compress-pdf-lib 1.0.2 → 1.0.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "compress-pdf-lib",
-  "version": "1.0.2",
+  "version": "1.0.3",
   "description": "Client-side PDF compression: pdf.js render + mozjpeg (WASM) encoding via a parallel worker pool + pdf-lib rebuild. No server, no UI. Ships as raw ESM source for Vite-based apps (React + Vite, Astro).",
   "type": "module",
   "main": "src/index.js",

package/src/compress.js

@@ -1,28 +1,53 @@
 /**
  * compress.js
  *
- * Full, UI-less PDF compression library: worker-pool + parallel page
- * rendering + mozjpeg (WASM) encoding + pdf-lib rebuild.
+ * PDF compression library with two strategies:
  *
- * Pairs with pdf-compressor.worker.js (same folder).
+ *   compressPDF(input, options)   — RECOMMENDED. Extracts each embedded raster
+ *                                    image from the PDF, decodes it, resizes
+ *                                    and recompresses it on its own (via the
+ *                                    worker pool + mozjpeg WASM), and writes
+ *                                    the result back into the exact same PDF
+ *                                    object slot. Text, fonts, and vector
+ *                                    drawing are never touched, so text-based
+ *                                    PDFs are not bloated by rasterization.
+ *
+ *   rasterizePDF(input, options)  — the older approach: renders every page to
+ *                                    a full-page image and rebuilds the PDF
+ *                                    from those images. Useful only for
+ *                                    PDFs that are already just scans (one
+ *                                    image per page) where per-image
+ *                                    extraction wouldn't find anything.
+ *
+ * Both use the same CompressionPool / pdf-compressor.worker.js for the
+ * actual JPEG encoding.
  *
  * Usage:
  *   import { compressPDF } from 'compress-pdf-lib';
  *
  *   const { file, stats } = await compressPDF(pdfFile, {
- *     quality: 70,
- *     resolution: 1600,
+ *     quality: 75,
+ *     scale: 1, // 1 = keep each image's native pixel size, just recompress
  *   });
  *
  *   console.log(stats);
  *   // {
- *   //   pages, originalBytes, compressedBytes, savedBytes,
- *   //   reduction, elapsed, workersUsed, perPage: [...]
+ *   //   pages, imagesFound, imagesCompressed, imagesSkipped,
+ *   //   originalBytes, compressedBytes, savedBytes, reduction, elapsed,
+ *   //   workersUsed, perImage: [...]
  *   // }
  */
 
 import * as pdfjsLib from "pdfjs-dist";
-import { PDFDocument } from "pdf-lib";
+import {
+  PDFDocument,
+  PDFName,
+  PDFRawStream,
+  PDFDict,
+  PDFArray,
+  PDFNumber,
+  decodePDFRawStream,
+} from "pdf-lib";
 import pdfWorkerUrl from "pdfjs-dist/build/pdf.worker.min.mjs?url";
 
 pdfjsLib.GlobalWorkerOptions.workerSrc = pdfWorkerUrl;
@@ -191,7 +216,7 @@ export class CompressionPool {
 }
 
 /* ============================================================
-   SAFE PDF PAGE CLEANUP
+   SHARED HELPERS
 ============================================================ */
 
 function safePageCleanup(page) {
@@ -204,10 +229,6 @@ function safePageCleanup(page) {
   }
 }
 
-/* ============================================================
-   INPUT NORMALIZATION
-============================================================ */
-
 async function normalizeInput(input) {
   let arrayBuffer;
   let fileName = "document.pdf";
@@ -234,11 +255,679 @@ async function normalizeInput(input) {
   return { arrayBuffer, fileName, fileType, originalBytes: arrayBuffer.byteLength };
 }
 
