Nano Banana Image Gen

---
name: nano-banana-image-gen
description: >-
  Generate images with Google's Nano Banana family (Gemini 2.5 Flash Image, Nano
  Banana 2, Nano Banana Pro) using character and style references. Use when the
  user mentions "nano banana", "gemini image", any "gemini-*-image*" model id,
  character consistency for image generation, style transfer with Gemini, or
  multi-reference image composition. NOT for Ideogram (use existing Ideogram
  scripts), Qwen on Apple Silicon (see ~/.claude/QWEN_IMAGE_GUIDE.md), generic
  prompt engineering, or non-image Gemini calls.
allowed-tools: Read,Write,Edit,Bash,Grep,Glob,WebFetch
argument-hint: '[scene description]'
metadata:
  category: AI & Machine Learning
  tags:
  - nano-banana
  - gemini
  - image-generation
  - character-reference
  - style-transfer
  pairs-with:
    - skill: image-generation-workflow-engine
      reason: For non-Gemini pipelines (FLUX, SD, ComfyUI) and multi-model orchestration.
    - skill: image-optimization-engineer
      reason: Convert Nano Banana PNG output to webp / size-optimized assets after generation.
---

# Nano Banana Image Gen

Generate editorial illustrations, marketing assets, and product mockups with Google's Gemini image-generation models, using character sheets and style templates as references.

## When to Use

✅ **Use for**:
- Generating with `gemini-2.5-flash-image`, `gemini-3.1-flash-image-preview`, `gemini-3-pro-image-preview`
- Character-consistent illustrations across multiple scenes (same person, new setting)
- Style transfer using a reference image as the visual template
- Multi-image composition ("character from img 1 wearing outfit from img 2 in environment from img 3")
- Iterative image editing with conversation-style refinement
- Anything mentioning "Nano Banana", "Nano Banana 2", or "Nano Banana Pro"

❌ **NOT for**:
- Ideogram V3 generation — use existing `next-app/scripts/gen-*.py` scripts and the `IDEOGRAM_API_KEY` pattern
- Qwen image generation on M4 Max — see `~/.claude/QWEN_IMAGE_GUIDE.md` (use v0.6.0)
- Generic prompt engineering tutorials
- Non-image Gemini API usage (text completion, embeddings, video)
- Image *classification* / object detection (use `computer-vision-pipeline`)

---

## Core Process

```mermaid
flowchart TD
  S0[Request] --> S1{Need character<br/>or style consistency?}
  S1 -->|No, simple scene| M25[Use gemini-2.5-flash-image]
  S1 -->|Yes| MPRO[Use gemini-3-pro-image-preview]
  M25 --> CFG
  MPRO --> CFG[Set image_config.aspect_ratio<br/>response_modalities=IMAGE]
  CFG --> REFS{How many<br/>references?}
  REFS -->|0| P1[Prompt only]
  REFS -->|1| P1
  REFS -->|2+| P2[Label each ref's role<br/>character / style / lighting / etc]
  P1 --> SCENE[Write scenario:<br/>Refs + Relationship + Scene]
  P2 --> SCENE
  SCENE --> POS[Use POSITIVE framing<br/>describe what SHOULD be there<br/>negatives do not work]
  POS --> CALL[POST to generateContent]
  CALL --> OUT{Image OK?}
  OUT -->|Drift / wrong aspect| FIX[See troubleshooting.md]
  OUT -->|Yes| DONE[Save + convert to webp]
  FIX --> CALL
```

### Step 1: Pick the right model

| Model ID | Marketing name | Status | Use when |
|---|---|---|---|
| `gemini-3-pro-image-preview` | **Nano Banana Pro** | Preview | **Default for character/style work**. Studio-quality, reliable text suppression, complex layouts. |
| `gemini-3.1-flash-image-preview` | Nano Banana 2 | Preview | High volume + extreme aspect ratios (1:4, 4:1, 1:8, 8:1). Up to 14 reference objects, 4 characters. |
| `gemini-2.5-flash-image` | Nano Banana | GA | Cheap/fast scratch generations. Simple non-character scenes. **Avoid for editorial illustration with strict style adherence — it drifts.** |

Full comparison: `references/model-selection.md`.

