diff --git a/Common/GPU/Vulkan/VulkanImage.cpp b/Common/GPU/Vulkan/VulkanImage.cpp
index 0361d66d4aec..f5406a569883 100644
--- a/Common/GPU/Vulkan/VulkanImage.cpp
+++ b/Common/GPU/Vulkan/VulkanImage.cpp
@@ -38,7 +38,7 @@ static bool IsDepthStencilFormat(VkFormat format) {
 	}
 }
 
-bool VulkanTexture::CreateDirect(VkCommandBuffer cmd, int w, int h, int depth, int numMips, VkFormat format, VkImageLayout initialLayout, VkImageUsageFlags usage, const VkComponentMapping *mapping) {
+bool VulkanTexture::CreateDirect(int w, int h, int depth, int numMips, VkFormat format, VkImageLayout initialLayout, VkImageUsageFlags usage, VulkanBarrierBatch *barrierBatch, const VkComponentMapping *mapping) {
 	if (w == 0 || h == 0 || numMips == 0) {
 		ERROR_LOG(G3D, "Can't create a zero-size VulkanTexture");
 		return false;
@@ -113,10 +113,10 @@ bool VulkanTexture::CreateDirect(VkCommandBuffer cmd, int w, int h, int depth, i
 			_assert_(false);
 			break;
 		}
-		TransitionImageLayout2(cmd, image_, 0, numMips, 1, VK_IMAGE_ASPECT_COLOR_BIT,
+		barrierBatch->TransitionImage(image_, 0, numMips, 1, VK_IMAGE_ASPECT_COLOR_BIT,
 			VK_IMAGE_LAYOUT_UNDEFINED, initialLayout,
-			VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, dstStage,
-			0, dstAccessFlags);
+			0, dstAccessFlags,
+			VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, dstStage);
 	}
 
 	// Create the view while we're at it.
diff --git a/Common/GPU/Vulkan/VulkanImage.h b/Common/GPU/Vulkan/VulkanImage.h
index 4adaf4f1a1e2..05e60391084f 100644
--- a/Common/GPU/Vulkan/VulkanImage.h
+++ b/Common/GPU/Vulkan/VulkanImage.h
@@ -28,7 +28,7 @@ class VulkanTexture {
 	// Fast uploads from buffer. Mipmaps supported.
 	// Usage must at least include VK_IMAGE_USAGE_TRANSFER_DST_BIT in order to use UploadMip.
 	// When using UploadMip, initialLayout should be VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL.
-	bool CreateDirect(VkCommandBuffer cmd, int w, int h, int depth, int numMips, VkFormat format, VkImageLayout initialLayout, VkImageUsageFlags usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT, const VkComponentMapping *mapping = nullptr);
+	bool CreateDirect(int w, int h, int depth, int numMips, VkFormat format, VkImageLayout initialLayout, VkImageUsageFlags usage, VulkanBarrierBatch *barrierBatch, const VkComponentMapping *mapping = nullptr);
 	void ClearMip(VkCommandBuffer cmd, int mip, uint32_t value);
 
 	// Can also be used to copy individual levels of a 3D texture.
diff --git a/Common/GPU/Vulkan/VulkanQueueRunner.cpp b/Common/GPU/Vulkan/VulkanQueueRunner.cpp
index 30a6e91a84d5..a0164c03ad72 100644
--- a/Common/GPU/Vulkan/VulkanQueueRunner.cpp
+++ b/Common/GPU/Vulkan/VulkanQueueRunner.cpp
@@ -1913,10 +1913,10 @@ void VulkanQueueRunner::PerformReadback(const VKRStep &step, VkCommandBuffer cmd
 		// We only take screenshots after the main render pass (anything else would be stupid) so we need to transition out of PRESENT,
 		// and then back into it.
 		// Regarding layers, backbuffer currently only has one layer.
-		TransitionImageLayout2(cmd, backbufferImage_, 0, 1, 1, VK_IMAGE_ASPECT_COLOR_BIT,
+		recordBarrier_.TransitionImage(backbufferImage_, 0, 1, 1, VK_IMAGE_ASPECT_COLOR_BIT,
 			VK_IMAGE_LAYOUT_PRESENT_SRC_KHR, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
-			VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
-			0, VK_ACCESS_TRANSFER_READ_BIT);
+			0, VK_ACCESS_TRANSFER_READ_BIT,
+			VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT);
 		copyLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
 		image = backbufferImage_;
 	} else {
@@ -1933,12 +1933,13 @@ void VulkanQueueRunner::PerformReadback(const VKRStep &step, VkCommandBuffer cmd
 
