Image Edit

Workflow

Phase 1: Input Validation

1. Verify the source image exists and is accessible
2. Check format (supported: jpg, jpeg, png, webp, gif, avif)
3. Note original dimensions for comparison
4. For local files: prepare base64 encoding or upload to FAL storage

Read references/image-to-image.md for handling local files.

Phase 2: Edit Intent

Determine what the user wants to change:

• Style transfer: Change the visual style (watercolor, vintage, anime, etc.)
• Background change: Replace or modify the background
• Subject modification: Change attributes of the subject (hair color, clothing, etc.)
• Text changes: Add, modify, or remove text in the image
• Color adjustments: Change color palette, temperature, saturation
• Composition changes: Reframe, crop, extend
• Variations: Generate similar images with variations

If unclear, ask the user what specific changes they want.

Phase 3: Model Selection

Read references/model-catalog.md for full details. Select based on edit type:

• Style transfer → Grok Imagine (cinematic) or Seedream v4.5 (general)
• Text editing → Nano Banana Pro (best text rendering)
• Fast iteration → FLUX.2 Flash
• Multi-reference compositing → Seedream v4.5 (up to 10 reference images)
• Fine control with negative prompt → FLUX.2 [klein]

Phase 4: Prompt + Parameter Setup

1. Spawn the prompt-architect agent with the edit description and target model
2. Set guidance_scale based on desired change magnitude:
   • 2-4: Subtle tweaks (color adjustments, minor style change)
   • 5-7: Moderate changes (style transfer, background swap)
   • 7-9: Significant transformation (major restyle)
   • 10+: Near-complete regeneration (use with caution — may cause artifacts)

Read references/image-to-image.md for the full strength/guidance tuning guide. Read references/prompt-engineering.md for edit-specific prompt strategies.

Phase 5: Generation

Execute the FAL API call:

./scripts/fal-submit.sh "{endpoint}" \
  '{"prompt":"[edit description]","image_urls":["[source_url]"],"guidance_scale":[value]}' \
  "./output/image-gen/edit-{model}-{timestamp}.png"

Read references/fal-auth.md to verify FAL_KEY is set.

Phase 6: Compare & Iterate

1. Present the original and edited images side by side
2. Describe what changed and what was preserved
3. Offer adjustments:
   • "Want more/less change? I can adjust the guidance scale."
   • "Want to try a different model?"
   • "Should I refine the edit prompt?"

Validation Gates

• Source image verified and accessible
• Edit intent is clear
• Model selected for the edit type
• Guidance scale set appropriately
• Edit prompt focuses on the desired change
• Image generated and downloaded
• Before/after comparison presented
• User satisfied or iteration planned

Important Notes

• Edit prompts should describe the change, not the entire scene
• Start with moderate guidance (5-7) and adjust based on results
• For Seedream, keep guidance_scale under 10 to avoid artifacts
• Always present original vs edited for comparison
• Save edited images with edit- prefix in the filename