diff --git a/panda/src/device/winInputDeviceManager.cxx b/panda/src/device/winInputDeviceManager.cxx
index 232c612186..2ca4540ae6 100644
--- a/panda/src/device/winInputDeviceManager.cxx
+++ b/panda/src/device/winInputDeviceManager.cxx
@@ -62,7 +62,9 @@ WinInputDeviceManager() :
   }
 
   // If we have threading enabled, start a thread with a message-only window
-  // loop to listen for input events.
+  // loop to listen for input events.  We can't actually just let this be
+  // handled by the main window loop, because the main window might actually
+  // have been created in a different thread.
 #ifdef HAVE_THREADS
   if (Thread::is_threading_supported()) {
     PT(Thread) thread = new InputThread(this);
@@ -462,6 +464,23 @@ destroy_message_loop() {
   }
 }
 
+/**
+ * Sends a signal to the thread input thread, asking it to shut itself down.
+ */
+void WinInputDeviceManager::
+stop_thread() {
+#ifdef HAVE_THREADS
+  WinInputDeviceManager *mgr = (WinInputDeviceManager *)_global_ptr;
+  if (mgr != nullptr) {
+    LightMutexHolder holder(mgr->_lock);
+    HWND hwnd = mgr->_message_hwnd;
+    if (hwnd) {
+      PostMessage(hwnd, WM_QUIT, 0, 0);
+    }
+  }
+#endif
+}
+
 /**
  * Implementation of the message loop.
  */
@@ -518,10 +537,26 @@ thread_main() {
   }
 
   MSG msg;
+#ifdef SIMPLE_THREADS
+  // In the simple threading case, we can't block the thread waiting for a
+  // message; we yield control back if there are no more messages.
+  while (true) {
+    if (PeekMessage(&msg, nullptr, 0, 0, PM_REMOVE)) {
+      if (msg.message == WM_QUIT) {
+        break;
+      }
+      TranslateMessage(&msg);
+      DispatchMessage(&msg);
+    } else {
+      Thread::force_yield();
+    }
+  }
+#else
   while (GetMessage(&msg, nullptr, 0, 0) > 0) {
     TranslateMessage(&msg);
     DispatchMessage(&msg);
   }
+#endif
 
   if (device_cat.is_debug()) {
     device_cat.debug()
diff --git a/panda/src/device/winInputDeviceManager.h b/panda/src/device/winInputDeviceManager.h
index 44eaa053dc..13cab3d7c0 100644
--- a/panda/src/device/winInputDeviceManager.h
+++ b/panda/src/device/winInputDeviceManager.h
@@ -44,6 +44,8 @@ public:
   HWND setup_message_loop();
   void destroy_message_loop();
 
+  static void stop_thread();
+
 private:
   // There are always exactly four of these in existence.
   XInputDevice _xinput_device0;
diff --git a/panda/src/display/graphicsEngine.cxx b/panda/src/display/graphicsEngine.cxx
index 671f12c6aa..39a1e73aa2 100644
--- a/panda/src/display/graphicsEngine.cxx
+++ b/panda/src/display/graphicsEngine.cxx
@@ -50,6 +50,10 @@
 #include "depthTestAttrib.h"
 #include "unionBoundingVolume.h"
 
+#if defined(_WIN32) && defined(HAVE_THREADS) && defined(SIMPLE_THREADS)
+#include "winInputDeviceManager.h"
+#endif
+
 #if defined(WIN32)
   #define WINDOWS_LEAN_AND_MEAN
   #include <WinSock2.h>
@@ -638,6 +642,12 @@ remove_all_windows() {
   PStatClient::get_global_pstats()->disconnect();
 #endif
 
+#if defined(_WIN32) && defined(HAVE_THREADS) && defined(SIMPLE_THREADS)
+  // Send a message to the input message pump asking it to shut itself down.
+  // If we don't do that, the next call will deadlock.
+  WinInputDeviceManager::stop_thread();
+#endif
+
   // Well, and why not clean up all threads here?
   Thread::prepare_for_exit();
 }
diff --git a/panda/src/glstuff/glGraphicsStateGuardian_src.cxx b/panda/src/glstuff/glGraphicsStateGuardian_src.cxx
index c913d1ff34..42b29f5944 100644
--- a/panda/src/glstuff/glGraphicsStateGuardian_src.cxx
+++ b/panda/src/glstuff/glGraphicsStateGuardian_src.cxx
@@ -275,6 +275,22 @@ uchar_bgr_to_rgb(unsigned char *dest, const unsigned char *source,
   }
 }
 
+/**
+ * Recopies the given array of pixels, converting from luminance to RGB
+ * arrangement.
+ */
+static void
+uchar_l_to_rgb(unsigned char *dest, const unsigned char *source,
+               int num_pixels) {
+  for (int i = 0; i < num_pixels; i++) {
+    dest[0] = source[0];
+    dest[1] = source[0];
+    dest[2] = source[0];
+    dest += 3;
+    source += 1;
+  }
+}
+
 /**
  * Recopies the given array of pixels, converting from BGRA to RGBA
  * arrangement.
@@ -292,6 +308,23 @@ uchar_bgra_to_rgba(unsigned char *dest, const unsigned char *source,
   }
 }
 
