diff --git a/CLTool/src/clunix.cpp b/CLTool/src/clunix.cpp
index aa39d52..a064c1f 100644
--- a/CLTool/src/clunix.cpp
+++ b/CLTool/src/clunix.cpp
@@ -53,6 +53,7 @@
 void PrintUsage() {
   fprintf(stderr, "Usage: tc [OPTIONS] imagefile\n");
   fprintf(stderr, "\n");
+  fprintf(stderr, "\t-f\t\tFormat to use. Either \"BPTC\" or \"PVRTC\". Default: BPTC\n");
   fprintf(stderr, "\t-l\t\tSave an output log.\n");
   fprintf(stderr, "\t-q <quality>\tSet compression quality level. Default: 50\n");
   fprintf(stderr, "\t-n <num>\tCompress the image num times and give the average time and PSNR. Default: 1\n");
@@ -92,7 +93,8 @@ int main(int argc, char **argv) {
   bool bUseSIMD = false;
   bool bSaveLog = false;
   bool bUseAtomics = false;
-  
+  ECompressionFormat format = eCompressionFormat_BPTC;
+
   bool knowArg = false;
   do {
     knowArg = false;
@@ -110,6 +112,23 @@ int main(int argc, char **argv) {
       continue;
     }
 
+    if(strcmp(argv[fileArg], "-f") == 0) {
+      fileArg++;
+
+      if(fileArg == argc) {
+        PrintUsage();
+        exit(1);
+      } else {
+        if(!strcmp(argv[fileArg], "PVRTC")) {
+          format = eCompressionFormat_PVRTC;
+        }
+      }
+
+      fileArg++;
+      knowArg = true;
+      continue;
+    }
+
     if(strcmp(argv[fileArg], "-l") == 0) {
       fileArg++;
       bSaveLog = true;
@@ -195,6 +214,7 @@ int main(int argc, char **argv) {
   }
   
   SCompressionSettings settings;
+  settings.format = format;
   settings.bUseSIMD = bUseSIMD;
   settings.bUseAtomics = bUseAtomics;
   settings.iNumThreads = numThreads;
diff --git a/Core/CMakeLists.txt b/Core/CMakeLists.txt
index 0e33516..21d1f2e 100644
--- a/Core/CMakeLists.txt
+++ b/Core/CMakeLists.txt
@@ -71,6 +71,9 @@ ELSE()
 	SET( LINK_FLAGS -lrt ${LINK_FLAGS} )
 ENDIF()
 
+INCLUDE_DIRECTORIES( ${FasTC_SOURCE_DIR} )
+INCLUDE_DIRECTORIES( ${FasTC_SOURCE_DIR}/PVRTCEncoder/include )
+
 INCLUDE_DIRECTORIES( ${FasTC_SOURCE_DIR}/BPTCEncoder/include )
 INCLUDE_DIRECTORIES( ${FasTC_BINARY_DIR}/BPTCEncoder/include )
 
@@ -149,6 +152,7 @@ ADD_LIBRARY( FasTCCore
 
 TARGET_LINK_LIBRARIES( FasTCCore FasTCIO )
 TARGET_LINK_LIBRARIES( FasTCCore BPTCEncoder )
+TARGET_LINK_LIBRARIES( FasTCCore PVRTCEncoder )
 
 IF( THREAD_API MATCHES "Boost" )
 	TARGET_LINK_LIBRARIES( FasTCCore ${Boost_LIBRARIES} )
diff --git a/Core/include/CompressedImage.h b/Core/include/CompressedImage.h
index 831a50c..8799d11 100644
--- a/Core/include/CompressedImage.h
+++ b/Core/include/CompressedImage.h
@@ -44,10 +44,13 @@
 #ifndef _COMPRESSED_IMAGE_H_
 #define _COMPRESSED_IMAGE_H_
 
+#include "TexCompTypes.h"
+
 enum ECompressionFormat {
   eCompressionFormat_DXT1,
   eCompressionFormat_DXT5,
   eCompressionFormat_BPTC,
+  eCompressionFormat_PVRTC,
 
   kNumCompressionFormats
 };
@@ -55,15 +58,15 @@ enum ECompressionFormat {
 class CompressedImage {
 
  private:
-  unsigned int m_Width;
-  unsigned int m_Height;
+  uint32 m_Width;
+  uint32 m_Height;
 
