image-to-image-character-generation

name	image-to-image-character-generation
description	Use this skill when an agent needs to generate consistent character images via kie.ai grok-imagine image-to-image from Supabase S3 or Google Drive to produce daily social media images of the same mascot/persona in different scenes, ensuring character identity stays consistent across posts. Includes reference selection (face anchors + pose refs by keyword score), dynamic prompt composition (scene + attire + mood + angle, anti-repeat deque), kie.ai API payload building, strength calibration, fallback model, and post-generation workflow. Supports both Supabase storage URLs and Google Drive file IDs. Authorized use only — never hardcode credentials.
version	1.0.0
author	Hermes Agent
license	MIT
compatibility	Cross-platform for Claude Code, OpenAI Codex, GitHub Copilot, Cursor, Windsurf, Gemini CLI, OpenClaw, Hermes Agent.
tags	["image-generation","character-consistency","kie-ai","image-to-image","social-media","brand"]
platforms	["claude-code","codex","cursor","gemini-cli","openclaw","hermes-agent"]
geo	{"primary_workflows":["daily_social_image_generation","character_consistent_image_generation","brand_mascot_imaging"],"target_roles":["content_creator","social_media_manager","brand_marketer"],"complexity_level":"intermediate"}

Image-to-Image Character Generation

Generate consistent images of the same character/mascot across different scenes using kie.ai's grok-imagine/image-to-image model, conditioned on reference photos from Supabase S3 storage or Google Drive. This is image-to-image, not text-to-image — the model generates from reference images, not from a prompt alone.

When to use

• You need daily social media images of the same recurring character/persona
• The character must be recognisable as "the same person" across posts
• You have a reference image set of the character in various poses/angles
• You want to generate image variants automatically for review + approval
• Do NOT use for one-off text-to-image generation (use a standard image gen skill)

Storage backends (choose one)

Supabase S3

Reference images are stored at a Supabase storage URL. Naming convention: {bucket_url}/{character}_reference_{nn}_{pose}.jpg

Example: https://dbbppsbtpsahdkmevxgm.supabase.co/storage/v1/object/public/instagram_images/isabelle_reference_01.jpg

Google Drive

Reference images are stored on Google Drive with file IDs. URL format: https://www.googleapis.com/drive/v3/files/{FILE_ID}?alt=media&key={GDRIVE_API_KEY}

Reference selection system

Face anchors (MANDATORY — always the first 2 refs)

The front-facing + 3/4 angle reference images. These keep the character's identity consistent across all generations. Never randomize or drop these.

Pose references (remaining 3-6 slots)

Each reference image has keyword tags (e.g. "sitting", "walking", "hand-near-face"). Score each pose ref by counting how many tags overlap with the scene description's keywords. Randomly sample from the top 14 scored refs to fill the remaining slots.

Provider cap

• grok-imagine/image-to-image: max 5 refs total via image_urls key
• flux-2/flex-image-to-image: max 8 refs total via input_urls key (fallback)

Character anchor

Define one canonical sentence describing the character's fixed appearance:

{Character name}, a {role} with {hair}, {skin}, {eyes}, {styling}

Prepend this to every prompt. If the anchor is missing from a prompt, add it.

Dynamic prompt composition

Build each prompt from combinatorial pools to avoid repetition:

Components