image-to-image-character-generation

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

1. Scene (~36 options) — e.g. "Mira walking along a misty lakeside path at sunrise"
2. Attire (~8 options) — the character's outfit for this scene
3. Detail (~10 options) — a small environmental detail (steam from a mug, pages fluttering)
4. Mood (~6 options) — lighting description (golden hour, overcast morning, blue hour)
5. Angle (~8 options) — camera framing (medium shot, close-up, full body, low angle)
6. Accent (~4 options) — colour palette (cool blue-grey, muted mauve, soft sand)

Format

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

Anti-repeat deque

Track the last 14 scenes used. Never reuse a scene until it falls out of the window.

kie.ai API call

Primary model

POST https://api.kie.ai/api/v1/jobs/createTask
{
  "model": "grok-imagine/image-to-image",
  "input": {
    "prompt": "<composed prompt>",
    "image_urls": ["<ref_url_1>", "<ref_url_2>", "<ref_url_3>", "<ref_url_4>", "<ref_url_5>"],
    "strength": 0.85,
    "aspect_ratio": "1:1",
    "resolution": "1K",
    "output_format": "jpg"
  }
}

Poll for result

GET https://api.kie.ai/api/v1/jobs/recordInfo?taskId={taskId}

Poll every 2 seconds until state == "success". Download result from resultUrls[0].

Fallback model

"model": "flux-2/flex-image-to-image"

Uses input_urls instead of image_urls, allows 8 refs.

Key parameters

• strength 0.85: Higher = more character-likeness (refs dominate). Lower = more

prompt-driven scene variation. If images look the same across days, lower to 0.6-0.7.

• aspect_ratio "1:1": Square format for social media. For video: "9:16".
• resolution "1K": Standard. For higher quality: "2K".
• grok caps at 5 refs with image_urls. flux caps at 8 with input_urls.
• The ref key MUST match the model or the createTask 500s:

ref_key = "image_urls" if "grok" in model else "input_urls"

Post-generation workflow

1. Image saved to a workspace directory ({workspace}/tool-image-generation/)
2. Variants + captions + platforms written to a review JSON
3. Delivered to the user for approval (Telegram, review UI, etc.)
4. On approval: posted to configured social platforms via platform-specific scripts

Character identity verification

Before declaring success, verify the character's identity is consistent:

• The face anchors are present in every reference selection (not just the first run)
• A test generation returns the character with recognisable face/features
• The character anchor text is identical in every prompt (no spelling drift)

Pitfalls

• Provider caps differ: grok-imagine accepts image_urls (max 5), flux-2/flex accepts

input_urls (max 8). Using the wrong key name causes a 500 error. Always model-aware.

• Strength 0.85 with same refs = same-looking output: if the image looks structurally

identical across days despite different scene prompts, the refs are dominating. Lower strength or add explicit shot-size framing contrast per variant.

• No-seed engines: grok-imagine has no seed parameter. Same prompt = same image.

Write genuinely different scene text per variant, not just re-run the same prompt.

• Credentials never hardcoded: read the kie.ai API key and storage credentials from a

credential source by label — never commit them to the skill or scripts.

• Face anchors must stay first: if face anchors are truncated by the 5-ref cap, identity

drifts. Always put the 2 face anchors at the start of the ref list.

Verification

• A test call returns the character with recognisable identity (same face, same features)
• The scene clearly differs from yesterday's (different setting, lighting, framing)
• The image is saved to the workspace and reviewable by the user before any posting