| 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. |
| type | capability-uplift |
Nano Banana 2 Image Generation (Capability Uplift)
This skill provides a Capability Uplift by overriding the model's native image generation biases (over-smoothing, dataset-averaging, "plastic" AI styling) using a strict parameterized JSON prompting schema.
🎯 The Objective
Neutralize default AI aesthetic biases and ensure raw, unretouched, hyper-realistic outputs via the generate_image or external Kie.ai pipeline.
🛠️ Execution Protocol (Dense Narrative Format)
When translating a user's image request into a prompt, construct a dense JSON string that includes specific camera physics, explicit physical flaws, lighting behavior, and strong negative commands.
{
"prompt": "A dense, ultra-descriptive narrative. Use specific camera math (85mm lens, f/1.8, ISO 200), explicit flaws (visible pores, mild redness, subtle freckles), lighting behavior (direct on-camera flash), and direct negative commands.",
"negative_prompt": "anatomy normalization, body proportion averaging, wide-angle distortion, beautification filters, skin smoothing, plastic skin",
"api_parameters": {
"resolution": "4K",
"output_format": "webp"
}
}
Best Practices:
- Camera Math: Define focal length, aperture, and ISO (e.g.,
85mm lens, f/2.0, ISO 200).
- Explicit Imperfections: Dictate flaws (e.g.,
mild redness, subtle scratches on aluminum).
- Negative Constraints: Always include the massive list of 'AI style' blockers in the negative space.
🚨 Trigger Tuning & Evals
- Trigger Check: ONLY invoke this skill when the user explicitly requests "highly detailed", "hyper realistic", "photography", or "unfiltered" imagery. Do not use for generic web assets or vector logos.
- Eval: Pass Rate: The payload must validly pass to the generation endpoint without JSON parsing errors.
- Eval: Outgrowth Check: If the LFM2-24B model's base image generation becomes naturally hyper-realistic without needing these extensive JSON guardrails, deprecate this skill.
