Cut out nsfwjs

package.json

@@ -38,8 +38,7 @@
 	"dependencies": {
 		"@bull-board/api": "^4.12.2",
 		"@bull-board/ui": "^4.12.2",
-		"@napi-rs/cli": "^2.15.0",
-		"@tensorflow/tfjs": "^3.21.0",
+		"@napi-rs/cli": "^2.15.2",
 		"js-yaml": "4.1.0",
 		"seedrandom": "^3.0.5"
 	},

packages/backend/package.json

@@ -90,7 +90,6 @@
 		"nested-property": "4.0.0",
 		"node-fetch": "3.3.0",
 		"nodemailer": "6.8.0",
-		"nsfwjs": "2.4.2",
 		"oauth": "^0.10.0",
 		"os-utils": "0.0.14",
 		"parse5": "7.1.2",

packages/backend/src/misc/get-file-info.ts

@@ -11,7 +11,6 @@ import probeImageSize from "probe-image-size";
 import { type predictionType } from "nsfwjs";
 import sharp from "sharp";
 import { encode } from "blurhash";
-import { detectSensitive } from "@/services/detect-sensitive.js";
 import { createTempDir } from "./create-temp.js";
 
 const pipeline = util.promisify(stream.pipeline);
@@ -126,23 +125,6 @@ export async function getFileInfo(
 	let sensitive = false;
 	let porn = false;
 
-	if (!opts.skipSensitiveDetection) {
-		await detectSensitivity(
-			path,
-			type.mime,
-			opts.sensitiveThreshold ?? 0.5,
-			opts.sensitiveThresholdForPorn ?? 0.75,
-			opts.enableSensitiveMediaDetectionForVideos ?? false,
-		).then(
-			(value) => {
-				[sensitive, porn] = value;
-			},
-			(error) => {
-				warnings.push(`detectSensitivity failed: ${error}`);
-			},
-		);
-	}
-
 	return {
 		size,
 		md5,
@@ -157,177 +139,6 @@ export async function getFileInfo(
 	};
 }
 
