diff --git a/Common/GPU/D3D11/thin3d_d3d11.cpp b/Common/GPU/D3D11/thin3d_d3d11.cpp
index 4192f7b3be..a911fa80ac 100644
--- a/Common/GPU/D3D11/thin3d_d3d11.cpp
+++ b/Common/GPU/D3D11/thin3d_d3d11.cpp
@@ -164,7 +164,7 @@ public:
 	void DrawIndexed(int indexCount, int offset) override;
 	void DrawUP(const void *vdata, int vertexCount) override;
 	void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount) override;
-	void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws) override;
+	void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws, const void *ub, size_t ubSize) override;
 
 	void Clear(int mask, uint32_t colorval, float depthVal, int stencilVal) override;
 
@@ -1407,7 +1407,13 @@ void D3D11DrawContext::DrawIndexedUP(const void *vdata, int vertexCount, const v
 	DrawIndexed(indexCount, 0);
 }
 
-void D3D11DrawContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws) {
+void D3D11DrawContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws, const void *ub, size_t ubSize) {
+	if (draws.is_empty() || !vertexCount || !indexCount) {
+		return;
+	}
+
+	curPipeline_ = (D3D11Pipeline *)draws[0].pipeline;
+
 	int vbyteSize = vertexCount * curPipeline_->input->stride;
 	int ibyteSize = indexCount * sizeof(u16);
 
@@ -1417,9 +1423,16 @@ void D3D11DrawContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCo
 	UpdateBuffer(upIBuffer_, (const uint8_t *)idata, 0, ibyteSize, Draw::UPDATE_DISCARD);
 	BindIndexBuffer(upIBuffer_, 0);
 
+	UpdateDynamicUniformBuffer(ub, ubSize);
 	ApplyCurrentState();
 
 	for (int i = 0; i < draws.size(); i++) {
+		if (draws[i].pipeline != curPipeline_) {
+			curPipeline_ = (D3D11Pipeline *)draws[i].pipeline;
+			ApplyCurrentState();
+			UpdateDynamicUniformBuffer(ub, ubSize);
+		}
+
 		if (draws[i].bindTexture) {
 			ID3D11ShaderResourceView *view = ((D3D11Texture *)draws[i].bindTexture)->View();
 			context_->PSSetShaderResources(0, 1, &view);
@@ -1427,6 +1440,8 @@ void D3D11DrawContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCo
 			ID3D11ShaderResourceView *view = ((D3D11Framebuffer *)draws[i].bindFramebufferAsTex)->colorSRView;
 			context_->PSSetShaderResources(0, 1, &view);
 		}
+		ID3D11SamplerState *sstate = ((D3D11SamplerState *)draws[i].samplerState)->ss;
+		context_->PSSetSamplers(0, 1, &sstate);
 		D3D11_RECT rc;
 		rc.left = draws[i].clipx;
 		rc.top = draws[i].clipy;
diff --git a/Common/GPU/D3D9/thin3d_d3d9.cpp b/Common/GPU/D3D9/thin3d_d3d9.cpp
index 733ec78c2f..72644baff9 100644
--- a/Common/GPU/D3D9/thin3d_d3d9.cpp
+++ b/Common/GPU/D3D9/thin3d_d3d9.cpp
@@ -574,7 +574,7 @@ public:
 	void DrawIndexed(int vertexCount, int offset) override;
 	void DrawUP(const void *vdata, int vertexCount) override;
 	void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount) override;
-	void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws) override;
+	void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws, const void *ub, size_t ubSize) override;
 
 	void Clear(int mask, uint32_t colorval, float depthVal, int stencilVal) override;
 
@@ -1203,16 +1203,29 @@ void D3D9Context::DrawIndexedUP(const void *vdata, int vertexCount, const void *
 		vdata, curPipeline_->inputLayout->GetStride());
 }
 
-void D3D9Context::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws) {
+void D3D9Context::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws, const void *ub, size_t ubSize) {
+	if (draws.is_empty() || !vertexCount || !indexCount) {
+		return;
+	}
+
+	BindPipeline(draws[0].pipeline);
 	curPipeline_->inputLayout->Apply(device_);
 	curPipeline_->Apply(device_, stencilRef_, stencilWriteMask_, stencilCompareMask_);
 	ApplyDynamicState();
+	UpdateDynamicUniformBuffer(ub, ubSize);
 
-	// Suboptimal!
+	// Suboptimal! Should dirty-track textures.
 	for (int i = 0; i < draws.size(); i++) {
+		if (draws[i].pipeline != curPipeline_) {
+			D3D9Pipeline *d3d9Pipeline = (D3D9Pipeline *)draws[i].pipeline;
+			d3d9Pipeline->Apply(device_, stencilRef_, stencilWriteMask_, stencilCompareMask_);
+			curPipeline_ = d3d9Pipeline;
+		}
+
 		if (draws[i].bindTexture) {
 			device_->SetTexture(0, ((D3D9Texture *)draws[i].bindTexture)->TexturePtr());
 		} else if (draws[i].bindFramebufferAsTex) {
+			// We ignore aspect in D3D9 :(
 			device_->SetTexture(0, ((D3D9Framebuffer *)draws[i].bindFramebufferAsTex)->tex.Get());
 		}
 