   ECompressionFormat m_Format;
 
-  unsigned char *m_Data;
-  unsigned int m_DataSz;
+  uint8 *m_Data;
+  uint32 m_DataSz;
 
-  void InitData(const unsigned char *withData);
+  void InitData(const uint8 *withData);
  public:
   CompressedImage();
 
@@ -71,23 +74,29 @@ class CompressedImage {
   // the passed format. The size of the data is expected to conform
   // to the width, height, and format specified.
   CompressedImage(
-    const unsigned int width, 
-    const unsigned int height, 
+    const uint32 width, 
+    const uint32 height, 
     const ECompressionFormat format, 
-    const unsigned char *data
+    const uint8 *data
   );
 
-  unsigned int GetHeight() const { return m_Height; }
-  unsigned int GetWidth() const { return m_Width; }
+  uint32 GetHeight() const { return m_Height; }
+  uint32 GetWidth() const { return m_Width; }
 
   CompressedImage( const CompressedImage &other );
   ~CompressedImage();
 
+  static uint32 GetCompressedSize(uint32 uncompressedSize, ECompressionFormat format);
+  static uint32 GetUncompressedSize(uint32 compressedSize, ECompressionFormat format) {
+    uint32 cmp = GetCompressedSize(compressedSize, format);
+    return compressedSize * (compressedSize / cmp);
+  }
+
   // Decompress the compressed image data into outBuf. outBufSz is expected
   // to be the proper size determined by the width, height, and format.
   // !FIXME! We should have a function to explicitly return the in/out buf
   // size for a given compressed image.
-  bool DecompressImage(unsigned char *outBuf, unsigned int outBufSz) const;
+  bool DecompressImage(uint8 *outBuf, uint32 outBufSz) const;
 };
 
 #endif // _COMPRESSED_IMAGE_H_
diff --git a/Core/include/TexComp.h b/Core/include/TexComp.h
index 1ad8c84..873a228 100644
--- a/Core/include/TexComp.h
+++ b/Core/include/TexComp.h
@@ -93,7 +93,8 @@ struct SCompressionSettings {
 
 extern bool CompressImageData(
   const unsigned char *data,
-  const unsigned int dataSz,
+  const unsigned int width,
+  const unsigned int height,
   unsigned char *cmpData,
   const unsigned int cmpDataSz,
   const SCompressionSettings &settings
diff --git a/Core/src/CompressedImage.cpp b/Core/src/CompressedImage.cpp
index 0f4f393..a30d9ba 100644
--- a/Core/src/CompressedImage.cpp
+++ b/Core/src/CompressedImage.cpp
@@ -50,6 +50,7 @@
 
 #include "TexCompTypes.h"
 #include "BC7Compressor.h"
+#include "PVRTCCompressor.h"
 
 CompressedImage::CompressedImage()
   : m_Width(0)
@@ -85,16 +86,7 @@ CompressedImage::CompressedImage(
 }
 
 void CompressedImage::InitData(const unsigned char *withData) {
-  m_DataSz = 0;
-  int uncompDataSz = m_Width * m_Height * 4;
-
-  switch(m_Format) {
-    default: assert(!"Not implemented!"); // Fall through V
-    case eCompressionFormat_DXT1: m_DataSz = uncompDataSz / 8; break;
-    case eCompressionFormat_DXT5: m_DataSz = uncompDataSz / 4; break;
-    case eCompressionFormat_BPTC: m_DataSz = uncompDataSz / 4; break;
-  }
-
+  m_DataSz = GetCompressedSize(m_Width * m_Height * 4, m_Format);
   if(m_DataSz > 0) {
     m_Data = new unsigned char[m_DataSz];
     memcpy(m_Data, withData, m_DataSz);
@@ -111,14 +103,7 @@ CompressedImage::~CompressedImage() {
 bool CompressedImage::DecompressImage(unsigned char *outBuf, unsigned int outBufSz) const {
 