-/* ============================================================
-   RENDER + COMPRESS ONE PAGE
-============================================================ */
+function jpegResultToBytes(item) {
+  if (item.jpeg instanceof ArrayBuffer) {
+    return new Uint8Array(item.jpeg);
+  }
+  if (item.jpeg instanceof Uint8Array) {
+    return item.jpeg;
+  }
+  if (ArrayBuffer.isView(item.jpeg)) {
+    return new Uint8Array(item.jpeg.buffer, item.jpeg.byteOffset, item.jpeg.byteLength);
+  }
+  throw new Error("Invalid JPEG encoder output");
+}
+
+/* ============================================================================
+   STRATEGY 1 (RECOMMENDED): EXTRACT EMBEDDED IMAGES, COMPRESS, REPLACE IN PLACE
+============================================================================ */
+
+/*
+ * ---- PDF filter name plumbing --------------------------------------------
+ */
+
+function filterNamesOf(dict) {
+  const filter = dict.get(PDFName.of("Filter"));
+
+  if (!filter) return [];
+
+  if (filter instanceof PDFName) {
+    return [filter.asString().replace(/^\//, "")];
+  }
+
+  if (filter instanceof PDFArray) {
+    return filter.asArray().map((f) => f.asString().replace(/^\//, ""));
+  }
+
+  return [];
+}
+
+const IMAGE_CODEC_FILTERS = new Set([
+  "DCTDecode",
+  "JPXDecode",
+  "CCITTFaxDecode",
+  "JBIG2Decode",
+]);
+
+/*
+ * ---- ColorSpace resolution --------------------------------------------
+ * Returns { kind, components, palette?, paletteComponents? }
+ */
+
+function resolveColorSpace(context, csObj, depth = 0) {
+  if (depth > 4 || !csObj) {
+    return { kind: "rgb", components: 3 };
+  }
+
+  const resolved = csObj instanceof PDFDict || csObj instanceof PDFArray || csObj instanceof PDFName
+    ? csObj
+    : context.lookupMaybe(csObj, PDFName) ||
+      context.lookupMaybe(csObj, PDFArray) ||
+      context.lookupMaybe(csObj, PDFDict) ||
+      csObj;
+
+  if (resolved instanceof PDFName) {
+    const name = resolved.asString().replace(/^\//, "");
+
+    if (name === "DeviceGray" || name === "CalGray" || name === "G") {
+      return { kind: "gray", components: 1 };
+    }
+    if (name === "DeviceCMYK" || name === "CMYK") {
+      return { kind: "cmyk", components: 4 };
+    }
+    // DeviceRGB, CalRGB, Lab (approximated as RGB), and anything unknown
+    return { kind: "rgb", components: 3 };
+  }
+
+  if (resolved instanceof PDFArray) {
+    const items = resolved.asArray();
+    const familyName = items[0] instanceof PDFName ? items[0].asString().replace(/^\//, "") : "";
+
+    if (familyName === "ICCBased") {
+      const stream = context.lookup(items[1]);
+      const n = stream?.dict?.get(PDFName.of("N"));
+      const components = n instanceof PDFNumber ? n.asNumber() : 3;
+
+      if (components === 1) return { kind: "gray", components: 1 };
+      if (components === 4) return { kind: "cmyk", components: 4 };
+      return { kind: "rgb", components: 3 };
+    }
+
+    if (familyName === "Indexed") {
+      const base = resolveColorSpace(context, items[1], depth + 1);
+      const lookupObj = context.lookup(items[2]) ?? items[2];
+
+      let paletteBytes;
+
+      if (lookupObj instanceof PDFRawStream || (lookupObj && lookupObj.dict && lookupObj.contents)) {
+        paletteBytes = decodePDFRawStream(lookupObj).decode();
+      } else if (lookupObj && typeof lookupObj.asBytes === "function") {
+        paletteBytes = lookupObj.asBytes();
+      } else if (lookupObj && typeof lookupObj.value === "string") {
+        paletteBytes = Uint8Array.from(lookupObj.value, (c) => c.charCodeAt(0));
+      } else {
+        paletteBytes = new Uint8Array(0);
+      }
+
+      return {
+        kind: "indexed",
+        components: 1,
+        palette: paletteBytes,
+        paletteComponents: base.components,
+        baseKind: base.kind,
+      };
+    }
+
+    if (familyName === "DeviceN" || familyName === "Separation") {
