Imagen

MCP server for generating pixel-art sprites with transparent backgrounds. Bring your own model — any diffusers-compatible text-to-image pipeline works.

The default configuration uses FLUX.2-klein-4B + pixel-art-lora, but you can swap in any model you like.

Features

• Bring your own model — any diffusers-compatible pipeline (FLUX, SDXL, SD3, etc.)

• Text-to-sprite generation — describe any character, get a pixel-art PNG

• Transparent background — automatic background removal via flood-fill

• Pixel-art effect — downscale/upscale with NEAREST interpolation

• Reproducible — optional seed for consistent results

• Batch generation — generate multiple sprites in one call

• MCP integration — works with any MCP-compatible client (opencode, Claude, etc.)

• Feedback loop — rate generated sprites, AI uses high-rated ones as reference

Related MCP server: aseprite-mcp

Quick Start

1. Install dependencies

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Download a model

You need a base text-to-image model. Optionally, a LoRA adapter for pixel-art style.

Example: FLUX.2-klein-4B + pixel-art-lora (default)

mkdir -p ~/models

# Base model (~23 GB)
huggingface-cli download black-forest-labs/FLUX.2-klein-4b \
    --local-dir ~/models/flux2-klein-4b

# LoRA adapter (~625 MB) — optional but recommended for pixel-art
huggingface-cli download Limbicnation/pixel-art-lora \
    --local-dir ~/models/pixel-art-lora

Other models that work:

Note: You may need to adjust LORA_SCALE and pipeline class in server.py depending on your model. See Configuration.

3. Configure paths

By default, models are expected at ~/models/. Override with environment variables:

export IMAGEGEN_MODEL_DIR=/path/to/your/base-model
export IMAGEGEN_LORA_DIR=/path/to/your/lora        # optional, set empty to disable
export IMAGEGEN_OUTPUT_DIR=/path/to/output

4. Run as MCP server

./venv/bin/python server.py

5. Review & rate sprites (GUI)

./venv/bin/pip install dearpygui pillow
./venv/bin/python review_gui.py

Features:

• Thumbnail preview of each generated sprite

• Star rating (1-5) with one click

• Text feedback per sprite

• Filter: all / unrated / top rated

• Export rated dataset as JSONL for fine-tuning

Or configure in your MCP client:

{
  "mcp": {
    "pixel-art": {
      "type": "local",
      "command": ["./venv/bin/python", "server.py"],
      "enabled": true
    }
  }
}

Tools

Generation

generate_sprite

Generate a single pixel-art sprite. Automatically saved to feedback DB (unrated).

Returns: output_path, db_id, generation_time, and other metadata.

batch_generate

Generate multiple sprites in one call. Each is saved to the feedback DB.

Feedback

rate_sprite

Rate a generated sprite 1-5 stars with optional feedback.

get_reference_sprites

Get highly-rated reference sprites for a prompt. The AI uses these as examples when generating similar sprites.

list_sprites

List sprites in the feedback database.

db_stats

Get database statistics: total sprites, rated, unrated, average rating.

Configuration

Environment variables

Swapping models

The server is configured for FLUX.2-klein by default. To use a different model, edit server.py:

1. Pipeline class — replace Flux2KleinPipeline with your model's pipeline (e.g. StableDiffusionXLPipeline for SDXL)

2. LoRA scale — adjust LORA_SCALE (rsLoRA needs ~0.1, regular LoRA typically 0.7-1.0)

3. Guidance scale — distilled models ignore it; standard models need 5-8

4. Steps — distilled models work at 4; standard models need 20-30

How It Works

1. Generation — text-to-image model generates a 512x512 image

2. Pixelation — downscale with LANCZOS, upscale with NEAREST → chunky pixel-art blocks

3. Background removal — detect border color, normalize to solid fill, flood-fill from edges → transparent PNG

Requirements

• GPU: NVIDIA with >= 8 GB VRAM (uses CPU offload automatically)

• Python: 3.12+

• CUDA: 12.0+

Credits

• Default model: FLUX.2-klein-4B by Black Forest Labs (Apache 2.0)

• Default LoRA: pixel-art-lora by Limbicnation (Apache 2.0)

• MCP SDK: modelcontextprotocol/python-sdk

License

MIT — see LICENSE

Model licenses are separate from this project. Check each model's license card for usage terms.