diff --git a/GPU/Common/FramebufferManagerCommon.cpp b/GPU/Common/FramebufferManagerCommon.cpp
index 8a4fc0fd9860..2444d04b2c98 100644
--- a/GPU/Common/FramebufferManagerCommon.cpp
+++ b/GPU/Common/FramebufferManagerCommon.cpp
@@ -1208,7 +1208,6 @@ void FramebufferManagerCommon::DrawPixels(VirtualFramebuffer *vfb, int dstX, int
 			vfb ? vfb->bufferHeight : g_display.pixel_yres,
 			u0, v0, u1, v1, ROTATION_LOCKED_HORIZONTAL, flags);
 
-		gpuStats.numUploads++;
 		draw_->Invalidate(InvalidationFlags::CACHED_RENDER_STATE);
 
 		gstate_c.Dirty(DIRTY_ALL_RENDER_STATE);
@@ -1324,6 +1323,19 @@ Draw::Texture *FramebufferManagerCommon::MakePixelTexture(const u8 *srcPixels, G
 		}
 	}
 
+	int bpp = BufferFormatBytesPerPixel(srcPixelFormat);
+	int srcStrideInBytes = srcStride * bpp;
+	int widthInBytes = width * bpp;
+
+	// Compute hash of contents.
+	XXH3_state_t *hashState = XXH3_createState();
+	XXH3_64bits_reset(hashState);
+	for (int y = 0; y < height; y++) {
+		XXH3_64bits_update(hashState, srcPixels + srcStrideInBytes, widthInBytes);
+	}
+	uint64_t imageHash = XXH3_64bits_digest(hashState);
+	XXH3_freeState(hashState);
+
 	// TODO: We can just change the texture format and flip some bits around instead of this.
 	// Could share code with the texture cache perhaps.
 	auto generateTexture = [&](uint8_t *data, const uint8_t *initData, uint32_t w, uint32_t h, uint32_t d, uint32_t byteStride, uint32_t sliceByteStride) {
@@ -1396,16 +1408,28 @@ Draw::Texture *FramebufferManagerCommon::MakePixelTexture(const u8 *srcPixels, G
 
 	int frameNumber = draw_->GetFrameCount();
 
-	// Look for a matching texture we can re-use.
+	// First look for an exact match (including contents hash) that we can re-use.
+	for (auto &iter : drawPixelsCache_) {
+		if (iter.contentsHash == imageHash && iter.tex->Width() == width && iter.tex->Height() == height && iter.tex->Format() == texFormat) {
+			iter.frameNumber = frameNumber;
+			gpuStats.numCachedUploads++;
+			return iter.tex;
+		}
+	}
+
+	// Then, look for an alternative one that's not been used recently that we can overwrite.
 	for (auto &iter : drawPixelsCache_) {
 		if (iter.frameNumber >= frameNumber - 3 || iter.tex->Width() != width || iter.tex->Height() != height || iter.tex->Format() != texFormat) {
 			continue;
 		}
 
 		// OK, current one seems good, let's use it (and mark it used).
+		gpuStats.numUploads++;
 		draw_->UpdateTextureLevels(iter.tex, &srcPixels, generateTexture, 1);
 		// NOTE: numFlips is no good - this is called every frame when paused sometimes!
 		iter.frameNumber = frameNumber;
+		// We need to update the hash for future matching.
+		iter.contentsHash = imageHash;
 		return iter.tex;
 	}
 
@@ -1435,8 +1459,9 @@ Draw::Texture *FramebufferManagerCommon::MakePixelTexture(const u8 *srcPixels, G
 
 	// INFO_LOG(G3D, "Creating drawPixelsCache texture: %dx%d", tex->Width(), tex->Height());
 
-	DrawPixelsEntry entry{ tex, frameNumber };
+	DrawPixelsEntry entry{ tex, imageHash, frameNumber };
 	drawPixelsCache_.push_back(entry);
+	gpuStats.numUploads++;
 	return tex;
 }
 
diff --git a/GPU/Common/FramebufferManagerCommon.h b/GPU/Common/FramebufferManagerCommon.h
index 20bf5a75ed08..76fa57852372 100644
--- a/GPU/Common/FramebufferManagerCommon.h
+++ b/GPU/Common/FramebufferManagerCommon.h
@@ -269,6 +269,7 @@ class DrawContext;
 
 struct DrawPixelsEntry {
 	Draw::Texture *tex;
+	uint64_t contentsHash;
 	int frameNumber;
 };
 
diff --git a/GPU/GPU.h b/GPU/GPU.h
index cbd148078cf0..7d4d4d1c0a07 100644
--- a/GPU/GPU.h
+++ b/GPU/GPU.h
@@ -92,6 +92,7 @@ struct GPUStatistics {
 		numBlockingReadbacks = 0;
 		numReadbacks = 0;
 		numUploads = 0;
+		numCachedUploads = 0;
 		numDepal = 0;
 		numClears = 0;
 		numDepthCopies = 0;
@@ -126,6 +127,7 @@ struct GPUStatistics {
 	int numBlockingReadbacks;
 	int numReadbacks;
 	int numUploads;
+	int numCachedUploads;
 	int numDepal;
 	int numClears;
 	int numDepthCopies;
diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
index 4491d08369b3..8170c0658d1a 100644
--- a/GPU/GPUCommonHW.cpp
+++ b/GPU/GPUCommonHW.cpp
@@ -1688,7 +1688,7 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 		"Vertices: %d drawn: %d\n"
 		"FBOs active: %d (evaluations: %d)\n"
 		"Textures: %d, dec: %d, invalidated: %d, hashed: %d kB\n"
-		"readbacks %d (%d non-block), uploads %d, depal %d\n"
+		"readbacks %d (%d non-block), upload %d (cached %d), depal %d\n"
 		"block transfers: %d\n"
 		"replacer: tracks %d references, %d unique textures\n"
 		"Cpy: depth %d, color %d, reint %d, blend %d, self %d\n"
@@ -1713,6 +1713,7 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 		gpuStats.numBlockingReadbacks,
 		gpuStats.numReadbacks,
 		gpuStats.numUploads,
+		gpuStats.numCachedUploads,
 		gpuStats.numDepal,
 		gpuStats.numBlockTransfers,
 		gpuStats.numReplacerTrackedTex,
diff --git a/GPU/ge_constants.h b/GPU/ge_constants.h
index cdb7cbce58af..e425dbeef791 100644
--- a/GPU/ge_constants.h
+++ b/GPU/ge_constants.h
@@ -460,10 +460,10 @@ inline bool IsTextureFormat16Bit(GETextureFormat tfmt) {
 
 inline int BufferFormatBytesPerPixel(GEBufferFormat format) {
 	switch (format) {
-	case GE_FORMAT_8888: return 4;  // applies to depth as well.
+	case GE_FORMAT_8888: return 4;
 	case GE_FORMAT_CLUT8: return 1;
 	default:
-		return 2;
+		return 2;  // works for depth as well as the 16-bit color formats.
 	}
 }