+      // Rare, hard to interpret correctly — fall back to gray-ish approximation
+      return { kind: "gray", components: 1 };
+    }
+
+    if (familyName === "CalRGB" || familyName === "Lab") {
+      return { kind: "rgb", components: 3 };
+    }
+
+    if (familyName === "CalGray") {
+      return { kind: "gray", components: 1 };
+    }
+  }
+
+  return { kind: "rgb", components: 3 };
+}
+
+/*
+ * ---- Raw sample bit-reader ------------------------------------------------
+ * PDF image rows are byte-aligned: each row starts on a new byte even if the
+ * previous row didn't end on one.
+ */
+
+function readSamples(rawBytes, width, height, bitsPerComponent, numComponents) {
+  const bitsPerPixel = bitsPerComponent * numComponents;
+  const rowBytes = Math.ceil((bitsPerPixel * width) / 8);
+  const samples = new Uint16Array(width * height * numComponents);
+
+  const maxVal = (1 << bitsPerComponent) - 1;
+
+  let sampleIndex = 0;
+
+  for (let y = 0; y < height; y++) {
+    const rowStart = y * rowBytes;
+    let bitPos = 0;
+
+    for (let x = 0; x < width * numComponents; x++) {
+      if (bitsPerComponent === 8) {
+        samples[sampleIndex++] = rawBytes[rowStart + x] || 0;
+      } else if (bitsPerComponent === 16) {
+        const byteOffset = rowStart + x * 2;
+        samples[sampleIndex++] =
+          ((rawBytes[byteOffset] || 0) << 8) | (rawBytes[byteOffset + 1] || 0);
+      } else {
+        // 1, 2, or 4 bits per component
+        const byteIndex = rowStart + (bitPos >> 3);
+        const bitOffsetInByte = bitPos & 7;
+        const byte = rawBytes[byteIndex] || 0;
+
+        const shift = 8 - bitOffsetInByte - bitsPerComponent;
+        const value = (byte >> Math.max(shift, 0)) & maxVal;
+
+        samples[sampleIndex++] = value;
+        bitPos += bitsPerComponent;
+      }
+    }
+  }
+
+  return { samples, maxVal };
+}
+
+function scaleSample(value, maxVal) {
+  return maxVal === 255 ? value : Math.round((value / maxVal) * 255);
+}
+
+/*
+ * ---- Convert decoded samples -> RGBA Uint8ClampedArray --------------------
+ */
+
+function samplesToRGBA(rawBytes, width, height, bitsPerComponent, colorSpaceInfo) {
+  const { kind, components } = colorSpaceInfo;
+  const { samples, maxVal } = readSamples(rawBytes, width, height, bitsPerComponent, components);
+
+  const rgba = new Uint8ClampedArray(width * height * 4);
+
+  for (let i = 0, s = 0; i < width * height; i++, s += components) {
+    let r, g, b;
+
+    if (kind === "gray") {
+      const v = scaleSample(samples[s], maxVal);
+      r = g = b = v;
+    } else if (kind === "cmyk") {
+      const c = samples[s] / maxVal;
+      const m = samples[s + 1] / maxVal;
+      const y = samples[s + 2] / maxVal;
+      const k = samples[s + 3] / maxVal;
+
+      r = 255 * (1 - c) * (1 - k);
+      g = 255 * (1 - m) * (1 - k);
+      b = 255 * (1 - y) * (1 - k);
+    } else if (kind === "indexed") {
+      const index = samples[s];
+      const { palette, paletteComponents, baseKind } = colorSpaceInfo;
+      const base = index * paletteComponents;
+
+      if (baseKind === "gray") {
+        r = g = b = palette[base] ?? 0;
+      } else if (baseKind === "cmyk") {
+        const c = (palette[base] ?? 0) / 255;
+        const m = (palette[base + 1] ?? 0) / 255;
+        const y = (palette[base + 2] ?? 0) / 255;
+        const k = (palette[base + 3] ?? 0) / 255;
+        r = 255 * (1 - c) * (1 - k);
+        g = 255 * (1 - m) * (1 - k);
+        b = 255 * (1 - y) * (1 - k);
+      } else {
+        r = palette[base] ?? 0;
+        g = palette[base + 1] ?? 0;
+        b = palette[base + 2] ?? 0;
+      }
+    } else {
+      // rgb
+      r = scaleSample(samples[s], maxVal);
+      g = scaleSample(samples[s + 1], maxVal);
+      b = scaleSample(samples[s + 2], maxVal);
+    }
+
+    const o = i * 4;
+    rgba[o] = r;