diff --git a/Common/GPU/OpenGL/thin3d_gl.cpp b/Common/GPU/OpenGL/thin3d_gl.cpp
index bfb61dbaf6..281eee3789 100644
--- a/Common/GPU/OpenGL/thin3d_gl.cpp
+++ b/Common/GPU/OpenGL/thin3d_gl.cpp
@@ -444,7 +444,7 @@ public:
 	void DrawIndexed(int vertexCount, int offset) override;
 	void DrawUP(const void *vdata, int vertexCount) override;
 	void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount) override;
-	void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws) override;
+	void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws, const void *ub, size_t ubSize) override;
 
 	void Clear(int mask, uint32_t colorval, float depthVal, int stencilVal) override;
 
@@ -1442,7 +1442,14 @@ void OpenGLContext::DrawIndexedUP(const void *vdata, int vertexCount, const void
 	renderManager_.DrawIndexed(curPipeline_->inputLayout->inputLayout_, vbuf, voffset, ibuf, ioffset, curPipeline_->prim, indexCount, GL_UNSIGNED_SHORT, 1);
 }
 
-void OpenGLContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws) {
+void OpenGLContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws, const void *ub, size_t ubSize) {
+	if (draws.is_empty() || !vertexCount || !indexCount) {
+		return;
+	}
+
+	BindPipeline(draws[0].pipeline);
+	UpdateDynamicUniformBuffer(ub, ubSize);
+
 	_assert_(curPipeline_->inputLayout != nullptr);
 	int stride = curPipeline_->inputLayout->stride;
 	uint32_t vdataSize = stride * vertexCount;
@@ -1463,6 +1470,12 @@ void OpenGLContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount
 
 	ApplySamplers();
 	for (auto &draw : draws) {
+		if (draw.pipeline != curPipeline_) {
+			OpenGLPipeline *glPipeline = (OpenGLPipeline *)draw.pipeline;
+			_dbg_assert_(glPipeline->inputLayout->stride == stride);
+			BindPipeline(glPipeline);  // this updated curPipeline_.
+			UpdateDynamicUniformBuffer(ub, ubSize);
+		}
 		if (draw.bindTexture) {
 			renderManager_.BindTexture(0, ((OpenGLTexture *)draw.bindTexture)->GetTex());
 		} else if (draw.bindFramebufferAsTex) {
diff --git a/Common/GPU/Vulkan/thin3d_vulkan.cpp b/Common/GPU/Vulkan/thin3d_vulkan.cpp
index c8f9194251..5d0a8b8c86 100644
--- a/Common/GPU/Vulkan/thin3d_vulkan.cpp
+++ b/Common/GPU/Vulkan/thin3d_vulkan.cpp
@@ -493,7 +493,7 @@ public:
 	void DrawUP(const void *vdata, int vertexCount) override;
 	void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount) override;
 	// Specialized for quick IMGUI drawing.
-	void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw>) override;
+	void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw>, const void *dynUniforms, size_t size) override;
 
 	void BindCurrentPipeline();
 	void ApplyDynamicState();
@@ -1557,13 +1557,15 @@ void VKContext::DrawIndexedUP(const void *vdata, int vertexCount, const void *id
 	renderManager_.DrawIndexed(descSetIndex, 1, &ubo_offset, vulkanVbuf, (int)vbBindOffset, vulkanIbuf, (int)ibBindOffset, indexCount, 1);
 }
 
-void VKContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws) {
+void VKContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws, const void *ub, size_t ubSize) {
 	_dbg_assert_(vertexCount >= 0);
 	_dbg_assert_(indexCount >= 0);
 	if (vertexCount <= 0 || indexCount <= 0 || draws.is_empty()) {
 		return;
 	}
 
+	curPipeline_ = (VKPipeline *)draws[0].pipeline;
+
 	VkBuffer vulkanVbuf, vulkanIbuf, vulkanUBObuf;
 	size_t vdataSize = vertexCount * curPipeline_->stride;
 	uint32_t vbBindOffset;
@@ -1579,18 +1581,28 @@ void VKContext::DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, co
 	_assert_(idataPtr != nullptr);
 	memcpy(idataPtr, idata, idataSize);
 
+	curPipeline_->SetDynamicUniformData(ub, ubSize);
+
 	uint32_t ubo_offset = (uint32_t)curPipeline_->PushUBO(push_, vulkan_, &vulkanUBObuf);
 