   // First make sure that we have enough data
-  uint32 dataSz = 0;
-  switch(m_Format) {
-    default: assert(!"Not implemented!"); // Fall through V
-    case eCompressionFormat_DXT1: dataSz = m_DataSz * 8; break;
-    case eCompressionFormat_DXT5: dataSz = m_DataSz * 4; break;
-    case eCompressionFormat_BPTC: dataSz = m_DataSz * 4; break;
-  }
-
+  uint32 dataSz = GetUncompressedSize(m_DataSz, m_Format);
   if(dataSz > outBufSz) {
     fprintf(stderr, "Not enough space to store entire decompressed image! "
                     "Got %d bytes, but need %d!\n", outBufSz, dataSz);
@@ -126,6 +111,13 @@ bool CompressedImage::DecompressImage(unsigned char *outBuf, unsigned int outBuf
   }
 
   switch(m_Format) {
+  case eCompressionFormat_PVRTC:
+    {
+      DecompressionJob dj (m_Data, outBuf, m_Width, m_Height);
+      PVRTCC::Decompress(dj);
+    }
+    break;
+
   case eCompressionFormat_BPTC: 
     { 
       DecompressionJob dj (m_Data, outBuf, m_Width, m_Height);
@@ -142,3 +134,26 @@ bool CompressedImage::DecompressImage(unsigned char *outBuf, unsigned int outBuf
 
   return true;
 }
+
+uint32 CompressedImage::GetCompressedSize(uint32 uncompressedSize, ECompressionFormat format) {
+  assert(uncompressedSize % 8 == 0);
+
+  uint32 cmpDataSzNeeded = 0;
+  switch(format) {
+  default:
+    assert(!"Not implemented!");
+    // Fall through V
+  case eCompressionFormat_DXT1:
+  case eCompressionFormat_PVRTC:
+    cmpDataSzNeeded = uncompressedSize / 8;
+    break;
+
+  case eCompressionFormat_DXT5:
+  case eCompressionFormat_BPTC:
+    cmpDataSzNeeded = uncompressedSize / 4;
+    break;
+  }
+
+  return cmpDataSzNeeded;
+}
+
diff --git a/Core/src/Image.cpp b/Core/src/Image.cpp
index 663469b..a918706 100644
--- a/Core/src/Image.cpp
+++ b/Core/src/Image.cpp
@@ -149,16 +149,10 @@ CompressedImage *Image::Compress(const SCompressionSettings &settings) const {
   assert(dataSz > 0);
 
   // Allocate data based on the compression method
-  int cmpDataSz = 0;
-  switch(settings.format) {
-    default: assert(!"Not implemented!"); // Fall Through V
-    case eCompressionFormat_DXT1: cmpDataSz = dataSz / 8; break;
-    case eCompressionFormat_DXT5: cmpDataSz = dataSz / 4; break;
-    case eCompressionFormat_BPTC: cmpDataSz = dataSz / 4; break;
-  }
+  int cmpDataSz = CompressedImage::GetCompressedSize(dataSz, settings.format);
 
   unsigned char *cmpData = new unsigned char[cmpDataSz];
-  CompressImageData(m_PixelData, dataSz, cmpData, cmpDataSz, settings);
+  CompressImageData(m_PixelData, GetWidth(), GetHeight(), cmpData, cmpDataSz, settings);
 
   outImg = new CompressedImage(GetWidth(), GetHeight(), settings.format, cmpData);
   
diff --git a/Core/src/TexComp.cpp b/Core/src/TexComp.cpp
index 3cf0bf5..3cc25ac 100644
--- a/Core/src/TexComp.cpp
+++ b/Core/src/TexComp.cpp
@@ -49,6 +49,7 @@
 #include <cstdio>
 #include <cassert>
 
+#include "PVRTCCompressor.h"
 #include "BC7Compressor.h"
 #include "Thread.h"
 #include "WorkerQueue.h"
@@ -68,6 +69,10 @@ static inline T sad(const T &a, const T &b) {
   return (a > b)? a - b : b - a;
 }
 