+    rgba[o + 1] = g;
+    rgba[o + 2] = b;
+    rgba[o + 3] = 255;
+  }
+
+  return rgba;
+}
+
+/*
+ * ---- Decode one PDF Image XObject into an ImageBitmap ---------------------
+ * Returns null if the image uses a codec we don't support (JPX/CCITT/JBIG2),
+ * is a stencil mask, or otherwise can't be safely handled.
+ */
+
+async function decodeImageXObject(context, ref) {
+  const stream = context.lookup(ref);
+
+  if (!stream || !(stream instanceof PDFRawStream)) return null;
+
+  const dict = stream.dict;
+
+  const subtype = dict.get(PDFName.of("Subtype"));
+  if (!subtype || subtype.asString().replace(/^\//, "") !== "Image") return null;
+
+  const isMask = dict.get(PDFName.of("ImageMask"));
+  if (isMask && isMask.constructor?.name === "PDFBool" && isMask.asBoolean?.()) return null;
+
+  // Color-key / stencil Mask (not SMask) changes meaning based on exact pixel
+  // values — recompressing would break it, so skip these entirely.
+  if (dict.get(PDFName.of("Mask"))) return null;
+
+  const filterNames = filterNamesOf(dict);
+  const lastFilter = filterNames[filterNames.length - 1];
+
+  if (lastFilter === "JPXDecode" || lastFilter === "CCITTFaxDecode" || lastFilter === "JBIG2Decode") {
+    return { unsupported: true, reason: lastFilter };
+  }
+
+  const width = dict.get(PDFName.of("Width"))?.asNumber?.();
+  const height = dict.get(PDFName.of("Height"))?.asNumber?.();
+
+  if (!width || !height) return null;
+
+  const originalBytes = stream.contents.length;
+
+  let bitmap;
+  let hasAlpha = false;
+
+  if (lastFilter === "DCTDecode") {
+    // Already a JPEG file (decodePDFRawStream only inverts general stream
+    // filters like Flate/LZW, it leaves the image codec itself alone).
+    const jpegBytes = decodePDFRawStream(stream).decode();
+    const blob = new Blob([jpegBytes], { type: "image/jpeg" });
+    bitmap = await createImageBitmap(blob);
+  } else {
+    // Raw samples (typically FlateDecode) — decode manually.
+    const bitsPerComponent = dict.get(PDFName.of("BitsPerComponent"))?.asNumber?.() || 8;
+    const colorSpaceObj = dict.get(PDFName.of("ColorSpace"));
+    const colorSpaceInfo = resolveColorSpace(context, colorSpaceObj);
+
+    const rawBytes = decodePDFRawStream(stream).decode();
+    const rgba = samplesToRGBA(rawBytes, width, height, bitsPerComponent, colorSpaceInfo);
+
+    const imageData = new ImageData(rgba, width, height);
+    bitmap = await createImageBitmap(imageData);
+  }
+
+  // Soft mask (alpha channel) — decode and merge in.
+  const smaskRef = dict.get(PDFName.of("SMask"));
+
+  if (smaskRef) {
+    try {
+      const smaskDecoded = await decodeImageXObject(context, smaskRef);
+
+      if (smaskDecoded && !smaskDecoded.unsupported && smaskDecoded.bitmap) {
+        hasAlpha = true;
+
+        const canvas = new OffscreenCanvas(width, height);
+        const ctx = canvas.getContext("2d");
+
+        // Draw color image
+        ctx.drawImage(bitmap, 0, 0, width, height);
+        const base = ctx.getImageData(0, 0, width, height);
+
+        // Draw alpha mask resized to match
+        const alphaCanvas = new OffscreenCanvas(width, height);
+        const alphaCtx = alphaCanvas.getContext("2d");
+        alphaCtx.drawImage(smaskDecoded.bitmap, 0, 0, width, height);
+        const alphaData = alphaCtx.getImageData(0, 0, width, height);
+
+        for (let i = 0; i < width * height; i++) {
+          base.data[i * 4 + 3] = alphaData.data[i * 4]; // gray channel -> alpha
+        }
+
+        bitmap.close();
+        bitmap = await createImageBitmap(base);
+      }
+    } catch (smaskError) {
+      console.warn("compressPDF: SMask decode failed, ignoring alpha:", smaskError);
+    }
+  }
+
+  return { bitmap, width, height, hasAlpha, originalBytes };
+}
+
+/*
+ * ---- Deflate (zlib) bytes using the native Compression Streams API --------
+ */
 
