification of existing images through two primary modes: image-to-image editing (global changes) and mask-based inpainting (local changes).

When To Use This Skill

• Modifying style, lighting, or mood of an existing image
• Replacing or adding specific elements within an image
• Iterating on generated images to reach desired output
• Making targeted local edits while preserving the rest

Modification Modes

Image-to-Image (Global Edits)

Changes the overall style or adds global elements while preserving core composition.

sh "${CLAUDE_PLUGIN_ROOT}/scripts/_python.sh" "${CLAUDE_PLUGIN_ROOT}/skills/imagery/scripts/provider.py" edit \
  --image ./source.png \
  --prompt "Same scene but with a dramatic sunset sky" \
  --output ./v2.png

Key control: The strength parameter determines how much the model deviates from the original.

• --strength 0.2-0.3 — Subtle changes (color grading, minor adjustments)
• --strength 0.4-0.6 — Moderate changes (style transfer, lighting shifts)
• --strength 0.7-0.9 — Dramatic changes (major style overhaul)

Inpainting (Local Edits)

Precise edits to specific regions using a binary mask (white = edit area, black = preserve).

sh "${CLAUDE_PLUGIN_ROOT}/scripts/_python.sh" "${CLAUDE_PLUGIN_ROOT}/skills/imagery/scripts/provider.py" inpaint \
  --image ./source.png \
  --mask ./mask.png \
  --prompt "Replace with floating lanterns" \
  --output ./v2_inpaint.png

Best practices for masks:

• Clean edges — avoid feathered or anti-aliased mask boundaries
• Include enough context area around the edit target
• For complex shapes, create the mask programmatically or use a drawing tool

Iterative Refinement Loop

When the output is close but not quite right:

1. DIAGNOSE  — What specifically is wrong? (color, composition, detail, artifact)
2. ADJUST    — Modify the right parameter for the issue:
   - Wrong elements → refine prompt, add negative prompt
   - Wrong style → adjust strength parameter
   - Wrong detail → fix seed, tweak guidance scale
   - Local issue → switch to inpainting with a targeted mask
3. REGENERATE — Run with adjusted parameters
4. REVIEW     — Use the review sub-skill to validate

Refinement Strategies

| Problem | Solution | |---------|----------| | Missing element | Increase weight in prompt with specific adjectives | | Unwanted artifact | Add to --negative-prompt | | Good composition, bad detail | Fix --seed, adjust --guidance-scale | | Almost perfect, one area wrong | Switch to inpainting for that region | | Too different from original | Lower --strength value | | Not different enough | Raise --strength value |

Final Polish

Use upscaling as the last step to bring a draft to production quality:

# Upscaling (cstudio only — other providers may not support this)
sh "${CLAUDE_PLUGIN_ROOT}/scripts/_python.sh" "${CLAUDE_PLUGIN_ROOT}/skills/imagery/scripts/provider.py" generate \
  --provider cstudio \
  --prompt "upscale" \
  --output ./final_hires.png

Detailed References

• Editing techniques, provider commands, and refinement strategies: refs/editing-reference.md

Integration With Other Sub-Skills

• Use review before altering to understand what needs to change
• Use review after altering to validate the changes
• If starting from scratch is better than editing, hand off to create
• Alteration is the "Execute" and "Refine" steps of the Analyze-Execute-Review-Refine loop