diff --git a/BPTCEncoder/src/Compressor.cpp b/BPTCEncoder/src/Compressor.cpp
index 0d2ff6b..2c8d928 100755
--- a/BPTCEncoder/src/Compressor.cpp
+++ b/BPTCEncoder/src/Compressor.cpp
@@ -1572,8 +1572,9 @@ void Compress(const FasTC::CompressionJob &cj, CompressionSettings settings) {
   const uint32 kBlockSz = GetBlockSize(FasTC::eCompressionFormat_BPTC);
   uint8 *outBuf = cj.OutBuf() + cj.CoordsToBlockIdx(cj.XStart(), cj.YStart()) * kBlockSz;
 
+  const uint32 endY = std::min(cj.YEnd(), cj.Height() - 4);
   uint32 startX = cj.XStart();
-  for(uint32 j = cj.YStart(); j <= cj.YEnd(); j += 4) {
+  for(uint32 j = cj.YStart(); j <= endY; j += 4) {
     const uint32 endX = j == cj.YEnd()? cj.XEnd() : cj.Width();
     for(uint32 i = startX; i < endX; i += 4) {
 
diff --git a/DXTEncoder/src/DXTCompressor.cpp b/DXTEncoder/src/DXTCompressor.cpp
index c98bc80..2f1ccd1 100755
--- a/DXTEncoder/src/DXTCompressor.cpp
+++ b/DXTEncoder/src/DXTCompressor.cpp
@@ -14,6 +14,7 @@
 // algorithms used in this code.
 
 #include "DXTCompressor.h"
+#include <algorithm>
 #include <cassert>
 #include <cstdlib>
 #include <cstring>
@@ -45,8 +46,9 @@ namespace DXTC
     uint8 *outBuf = cj.OutBuf() + startBlock * kBlockSz;
 
     const uint32 *inPixels = reinterpret_cast<const uint32 *>(cj.InBuf());
+    uint32 endY = std::min(cj.YEnd(), cj.Height() - 4);
     uint32 startX = cj.XStart();
-    for(uint32 j = cj.YStart(); j <= cj.YEnd(); j += 4) {
+    for(uint32 j = cj.YStart(); j <= endY; j += 4) {
       const uint32 endX = j == cj.YEnd()? cj.XEnd() : cj.Width();
       for(uint32 i = startX; i < endX; i += 4) {
 
@@ -75,10 +77,11 @@ namespace DXTC
     uint8 *outBuf = cj.OutBuf() + startBlock * kBlockSz;
     
     const uint32 *inPixels = reinterpret_cast<const uint32 *>(cj.InBuf());
+    uint32 endY = std::min(cj.YEnd(), cj.Height() - 4);
     uint32 startX = cj.XStart();
-    bool done = false;
-    for(uint32 j = cj.YStart(); !done; j += 4) {
-      for(uint32 i = startX; !done && i < cj.Width(); i += 4) {
+    for(uint32 j = cj.YStart(); j <= endY; j += 4) {
+      const uint32 endX = j == cj.YEnd()? cj.XEnd() : cj.Width();
+      for(uint32 i = startX; i < endX; i += 4) {
 
         const uint32 kOffset = j*cj.Width() + i;
         ExtractBlock(inPixels + kOffset, cj.Width(), block);
@@ -89,8 +92,8 @@ namespace DXTC
         EmitWord(outBuf, ColorTo565(maxColor));
         EmitWord(outBuf, ColorTo565(minColor));
         EmitColorIndices(block, outBuf, minColor, maxColor);
-        done = i+4 >= cj.XEnd() && j+(i+4 == cj.Width()? 4 : 0) >= cj.YEnd();
       }
+      startX = 0;
     }
   }
 
diff --git a/ETCEncoder/src/Compressor.cpp b/ETCEncoder/src/Compressor.cpp
index 90b4596..5afbb14 100644
--- a/ETCEncoder/src/Compressor.cpp
+++ b/ETCEncoder/src/Compressor.cpp
@@ -52,6 +52,7 @@
 
 #include "rg_etc1.h"
 #include "ETCCompressor.h"
+#include <algorithm>
 #include <cstring>
 
 namespace ETCC {
@@ -62,11 +63,13 @@ namespace ETCC {
     params.m_quality = rg_etc1::cLowQuality;
     rg_etc1::pack_etc1_block_init();
 
-    const uint32 kBlockSz = GetBlockSize(FasTC::eCompressionFormat_ETC1);
+    uint32 kBlockSz = GetBlockSize(FasTC::eCompressionFormat_ETC1);
     const uint32 startBlock = cj.CoordsToBlockIdx(cj.XStart(), cj.YStart());
     uint8 *outBuf = cj.OutBuf() + startBlock * kBlockSz;
+
+    const uint32 endY = std::min(cj.YEnd(), cj.Height() - 4);
     uint32 startX = cj.XStart();
-    for(uint32 j = cj.YStart(); j <= cj.YEnd(); j += 4) {
+    for(uint32 j = cj.YStart(); j <= endY; j += 4) {
       const uint32 endX = j == cj.YEnd()? cj.XEnd() : cj.Width();
       for(uint32 i = startX; i < endX; i += 4) {