From cba15a62829a0fe57a03c8031f981e8f790e2efc Mon Sep 17 00:00:00 2001 From: tanaka takashi Date: Sat, 12 Sep 2026 18:04:56 +0900 Subject: feat: Stable Diffusion 1.x MCPサーバを実装 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 仕様書・設計書に基づき、ローカルCUDA GPU上のStable Diffusion 1.x互換モデルをMCP経由で利用するサーバを新規実装。 実装内容: - generate_image、generate_guided_image、inpaint_imageの3ツールをMCP SDK v2で公開 - TOMLからcheckpoint、VAE、複数LoRAと重み、ControlNet、生成設定を読み込み - モデル種別ごとの固定ディレクトリ解決と起動時の設定・ファイル存在検証を追加 - Depth/MiDaS、OpenPose body/full、ControlNet Inpaintの前処理とパイプライン切替を実装 - FP16/CUDA、1枚固定生成、ControlNetオンデマンドロード、単一GPU直列実行を実装 - Base64 PNG/JPEG入力の形式・サイズ検証と任意のdata URI入力に対応 - seed、サイズ、モード、時刻、処理時間をPNG metadataへ記録 - 任意のローカルPNG保存、生成エラー分類、CUDA OOM後のキャッシュ回復を追加 - --configと--versionを備えたstdio CLIエントリポイントを追加 開発・運用: - Python 3.11およびuvによる依存・lock管理を構成 - config.example.tomlとセットアップ・運用手順をREADMEへ追加 - Ruff、Pyright、pytestの設定を追加 - 設定、パス解決、schema、画像検証、サービスルーティング、排他制御、MCP画像返却をテスト 検証結果: - pytest: 30 passed - Ruff lint/format: passed - Pyright: 0 errors, 0 warnings - CLI versionおよびMCPクライアント経由のPNG返却を確認 - 実GPU smoke testはCUDA GPUとモデルファイルがないため対象外 --- src/mcp_sd15/schema.py | 142 +++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 142 insertions(+) create mode 100644 src/mcp_sd15/schema.py (limited to 'src/mcp_sd15/schema.py') diff --git a/src/mcp_sd15/schema.py b/src/mcp_sd15/schema.py new file mode 100644 index 0000000..e22450d --- /dev/null +++ b/src/mcp_sd15/schema.py @@ -0,0 +1,142 @@ +"""Validated request and response types independent of MCP and Diffusers.""" + +from __future__ import annotations + +from dataclasses import dataclass +from enum import StrEnum +from typing import Any + +from .errors import InvalidRequestError + + +class ImageSize(StrEnum): + SQUARE_512 = "square_512" + LANDSCAPE_512 = "landscape_512" + PORTRAIT_512 = "portrait_512" + + @property + def dimensions(self) -> tuple[int, int]: + return { + ImageSize.SQUARE_512: (512, 512), + ImageSize.LANDSCAPE_512: (512, 384), + ImageSize.PORTRAIT_512: (384, 512), + }[self] + + +class ControlMode(StrEnum): + DEPTH = "depth" + OPENPOSE = "openpose" + OPENPOSE_FULL = "openpose_full" + + +def _prompt(value: Any, label: str, *, required: bool) -> str: + if value is None and not required: + return "" + if not isinstance(value, str): + raise InvalidRequestError(f"{label} must be a string") + value = value.strip() + if required and not value: + raise InvalidRequestError(f"{label} must not be empty") + if len(value) > 10_000: + raise InvalidRequestError(f"{label} is too long") + return value + + +def _size(value: str | ImageSize | None, default: str) -> ImageSize: + try: + return ImageSize(value or default) + except ValueError as exc: + raise InvalidRequestError(f"unsupported size: {value}") from exc + + +@dataclass(frozen=True) +class TextToImageRequest: + positive_prompt: str + negative_prompt: str + size: ImageSize + + @classmethod + def create( + cls, + positive_prompt: str, + negative_prompt: str | None = None, + size: str | ImageSize | None = None, + *, + default_size: str = ImageSize.SQUARE_512, + ) -> TextToImageRequest: + return cls( + _prompt(positive_prompt, "positive_prompt", required=True), + _prompt(negative_prompt, "negative_prompt", required=False), + _size(size, default_size), + ) + + +@dataclass(frozen=True) +class GuidedImageRequest: + positive_prompt: str + negative_prompt: str + size: ImageSize + control_mode: ControlMode + input_image: str + + @classmethod + def create( + cls, + positive_prompt: str, + control_mode: str | ControlMode, + input_image: str, + negative_prompt: str | None = None, + size: str | ImageSize | None = None, + *, + default_size: str = ImageSize.SQUARE_512, + ) -> GuidedImageRequest: + base = TextToImageRequest.create( + positive_prompt, negative_prompt, size, default_size=default_size + ) + try: + mode = ControlMode(control_mode) + except ValueError as exc: + raise InvalidRequestError(f"unsupported control_mode: {control_mode}") from exc + if not isinstance(input_image, str) or not input_image.strip(): + raise InvalidRequestError("input_image must be a non-empty Base64 string") + return cls(base.positive_prompt, base.negative_prompt, base.size, mode, input_image) + + +@dataclass(frozen=True) +class InpaintImageRequest: + positive_prompt: str + negative_prompt: str + size: ImageSize + input_image: str + mask_image: str + + @classmethod + def create( + cls, + positive_prompt: str, + input_image: str, + mask_image: str, + negative_prompt: str | None = None, + size: str | ImageSize | None = None, + *, + default_size: str = ImageSize.SQUARE_512, + ) -> InpaintImageRequest: + base = TextToImageRequest.create( + positive_prompt, negative_prompt, size, default_size=default_size + ) + for label, value in (("input_image", input_image), ("mask_image", mask_image)): + if not isinstance(value, str) or not value.strip(): + raise InvalidRequestError(f"{label} must be a non-empty Base64 string") + return cls( + base.positive_prompt, + base.negative_prompt, + base.size, + input_image, + mask_image, + ) + + +@dataclass(frozen=True) +class GenerationResult: + png_data: bytes + metadata: dict[str, str | int | float] -- cgit v1.2.3