-async function processPage(pdf, pool, pageNumber, totalPages, { quality, resolution, scale: fixedScale }) {
+async function deflateBytes(bytes) {
+  if (typeof CompressionStream === "undefined") {
+    throw new Error("CompressionStream API unavailable — cannot write alpha/PNG-style images");
+  }
+
+  const cs = new CompressionStream("deflate");
+  const writer = cs.writable.getWriter();
+  writer.write(bytes);
+  writer.close();
+
+  const buffer = await new Response(cs.readable).arrayBuffer();
+  return new Uint8Array(buffer);
+}
+
+/*
+ * ---- Replace one image object in place ------------------------------------
+ */
+
+async function replaceImageWithJpeg(context, ref, jpegBytes, width, height) {
+  const dict = context.obj({
+    Type: "XObject",
+    Subtype: "Image",
+    Width: width,
+    Height: height,
+    ColorSpace: "DeviceRGB",
+    BitsPerComponent: 8,
+    Filter: "DCTDecode",
+  });
+
+  context.assign(ref, PDFRawStream.of(dict, jpegBytes));
+}
+
+async function replaceImageWithFlateRGBA(context, ref, rgbaBitmap, width, height) {
+  const canvas = new OffscreenCanvas(width, height);
+  const ctx = canvas.getContext("2d");
+  ctx.drawImage(rgbaBitmap, 0, 0, width, height);
+  const imageData = ctx.getImageData(0, 0, width, height);
+
+  const rgbBytes = new Uint8Array(width * height * 3);
+  const alphaBytes = new Uint8Array(width * height);
+
+  for (let i = 0; i < width * height; i++) {
+    rgbBytes[i * 3] = imageData.data[i * 4];
+    rgbBytes[i * 3 + 1] = imageData.data[i * 4 + 1];
+    rgbBytes[i * 3 + 2] = imageData.data[i * 4 + 2];
+    alphaBytes[i] = imageData.data[i * 4 + 3];
+  }
+
+  const [rgbDeflated, alphaDeflated] = await Promise.all([
+    deflateBytes(rgbBytes),
+    deflateBytes(alphaBytes),
+  ]);
+
+  const smaskDict = context.obj({
+    Type: "XObject",
+    Subtype: "Image",
+    Width: width,
+    Height: height,
+    ColorSpace: "DeviceGray",
+    BitsPerComponent: 8,
+    Filter: "FlateDecode",
+  });
+
+  const smaskRef = context.register(PDFRawStream.of(smaskDict, alphaDeflated));
+
+  const mainDict = context.obj({
+    Type: "XObject",
+    Subtype: "Image",
+    Width: width,
+    Height: height,
+    ColorSpace: "DeviceRGB",
+    BitsPerComponent: 8,
+    Filter: "FlateDecode",
+    SMask: smaskRef,
+  });
+
+  context.assign(ref, PDFRawStream.of(mainDict, rgbDeflated));
+
+  return rgbDeflated.length + alphaDeflated.length;
+}
+
+/*
+ * ---- Find every Image XObject in the document ----------------------------
+ */
+
+function findImageRefs(pdfDoc) {
+  const refs = [];
+
+  for (const [ref, obj] of pdfDoc.context.enumerateIndirectObjects()) {
+    if (!(obj instanceof PDFRawStream)) continue;
+
+    const subtype = obj.dict.get(PDFName.of("Subtype"));
+    if (subtype && subtype.asString?.().replace(/^\//, "") === "Image") {
+      refs.push(ref);
+    }
+  }
+
+  return refs;
+}
+
+/*
+ * ---- compressPDF: the public, recommended API -----------------------------
+ */
+
+/**
+ * Compress a PDF by extracting each embedded raster image, resizing and
+ * recompressing it individually (via a parallel worker pool running mozjpeg
+ * WASM), and writing the result back into the same PDF object slot. Text,
+ * fonts, and vector drawing instructions are left completely untouched.
+ *
+ * @param {File|Blob|ArrayBuffer|ArrayBufferView} input
+ * @param {Object} [options]
+ * @param {number} [options.quality=75] - JPEG quality, 0-100 (only applies to
+ *   opaque images; images with transparency are re-encoded losslessly as
+ *   Flate-compressed raw pixels + an SMask, since JPEG has no alpha channel)
+ * @param {number} [options.scale=1] - resize factor applied to each image's
+ *   OWN native pixel dimensions (not the page). 1 = keep native size and only
+ *   recompress; 0.5 = half width/height; etc. Clamped to 0.05–1.
+ * @param {number} [options.minImageBytes=2048] - skip images already smaller
+ *   than this (not worth the re-encode overhead)
+ * @param {boolean} [options.onlyIfSmaller=true] - keep the original image
+ *   bytes if the recompressed version would end up bigger
+ * @param {number} [options.workers] - override auto-detected worker count
+ * @param {(update: object) => void} [options.onProgress] - progress callback
+ * @returns {Promise<{file: File|Blob, stats: object}>}
+ */
+export async function compressPDF(input, options = {}) {
+  const {
+    quality = 75,
+    scale = 1,
+    minImageBytes = 2048,
+    onlyIfSmaller = true,
+    workers: workerOverride,
+    onProgress,
+  } = options;
+
+  const clampedScale = Math.max(0.05, Math.min(scale, 1));
+
+  const startedAt = performance.now();
+
+  const { arrayBuffer, fileName, fileType, originalBytes } = await normalizeInput(input);
+
+  const pdfDoc = await PDFDocument.load(arrayBuffer, {
+    updateMetadata: false,
+    ignoreEncryption: true,
+  });
+
+  const context = pdfDoc.context;
+  const imageRefs = findImageRefs(pdfDoc);
+
+  const power = getClientPower();
+  const workerCount = workerOverride || power.workers;
+
+  const pool = new CompressionPool();
+  pool.init(workerCount);
+
+  const perImage = [];
+  let imagesCompressed = 0;
+  let imagesSkipped = 0;
+
+  try {
+    let nextIndex = 0;
+    let completed = 0;
+
+    async function runner() {
+      while (true) {
+        const index = nextIndex++;
+        if (index >= imageRefs.length) return;
+
+        const ref = imageRefs[index];
+
+        let entry = {
+          ref: ref.toString(),
+          skipped: null,
+        };
+
+        try {