 	BindCurrentPipeline();
 	ApplyDynamicState();
 
 	for (auto &draw : draws) {
+		if (draw.pipeline != curPipeline_) {
+			VKPipeline *vkPipe = (VKPipeline *)draw.pipeline;
+			renderManager_.BindPipeline(vkPipe->pipeline, vkPipe->flags, pipelineLayout_);
+			curPipeline_ = (VKPipeline *)draw.pipeline;
+			curPipeline_->SetDynamicUniformData(ub, ubSize);
+		}
 		// TODO: Dirty-check these.
 		if (draw.bindTexture) {
 			BindTexture(0, draw.bindTexture);
 		} else if (draw.bindFramebufferAsTex) {
 			BindFramebufferAsTexture(draw.bindFramebufferAsTex, 0, FBChannel::FB_COLOR_BIT, 0);
 		}
+		Draw::SamplerState *sstate = draw.samplerState;
+		BindSamplerStates(0, 1, &sstate);
 		int descSetIndex;
 		PackedDescriptor *descriptors = renderManager_.PushDescriptorSet(4, &descSetIndex);
 		BindDescriptors(vulkanUBObuf, descriptors);
diff --git a/Common/GPU/thin3d.h b/Common/GPU/thin3d.h
index 3e2076cc8c..c1210c7d1c 100644
--- a/Common/GPU/thin3d.h
+++ b/Common/GPU/thin3d.h
@@ -705,6 +705,8 @@ struct ClippedDraw {
 	s16 cliph;
 	Draw::Texture *bindTexture;
 	Draw::Framebuffer *bindFramebufferAsTex;
+	Draw::SamplerState *samplerState;
+	Draw::Pipeline *pipeline;
 };
 
 class DrawContext {
@@ -843,7 +845,7 @@ public:
 	virtual void DrawUP(const void *vdata, int vertexCount) = 0;
 	virtual void DrawIndexedUP(const void *vdata, int vertexCount, const void *idata, int indexCount) = 0;
 	// Intended for ImGui display lists, easier to do optimally this way.
-	virtual void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws) = 0;
+	virtual void DrawIndexedClippedBatchUP(const void *vdata, int vertexCount, const void *idata, int indexCount, Slice<ClippedDraw> draws, const void *dynUniforms, size_t size) = 0;
 
 	// Frame management (for the purposes of sync and resource management, necessary with modern APIs). Default implementations here.
 	virtual void BeginFrame(DebugFlags debugFlags) = 0;
diff --git a/GPU/Common/FramebufferManagerCommon.cpp b/GPU/Common/FramebufferManagerCommon.cpp
index bc3674a901..c06e73b7c6 100644
--- a/GPU/Common/FramebufferManagerCommon.cpp
+++ b/GPU/Common/FramebufferManagerCommon.cpp
@@ -3667,13 +3667,12 @@ static void ApplyKillzoneFramebufferSplit(FramebufferHeuristicParams *params, in
 
 void FramebufferManagerCommon::DrawImGuiDebug(int &selected) const {
 	ImGui::BeginTable("framebuffers", 4);
-	ImGui::TableSetupColumn("Tag");
-	ImGui::TableSetupColumn("Color Addr");
-	ImGui::TableSetupColumn("Depth Addr");
-	ImGui::TableSetupColumn("Size");
+	ImGui::TableSetupColumn("Tag", ImGuiTableColumnFlags_WidthFixed);
+	ImGui::TableSetupColumn("Color Addr", ImGuiTableColumnFlags_WidthFixed);
+	ImGui::TableSetupColumn("Depth Addr", ImGuiTableColumnFlags_WidthFixed);
+	ImGui::TableSetupColumn("Size", ImGuiTableColumnFlags_WidthFixed);
 
 	ImGui::TableHeadersRow();
-	ImGui::TableSetColumnIndex(0);
 
 	for (int i = 0; i < (int)vfbs_.size(); i++) {
 		ImGui::TableNextRow();
@@ -3705,10 +3704,15 @@ void FramebufferManagerCommon::DrawImGuiDebug(int &selected) const {
 	}
 	ImGui::EndTable();
 
+	// Fix out-of-bounds issues when framebuffers are removed.
+	if (selected >= vfbs_.size()) {
+		selected = -1;
+	}
+
 	if (selected != -1) {
 		// Now, draw the image of the selected framebuffer.
 		Draw::Framebuffer *fb = vfbs_[selected]->fbo;
-		ImTextureID texId = ImGui_ImplThin3d_AddFBAsTextureTemp(fb);
+		ImTextureID texId = ImGui_ImplThin3d_AddFBAsTextureTemp(fb, Draw::FB_COLOR_BIT, ImGuiPipeline::TexturedOpaque);
 		ImGui::Image(texId, ImVec2(fb->Width(), fb->Height()));
 	}
 }
diff --git a/UI/ImDebugger/ImDebugger.cpp b/UI/ImDebugger/ImDebugger.cpp
index 5607252c2b..6742423dee 100644
--- a/UI/ImDebugger/ImDebugger.cpp
+++ b/UI/ImDebugger/ImDebugger.cpp
@@ -4,6 +4,7 @@
 #include "ext/imgui/imgui_internal.h"
 
