diff --git a/rpcs3/Emu/RSX/VK/VKFrameGen.cpp b/rpcs3/Emu/RSX/VK/VKFrameGen.cpp
index 1c8d135ba..31e4ec773 100644
--- a/rpcs3/Emu/RSX/VK/VKFrameGen.cpp
+++ b/rpcs3/Emu/RSX/VK/VKFrameGen.cpp
@@ -1382,10 +1382,19 @@ namespace vk::frame_gen
 		to_read.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
 		to_read.image = g_source_image;
 		to_read.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
-		to_read.srcAccessMask = VK_ACCESS_MEMORY_READ_BIT;
+		// The frame's WRITES are what has to become visible here, not a read.
+		//
+		// This used TOP_OF_PIPE with MEMORY_READ, and TOP_OF_PIPE in srcStageMask specifies no
+		// stage at all -- an empty first scope, so no availability operation happened for the
+		// colour-attachment and transfer writes the frame made into this image in the previous
+		// submission. Submission order on a queue is not a memory dependency. The interpolation
+		// could therefore read a partially-flushed frame, which looks like smearing and does not
+		// respond to any interpolation setting.
+		to_read.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT | VK_ACCESS_TRANSFER_WRITE_BIT;
 		to_read.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_SHADER_READ_BIT;
 
-		vkCmdPipelineBarrier(g_native.cmd[slot], VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+		vkCmdPipelineBarrier(g_native.cmd[slot],
+			VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT,
 			VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
 			0, 0, nullptr, 0, nullptr, 1, &to_read);
 
@@ -1398,12 +1407,14 @@ namespace vk::frame_gen
 		VkImageMemoryBarrier from_read = to_read;
 		from_read.oldLayout = VK_IMAGE_LAYOUT_GENERAL;
 		from_read.newLayout = g_source_layout;
+		// BOTTOM_OF_PIPE in dstStageMask is likewise an empty second scope. The presentation
+		// engine is the consumer, so the release has to be visible to it.
 		from_read.srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_SHADER_READ_BIT;
 		from_read.dstAccessMask = VK_ACCESS_MEMORY_READ_BIT;
 
 		vkCmdPipelineBarrier(g_native.cmd[slot],
 			VK_PIPELINE_STAGE_TRANSFER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
-			VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, 0, nullptr, 0, nullptr, 1, &from_read);
+			VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0, 0, nullptr, 0, nullptr, 1, &from_read);
 
 		// Plan straight after Process, as the ARMSX2 driver does: Process is what updates the
 		// pacer's view of the frame, and GeneratedFrameCount() only answers once it has warmed up.
diff --git a/rpcs3/Emu/RSX/VK/vkutils/swapchain.cpp b/rpcs3/Emu/RSX/VK/vkutils/swapchain.cpp
index f09eea7af..9b47ad7ad 100644
--- a/rpcs3/Emu/RSX/VK/vkutils/swapchain.cpp
+++ b/rpcs3/Emu/RSX/VK/vkutils/swapchain.cpp
@@ -410,6 +410,26 @@ namespace vk
 		swap_info.imageColorSpace = m_color_space;
 
 		swap_info.imageUsage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
+
+		// Frame generation READS the presented image: FrameGen::Process copies it into the
+		// interpolation chain as TRANSFER_SRC_OPTIMAL. Both that layout and vkCmdCopyImage
+		// require the image to have been created with TRANSFER_SRC, and nothing else in the
+		// renderer ever reads a WSI image, so this was never needed before and was never asked
+		// for -- the copy was reading an image the driver had not prepared for reading.
+		//
+		// Requested only where the surface offers it, since TRANSFER_SRC is not guaranteed on
+		// Android surfaces. Where it is missing, frame generation captures nothing valid, which
+		// is a better failure than an invalid one.
+		if (g_cfg.video.frame_generation != frame_generation_mode::off &&
+			(surface_descriptors.supportedUsageFlags & VK_IMAGE_USAGE_TRANSFER_SRC_BIT))
+		{
+			swap_info.imageUsage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+		}
+		else if (g_cfg.video.frame_generation != frame_generation_mode::off)
+		{
+			rsx_log.warning("Swapchain: surface does not support TRANSFER_SRC; frame generation"
+				" cannot read presented frames on this surface.");
+		}
 		swap_info.preTransform = pre_transform;
 		swap_info.compositeAlpha = VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR;
 		swap_info.imageArrayLayers = 1;