Skill v1.0.1

currentAutomated scan100/100

eyecrackcodes/roli/docs

+2 new

──Details

PublishedJune 5, 2026 at 02:54 PM

Content Hashsha256:9e0977df563c7077...

Git SHAc81fc3ed47db

Bump Typepatch

Compare with v1.0.0

──Files

Files (1 file, 4.7 KB)

SKILL.md4.7 KBactive

SKILL.md · 127 lines · 4.7 KB

version: "1.0.1" name: Nano Banana 2 Image Generation Master description: A formalized skill for generating hyper-realistic, highly-controlled images using the Nano Banana 2 (Gemini 3.1 Flash) model through parameterized JSON prompting via the Kie.ai API.

Nano Banana 2 Image Generation Master

Goal

Provide a standardized, highly controlled method for generating images using Nano Banana 2 (Gemini 3.1 Flash via Kie.ai). Strictly enforcing a structured JSON parameter schema neutralizes native model biases (over-smoothing, dataset-averaging, "plastic" AI styling) and ensures raw, unretouched, hyper-realistic outputs.

Prerequisites

KIE_AI_API_KEY set in .env at project root
Python installed
Scripts at scripts/generate_kie.py and scripts/get_kie_image.py

Workflow

Step 1 — Clarify the request

Ask the user for:

Subject: Who or what (if a person, get physical description, outfit, pose)
Environment: Location, background, lighting
Output: Single image or collage, aspect ratio (e.g. 4:5, 16:9, 3:4)
Style: Candid/realistic, product shot, documentary, etc.

Step 2 — Build the Dense Narrative Prompt (Paradigm 2)

Construct a JSON file saved to /prompts/ using the Dense Narrative Format:

json

{
  "prompt": "Ultra-descriptive narrative. Include: exact camera specs (85mm lens, f/2.0, ISO 200), explicit imperfections (visible pores, mild redness, subtle freckles), lighting behavior (direct on-camera flash creating sharp highlights on skin), direct negative commands inside the prompt (Do not beautify or alter facial features. No makeup styling.).",
  "negative_prompt": "no plastic skin, no CGI, no skin smoothing, no anatomy normalization, no body proportion averaging, no beautification filters, no studio lighting, no airbrushed texture, no stylized realism, no editorial fashion proportions, no high-end retouching",
  "image_input": [],
  "api_parameters": {
    "google_search": false,
    "resolution": "2K",
    "output_format": "png",
    "aspect_ratio": "4:5"
  },
  "settings": {
    "resolution": "2K",
    "style": "documentary realism",
    "lighting": "natural ambient",
    "camera_angle": "eye level",
    "depth_of_field": "shallow, subject sharp",
    "quality": "high detail, unretouched"
  }
}

Step 3 — Execute via Python

powershell

python scripts/generate_kie.py prompts/your_prompt.json images/output.png "4:5"

Step 4 — Deliver result

Tell the user where the image was saved and offer to iterate.

Core Schema (Paradigm 1 — Structured JSON for complex builds)

Use this when you need maximum control over multi-panel collages or product shots:

json

{
  "task": "string - e.g., 'single_macro_portrait'",
  "output": {
    "type": "single_image or 4-panel_collage",
    "layout": "1x1 or 2x2_grid",
    "aspect_ratio": "4:5",
    "resolution": "ultra_high",
    "camera_style": "smartphone_front_camera or professional_dslr"
  },
  "image_quality_simulation": {
    "sharpness": "tack_sharp",
    "noise": "unfiltered_sensor_grain",
    "compression_artifacts": false,
    "dynamic_range": "hdr_capable",
    "white_balance": "slightly_warm",
    "lens_imperfections": ["subtle chromatic aberration", "minor lens distortion"]
  },
  "subject": {
    "type": "human_portrait",
    "human_details": {
      "identity": "string",
      "appearance": "extremely specific — visible pores, mild redness, asymmetrical features",
      "outfit": "string"
    }
  },
  "environment": {
    "location": "string",
    "background": "string",
    "lighting": {
      "type": "natural or overhead",
      "quality": "uneven, realistic, non-studio"
    }
  },
  "explicit_restrictions": {
    "no_professional_retouching": true,
    "no_studio_lighting": true,
    "no_ai_beauty_filters": true,
    "no_high_end_camera_look": false
  },
  "negative_prompt": {
    "forbidden_elements": [
      "anatomy normalization", "body proportion averaging", "dataset-average anatomy",
      "beautification filters", "skin smoothing", "plastic skin", "airbrushed texture",
      "stylized realism", "editorial fashion proportions", "depth flattening",
      "mirror selfies", "reflections"
    ]
  }
}

Best Practices

Camera Mathematics — Always specify focal length, aperture, ISO: 85mm, f/2.0, ISO 200
Explicit Imperfections — Name the flaws: mild redness, subtle freckles, peach fuzz
Direct Commands — Embed negatives in the positive prompt: Do not beautify. No makeup styling.
Lighting Behavior — Describe what the light does, not just its type
Noise Trap — Keep ISO below 800; use physical imperfections for realism, not heavy grain
Negative Stack — Always include the full forbidden_elements list

Cost

~$0.04–$0.09 per image via Kie.ai API

← v1.0.0 All versions