 		if (srcImage->layout != VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL) {
 			SetupTransitionToTransferSrc(*srcImage, step.readback.aspectMask, &recordBarrier_);
-			recordBarrier_.Flush(cmd);
 		}
 		image = srcImage->image;
 		copyLayout = srcImage->layout;
 	}
 
+	recordBarrier_.Flush(cmd);
+
 	// TODO: Handle different readback formats!
 	u32 readbackSizeInBytes = sizeof(uint32_t) * step.readback.srcRect.extent.width * step.readback.srcRect.extent.height;
 
@@ -1980,10 +1981,11 @@ void VulkanQueueRunner::PerformReadback(const VKRStep &step, VkCommandBuffer cmd
 		// We only take screenshots after the main render pass (anything else would be stupid) so we need to transition out of PRESENT,
 		// and then back into it.
 		// Regarding layers, backbuffer currently only has one layer.
-		TransitionImageLayout2(cmd, backbufferImage_, 0, 1, 1, VK_IMAGE_ASPECT_COLOR_BIT,
+		recordBarrier_.TransitionImage(backbufferImage_, 0, 1, 1, VK_IMAGE_ASPECT_COLOR_BIT,
 			VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_IMAGE_LAYOUT_PRESENT_SRC_KHR,
-			VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
-			VK_ACCESS_TRANSFER_READ_BIT, 0);
+			VK_ACCESS_TRANSFER_READ_BIT, 0,
+			VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
+		recordBarrier_.Flush(cmd);  // probably not needed
 		copyLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
 	}
 }
@@ -2012,11 +2014,12 @@ void VulkanQueueRunner::PerformReadbackImage(const VKRStep &step, VkCommandBuffe
 	vkCmdCopyImageToBuffer(cmd, step.readback_image.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, syncReadback_.buffer, 1, &region);
 
 	// Now transfer it back to a texture.
-	TransitionImageLayout2(cmd, step.readback_image.image, 0, 1, 1,  // I don't think we have any multilayer cases for regular textures. Above in PerformReadback, though..
+	recordBarrier_.TransitionImage(step.readback_image.image, 0, 1, 1,  // I don't think we have any multilayer cases for regular textures. Above in PerformReadback, though..
 		VK_IMAGE_ASPECT_COLOR_BIT,
 		VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
-		VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
-		VK_ACCESS_TRANSFER_READ_BIT, VK_ACCESS_SHADER_READ_BIT);
+		VK_ACCESS_TRANSFER_READ_BIT, VK_ACCESS_SHADER_READ_BIT,
+		VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT);
+	recordBarrier_.Flush(cmd);  // probably not needed
 
 	// NOTE: Can't read the buffer using the CPU here - need to sync first.
 	// Doing that will also act like a heavyweight barrier ensuring that device writes are visible on the host.
diff --git a/Common/GPU/Vulkan/thin3d_vulkan.cpp b/Common/GPU/Vulkan/thin3d_vulkan.cpp
index eaace00be079..8061ba1cf760 100644
--- a/Common/GPU/Vulkan/thin3d_vulkan.cpp
+++ b/Common/GPU/Vulkan/thin3d_vulkan.cpp
@@ -699,8 +699,10 @@ VulkanTexture *VKContext::GetNullTexture() {
 		nullTexture_ = new VulkanTexture(vulkan_, "Null");
 		int w = 8;
 		int h = 8;
-		nullTexture_->CreateDirect(cmdInit, w, h, 1, 1, VK_FORMAT_A8B8G8R8_UNORM_PACK32, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
-			VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT);
+		VulkanBarrierBatch barrier;
+		nullTexture_->CreateDirect(w, h, 1, 1, VK_FORMAT_A8B8G8R8_UNORM_PACK32, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+			VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT, &barrier);
+		barrier.Flush(cmdInit);
 		uint32_t bindOffset;
 		VkBuffer bindBuf;
 		uint32_t *data = (uint32_t *)push_->Allocate(w * h * 4, 4, &bindBuf, &bindOffset);
@@ -791,10 +793,12 @@ bool VKTexture::Create(VkCommandBuffer cmd, VulkanBarrierBatch *postBarriers, Vu
 