 #include "Common/StringUtils.h"
+#include "Common/Data/Format/IniFile.h"
 #include "Core/Config.h"
 #include "Core/System.h"
 #include "Core/RetroAchievements.h"
@@ -230,7 +231,7 @@ static void DrawFilesystemBrowser(ImConfig &cfg) {
 }
 
 static void DrawKernelObjects(ImConfig &cfg) {
-	if (!ImGui::Begin("Kernel Objects", &cfg.filesystemBrowserOpen)) {
+	if (!ImGui::Begin("Kernel Objects", &cfg.kernelObjectsOpen)) {
 		ImGui::End();
 		return;
 	}
@@ -691,6 +692,7 @@ void ImDebugger::Frame(MIPSDebugInterface *mipsDebug, GPUDebugInterface *gpuDebu
 			ImGui::EndMenu();
 		}
 		if (ImGui::BeginMenu("Ge (GPU)")) {
+			ImGui::MenuItem("Display Output", nullptr, &cfg_.displayOpen);
 			ImGui::MenuItem("Framebuffers", nullptr, &cfg_.framebuffersOpen);
 			// More to come here...
 			ImGui::EndMenu();
@@ -762,6 +764,10 @@ void ImDebugger::Frame(MIPSDebugInterface *mipsDebug, GPUDebugInterface *gpuDebu
 		DrawFramebuffersWindow(cfg_, gpuDebug->GetFramebufferManagerCommon());
 	}
 
+	if (cfg_.displayOpen) {
+		DrawDisplayWindow(cfg_, gpuDebug->GetFramebufferManagerCommon());
+	}
+
 	if (cfg_.structViewerOpen) {
 		structViewer_.Draw(mipsDebug, &cfg_.structViewerOpen);
 	}
@@ -914,3 +920,7 @@ void ImDisasmWindow::Draw(MIPSDebugInterface *mipsDebug, bool *open, CoreState c
 	}
 	ImGui::End();
 }
+
+void ImDebugger::LoadConfig() {
+	IniFile ini;
+}
diff --git a/UI/ImDebugger/ImDebugger.h b/UI/ImDebugger/ImDebugger.h
index a1ba864dd9..ee2ed91ccc 100644
--- a/UI/ImDebugger/ImDebugger.h
+++ b/UI/ImDebugger/ImDebugger.h
@@ -37,6 +37,7 @@ public:
 	void DirtySymbolMap() {
 		symsDirty_ = true;
 	}
+
 private:
 	// We just keep the state directly in the window. Can refactor later.
 
@@ -56,11 +57,6 @@ private:
 	char searchTerm_[64]{};
 };
 
-class ImLuaConsole {
-public:
-	// Stub
-};
-
 struct ImConfig {
 	bool disasmOpen = true;
 	bool demoOpen  = false;
@@ -73,6 +69,7 @@ struct ImConfig {
 	bool atracOpen = true;
 	bool structViewerOpen = false;
 	bool framebuffersOpen = false;
+	bool displayOpen = false;
 	bool styleEditorOpen = false;
 	bool filesystemBrowserOpen = false;
 	bool kernelObjectsOpen = false;
@@ -86,6 +83,8 @@ struct ImConfig {
 	int selectedFramebuffer = -1;
 	int selectedBreakpoint = -1;
 	int selectedMemCheck = -1;
+
+	bool displayLatched = false;
 };
 
 enum ImUiCmd {
@@ -96,14 +95,20 @@ struct ImUiCommand {
 	ImUiCmd cmd;
 };
 
-struct ImDebugger {
+class ImDebugger {
+public:
 	ImDebugger();
+
 	void Frame(MIPSDebugInterface *mipsDebug, GPUDebugInterface *gpuDebug);
 
+private:
+	// We use a separate ini file from the main PPSSPP config.
+	void LoadConfig();
+	void SaveConfig();
+
 	RequesterToken reqToken_;
 
 	ImDisasmWindow disasm_;
-	ImLuaConsole luaConsole_;
 	ImStructViewer structViewer_;
 
 	// Open variables.
diff --git a/UI/ImDebugger/ImGe.cpp b/UI/ImDebugger/ImGe.cpp
index 0945030ca0..3b096b650c 100644
--- a/UI/ImDebugger/ImGe.cpp
+++ b/UI/ImDebugger/ImGe.cpp
@@ -1,7 +1,11 @@
+#include "ext/imgui/imgui.h"
+#include "ext/imgui/imgui_impl_thin3d.h"
 #include "UI/ImDebugger/ImGe.h"
 #include "UI/ImDebugger/ImDebugger.h"
 #include "GPU/Common/FramebufferManagerCommon.h"
 
+#include "Core/HLE/sceDisplay.h"
+
 void DrawFramebuffersWindow(ImConfig &cfg, FramebufferManagerCommon *framebufferManager) {
 	if (!ImGui::Begin("Framebuffers", &cfg.framebuffersOpen)) {
 		ImGui::End();
@@ -12,3 +16,30 @@ void DrawFramebuffersWindow(ImConfig &cfg, FramebufferManagerCommon *framebuffer
 
