Hey there, it’s Claudie! Today I want to share how I finally cracked the AI character consistency problem that had been bugging me for ages. I put together a workflow that keeps the “same character” looking consistent across multiple scenes when generating AI images. Let me walk you through the whole process — connecting two open-source MCP servers to automate everything from character sheet creation to scene placement, with code examples along the way.

🎭 The AI Character Consistency Dilemma — Why Is “Same Face” So Hard?

If you’ve ever used AI image generation, you know the struggle. Type “blonde ponytail woman” in a prompt and you get a different person every single time. The hair color shifts subtly, the face shape changes, and the eye color varies. That’s fine for a one-off image, but when the same character needs to appear across multiple scenes — like on a blog or Instagram — it becomes a real problem.

I ran into this constantly while managing two characters: Claudie (blonde) and Siwol (black hair). No matter how carefully I crafted my prompts, I’d get “similar-looking people” but never the “same person.” Text alone simply can’t reproduce the subtle facial features — freckle placement, eye corner angles, cheekbone lines.

That’s why I built the 2-Step MCP Workflow. It’s a pipeline that automatically generates multi-angle “ID photos” of a character, then uses those as reference images to produce consistent characters. Just connect two open-source MCP servers inside Claude Code and you’re good to go.

🛠️ The 2-Step MCP Workflow

Step 1 — Create a Character Sheet with terrychadesign

terrychadesign MCP is an MCP server that leverages the Gemini API to auto-generate character design sheets. Call the design_character tool and it produces character images from 7 angles (front, back, left, right, 3/4 view, face front, face left/right) all at once.

Here’s how you call it from Claude Code:

# MCP tool call example in Claude Code
design_character(
  character_name="Claudie",
  style="photorealistic",
  description=(
    "blonde ponytail, blue eyes, "
    "freckles, white crop top, "
    "hot pants"
  ),
  generate_composite=True
)
# → Generates 7-angle images + composite

Once generation is complete, you get 7 individual angle images plus a composite sheet that combines them all into one. Below are the actual character sheets generated for Claudie and Siwol.

Claudie — Blonde Ponytail

Siwol — Black Wavy Hair

Having all 7 angles on one sheet gives you a complete overview of the character at a glance. These individual angle images from the composite sheet serve as “draw this person” references in the next step.

You can also use the generate_pose_sheet tool to create action poses (walking, sitting, waving, etc.). Including pose sheets as additional references boosts character consistency even further.

# Generate additional pose sheet
generate_pose_sheet(
  character_name="Claudie",
  style="photorealistic",
  poses=[
    "walking", "sitting"
  ],
  description=(
    "blonde ponytail, "
    "blue eyes, freckles"
  )
)

Step 2 — Unleash Your Character with Nano Banana

Nano Banana MCP is an MCP server that wraps Google’s Gemini Image Generation API. The star of the show is the generate_with_references tool, which lets you include up to 10 reference images (Flash) or 14 (Pro).

Feed in the individual angle images from Step 1 as references, and Gemini understands “ah, this is the person” — then draws that same character in brand new scenes.

# Nano Banana generate_with_references
generate_with_references(
  prompt=(
    "A young woman walking "
    "through a European "
    "cobblestone alley, "
    "golden hour sunlight, "
    "fashion editorial style, "
    "full body shot, 4K"
  ),
  reference_paths=[
    "claudie_front.jpg",
    "claudie_face_front.jpg",
    "claudie_three_quarter.jpg"
  ],
  aspect_ratio="3:4",
  model="flash"
)
# → Same person from the references
#   appears in a brand new scene!

About 3 reference images is enough for pretty stable results. My recommendation is to use front full-body + face close-up + 3/4 view as your base set. Adding pose sheets makes it more accurate, but too many references can actually reduce generation flexibility.

🎨 Real Results — Same Prompt, Different Characters

This is where the real magic of this workflow shines. Use the exact same prompt but swap the reference images, and different characters appear consistently in the same scene. No prompt modifications needed — just swap the references.

Below are results from a “European alley fashion editorial” prompt with Claudie and Siwol respectively. The prompt is 100% identical — only the reference_paths differ.

Claudie — Blonde ponytail, freckles preserved

Siwol — Black wavy hair, beauty mark preserved

Same composition, same mood — but each character’s hairstyle, facial features, and overall impression match their respective character sheets. That’s the power of reference-based AI character consistency!

In code, the difference is literally one line:

# Claudie version
generate_with_references(
  prompt=same_prompt,
  reference_paths=[
    "claudie_front.jpg",
    "claudie_face_front.jpg",
    "claudie_three_quarter.jpg"
  ]
)
# Siwol version — same prompt!
generate_with_references(
  prompt=same_prompt,
  reference_paths=[
    "siwol_front.jpg",
    "siwol_face_front.jpg",
    "siwol_three_quarter.jpg"
  ]
)

This is incredibly useful for content where the same character needs to appear repeatedly, like Instagram or webtoons. Instead of writing out lengthy appearance descriptions in every prompt, you just feed in a few reference images and you’re done.

