diff --git a/win32ss/gdi/eng/surface.c b/win32ss/gdi/eng/surface.c
index 3bfe6f89276..524ec7921ec 100644
--- a/win32ss/gdi/eng/surface.c
+++ b/win32ss/gdi/eng/surface.c
@@ -127,6 +127,10 @@ SURFACE_AllocSurface(
     PSURFACE psurf;
     SURFOBJ *pso;
     PVOID pvSection;
+    BOOLEAN bExtraBitsNeeded = FALSE;
+
+    DPRINT("cy '%ld' iFormat '%ld' cjWidth '%ld' cjBufSize '%ld' pvBits '%p'\n",
+           cy, iFormat, cjWidth, cjBufSize, pvBits);
 
     NT_ASSERT(!pvBits || (iType == STYPE_BITMAP));
     NT_ASSERT((iFormat <= BMF_32BPP) || (cjBufSize != 0));
@@ -171,9 +175,9 @@ SURFACE_AllocSurface(
             /* Make sure the buffer is large enough */
             if (cjBufSize < cjBits)
             {
-                DPRINT1("Buffer is too small, required: %lu, got %lu\n",
-                        cjBits, cjBufSize);
-                return NULL;
+                DPRINT("Buffer is too small, required: %lu, got %lu\n",
+                       cjBits, cjBufSize);
+                bExtraBitsNeeded = TRUE;
             }
         }
     }
@@ -185,8 +189,9 @@ SURFACE_AllocSurface(
     }
 
     /* Check if we need an extra large object */
-    if ((iType == STYPE_BITMAP) && (pvBits == NULL) &&
-        !(fjBitmap & BMF_USERMEM) && !(fjBitmap & BMF_KMSECTION))
+    if (((iType == STYPE_BITMAP) && (pvBits == NULL) &&
+        !(fjBitmap & BMF_USERMEM) && !(fjBitmap & BMF_KMSECTION)) ||
+        bExtraBitsNeeded)
     {
         /* Allocate an object large enough to hold the bits */
         cjObject = sizeof(SURFACE) + cjBits;
diff --git a/win32ss/gdi/gdi32/objects/bitmap.c b/win32ss/gdi/gdi32/objects/bitmap.c
index b029d4ad2df..c858710aeac 100644
--- a/win32ss/gdi/gdi32/objects/bitmap.c
+++ b/win32ss/gdi/gdi32/objects/bitmap.c
@@ -577,6 +577,9 @@ SetDIBits(
     INT LinesCopied = 0;
     BOOL newDC = FALSE;
 
+    if (fuColorUse != DIB_RGB_COLORS && fuColorUse != DIB_PAL_COLORS)
+        return 0;
+
     if (!lpvBits || (GDI_HANDLE_GET_TYPE(hBitmap) != GDI_OBJECT_TYPE_BITMAP))
         return 0;
 
@@ -590,6 +593,16 @@ SetDIBits(
         }
     }
 
+    if (lpbmi->bmiHeader.biCompression == BI_BITFIELDS)
+    {
+        DWORD *masks = (DWORD *)lpbmi->bmiColors;
+        if (!masks[0] || !masks[1] || !masks[2])
+        {
+            SetLastError( ERROR_INVALID_PARAMETER );
+            return 0;
+        }
+    }
+
     hDCc = NtGdiGetDCforBitmap(hBitmap); // hDC can be NULL, so, get it from the bitmap.
     SavehDC = hDCc;
     if (!hDCc) // No DC associated with bitmap, Clone or Create one.
@@ -666,10 +679,28 @@ SetDIBitsToDevice(
     UINT cjBmpScanSize = 0;
     BOOL Hit = FALSE;
     PVOID pvSafeBits = (PVOID) Bits;
+    UINT bmiHeight;
+    BOOL top_down;
+    INT src_y = 0;
+
+    #define MaxScanLines 0x2000
+    #define aligned(Address, Count) \
+        (((ULONG_PTR)(void *)(Address)) % (Count) == 0)
 
     if (!ScanLines || !lpbmi || !Bits)
         return 0;
 
