Image Generation

Image Generation Mega Skill

Models Supported

Default: Nano Banana Pro via Gemini API — this is what we use in Claude Code.

Three Modes

Mode 1: GENERATE (From Scratch)

Create an image from a text description. No reference needed.

When to use: Client needs a new image for a scene that doesn't exist yet.

Prompt Formula (Nano Banana Pro):

[Identity Anchor] + [Scene Description] + [Wardrobe] + [Action/Pose] +
[Lighting] + [Background] + [Framing] + [Style/Mood]

The Six Variables (Higgsfield Official):

1. Subject — Who/what specifically (not generic)
2. Composition — Camera angle, lens, framing
3. Action — What's happening (movement, gesture, expression)
4. Location — Where, with atmospheric detail
5. Style — Art direction (cinematic, editorial, raw UGC, etc.)
6. Negative guidance — What to exclude (blur, artifacts, distortion)

Mode 2: REFERENCE (Match a Screenshot)

User drops a reference image. Extract its DNA and recreate for a different subject.

When to use: "I want this same vibe/composition/style but for [different person/scene]"

Process:

1. Read the reference image visually
2. Extract: composition, camera angle, lighting setup, color palette, mood, wardrobe style, background, framing
3. Rebuild as a prompt with the new subject inserted
4. Ask user: "Keep everything or change something?"
5. Generate

Extraction Template:

### Mode 3: STYLE-LOCK (Saved Client Style)

Use a pre-configured client style. Just describe the scene — the style is already defined.

**When to use:** "Give me another Claudia podcast shot" or "Ryan style thumbnail"

**Process:**
1. Load client config from `configs/[client].md`
2. User only needs to describe: scene, pose, wardrobe change (if any)
3. Config provides: lighting, background, camera, mood, framing, identity anchor
4. Merge and generate

---

### Core Pattern
```python
import os
from google import genai
from google.genai import types
from PIL import Image

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

# Text-to-image
response = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents=["Your prompt here"],
    config=types.GenerateContentConfig(
        response_modalities=['TEXT', 'IMAGE'],
        image_config=types.ImageConfig(
            aspect_ratio="9:16",  # Vertical for reels
            image_size="2K"       # Quality
        ),
    ),
)

for part in response.parts:
    if part.text:
        print(part.text)
    elif part.inline_data:
        image = part.as_image()
        image.save("output.jpg")  # ALWAYS .jpg — Gemini returns JPEG

Multi-Turn Refinement (Chat Mode)

chat = client.chats.create(
    model="gemini-3-pro-image-preview",
    config=types.GenerateContentConfig(
        response_modalities=['TEXT', 'IMAGE'],
        image_config=types.ImageConfig(
            aspect_ratio="1:1",
            image_size="2K"
        ),
    )
)

# First generation
response = chat.send_message(["Generate...", ref_img])
# Save image...

# Refine
response = chat.send_message("Make the lighting warmer and add more bokeh")
# Save refined image...

CRITICAL: File Format

Gemini returns JPEG. Always save as .jpg. Saving as .png creates a JPEG with a PNG extension and causes downstream errors.

# CORRECT
image.save("output.jpg")

# WRONG
image.save("output.png")  # JPEG data with PNG extension!

# If you NEED PNG:
image.save("output.png", format="PNG")

Prompt Engineering (Nano Banana Pro Specific)

Seed Locking

Reuse successful seeds for series consistency. When you get a good result, note the conversation context — multi-turn chat preserves the style.

Identity Anchor Pattern (For Character Consistency)

When generating images of a known character (Claudia, Ryan, etc.), always start with:

"Generate a new photorealistic image of THIS EXACT [PERSON] from the reference photos.
[He/She] must have the SAME face, same [hair description], same facial features,
same bone structure. NEW SCENE: [scene description]"

This is the proven pattern from Claudia's generations. Upload the CANON reference image alongside this prompt.

Client Style Configs

Each client gets a config file at configs/[name].md that defines their locked style.

Tattoo Handling (For Characters With Tattoos)

Soul ID (Best Method)

Upload 15-20 photos with tattoos clearly visible. Soul ID bakes them into the identity. Placement stays consistent. Fine detail varies but presence is maintained.

Higgsfield Prompting Guide (When Using Web UI)

Nano Banana Pro (via Higgsfield)

Same as Gemini API prompting but through the Higgsfield web interface. Select "Nano Banana Pro" model in the image creation dashboard.

Soul Model (Higgsfield's Own)

Better for fashion-grade, editorial aesthetic. Select "Higgsfield Soul" in model dropdown.

Full Specification

Complete details, decision trees, protocols, and implementation specs: references/full-details.md

Change Log

• [2026-03-13]: Context Diet — split 429 lines into routing card (195 lines) + references/full-details.md (273 lines). Zero content deleted. By skill-diet.py.

• 2026-02-06: Created. Source: Luis Carrillo session. Consolidates Nano Banana Pro (Gemini API), Higgsfield Soul/Soul ID, and reference image workflows into one skill. Client configs for Claudia (proven podcast style). Tattoo handling documented. Mandatory video script image+video gen prompts for clone era.