 	ImGui::End();
 }
+
+void DrawDisplayWindow(ImConfig &cfg, FramebufferManagerCommon *framebufferManager) {
+	if (!ImGui::Begin("Display", &cfg.framebuffersOpen)) {
+		ImGui::End();
+		return;
+	}
+
+	ImGui::Checkbox("Display latched", &cfg.displayLatched);
+
+	PSPPointer<u8> topaddr;
+	u32 linesize;
+	u32 pixelFormat;
+
+	__DisplayGetFramebuf(&topaddr, &linesize, &pixelFormat, cfg.displayLatched);
+
+	VirtualFramebuffer *fb = framebufferManager->GetVFBAt(topaddr.ptr);
+	if (fb && fb->fbo) {
+		ImTextureID texId = ImGui_ImplThin3d_AddFBAsTextureTemp(fb->fbo, Draw::FB_COLOR_BIT, ImGuiPipeline::TexturedOpaque);
+		ImGui::Image(texId, ImVec2(fb->width, fb->height));
+		ImGui::Text("%s - %08x", fb->fbo->Tag(), topaddr.ptr);
+	} else {
+		// TODO: Sometimes we should display RAM here.
+		ImGui::Text("Framebuffer not available to display");
+	}
+
+	ImGui::End();
+}
diff --git a/UI/ImDebugger/ImGe.h b/UI/ImDebugger/ImGe.h
index 958465ba17..08868b7f0a 100644
--- a/UI/ImDebugger/ImGe.h
+++ b/UI/ImDebugger/ImGe.h
@@ -7,6 +7,7 @@ struct ImConfig;
 class FramebufferManagerCommon;
 
 void DrawFramebuffersWindow(ImConfig &cfg, FramebufferManagerCommon *framebufferManager);
+void DrawDisplayWindow(ImConfig &cfg, FramebufferManagerCommon *framebufferManager);
 
 class ImGeDebugger {
 public:
diff --git a/ext/imgui/imgui_impl_thin3d.cpp b/ext/imgui/imgui_impl_thin3d.cpp
index ac0fc0a694..18c0a4fef4 100644
--- a/ext/imgui/imgui_impl_thin3d.cpp
+++ b/ext/imgui/imgui_impl_thin3d.cpp
@@ -16,14 +16,18 @@ struct RegisteredTexture {
 	bool isFramebuffer;
 	union {
 		Draw::Texture *texture;
-		Draw::Framebuffer *framebuffer;
+		struct {
+			Draw::Framebuffer *framebuffer;
+			Draw::FBChannel aspect;
+		};
 	};
+	ImGuiPipeline pipeline;
 };
 
 struct BackendData {
 	Draw::SamplerState *fontSampler = nullptr;
 	Draw::Texture *fontImage = nullptr;
-	Draw::Pipeline *pipeline = nullptr;
+	Draw::Pipeline *pipelines[2]{};
 	std::vector<RegisteredTexture> tempTextures;
 };
 
@@ -36,39 +40,6 @@ static BackendData *ImGui_ImplThin3d_GetBackendData() {
 	return ImGui::GetCurrentContext() ? (BackendData *)ImGui::GetIO().BackendRendererUserData : nullptr;
 }
 
-static void ImGui_ImplThin3d_SetupRenderState(Draw::DrawContext *draw, ImDrawData* drawData, Draw::Pipeline *pipeline, int fb_width, int fb_height) {
-	BackendData *bd = ImGui_ImplThin3d_GetBackendData();
-
-	// Bind pipeline and texture
-	draw->BindPipeline(pipeline);
-	draw->BindSamplerStates(0, 1, &bd->fontSampler);
-
-	// Setup viewport
-	{
-		Draw::Viewport viewport;
-		viewport.TopLeftX = 0;
-		viewport.TopLeftY = 0;
-		viewport.Width = (float)fb_width;
-		viewport.Height = (float)fb_height;
-		viewport.MinDepth = 0.0f;
-		viewport.MaxDepth = 1.0f;
-		draw->SetViewport(viewport);
-	}
-
-	// Setup scale and translation:
-	// Our visible imgui space lies from drawData->DisplayPps (top left) to drawData->DisplayPos + drawData->DisplaySize (bottom right).
-	// DisplayPos is (0,0) for single viewport apps. We currently ignore DisplayPos.
-	// We probably only need to do this at the start of the frame.
-	{
-		Lin::Matrix4x4 mtx = ComputeOrthoMatrix(drawData->DisplaySize.x, drawData->DisplaySize.y, draw->GetDeviceCaps().coordConvention);
-
-		Draw::VsTexColUB ub{};
-		memcpy(ub.WorldViewProj, mtx.getReadPtr(), sizeof(Lin::Matrix4x4));
-		ub.saturation = 1.0f;
-		draw->UpdateDynamicUniformBuffer(&ub, sizeof(ub));
-	}
-}
-
 // Render function
 void ImGui_ImplThin3d_RenderDrawData(ImDrawData* draw_data, Draw::DrawContext *draw) {
 	// Avoid rendering when minimized, scale coordinates for retina displays (screen coordinates != framebuffer coordinates)
@@ -79,9 +50,23 @@ void ImGui_ImplThin3d_RenderDrawData(ImDrawData* draw_data, Draw::DrawContext *d
 	}
 
 	BackendData* bd = ImGui_ImplThin3d_GetBackendData();
+	draw->BindSamplerStates(0, 1, &bd->fontSampler);
 
