mcpbytes-pixel-sprite

Generate pixel-art character sprite sheets on your own NVIDIA GPU, from the command line or from any MCP client (Claude, Cursor, VS Code, …). It is the open-source engine behind MCPBytes' hosted create_pixel_sprite tool, and can also send jobs to the hosted tool when you have no GPU.

Each candidate is a 128×128 PNG of sixteen 32×32 cells. Rows face down, left, right and up; columns 0–2 are walk poses and column 3 holds four still poses (arms raised, jump left, jump right, lying). Every candidate also comes with frame PNGs, looping walk GIFs, QA metadata, a Godot 4 scene and a Unity 6 importer.

Requirements

• An NVIDIA GPU with a CUDA 12.8 driver (R570 or newer) on Windows or Linux. The memory mode is picked from the GPU's size; you can override it with --offload:

  --offload	GPU memory used	Time per candidate*
  none (picked for GPUs with 20 GB or more)	16.2 GiB	2.3 s
  model (picked for 10–20 GB)	7.8 GiB	7.4 s
  sequential (picked for less than 10 GB)	1.0 GiB	34 s

  *Measured on an RTX PRO 6000 with the models already loaded. Your times will differ.

• About 16 GB of disk for the model files, downloaded once into the Hugging Face cache.

• uv (or Python 3.11–3.13 with pip).

• No NVIDIA GPU (an Apple Silicon Mac, for example)? Use --cloud with an MCPBytes API key; see below.

Install and run

uvx --torch-backend=auto mcpbytes-pixel-sprite download
uvx --torch-backend=auto mcpbytes-pixel-sprite generate "A forest ranger with a green hood and a brown satchel" --seed 42

--torch-backend=auto installs the CUDA build of PyTorch that matches your driver. Without it, PyTorch from PyPI is CPU-only on Windows. With pip: pip install mcpbytes-pixel-sprite --extra-index-url https://download.pytorch.org/whl/cu128.

download fetches the pinned files listed below (about 16 GB, resumable). generate also downloads anything that is missing, then writes to pixel-sprites/<timestamp>/ (or to -o <new or empty folder>):

Options: --variants 1-4, --reference image.png (a PNG or JPEG of your character), --opaque, --palette-colors 16|32|64, --palette "#rrggbb,#rrggbb,…", --frame-duration-ms 40-1000, --seed. The same seed, inputs, hardware and --offload mode reproduce a result byte for byte. Seeds are derived the same way as in the hosted tool, but individual pixels can still differ between GPUs.

qa_status is unreviewed or needs_review; it is never an approval. Check anatomy, facing direction, identity across views and transparency yourself. Walk previews play poses [0, 1, 2, 1]: three drawn poses, not four.

MCP server

Add this to your MCP client's configuration (for example claude_desktop_config.json, .cursor/mcp.json):

{
  "mcpServers": {
    "pixel-sprite": {
      "command": "uvx",
      "args": ["--torch-backend=auto", "mcpbytes-pixel-sprite", "mcp"]
    }
  }
}

Run mcpbytes-pixel-sprite download once first. A 16 GB download would outlast any client's timeout, so the tool asks for it instead of starting it. The server provides one tool, create_pixel_sprite, with the options above plus output_dir. It writes the files, returns their summary and shows each sheet enlarged 4× so the model can look at it. Files go to ~/mcpbytes/pixel-sprites/<timestamp>/ unless the call passes output_dir, or you set MCPBYTES_OUTPUT_DIR. The models load on the first call and stay in memory.

Cloud mode (optional)

--cloud sends the job to MCPBytes GPUs instead and charges prepaid credits (prices at mcpbytes.com/pricing). Create a key at console.mcpbytes.com/keys:

export MCPBYTES_API_KEY=...
uvx mcpbytes-pixel-sprite generate "A forest ranger" --cloud
uvx mcpbytes-pixel-sprite mcp --cloud      # the MCP server, backed by the hosted tool

The files are the same. The run prints a request key: if it is interrupted, run it again with --idempotency-key <key> to resume the same job without a second charge. Nothing is sent to MCPBytes unless you pass --cloud.

Models

Nothing is bundled. These files are downloaded from Hugging Face at pinned commits and licensed to you by their publishers (all Apache-2.0, no account needed):

• black-forest-labs/FLUX.2-klein-base-4B — transformer, VAE, tokenizer
• svntax-dev/pixel_spritesheet_4walk_small_lora_v1 — the sprite-sheet LoRA
• Comfy-Org/z_image_turbo — the Qwen3 4B text encoder

HF_HOME, HF_ENDPOINT, HF_TOKEN and HF_HUB_OFFLINE work as usual. See NOTICE for attributions.

Licence

GNU AGPL v3.0 only. The files copied into every export (IMPORTING.md and the Unity importer script) are MIT-0, so you can ship them in a game under any terms. The sprites you generate are yours. See NOTICE.