+          const decoded = await decodeImageXObject(context, ref);
+
+          if (!decoded) {
+            entry.skipped = "unreadable-or-stencil";
+            imagesSkipped++;
+          } else if (decoded.unsupported) {
+            entry.skipped = `unsupported-codec:${decoded.reason}`;
+            imagesSkipped++;
+          } else if (decoded.originalBytes < minImageBytes) {
+            decoded.bitmap.close();
+            entry.skipped = "below-min-size";
+            entry.originalBytes = decoded.originalBytes;
+            imagesSkipped++;
+          } else {
+            const { bitmap, width, height, hasAlpha, originalBytes: origImgBytes } = decoded;
+
+            const newWidth = Math.max(1, Math.round(width * clampedScale));
+            const newHeight = Math.max(1, Math.round(height * clampedScale));
+
+            entry.width = width;
+            entry.height = height;
+            entry.newWidth = newWidth;
+            entry.newHeight = newHeight;
+            entry.originalBytes = origImgBytes;
+
+            let resizedBitmap = bitmap;
+
+            if (newWidth !== width || newHeight !== height) {
+              const resizeCanvas = new OffscreenCanvas(newWidth, newHeight);
+              const resizeCtx = resizeCanvas.getContext("2d");
+              resizeCtx.drawImage(bitmap, 0, 0, newWidth, newHeight);
+              bitmap.close();
+              resizedBitmap = await createImageBitmap(resizeCanvas);
+            }
+
+            if (hasAlpha) {
+              const newBytes = await replaceImageWithFlateRGBA(
+                context,
+                ref,
+                resizedBitmap,
+                newWidth,
+                newHeight
+              );
+
+              resizedBitmap.close();
+
+              entry.format = "flate+smask";
+              entry.compressedBytes = newBytes;
+              imagesCompressed++;
+            } else {
+              const compressed = await pool.run(
+                {
+                  type: "compress-image",
+                  bitmap: resizedBitmap,
+                  quality,
+                  pageNumber: index + 1,
+                  totalPages: imageRefs.length,
+                },
+                [resizedBitmap]
+              );
+
+              const jpegBytes = jpegResultToBytes(compressed);
+
+              if (onlyIfSmaller && jpegBytes.length >= origImgBytes) {
+                entry.skipped = "recompressed-not-smaller";
+                entry.compressedBytes = origImgBytes;
+                imagesSkipped++;
+              } else {
+                await replaceImageWithJpeg(context, ref, jpegBytes, newWidth, newHeight);
+                entry.format = "jpeg";
+                entry.compressedBytes = jpegBytes.length;
+                imagesCompressed++;
+              }
+            }
+          }
+        } catch (error) {
+          console.warn(`compressPDF: skipping image (${ref.toString()}):`, error);
+          entry.skipped = "error";
+          entry.error = error?.message || String(error);
+          imagesSkipped++;
+        }
+
+        perImage.push(entry);
+        completed++;
+
+        if (typeof onProgress === "function") {
+          onProgress({
+            stage: "compressing",
+            imageIndex: index + 1,
+            totalImages: imageRefs.length,
+            completed,
+            progress: imageRefs.length
+              ? Math.round((completed / imageRefs.length) * 90)
+              : 90,
+          });
+        }
+      }
+    }
+
+    const concurrency = Math.max(1, Math.min(workerCount, 4));
+    const runnerCount = Math.min(concurrency, Math.max(imageRefs.length, 1));
+
+    await Promise.all(Array.from({ length: runnerCount }, () => runner()));
+
+    if (typeof onProgress === "function") {
+      onProgress({ stage: "saving", progress: 95 });
+    }
+
+    const pdfBytes = await pdfDoc.save({
+      useObjectStreams: true,
+      addDefaultPage: false,
+    });
+
+    const blob = new Blob([pdfBytes], { type: "application/pdf" });
+    const compressedBytes = blob.size;
+    const savedBytes = Math.max(0, originalBytes - compressedBytes);
+    const reduction = originalBytes > 0 ? (savedBytes / originalBytes) * 100 : 0;
+    const elapsed = performance.now() - startedAt;
+
+    const stats = {
+      pages: pdfDoc.getPageCount(),
+      imagesFound: imageRefs.length,
+      imagesCompressed,
+      imagesSkipped,
+      originalBytes,
+      compressedBytes,
+      savedBytes,
+      reduction,
+      elapsed,
+      workersUsed: workerCount,
+      perImage,
+    };
+
+    if (typeof onProgress === "function") {
+      onProgress({ stage: "complete", progress: 100 });
+    }
+
+    const file =
+      typeof File !== "undefined"
+        ? new File([blob], fileName, { type: fileType })
+        : blob;
+
+    return { file, stats };
+  } finally {
+    pool.destroy();
+  }
+}
+
+/* ============================================================================
+   STRATEGY 2 (LEGACY): FULL-PAGE RASTERIZATION
+   Kept for cases where the whole page really is a single scanned image and
+   there's nothing for compressPDF's per-image extraction to find separately.
+============================================================================ */
+
+async function rasterizeProcessPage(pdf, pool, pageNumber, totalPages, { quality, resolution, scale: fixedScale }) {
   let page = null;
   let canvas = null;
   let bitmap = null;
@@ -255,19 +944,8 @@ async function processPage(pdf, pool, pageNumber, totalPages, { quality, resolut
     let scale;
 