-async function detectSensitivity(
-	source: string,
-	mime: string,
-	sensitiveThreshold: number,
-	sensitiveThresholdForPorn: number,
-	analyzeVideo: boolean,
-): Promise<[sensitive: boolean, porn: boolean]> {
-	let sensitive = false;
-	let porn = false;
-
-	function judgePrediction(
-		result: readonly predictionType[],
-	): [sensitive: boolean, porn: boolean] {
-		let sensitive = false;
-		let porn = false;
-
-		if (
-			(result.find((x) => x.className === "Sexy")?.probability ?? 0) >
-			sensitiveThreshold
-		)
-			sensitive = true;
-		if (
-			(result.find((x) => x.className === "Hentai")?.probability ?? 0) >
-			sensitiveThreshold
-		)
-			sensitive = true;
-		if (
-			(result.find((x) => x.className === "Porn")?.probability ?? 0) >
-			sensitiveThreshold
-		)
-			sensitive = true;
-
-		if (
-			(result.find((x) => x.className === "Porn")?.probability ?? 0) >
-			sensitiveThresholdForPorn
-		)
-			porn = true;
-
-		return [sensitive, porn];
-	}
-
-	if (["image/jpeg", "image/png", "image/webp"].includes(mime)) {
-		const result = await detectSensitive(source);
-		if (result) {
-			[sensitive, porn] = judgePrediction(result);
-		}
-	} else if (
-		analyzeVideo &&
-		(mime === "image/apng" || mime.startsWith("video/"))
-	) {
-		const [outDir, disposeOutDir] = await createTempDir();
-		try {
-			const command = FFmpeg()
-				.input(source)
-				.inputOptions([
-					"-skip_frame",
-					"nokey", // 可能ならキーフレームのみを取得してほしいとする（そうなるとは限らない）
-					"-lowres",
-					"3", // 元の画質でデコードする必要はないので 1/8 画質でデコードしてもよいとする（そうなるとは限らない）
-				])
-				.noAudio()
-				.videoFilters([
-					{
-						filter: "select", // フレームのフィルタリング
-						options: {
-							e: "eq(pict_type,PICT_TYPE_I)", // I-Frame のみをフィルタする（VP9 とかはデコードしてみないとわからないっぽい）
-						},
-					},
-					{
-						filter: "blackframe", // 暗いフレームの検出
-						options: {
-							amount: "0", // 暗さに関わらず全てのフレームで測定値を取る
-						},
-					},
-					{
-						filter: "metadata",
-						options: {
-							mode: "select", // フレーム選択モード
-							key: "lavfi.blackframe.pblack", // フレームにおける暗部の百分率（前のフィルタからのメタデータを参照する）
-							value: "50",
-							function: "less", // 50% 未満のフレームを選択する（50% 以上暗部があるフレームだと誤検知を招くかもしれないので）
-						},
-					},
-					{
-						filter: "scale",
-						options: {
-							w: 299,
-							h: 299,
-						},
-					},
-				])
-				.format("image2")
-				.output(join(outDir, "%d.png"))
-				.outputOptions(["-vsync", "0"]); // 可変フレームレートにすることで穴埋めをさせない
-			const results: ReturnType<typeof judgePrediction>[] = [];
-			let frameIndex = 0;
-			let targetIndex = 0;
-			let nextIndex = 1;
-			for await (const path of asyncIterateFrames(outDir, command)) {
-				try {
-					const index = frameIndex++;
-					if (index !== targetIndex) {
-						continue;
-					}
-					targetIndex = nextIndex;
-					nextIndex += index; // fibonacci sequence によってフレーム数制限を掛ける
-					const result = await detectSensitive(path);
-					if (result) {
-						results.push(judgePrediction(result));
-					}
-				} finally {
-					fs.promises.unlink(path);
-				}
-			}
-			sensitive =
-				results.filter((x) => x[0]).length >=
-				Math.ceil(results.length * sensitiveThreshold);
-			porn =
-				results.filter((x) => x[1]).length >=
-				Math.ceil(results.length * sensitiveThresholdForPorn);
-		} finally {
-			disposeOutDir();
-		}
-	}
-
-	return [sensitive, porn];
-}
-
-async function* asyncIterateFrames(
-	cwd: string,
-	command: FFmpeg.FfmpegCommand,
-): AsyncGenerator<string, void> {
-	const watcher = new FSWatcher({
-		cwd,
-		disableGlobbing: true,
-	});
-	let finished = false;
-	command.once("end", () => {
-		finished = true;
-		watcher.close();
-	});
-	command.run();
-	for (let i = 1; true; i++) {
-		const current = `${i}.png`;
-		const next = `${i + 1}.png`;
-		const framePath = join(cwd, current);
-		if (await exists(join(cwd, next))) {
-			yield framePath;
-		} else if (!finished) {
-			watcher.add(next);
-			await new Promise<void>((resolve, reject) => {
-				watcher.on("add", function onAdd(path) {
-					if (path === next) {
-						// 次フレームの書き出しが始まっているなら、現在フレームの書き出しは終わっている
-						watcher.unwatch(current);
-						watcher.off("add", onAdd);
-						resolve();
-					}
-				});
-				command.once("end", resolve); // 全てのフレームを処理し終わったなら、最終フレームである現在フレームの書き出しは終わっている
-				command.once("error", reject);
-			});
-			yield framePath;
-		} else if (await exists(framePath)) {
-			yield framePath;
-		} else {
-			return;
-		}
-	}
-}
-
 function exists(path: string): Promise<boolean> {
 	return fs.promises.access(path).then(
 		() => true,

packages/backend/src/services/detect-sensitive.ts

@@ -1,55 +0,0 @@
-import * as fs from "node:fs";
-import { fileURLToPath } from "node:url";
-import { dirname } from "node:path";
-import * as nsfw from "nsfwjs";
-import si from "systeminformation";
-
-const _filename = fileURLToPath(import.meta.url);
-const _dirname = dirname(_filename);
-
-const REQUIRED_CPU_FLAGS = ["avx2", "fma"];
-let isSupportedCpu: undefined | boolean = undefined;
-
-let model: nsfw.NSFWJS;
-
-export async function detectSensitive(
-	path: string,
-): Promise<nsfw.predictionType[] | null> {
-	try {
-		if (isSupportedCpu === undefined) {
-			const cpuFlags = await getCpuFlags();
-			isSupportedCpu = REQUIRED_CPU_FLAGS.every((required) =>
-				cpuFlags.includes(required),
-			);
-		}
-
-		if (!isSupportedCpu) {
-			console.error("This CPU cannot use TensorFlow.");
-			return null;
-		}
-
-		const tf = await import("@tensorflow/tfjs-node");
-
-		if (model == null)
-			model = await nsfw.load(`file://${_dirname}/../../nsfw-model/`, {
-				size: 299,
-			});
-
-		const buffer = await fs.promises.readFile(path);
-		const image = (await tf.node.decodeImage(buffer, 3)) as any;
-		try {
-			const predictions = await model.classify(image);
-			return predictions;
-		} finally {
-			image.dispose();
-		}
-	} catch (err) {
-		console.error(err);
-		return null;
-	}
-}
-
-async function getCpuFlags(): Promise<string[]> {
-	const str = await si.cpuFlags();
-	return str.split(/\s+/);
-}