+    DPRINT("StartScan '%d' ScanLines '%d' Bits '%p'\n"
+           "    lpbmi '%p' ColorUse '%d' SizeImage '%d'\n"
+           "    biHeight '%d' biWidth '%d' biBitCount '%d'\n",
+           StartScan, ScanLines, Bits, lpbmi, ColorUse,
+           lpbmi->bmiHeader.biSizeImage,
+           lpbmi->bmiHeader.biHeight, lpbmi->bmiHeader.biWidth,
+           lpbmi->bmiHeader.biBitCount);
+
+    if (lpbmi->bmiHeader.biWidth < 0)
+        return 0;
+
     if (ColorUse && ColorUse != DIB_PAL_COLORS && ColorUse != DIB_PAL_COLORS + 1)
         return 0;
 
@@ -677,6 +708,81 @@ SetDIBitsToDevice(
     if (!pConvertedInfo)
         return 0;
 
+    bmiHeight = abs(pConvertedInfo->bmiHeader.biHeight);
+    top_down = (pConvertedInfo->bmiHeader.biHeight < 0);
+    if ((StartScan > bmiHeight) && (ScanLines > bmiHeight))
+    {
+        DPRINT("Returning ScanLines of '%d'.\n", ScanLines);
+        LinesCopied = ScanLines;
+        goto Exit;
+    }
+
+    if (pConvertedInfo->bmiHeader.biHeight == 0)
+    {
+        LinesCopied = 0;
+        goto Exit;
+    }
+
+    if (!aligned(Bits, 4) && (ScanLines > Height))
+    {
+        LinesCopied = 0;
+        goto Exit;
+    }
+
+    /* Below code modeled after Wine's nulldrv_SetDIBitsToDevice */
+    if (StartScan <= YSrc + bmiHeight)
+    {
+        if ((pConvertedInfo->bmiHeader.biCompression == BI_RLE8) ||
+            (pConvertedInfo->bmiHeader.biCompression == BI_RLE4))
+        {
+            StartScan = 0;
+            ScanLines = bmiHeight;
+        }
+        else
+        {
+            if (StartScan >= bmiHeight)
+            {
+                LinesCopied = 0;
+                goto Exit;
+            }
+            if (!top_down && ScanLines > bmiHeight - StartScan)
+            {
+                ScanLines = bmiHeight - StartScan;
+            }
+
+            src_y = StartScan + ScanLines - (YSrc + Height);
+            if (!top_down)
+            {
+                /* get rid of unnecessary lines */
+                if ((src_y >= (INT)ScanLines || src_y < 0) &&
+                   pConvertedInfo->bmiHeader.biCompression != BI_BITFIELDS)
+                {
+                    LinesCopied = ScanLines;
+                    goto Exit;
+                }
+                if (YDest >= 0)
+                {
+                    LinesCopied = ScanLines + StartScan;
+                    ScanLines -= src_y;
+                }
+                else
+                {
+                    LinesCopied = ScanLines - src_y;
+                }
+            }
+            else if (src_y >= ScanLines)
+            {
+                if ((lpbmi->bmiHeader.biHeight < 0) && (StartScan > MaxScanLines))
+                    ScanLines = lpbmi->bmiHeader.biHeight - StartScan;
+                DPRINT("Returning ScanLines of '%d'.\n", ScanLines);
+                LinesCopied = ScanLines;
+                goto Exit;
+            }
+            if (top_down)
+                LinesCopied = ScanLines;
+        }
+    }
+
     HANDLE_METADC(INT,
                   SetDIBitsToDevice,
                   0,
@@ -769,11 +875,35 @@ SetDIBitsToDevice(
      pConvertedInfo->bmiHeader.biCompression  == BI_PNG )) )*/
     {
         LinesCopied = NtGdiSetDIBitsToDeviceInternal(hdc, XDest, YDest, Width, Height, XSrc, YSrc,
-            StartScan, ScanLines, (LPBYTE) pvSafeBits, (LPBITMAPINFO) pConvertedInfo, ColorUse,
+            StartScan, max(ScanLines, LinesCopied), (LPBYTE) pvSafeBits, (LPBITMAPINFO) pConvertedInfo, ColorUse,
             cjBmpScanSize, ConvertedInfoSize,
             TRUE,
             NULL);
     }
+
+    if (bmiHeight > MaxScanLines)
+    {
+        LinesCopied = ScanLines;
+    }
+    if (YDest < 0)
+    {
+        if (top_down)
+            LinesCopied = ScanLines;
+        else
+            LinesCopied = ScanLines - src_y;
+    }
+
+    if (pConvertedInfo->bmiHeader.biCompression == BI_RLE8 ||
+        pConvertedInfo->bmiHeader.biCompression == BI_RLE4)
+    {
+        LinesCopied = bmiHeight;
+    }
+
+    if (pConvertedInfo->bmiHeader.biHeight < 0)
+    {
+        LinesCopied = ScanLines;
+    }
+
 Exit:
     if (Bits != pvSafeBits)
         RtlFreeHeap(RtlGetProcessHeap(), 0, pvSafeBits);
diff --git a/win32ss/gdi/ntgdi/dibobj.c b/win32ss/gdi/ntgdi/dibobj.c
index ec5fa69a469..430563ed2d7 100644
--- a/win32ss/gdi/ntgdi/dibobj.c
+++ b/win32ss/gdi/ntgdi/dibobj.c
@@ -501,6 +501,17 @@ NtGdiSetDIBitsToDeviceInternal(
 