     if (typeof fixedScale === "number") {
-      /*
-       * Direct scale multiplier — same units as pdf.js's own
-       * page.getViewport({ scale }). 1 = native ~72 DPI, 2 = ~144 DPI,
-       * 3 = ~216 DPI (max allowed). Proportional across any page size,
-       * unlike a fixed pixel target.
-       */
       scale = Math.max(0.25, Math.min(fixedScale, 3));
     } else if (resolution === "original" || resolution === Infinity) {
-      /*
-       * No downscaling — render at the max allowed multiplier so page
-       * quality is limited only by the JPEG quality setting, not by
-       * resizing.
-       */
       scale = 3;
     } else {
       scale = resolution / largestDimension;
@@ -341,11 +1019,7 @@ async function processPage(pdf, pool, pageNumber, totalPages, { quality, resolut
   }
 }
 
-/* ============================================================
-   PARALLEL PAGE PIPELINE
-============================================================ */
-
-async function processPages(pdf, pool, totalPages, { quality, resolution, scale, workers, onProgress }) {
+async function rasterizeProcessPages(pdf, pool, totalPages, { quality, resolution, scale, workers, onProgress }) {
   const results = new Array(totalPages);
 
   const renderConcurrency = Math.max(1, Math.min(workers, 4));
@@ -361,7 +1035,7 @@ async function processPages(pdf, pool, totalPages, { quality, resolution, scale,
         return;
       }
 