-	// Setup desired Vulkan state
-	ImGui_ImplThin3d_SetupRenderState(draw, draw_data, bd->pipeline, fb_width, fb_height);
+	// Setup viewport
+	Draw::Viewport viewport;
+	viewport.TopLeftX = 0;
+	viewport.TopLeftY = 0;
+	viewport.Width = (float)fb_width;
+	viewport.Height = (float)fb_height;
+	viewport.MinDepth = 0.0f;
+	viewport.MaxDepth = 1.0f;
+	draw->SetViewport(viewport);
+
+	Lin::Matrix4x4 mtx = ComputeOrthoMatrix(draw_data->DisplaySize.x, draw_data->DisplaySize.y, draw->GetDeviceCaps().coordConvention);
+
+	Draw::VsTexColUB ub{};
+	memcpy(ub.WorldViewProj, mtx.getReadPtr(), sizeof(Lin::Matrix4x4));
+	ub.saturation = 1.0f;
 
 	// Will project scissor/clipping rectangles into framebuffer space
 	ImVec2 clip_off = draw_data->DisplayPos;         // (0,0) unless using multi-viewports
@@ -94,6 +79,8 @@ void ImGui_ImplThin3d_RenderDrawData(ImDrawData* draw_data, Draw::DrawContext *d
 	std::vector<Draw::ClippedDraw> draws;
 	Draw::Texture *boundTexture;
 	Draw::Framebuffer *boundFBAsTexture;
+	Draw::Pipeline *boundPipeline = bd->pipelines[0];
+	Draw::SamplerState *boundSampler = bd->fontSampler;
 
 	// Render command lists
 	for (int n = 0; n < draw_data->CmdListsCount; n++) {
@@ -101,55 +88,56 @@ void ImGui_ImplThin3d_RenderDrawData(ImDrawData* draw_data, Draw::DrawContext *d
 		draws.clear();
 		for (int cmd_i = 0; cmd_i < cmd_list->CmdBuffer.Size; cmd_i++) {
 			const ImDrawCmd* pcmd = &cmd_list->CmdBuffer[cmd_i];
-			if (pcmd->UserCallback != nullptr) {
-				// User callback, registered via ImDrawList::AddCallback()
-				// (ImDrawCallback_ResetRenderState is a special callback value used by the user to request the renderer to reset render state.)
-				if (pcmd->UserCallback == ImDrawCallback_ResetRenderState) {
-					ImGui_ImplThin3d_SetupRenderState(draw, draw_data, bd->pipeline, fb_width, fb_height);
-				} else {
-					pcmd->UserCallback(cmd_list, pcmd);
-				}
+			// We don't use the callback mechanism.
+			_dbg_assert_(pcmd->UserCallback == nullptr);
+
+			// Update the texture pointers.
+			if (!pcmd->TextureId) {
+				// Default
+				boundTexture = bd->fontImage;
+				boundFBAsTexture = nullptr;
+				boundPipeline = bd->pipelines[0];
+				boundSampler = bd->fontSampler;
 			} else {
-				// Update the texture pointers.
-				if (!pcmd->TextureId) {
-					boundTexture = bd->fontImage;
-					boundFBAsTexture = nullptr;
+				size_t index = (size_t)pcmd->TextureId - TEX_ID_OFFSET;
+				_dbg_assert_(index < bd->tempTextures.size());
+				if (bd->tempTextures[index].framebuffer) {
+					boundFBAsTexture = bd->tempTextures[index].framebuffer;
+					boundTexture = nullptr;
 				} else {
-					size_t index = (size_t)pcmd->TextureId - TEX_ID_OFFSET;
-					_dbg_assert_(index < bd->tempTextures.size());
-					if (bd->tempTextures[index].framebuffer) {
-						boundFBAsTexture = bd->tempTextures[index].framebuffer;
-						boundTexture = nullptr;
-					} else {
-						boundTexture = bd->tempTextures[index].texture;
-						boundFBAsTexture = nullptr;
-					}
+					boundTexture = bd->tempTextures[index].texture;
+					boundFBAsTexture = nullptr;
 				}
-				// Project scissor/clipping rectangles into framebuffer space
-				ImVec2 clip_min((pcmd->ClipRect.x - clip_off.x) * clip_scale.x, (pcmd->ClipRect.y - clip_off.y) * clip_scale.y);
-				ImVec2 clip_max((pcmd->ClipRect.z - clip_off.x) * clip_scale.x, (pcmd->ClipRect.w - clip_off.y) * clip_scale.y);
-
-				// Clamp to viewport as vkCmdSetScissor() won't accept values that are off bounds
-				if (clip_min.x < 0.0f) { clip_min.x = 0.0f; }
-				if (clip_min.y < 0.0f) { clip_min.y = 0.0f; }
-				if (clip_max.x > fb_width) { clip_max.x = (float)fb_width; }
-				if (clip_max.y > fb_height) { clip_max.y = (float)fb_height; }
-				if (clip_max.x <= clip_min.x || clip_max.y <= clip_min.y)
-					continue;
-
-				Draw::ClippedDraw clippedDraw;
-				clippedDraw.bindTexture = boundTexture;
-				clippedDraw.bindFramebufferAsTex = boundFBAsTexture;
-				clippedDraw.clipx = clip_min.x;
-				clippedDraw.clipy = clip_min.y;
-				clippedDraw.clipw = clip_max.x - clip_min.x;
-				clippedDraw.cliph = clip_max.y - clip_min.y;
-				clippedDraw.indexCount = pcmd->ElemCount;
-				clippedDraw.indexOffset = pcmd->IdxOffset;
-				draws.push_back(clippedDraw);
+				boundPipeline = bd->pipelines[(int)bd->tempTextures[index].pipeline];
+				boundSampler = bd->fontSampler;
 			}
+
+			// Project scissor/clipping rectangles into framebuffer space
+			ImVec2 clip_min((pcmd->ClipRect.x - clip_off.x) * clip_scale.x, (pcmd->ClipRect.y - clip_off.y) * clip_scale.y);
+			ImVec2 clip_max((pcmd->ClipRect.z - clip_off.x) * clip_scale.x, (pcmd->ClipRect.w - clip_off.y) * clip_scale.y);
+
+			// Clamp to viewport as vkCmdSetScissor() won't accept values that are off bounds
+			if (clip_min.x < 0.0f) { clip_min.x = 0.0f; }
+			if (clip_min.y < 0.0f) { clip_min.y = 0.0f; }
+			if (clip_max.x > fb_width) { clip_max.x = (float)fb_width; }
+			if (clip_max.y > fb_height) { clip_max.y = (float)fb_height; }
+			if (clip_max.x <= clip_min.x || clip_max.y <= clip_min.y)
+				continue;
+
+			Draw::ClippedDraw clippedDraw;
+			clippedDraw.pipeline = boundPipeline;
+			clippedDraw.bindTexture = boundTexture;
+			clippedDraw.bindFramebufferAsTex = boundFBAsTexture;
+			clippedDraw.samplerState = boundSampler;
+			clippedDraw.clipx = clip_min.x;
+			clippedDraw.clipy = clip_min.y;
+			clippedDraw.clipw = clip_max.x - clip_min.x;
+			clippedDraw.cliph = clip_max.y - clip_min.y;
+			clippedDraw.indexCount = pcmd->ElemCount;
+			clippedDraw.indexOffset = pcmd->IdxOffset;
+			draws.push_back(clippedDraw);
 		}
-		draw->DrawIndexedClippedBatchUP(cmd_list->VtxBuffer.Data, cmd_list->VtxBuffer.size(), cmd_list->IdxBuffer.Data, cmd_list->IdxBuffer.size(), draws);
+		draw->DrawIndexedClippedBatchUP(cmd_list->VtxBuffer.Data, cmd_list->VtxBuffer.size(), cmd_list->IdxBuffer.Data, cmd_list->IdxBuffer.size(), draws, &ub, sizeof(ub));
 	}
 