+static void CompressPVRTC(const CompressionJob &cj) {
+  PVRTCC::Compress(cj);
+}
+
 SCompressionSettings:: SCompressionSettings()
   : format(eCompressionFormat_BPTC)
   , bUseSIMD(false)
@@ -86,6 +91,12 @@ static  CompressionFuncWithStats ChooseFuncFromSettingsWithStats(const SCompress
     }
     break;
 
+    case eCompressionFormat_PVRTC:
+    {
+      // !FIXME! actually implement one of these methods...
+      return NULL;
+    }
+
     default:
     {
       assert(!"Not implemented!");
@@ -109,6 +120,11 @@ static CompressionFunc ChooseFuncFromSettings(const SCompressionSettings &s) {
     }
     break;
 
+    case eCompressionFormat_PVRTC:
+    {
+      return CompressPVRTC;
+    }
+
     default:
     {
       assert(!"Not implemented!");
@@ -123,8 +139,9 @@ static void ReportError(const char *msg) {
 }
 
 static double CompressImageInSerial(
-  const unsigned char *imgData,
-  const unsigned int imgDataSz,
+  const uint8 *imgData,
+  const uint32 imgWidth,
+  const uint32 imgHeight,
   const SCompressionSettings &settings,
   unsigned char *outBuf
 ) {
@@ -133,14 +150,14 @@ static double CompressImageInSerial(
 
   double cmpTimeTotal = 0.0;
 
+  StopWatch stopWatch = StopWatch();
   for(int i = 0; i < settings.iNumCompressions; i++) {
 
-    StopWatch stopWatch = StopWatch();
     stopWatch.Reset();
     stopWatch.Start();
 
     // !FIXME! We're assuming that we have 4x4 blocks here...
-    CompressionJob cj (imgData, outBuf, imgDataSz / 16, 4);
+    CompressionJob cj (imgData, outBuf, imgWidth, imgHeight);
     if(fStats && settings.pStatManager) {
       (*fStats)(cj, *(settings.pStatManager));
     }
@@ -332,12 +349,15 @@ static double CompressImageWithWorkerQueue(
 
 bool CompressImageData(
   const unsigned char *data, 
-  const unsigned int dataSz,
+  const unsigned int width,
+  const unsigned int height,
   unsigned char *cmpData,
   const unsigned int cmpDataSz,
   const SCompressionSettings &settings
 ) { 
 
+  uint32 dataSz = width * height * 4;
+
   // Make sure that platform supports SSE if they chose this
   // option...
   #ifndef HAS_SSE_41
@@ -352,15 +372,23 @@ bool CompressImageData(
     return false;
   }
 
-  // Allocate data based on the compression method
-  uint32 cmpDataSzNeeded = 0;
-  switch(settings.format) {
-    default: assert(!"Not implemented!"); // Fall through V
-    case eCompressionFormat_DXT1: cmpDataSzNeeded = dataSz / 8; break;
-    case eCompressionFormat_DXT5: cmpDataSzNeeded = dataSz / 4; break;
-    case eCompressionFormat_BPTC: cmpDataSzNeeded = dataSz / 4; break;
+  uint32 numThreads = settings.iNumThreads;
+  if(settings.format == eCompressionFormat_PVRTC &&
+     (settings.iNumThreads > 1 || settings.pStatManager)) {
+    if(settings.iNumThreads > 1) {
+      ReportError("WARNING - PVRTC compressor does not support multithreading.");
+      numThreads = 1;
+    }
+
+    if(settings.pStatManager) {
+      ReportError("WARNING - PVRTC compressor does not support stat collection.");
+    }
   }
 
+  // Allocate data based on the compression method
+  uint32 cmpDataSzNeeded =
+    CompressedImage::GetCompressedSize(dataSz, settings.format);
+
   if(cmpDataSzNeeded == 0) {
     ReportError("Unknown compression format");
     return false;
@@ -375,21 +403,17 @@ bool CompressImageData(
 
     double cmpMSTime = 0.0;
 