### Step 2: Build the request correctly

The aspect ratio MUST be set in the generation config. Putting "16:9" in the prompt does not work.

```python
config = {
    "responseModalities": ["IMAGE"],
    "imageConfig": {"aspectRatio": "16:9"},  # ← non-negotiable
}
```

Supported ratios: `1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9`. Nano Banana 2 adds `1:4, 4:1, 1:8, 8:1`.

Endpoint: `POST https://generativelanguage.googleapis.com/v1beta/models/{MODEL}:generateContent` with header `x-goog-api-key: $GEMINI_API_KEY`.

Reference images attach as inline-data parts in `contents[0].parts`:

```python
parts = [
    {"inlineData": {"mimeType": "image/png", "data": base64_of_char_sheet}},
    {"inlineData": {"mimeType": "image/png", "data": base64_of_style_ref}},
    {"text": prompt_with_role_labels},
]
```

A complete runnable example: `scripts/generate.py`.

### Step 3: Use the official prompt formula

```
[Reference images] + [Relationship instruction] + [New scenario]
```

The **relationship instruction** is the load-bearing piece — explicitly label what role each reference plays. Without role labels, the model partly absorbs and partly ignores the reference.

Worked example for an editorial hero illustration:

```
Reference image 1 is the CHARACTER SHEET — preserve the man's face,
build, hair, and teal hoodie exactly.
Reference image 2 is the STYLE TEMPLATE — match its graphic-novel rendering,
warm leather-and-amber palette, atmospheric lighting, and cinematic composition.

Generate a new scene where: [scene description].

Wide cinematic 16:9 framing, character occupies the left third.
Laptop screens display abstract glowing UI shapes with no readable typography,
indistinct soft-blurred interface.
```

Deep dive on the formula and more variants: `references/prompt-formula.md`.

### Step 4: Layer character consistency

Three techniques, in order of impact:

1. **Words + reference together** — describe identifying details in text *and* attach the reference. "Young Latino man, teal hoodie, dark curly hair" + image. Image alone drifts.
2. **Iterative re-feeding** — for multi-scene consistency, attach the previous generation in the next request alongside the original sheet. Identity tightens with each pass.
3. **Annotated references** — Nano Banana reads labels and captions inside reference images. A labeled character sheet outperforms a clean illustration.

Details + edge cases: `references/character-consistency.md`.

### Step 5: Style references are separate from character refs

For matching an existing visual style, attach **two** images with explicit roles:

- Ref 1: the character (face/build/clothing to preserve)
- Ref 2: a successful piece in the target style (rendering, palette, lighting, composition to match)

Single-reference style transfer drifts. Dual-reference with role labels is the documented pattern. See `references/style-references.md` for the full pattern and when to use 3+ refs.

### Step 6: Suppress unwanted text/labels with positive framing only

There is no negative prompt. Two things that work:

- **Positive framing**: describe what *should* occupy the space. Not "no text on the laptop screens" → "the laptop screens display abstract glowing UI shapes, indistinct soft-blurred interface, no readable typography."
- **Use Nano Banana Pro**: its text-rendering architecture is also good at *not* rendering text when told to. The 2.5-flash model is unreliable here.

---

## Anti-Patterns

### Anti-Pattern: Aspect ratio in the prompt

**Novice**: "I asked for a 16:9 wide editorial illustration in the prompt — Gemini gave me a square."
**Expert**: Aspect ratio belongs in `generationConfig.imageConfig.aspectRatio`, never in prompt text. The model ignores prompt-level dimensions. The default is square.
**Timeline**: This config field landed with the 2.5-flash-image GA release in late 2025. Earlier docs and tutorials often omit it; verify against the current docs.