     if (!Bits) return 0;
 
+    DPRINT("StartScan '%d' ScanLines '%d' Bits '%p'\n"
+           "    bmi '%p' ColorUse '%d' SizeImage '%d'\n"
+           "    Height %d Width %d\n"
+           "    biHeight '%d' biWidth '%d' biBitCount '%d'\n"
+           "    XSrc '%d' YSrc '%d' xDext '%d' yDest '%d'\n",
+           StartScan, ScanLines, Bits, bmi, ColorUse,
+           Height, Width,
+           bmi->bmiHeader.biSizeImage, bmi->bmiHeader.biHeight,
+           bmi->bmiHeader.biWidth, bmi->bmiHeader.biBitCount,
+           XSrc, YSrc, XDest, YDest);
+
     pbmiSafe = ExAllocatePoolWithTag(PagedPool, cjMaxInfo, 'pmTG');
     if (!pbmiSafe) return 0;
 
@@ -518,7 +529,17 @@ NtGdiSetDIBitsToDeviceInternal(
     }
     _SEH2_END;
 
-    ScanLines = min(ScanLines, abs(bmi->bmiHeader.biHeight) - StartScan);
+    if (0 <= YDest && YDest <= Height)
+    {
+        ScanLines = min(abs(Height), ScanLines);
+        if (YSrc > 0 && YDest <= Height)
+            ScanLines += YSrc;
+    }
+    else
+    {
+        ScanLines = min(ScanLines, abs(bmi->bmiHeader.biHeight) - StartScan);
+    }
+
     if (ScanLines == 0)
     {
         DPRINT1("ScanLines == 0\n");
@@ -562,6 +583,11 @@ NtGdiSetDIBitsToDeviceInternal(
 
     SourceSize.cx = bmi->bmiHeader.biWidth;
     SourceSize.cy = ScanLines;
+    if (0 < YDest && YDest <= Height)
+    {
+        if (YSrc > 0)
+            ScanLines += YSrc;
+    }
 
     //DIBWidth = WIDTH_BYTES_ALIGN32(SourceSize.cx, bmi->bmiHeader.biBitCount);