+/**
+ * Recopies the given array of pixels, converting from luminance-alpha to RGBA
+ * arrangement.
+ */
+static void
+uchar_la_to_rgba(unsigned char *dest, const unsigned char *source,
+                 int num_pixels) {
+  for (int i = 0; i < num_pixels; i++) {
+    dest[0] = source[0];
+    dest[1] = source[0];
+    dest[2] = source[0];
+    dest[3] = source[1];
+    dest += 4;
+    source += 2;
+  }
+}
+
 /**
  * Recopies the given array of pixels, converting from BGR to RGB arrangement.
  */
@@ -307,6 +340,22 @@ ushort_bgr_to_rgb(unsigned short *dest, const unsigned short *source,
   }
 }
 
+/**
+ * Recopies the given array of pixels, converting from luminance to RGB
+ * arrangement.
+ */
+static void
+ushort_l_to_rgb(unsigned short *dest, const unsigned short *source,
+                int num_pixels) {
+  for (int i = 0; i < num_pixels; i++) {
+    dest[0] = source[0];
+    dest[1] = source[0];
+    dest[2] = source[0];
+    dest += 3;
+    source += 1;
+  }
+}
+
 /**
  * Recopies the given array of pixels, converting from BGRA to RGBA
  * arrangement.
@@ -324,6 +373,23 @@ ushort_bgra_to_rgba(unsigned short *dest, const unsigned short *source,
   }
 }
 
+/**
+ * Recopies the given array of pixels, converting from luminance-alpha to RGBA
+ * arrangement.
+ */
+static void
+ushort_la_to_rgba(unsigned short *dest, const unsigned short *source,
+                  int num_pixels) {
+  for (int i = 0; i < num_pixels; i++) {
+    dest[0] = source[0];
+    dest[1] = source[0];
+    dest[2] = source[0];
+    dest[3] = source[1];
+    dest += 4;
+    source += 2;
+  }
+}
+
 /**
  * Reverses the order of the components within the image, to convert (for
  * instance) GL_BGR to GL_RGB. Returns the byte pointer representing the
@@ -343,6 +409,12 @@ fix_component_ordering(PTA_uchar &new_image,
 
   switch (external_format) {
   case GL_RGB:
+    if (tex->get_num_components() == 1) {
+      new_image = PTA_uchar::empty_array(orig_image_size * 3);
+      uchar_l_to_rgb(new_image, orig_image, orig_image_size);
+      result = new_image;
+      break;
+    }
     switch (tex->get_component_type()) {
     case Texture::T_unsigned_byte:
     case Texture::T_byte:
@@ -366,6 +438,12 @@ fix_component_ordering(PTA_uchar &new_image,
     break;
 
   case GL_RGBA:
+    if (tex->get_num_components() == 2) {
+      new_image = PTA_uchar::empty_array(orig_image_size * 2);
+      uchar_la_to_rgba(new_image, orig_image, orig_image_size / 2);
+      result = new_image;
+      break;
+    }
     switch (tex->get_component_type()) {
     case Texture::T_unsigned_byte:
     case Texture::T_byte:
@@ -9450,7 +9528,6 @@ get_external_image_format(Texture *tex) const {
     return _supports_bgr ? GL_BGRA : GL_RGBA;
 
   case Texture::F_luminance:
-  case Texture::F_sluminance:
 #ifdef OPENGLES
     return GL_LUMINANCE;
 #else
@@ -9458,13 +9535,25 @@ get_external_image_format(Texture *tex) const {
 #endif
   case Texture::F_luminance_alphamask:
   case Texture::F_luminance_alpha:
-  case Texture::F_sluminance_alpha:
 #ifdef OPENGLES
     return GL_LUMINANCE_ALPHA;
 #else
     return _supports_luminance_texture ? GL_LUMINANCE_ALPHA : GL_RG;
 #endif
 
+  case Texture::F_sluminance:
+#ifdef OPENGLES
+    return GL_LUMINANCE;
+#else
+    return _supports_luminance_texture ? GL_LUMINANCE : GL_RGB;
+#endif
+  case Texture::F_sluminance_alpha:
+#ifdef OPENGLES
+    return GL_LUMINANCE_ALPHA;
+#else
+    return _supports_luminance_texture ? GL_LUMINANCE_ALPHA : GL_RGBA;
+#endif
+
 #ifndef OPENGLES_1
   case Texture::F_r8i:
   case Texture::F_r16i:
@@ -10200,9 +10289,9 @@ get_internal_image_format(Texture *tex, bool force_sized) const {
 #endif
 #ifndef OPENGLES
   case Texture::F_sluminance:
-    return GL_SLUMINANCE8;
+    return _core_profile ? GL_SRGB8 : GL_SLUMINANCE8;
   case Texture::F_sluminance_alpha:
-    return GL_SLUMINANCE8_ALPHA8;
+    return _core_profile ? GL_SRGB8_ALPHA8 : GL_SLUMINANCE8_ALPHA8;
 #endif
 
 #ifndef OPENGLES