Betterlife Image Generation

BetterLife Daily Social Image Generation (kie.ai image-to-image)

Generate the daily BetterLife social media image using kie.ai's grok-imagine/image-to-image model, conditioned on Mira reference photos from Google Drive. This is image-to-image, not text-to-image — the model generates from reference images, not from a prompt alone.

Architecture

Compose Prompt (scene + attire + mood + angle)
  → Select 8 Mira reference images (5 for grok) from 20 GDrive photos
    → Send to kie.ai grok-imagine/image-to-image (strength 0.85)
      → Fallback: flux-2/flex-image-to-image on failure
        → Save to ~/workspace/tool-image-generation/

Credentials

• KIE_API_KEY — for the kie.ai API
• Read from credential source by label — never hardcode.

Reference images (Supabase S3 storage)

Reference images are stored in Supabase storage at: https://dbbppsbtpsahdkmevxgm.supabase.co/storage/v1/object/public/instagram_images/

File naming: {character}_reference_{nn}_{pose}.jpg

Reference selection

• Face anchors (MANDATORY): front + 3/4 angle references. Always in the first 2 slots.
• Pose refs (3-6 remaining): scored by tag overlap with scene keywords, randomly sampled from top 14. Grok caps at 5 total refs, flux at 8.

Character anchor

Mira, a wellness coach with natural beauty, shoulder-length dark hair,
warm olive skin, kind eyes, minimal makeup

Prompt composition

Format: {anchor}, {scene}. {attire}. {detail}. {mood} {accent}. Premium editorial wellness photography, {angle}, no text, no watermarks.

Components picked combinatorially to avoid repeats: 36 scenes, 8 attires, 10 details, 6 moods, 8 angles, 4 accents.

kie.ai API call

POST https://api.kie.ai/api/v1/jobs/createTask
{
  "model": "grok-imagine/image-to-image",
  "input": {
    "prompt": "<composed prompt>",
    "image_urls": ["<5 ref URLs>"],
    "strength": 0.85,
    "aspect_ratio": "1:1", "resolution": "1K", "output_format": "jpg"
  }
}

Poll GET /api/v1/jobs/recordInfo?taskId={id} every 2s until state=="success".

Fallback: flux-2/flex-image-to-image (uses input_urls, allows 8 refs).

Key parameters

• strength 0.85 — higher = more Mira-likeness, lower = more prompt-driven scene
• Known limitation: at strength 0.85 with the same 20 refs, output is structurally dominated by the refs. Scene words shift background but not composition. For genuine variety, lower strength or add shot-size/framing contrast per variant.

Post-generation

1. Image saved to ~/workspace/tool-image-generation/betterlife-{theme}-{date}-v{n}.jpg
2. Review JSON at /tmp/bl-social-review.json
3. Posted via bl_social_post.py on approval