-      const result = await processPage(pdf, pool, pageNumber, totalPages, {
+      const result = await rasterizeProcessPage(pdf, pool, pageNumber, totalPages, {
         quality,
         resolution,
         scale,
@@ -389,11 +1063,7 @@ async function processPages(pdf, pool, totalPages, { quality, resolution, scale,
   return results;
 }
 
-/* ============================================================
-   BUILD FINAL PDF
-============================================================ */
-
-async function buildPdf(compressedPages, { originalBytes, startedAt, workersUsed, onProgress }) {
+async function rasterizeBuildPdf(compressedPages, { originalBytes, startedAt, workersUsed, onProgress }) {
   if (typeof onProgress === "function") {
     onProgress({ stage: "building", progress: 92 });
   }
@@ -409,17 +1079,7 @@ async function buildPdf(compressedPages, { originalBytes, startedAt, workersUsed
       throw new Error(`Missing compressed page ${i + 1}`);
     }
 
-    let jpegBytes;
-
-    if (item.jpeg instanceof ArrayBuffer) {
-      jpegBytes = new Uint8Array(item.jpeg);
-    } else if (item.jpeg instanceof Uint8Array) {
-      jpegBytes = item.jpeg;
-    } else if (ArrayBuffer.isView(item.jpeg)) {
-      jpegBytes = new Uint8Array(item.jpeg.buffer, item.jpeg.byteOffset, item.jpeg.byteLength);
-    } else {
-      throw new Error(`Invalid JPEG for page ${i + 1}`);
-    }
+    const jpegBytes = jpegResultToBytes(item);
 
     const image = await outputPdf.embedJpg(jpegBytes);
 
@@ -484,31 +1144,19 @@ async function buildPdf(compressedPages, { originalBytes, startedAt, workersUsed
   return { blob, stats };
 }
 
-/* ============================================================
-   PUBLIC API
-============================================================ */
-
 /**
- * Compress a PDF using the same render -> mozjpeg WASM -> pdf-lib rebuild
- * pipeline as the original app, run across a parallel worker pool.
+ * Legacy full-page rasterization strategy. Renders every page to a bitmap and
+ * rebuilds the PDF from those images — this DOES discard text/vector content
+ * in favor of pictures of it, so prefer compressPDF() unless you specifically
+ * want this (e.g. flattening a document, or it's already scan-only).
+ *
+ * Same options as before: quality, resolution ("original" to skip
+ * downscaling), scale (direct render multiplier, overrides resolution),
+ * workers, onProgress.
  *
- * @param {File|Blob|ArrayBuffer|ArrayBufferView} input
- * @param {Object} [options]
- * @param {number} [options.quality=65] - JPEG quality, 0-100
- * @param {number|"original"} [options.resolution=1600] - max px on the longest
- *   page side. Pass "original" (or Infinity) to skip downscaling entirely and
- *   rely on `quality` alone — renders at the max supported multiplier (3x).
- *   Ignored if `scale` is set.
- * @param {number} [options.scale] - direct render-scale multiplier (same units
- *   as pdf.js's `page.getViewport({ scale })`), clamped to 0.25–3.
- *   1 = native ~72 DPI, 2 ≈ 144 DPI, 3 ≈ 216 DPI (max). Proportional across
- *   any page size, unlike `resolution`'s fixed pixel target. Takes
- *   precedence over `resolution` when provided.
- * @param {number} [options.workers] - override auto-detected worker count
- * @param {(update: object) => void} [options.onProgress] - optional progress callback
  * @returns {Promise<{file: File|Blob, stats: object}>}
  */
-export async function compressPDF(input, options = {}) {
+export async function rasterizePDF(input, options = {}) {
   const {
     quality = 65,
     resolution = 1600,
@@ -537,7 +1185,7 @@ export async function compressPDF(input, options = {}) {
     const pdf = await loadingTask.promise;
     const totalPages = pdf.numPages;
 
-    const compressedPages = await processPages(pdf, pool, totalPages, {
+    const compressedPages = await rasterizeProcessPages(pdf, pool, totalPages, {
       quality,
       resolution,
       scale,
@@ -545,7 +1193,7 @@ export async function compressPDF(input, options = {}) {
       onProgress,
     });
 
-    const { blob, stats } = await buildPdf(compressedPages, {
+    const { blob, stats } = await rasterizeBuildPdf(compressedPages, {
       originalBytes,
       startedAt,
       workersUsed: workerCount,
@@ -564,7 +1212,7 @@ export async function compressPDF(input, options = {}) {
         await loadingTask.destroy();
       }
     } catch (cleanupError) {
-      console.warn("compressPDF: loading task cleanup skipped:", cleanupError);
+      console.warn("rasterizePDF: loading task cleanup skipped:", cleanupError);
     }
 
     pool.destroy();

package/src/index.js

@@ -1 +1,6 @@
-export { compressPDF, getClientPower, CompressionPool } from "./compress.js";
+export {
+  compressPDF,
+  rasterizePDF,
+  getClientPower,
+  CompressionPool,
+} from "./compress.js";