 	draw->SetScissorRect(0, 0, fb_width, fb_height);
@@ -174,7 +162,7 @@ bool ImGui_ImplThin3d_CreateDeviceObjects(Draw::DrawContext *draw) {
 		bd->fontSampler = draw->CreateSamplerState(desc);
 	}
 
-	if (!bd->pipeline) {
+	if (!bd->pipelines[0]) {
 		BackendData* bd = ImGui_ImplThin3d_GetBackendData();
 
 		using namespace Draw;
@@ -193,6 +181,7 @@ bool ImGui_ImplThin3d_CreateDeviceObjects(Draw::DrawContext *draw) {
 			BlendFactor::SRC_ALPHA, BlendFactor::ONE_MINUS_SRC_ALPHA, BlendOp::ADD,
 			BlendFactor::ONE, BlendFactor::ONE_MINUS_SRC_ALPHA, BlendOp::ADD,
 			});
+		BlendState *blendOpaque = draw->CreateBlendState({ false, 0xF });
 
 		DepthStencilStateDesc dsDesc{};
 		DepthStencilState *depthStencil = draw->CreateDepthStencilState(dsDesc);
@@ -211,7 +200,15 @@ bool ImGui_ImplThin3d_CreateDeviceObjects(Draw::DrawContext *draw) {
 			&vsTexColBufDesc
 		};
 
-		bd->pipeline = draw->CreateGraphicsPipeline(pipelineDesc, "imgui-pipeline");
+		bd->pipelines[0] = draw->CreateGraphicsPipeline(pipelineDesc, "imgui-pipeline");
+		pipelineDesc.blend = blendOpaque;
+		bd->pipelines[1] = draw->CreateGraphicsPipeline(pipelineDesc, "imgui-pipeline-opaque");
+
+		inputLayout->Release();
+		blend->Release();
+		blendOpaque->Release();
+		depthStencil->Release();
+		rasterNoCull->Release();
 	}
 
 	if (!bd->fontImage) {
@@ -247,9 +244,11 @@ void ImGui_ImplThin3d_DestroyDeviceObjects() {
 		bd->fontImage = nullptr;
 		io.Fonts->SetTexID(0);
 	}
-	if (bd->pipeline) {
-		bd->pipeline->Release();
-		bd->pipeline = nullptr;
+	for (int i = 0; i < ARRAY_SIZE(bd->pipelines); i++) {
+		if (bd->pipelines[i]) {
+			bd->pipelines[i]->Release();
+			bd->pipelines[i] = nullptr;
+		}
 	}
 	if (bd->fontSampler) {
 		bd->fontSampler->Release();
@@ -310,21 +309,24 @@ void ImGui_ImplThin3d_NewFrame(Draw::DrawContext *draw, Lin::Matrix4x4 drawMatri
 	g_drawMatrix = drawMatrix;
 }
 