-    if(settings.iNumThreads > 1) {
+    if(numThreads > 1) {
       if(settings.bUseAtomics) {
-        //!KLUDGE!
-        unsigned int height = 4;
-        unsigned int width = dataSz / 16;
-
         cmpMSTime = CompressImageWithAtomics(data, width, height, settings, cmpData);
-      }
-      else if(settings.iJobSize > 0)
+      } else if(settings.iJobSize > 0) {
         cmpMSTime = CompressImageWithWorkerQueue(data, dataSz, settings, cmpData);
-      else
+      } else {
         cmpMSTime = CompressImageWithThreads(data, dataSz, settings, cmpData);
+      }
     }
     else {
-      cmpMSTime = CompressImageInSerial(data, dataSz, settings, cmpData);
+      cmpMSTime = CompressImageInSerial(data, width, height, settings, cmpData);
     }
 
     // Report compression time
diff --git a/PVRTCEncoder/include/PVRTCCompressor.h b/PVRTCEncoder/include/PVRTCCompressor.h
index c9001cd..657c71a 100644
--- a/PVRTCEncoder/include/PVRTCCompressor.h
+++ b/PVRTCEncoder/include/PVRTCCompressor.h
@@ -76,7 +76,7 @@ namespace PVRTCC {
   // Takes a stream of uncompressed RGBA8 data and compresses it into PVRTC
   // version one. The width and height must be specified in order to properly
   // decompress the data.
-  void Compress(const DecompressionJob &,
+  void Compress(const CompressionJob &,
                 bool bTwoBitMode = false,
                 const EWrapMode wrapMode = eWrapMode_Wrap);
 
diff --git a/PVRTCEncoder/src/Compressor.cpp b/PVRTCEncoder/src/Compressor.cpp
index b17484f..b26c51e 100644
--- a/PVRTCEncoder/src/Compressor.cpp
+++ b/PVRTCEncoder/src/Compressor.cpp
@@ -57,20 +57,52 @@
 
 namespace PVRTCC {
 
-  void Compress(const DecompressionJob &dcj,
+  void Compress(const CompressionJob &dcj,
                 bool bTwoBitMode,
                 const EWrapMode wrapMode) {
     Image img(dcj.height, dcj.width);
-    for(uint32 j = 0; j < dcj.height; j++) {
-      for(uint32 i = 0; i < dcj.width; i++) {
-        const uint32 *pixels = reinterpret_cast<const uint32 *>(dcj.inBuf);
-        img(i, j).UnpackRGBA(pixels[j * dcj.width + i]);
-      }
+    uint32 nPixels = dcj.height * dcj.width;
+    for(uint32 i = 0; i < nPixels; i++) {
+      // Assume block stream order (whyyyy)
+      uint32 blockIdx = i / 16;
+      uint32 blockWidth = dcj.width / 4;
+      uint32 blockX = blockIdx % blockWidth;
+      uint32 blockY = blockIdx / blockWidth;
+
+      uint32 x = blockX * 4 + (i % 4);
+      uint32 y = blockY * 4 + (i % 16) / 4;
+
+      const uint32 *pixels = reinterpret_cast<const uint32 *>(dcj.inBuf);
+      img(x, y).UnpackRGBA(pixels[i]);      
     }
 
+    Image original = img;
+    img.DebugOutput("Original");
+
     // Downscale it using anisotropic diffusion based scheme in order to preserve
     // image features, then reupscale and compute deltas. Use deltas to generate
     // initial A & B images followed by modulation data.
+    img.ContentAwareDownscale(1, 1, eWrapMode_Wrap, true);
+    img.DebugOutput("DownscaledOnce");
+    img.ContentAwareDownscale(1, 1, eWrapMode_Wrap, false);
+    img.DebugOutput("DownscaledTwice");
+
+    Image downscaled = img;
+
+    // Upscale it again
+    img.BilinearUpscale(2, 2, eWrapMode_Wrap);
+
+    img.DebugOutput("Reconstruction");
+
+    // Compute difference...
+    Image difference = img;
+    for(uint32 j = 0; j < dcj.height; j++) {
+      for(uint32 i = 0; i < dcj.width; i++) {
+        for(uint32 c = 0; c < 4; c++) {
+          difference(i, j).Component(c) -= img(i, j).Component(c);
+        }
+      }
+    }
   }
 
 }  // namespace PVRTCC