### Anti-Pattern: Single reference, expecting both character AND style fidelity

**Novice**: "I attached the character sheet and asked for the same style — Gemini gave me 16-bit pixel art."
**Expert**: A single illustrated reference is ambiguous about role. The model treats it as a loose mood board and may hallucinate aesthetic interpretations that aren't in the source. Attach *two* references with explicit role labels (`Reference 1 is the CHARACTER`, `Reference 2 is the STYLE TEMPLATE`). Character fidelity and style fidelity are different jobs.
**Timeline**: This became necessary with the move from Imagen-style "one ref + style flag" pipelines (Ideogram V3) to Gemini's reference-as-context model. Pre-2026 patterns don't transfer.

### Anti-Pattern: Negative prompts ("no text", "no watermark")

**Novice**: "I added 'no text, no labels, no watermarks' but Gemini still wrote 'CONCERT TICKETS' across the screen."
**Expert**: Gemini does not honor negative prompts the way Ideogram or Stable Diffusion do. Describe what *should* be there instead — abstract shapes, blurred interface, plain surface. Or upgrade to Nano Banana Pro, whose text architecture is precise enough to reliably *not* generate text on command.
**Timeline**: This is a structural property of Gemini's image head, not a temporary limitation. It applies to all current models in the family.

### Anti-Pattern: Using `gemini-2.5-flash-image` for character work

**Novice**: "I'll save money by using 2.5-flash for the editorial hero illustration."
**Expert**: 2.5-flash is the weakest character-consistency model in the family and unreliable for strict style adherence. Pro is the right tool for character/style work; 2.5-flash is for cheap simple scenes. The cost difference is rarely worth the regeneration loop.
**Timeline**: Pro became available (preview) in early 2026. Before that, the choice was between 2.5-flash and external tools like Ideogram.

---

## References

Consult these for deep dives — they are NOT loaded by default:

| File | Consult When |
|------|-------------|
| `references/model-selection.md` | Choosing between 2.5-flash, 3.1-flash, and 3-pro for a given task |
| `references/prompt-formula.md` | Writing reference-grounded prompts, role-labeling patterns, edge cases |
| `references/character-consistency.md` | Multi-scene generation with the same character; iterative re-feeding |
| `references/style-references.md` | Style transfer pattern, dual-ref vs triple-ref, matching existing illustrated work |
| `references/troubleshooting.md` | Output is square, contains text, drifted style, ignored references |
| `references/comparison-ideogram-qwen.md` | When to use Nano Banana vs Ideogram V3 vs Qwen on M4 Max |
| `scripts/generate.py` | Runnable Python example with two refs + aspect ratio + role-labeled prompt |

---

## Quick Reference: Repo Conventions (jbuds4life)

When this skill is invoked inside the `jbuds4life` repo:

- Character sheets live in `experiments/blog-image-styles/character-sheets/` (alex, jordan, marcus, mei, sofia — all `*-v3-sheet.png`)
- Existing successful heroes (use as STYLE TEMPLATE refs) live in `next-app/public/images/blog/*.webp`
- API key in `next-app/.env.local` as `GEMINI_API_KEY`
- Output convention: save timestamped PNG first, then convert to webp via `cwebp -q 85`, then place at `next-app/public/images/blog/{slug}-hero.webp`
- The blog frontmatter `og_image_url: /images/blog/{slug}-hero.webp` is what's served as the OG image

---

## Sources

- [Gemini API: image generation](https://ai.google.dev/gemini-api/docs/image-generation)
- [Gemini models reference](https://ai.google.dev/gemini-api/docs/models)
- [Google Cloud — Ultimate prompting guide for Nano Banana](https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-nano-banana)
- [Google Cloud — Gemini 2.5 Flash Image on Vertex AI](https://cloud.google.com/blog/products/ai-machine-learning/gemini-2-5-flash-image-on-vertex-ai)
- [Google Developers — Introducing Gemini 2.5 Flash Image](https://developers.googleblog.com/en/introducing-gemini-2-5-flash-image/)