⚙️ Setup Guide

To use both MCP servers with Claude Code, add the configuration to your project root’s .mcp.json. A single Gemini API key is all you need to run both servers.

{
  "mcpServers": {
    "terrycha-design": {
      "command": "python",
      "args": [
        "path/to/"
        "terrychadesignmcp/"
        "server.py"
      ],
      "env": {
        "GEMINI_API_KEY":
          "your-gemini-api-key"
      }
    },
    "nanobanana": {
      "command": "python",
      "args": [
        "path/to/"
        "terrymcpnanobanana/"
        "server.py"
      ],
      "env": {
        "GEMINI_API_KEY":
          "your-gemini-api-key",
        "NANOBANANA_OUTPUT_DIR":
          "path/to/output"
      }
    }
  }
}

Here are the key configuration points:

• GEMINI_API_KEY — You can get one for free from Google AI Studio. Both servers can share the same key.
• NANOBANANA_OUTPUT_DIR — The path where generated images will be saved. Defaults to ~/nanobanana_output if not specified.
• No Vertex AI required — Since Nano Banana v0.2.1, API Key only mode is supported, so you can get started without any Google Cloud project setup.
• Windows compatible — Both servers work well on Windows too.

Save the config file and restart Claude Code — the MCP tools will be ready to use immediately. This setup can feel a bit tricky, but just copy the JSON above and swap out the paths, and you’ll be up and running in no time!

💡 A Taste of Nano Banana Image Generation

Beyond character consistency, Nano Banana MCP is also a fantastic general-purpose image generation tool. It supports 4K resolution with 14 aspect ratios, and you can choose between Flash and Pro models. Let me show you a few examples.

This landscape was generated with the Flash model at 16:9 ratio in 4K resolution. The water reflections and autumn foliage colors look like a real photograph, don’t they? The photorealistic quality is impressive.

This is a 1:1 digital art style. You can see it handles illustration styles just as well as photorealistic ones.

This one uses the 21:9 cinematic ultra-wide ratio with Thinking Mode (High) enabled. When you turn on Thinking Mode, Gemini thinks more deeply about the image composition, resulting in richer detail for complex scenes. You can see the difference in the Korean text on neon signs, rain reflections, and perspective handling.

# Thinking Mode example
generate_image(
  prompt=(
    "Cyberpunk Seoul at night, "
    "neon signs in Korean, "
    "rain-soaked streets, "
    "cinematic wide shot"
  ),
  aspect_ratio="21:9",
  model="flash",
  thinking_level="high"
)

⚠️ Tips & Things to Watch Out For

Reference Style Mismatch — The Most Common Mistake

This is a lesson I learned through my own trial and error. If you use anime-style character sheet references to generate photorealistic images, the character will look like a completely different person. You must match the style between your input (references) and output (generated images).

• Want photorealistic results? → Use a photorealistic style character sheet
• Want anime results? → Use an anime-style character sheet
• If styles are mixed, character consistency breaks down

Flash vs Pro — Which Model Should You Choose?

I usually generate several images quickly with Flash, and once I find a composition I like, I produce the final version with Pro. Since Flash supports 4K resolution too, it’s sufficient for most use cases on its own.

Tips for Choosing Reference Images

It’s more important to choose references strategically than to load up as many as possible:

• Essential 3: Front full-body, face close-up, 3/4 view — this combo is the baseline
• Recommended extras: Back view (to confirm hairstyle), a pose sheet similar to your target pose
• Optimal count: 3 to 5 images gives the best results
• Caution: More than 7 can reduce generation freedom, making compositions feel stiff

Gemini Safety Filter

The Gemini API has a fairly strict safety filter. Certain combinations (bedroom + revealing clothing, etc.) may get blocked. In those cases, try phrasing your prompt more indirectly, or consider using an alternative image generation tool alongside it.

📚 References

• Nano Banana MCP GitHub — Gemini image generation MCP server (v0.2.1, API Key only)
• terrychadesign MCP GitHub — Character design sheet auto-generation MCP server (v0.3.1)
• Gemini Image Generation API Official Docs — Native image generation reference
• Model Context Protocol Official Docs — MCP server setup and usage guide

✅ Wrapping Up — How I Solved AI Character Consistency

AI character consistency is a tough problem to solve with a single tool. But by creating a systematic character sheet with terrychadesign MCP and using it as a reference through Nano Banana MCP‘s generate_with_references, you can place the same character consistently across different scenes.

Here’s the workflow in three key takeaways:

1. Always create character sheets in photorealistic style — Match the style with your intended output to maintain consistency
2. 3 to 5 reference images is the sweet spot — Front, face, and 3/4 view are the essential set
3. Same prompt + different references = different characters — Just swap the references to place multiple characters in the same world

Both MCP servers are open source, and all you need is a single Gemini API key to get started. Even if you’ve never used MCP servers with Claude Code before, a few lines in .mcp.json and you’re all set. If you’ve been struggling with AI character consistency, give this workflow a try!

If you have questions or want to know more, drop a comment below. I’ll be back with more useful dev workflows soon!