-ImTextureID ImGui_ImplThin3d_AddTextureTemp(Draw::Texture *texture) {
+ImTextureID ImGui_ImplThin3d_AddTextureTemp(Draw::Texture *texture, ImGuiPipeline pipeline) {
 	BackendData* bd = ImGui_ImplThin3d_GetBackendData();
 
 	RegisteredTexture tex{ false };
 	tex.texture = texture;
+	tex.pipeline = pipeline;
 
 	bd->tempTextures.push_back(tex);
 	return (ImTextureID)(uint64_t)(TEX_ID_OFFSET + bd->tempTextures.size() - 1);
 }
 
-ImTextureID ImGui_ImplThin3d_AddFBAsTextureTemp(Draw::Framebuffer *framebuffer) {
+ImTextureID ImGui_ImplThin3d_AddFBAsTextureTemp(Draw::Framebuffer *framebuffer, Draw::FBChannel aspect, ImGuiPipeline pipeline) {
 	BackendData* bd = ImGui_ImplThin3d_GetBackendData();
 
 	RegisteredTexture tex{ true };
 	tex.framebuffer = framebuffer;
+	tex.aspect = aspect;
+	tex.pipeline = pipeline;
 
 	bd->tempTextures.push_back(tex);
 	return (ImTextureID)(uint64_t)(TEX_ID_OFFSET + bd->tempTextures.size() - 1);
diff --git a/ext/imgui/imgui_impl_thin3d.h b/ext/imgui/imgui_impl_thin3d.h
index 98cad60f64..11096c9dfd 100644
--- a/ext/imgui/imgui_impl_thin3d.h
+++ b/ext/imgui/imgui_impl_thin3d.h
@@ -34,28 +34,32 @@
 #include "Common/Math/lin/matrix4x4.h"
 
 // Called by user code. Takes ownership of the font buffer and later deletes it.
-IMGUI_IMPL_API bool         ImGui_ImplThin3d_Init(Draw::DrawContext *draw, const uint8_t *ttf_font, size_t size);
-IMGUI_IMPL_API void         ImGui_ImplThin3d_Shutdown();
-IMGUI_IMPL_API void         ImGui_ImplThin3d_NewFrame(Draw::DrawContext *draw, Lin::Matrix4x4 drawMatrix);
-IMGUI_IMPL_API void         ImGui_ImplThin3d_RenderDrawData(ImDrawData* draw_data, Draw::DrawContext *draw);
-IMGUI_IMPL_API bool         ImGui_ImplThin3d_CreateDeviceObjects(Draw::DrawContext *draw);
-IMGUI_IMPL_API void         ImGui_ImplThin3d_DestroyDeviceObjects();
+IMGUI_IMPL_API bool ImGui_ImplThin3d_Init(Draw::DrawContext *draw, const uint8_t *ttf_font, size_t size);
+IMGUI_IMPL_API void ImGui_ImplThin3d_Shutdown();
+IMGUI_IMPL_API void ImGui_ImplThin3d_NewFrame(Draw::DrawContext *draw, Lin::Matrix4x4 drawMatrix);
+IMGUI_IMPL_API void ImGui_ImplThin3d_RenderDrawData(ImDrawData* draw_data, Draw::DrawContext *draw);
+IMGUI_IMPL_API bool ImGui_ImplThin3d_CreateDeviceObjects(Draw::DrawContext *draw);
+IMGUI_IMPL_API void ImGui_ImplThin3d_DestroyDeviceObjects();
+
+enum class ImGuiPipeline {
+	TexturedAlphaBlend = 0,
+	TexturedOpaque = 1,
+};
 
 // These register a texture for imgui drawing, but just for the current frame.
 // Textures are unregistered again in RenderDrawData. This is just simpler.
-IMGUI_IMPL_API ImTextureID ImGui_ImplThin3d_AddTextureTemp(Draw::Texture *texture);
-IMGUI_IMPL_API ImTextureID ImGui_ImplThin3d_AddFBAsTextureTemp(Draw::Framebuffer *framebuffer);
+IMGUI_IMPL_API ImTextureID ImGui_ImplThin3d_AddTextureTemp(Draw::Texture *texture, ImGuiPipeline pipeline = ImGuiPipeline::TexturedAlphaBlend);
+IMGUI_IMPL_API ImTextureID ImGui_ImplThin3d_AddFBAsTextureTemp(Draw::Framebuffer *framebuffer, Draw::FBChannel aspect = Draw::FB_COLOR_BIT, ImGuiPipeline pipeline = ImGuiPipeline::TexturedAlphaBlend);
 
 void ImGui_PushFixedFont();
 void ImGui_PopFont();
 
 // Helper structure to hold the data needed by one rendering context into one OS window
 // (Used by example's main.cpp. Used by multi-viewport features. Probably NOT used by your own engine/app.)
-struct ImGui_ImplThin3dH_Window
-{
-	int                 Width = 0;
-	int                 Height = 0;
-	bool                ClearEnable = true;
+struct ImGui_ImplThin3dH_Window {
+	int  Width = 0;
+	int  Height = 0;
+	bool ClearEnable = true;
 };
 
 #endif // #ifndef IMGUI_DISABLE