 	VkComponentMapping r8AsAlpha[4] = { VK_COMPONENT_SWIZZLE_ONE, VK_COMPONENT_SWIZZLE_ONE, VK_COMPONENT_SWIZZLE_ONE, VK_COMPONENT_SWIZZLE_R };
 
-	if (!vkTex_->CreateDirect(cmd, width_, height_, 1, mipLevels_, vulkanFormat, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, usageBits, desc.swizzle == TextureSwizzle::R8_AS_ALPHA ? r8AsAlpha : nullptr)) {
+	VulkanBarrierBatch barrier;
+	if (!vkTex_->CreateDirect(width_, height_, 1, mipLevels_, vulkanFormat, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, usageBits, &barrier, desc.swizzle == TextureSwizzle::R8_AS_ALPHA ? r8AsAlpha : nullptr)) {
 		ERROR_LOG(G3D,  "Failed to create VulkanTexture: %dx%dx%d fmt %d, %d levels", width_, height_, depth_, (int)vulkanFormat, mipLevels_);
 		return false;
 	}
+	barrier.Flush(cmd);
 	VkImageLayout layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
 	if (desc.initData.size()) {
 		UpdateInternal(cmd, pushBuffer, desc.initData.data(), desc.initDataCallback, (int)desc.initData.size());
diff --git a/GPU/Vulkan/TextureCacheVulkan.cpp b/GPU/Vulkan/TextureCacheVulkan.cpp
index 08efbc62ed0c..d12b3267fcbb 100644
--- a/GPU/Vulkan/TextureCacheVulkan.cpp
+++ b/GPU/Vulkan/TextureCacheVulkan.cpp
@@ -512,7 +512,10 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 	snprintf(texName, sizeof(texName), "tex_%08x_%s_%s", entry->addr, GeTextureFormatToString((GETextureFormat)entry->format, gstate.getClutPaletteFormat()), gstate.isTextureSwizzled() ? "swz" : "lin");
 	entry->vkTex = new VulkanTexture(vulkan, texName);
 	VulkanTexture *image = entry->vkTex;
-	bool allocSuccess = image->CreateDirect(cmdInit, plan.createW, plan.createH, plan.depth, plan.levelsToCreate, actualFmt, imageLayout, usage, mapping);
+
+	VulkanBarrierBatch barrier;
+	bool allocSuccess = image->CreateDirect(plan.createW, plan.createH, plan.depth, plan.levelsToCreate, actualFmt, imageLayout, usage, &barrier, mapping);
+	barrier.Flush(cmdInit);
 	if (!allocSuccess && !lowMemoryMode_) {
 		WARN_LOG_REPORT(G3D, "Texture cache ran out of GPU memory; switching to low memory mode");
 		lowMemoryMode_ = true;
@@ -537,7 +540,8 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
 		plan.scaleFactor = 1;
 		actualFmt = dstFmt;
 
-		allocSuccess = image->CreateDirect(cmdInit, plan.createW, plan.createH, plan.depth, plan.levelsToCreate, actualFmt, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT, mapping);
+		allocSuccess = image->CreateDirect(plan.createW, plan.createH, plan.depth, plan.levelsToCreate, actualFmt, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT, &barrier, mapping);
+		barrier.Flush(cmdInit);
 	}
 
 	if (!allocSuccess) {