diff --git a/src/client/mapblock_mesh.cpp b/src/client/mapblock_mesh.cpp
index e50a1f5b78..c4c45d9e24 100644
--- a/src/client/mapblock_mesh.cpp
+++ b/src/client/mapblock_mesh.cpp
@@ -16,7 +16,9 @@
 #include "client/meshgen/collector.h"
 #include "client/renderingengine.h"
 #include <array>
+#include <algorithm>
 #include <cmath>
+#include <cassert>
 #include "client/texturesource.h"
 #include <SMesh.h>
 #include <IMeshBuffer.h>
@@ -585,6 +587,47 @@ void PartialMeshBuffer::draw(video::IVideoDriver *driver) const
 	MapBlockMesh
 */
 
+static void applyColorAndMerge(std::vector<PreMeshBuffer> &prebuffers)
+{
+	// TODO: we should change the meshgen so it already applies the tile color
+	// so that we don't need to this extra step.
+	// However currently the CAO code relies on the ability to erase the vertex
+	// colors (light data) before applying the tile colors.
+
+	for (auto &p : prebuffers) {
+		// bake color into vertices
+		p.applyTileColor();
+		// erase color information for later comparisons
+		p.layer.has_color = false;
+		p.layer.color = 0;
+	}
+
+	std::unordered_map<TileLayer, size_t> seen;
+	for (size_t i = 0; i < prebuffers.size(); i++) {
+		PreMeshBuffer &p = prebuffers[i];
+		auto it = seen.find(p.layer);
+		if (it == seen.end()) { // first time
+			seen[p.layer] = i;
+			continue;
+		}
+		// merge
+		auto &dst = prebuffers[it->second];
+		assert(p.layer == dst.layer);
+		if (dst.append(p)) {
+			p = PreMeshBuffer();
+		} else {
+			// other buffer full, this one becomes the new target
+			it->second = i;
+		}
+	}
+
+	// remove all empty buffers
+	prebuffers.erase(std::remove_if(prebuffers.begin(), prebuffers.end(),
+		[] (const PreMeshBuffer &p) {
+		return p.empty();
+	}), prebuffers.end());
+}
+
 MapBlockMesh::MapBlockMesh(Client *client, MeshMakeData *data):
 	m_tsrc(client->getTextureSource()),
 	m_shdrsrc(client->getShaderSource()),
@@ -638,11 +681,12 @@ MapBlockMesh::MapBlockMesh(Client *client, MeshMakeData *data):
 	for (int layer = 0; layer < MAX_TILE_LAYERS; layer++) {
 		scene::SMesh *mesh = static_cast<scene::SMesh *>(m_mesh[layer].get());
 
-		for(u32 i = 0; i < collector.prebuffers[layer].size(); i++)
-		{
-			PreMeshBuffer &p = collector.prebuffers[layer][i];
+		applyColorAndMerge(collector.prebuffers[layer]);
 
-			p.applyTileColor();
+		for (size_t i = 0; i < collector.prebuffers[layer].size(); i++) {
+			PreMeshBuffer &p = collector.prebuffers[layer][i];
+			// Note that the buffer index matters, so 'continue' is forbidden here.
+			assert(!p.empty());
 
 			// Generate animation data
 			if (p.layer.material_flags & MATERIAL_FLAG_ANIMATION) {
diff --git a/src/client/meshgen/collector.cpp b/src/client/meshgen/collector.cpp
index 0c0c73c845..309fd10d12 100644
--- a/src/client/meshgen/collector.cpp
+++ b/src/client/meshgen/collector.cpp
@@ -4,6 +4,21 @@
 
 #include "collector.h"
 #include <stdexcept>
+#include <cassert>
+
+bool PreMeshBuffer::append(const PreMeshBuffer &other)
+{
+	const size_t nv = vertices.size();
+	const size_t ni = indices.size();
+	if (nv + other.vertices.size() > U16_MAX)
+		return false;
+
+	vertices.insert(vertices.end(), other.vertices.begin(), other.vertices.end());
+	indices.insert(indices.end(), other.indices.begin(), other.indices.end());
+	for (size_t i = ni; i < indices.size(); i++)
+		indices[i] += nv;
+	return true;
+}
 
 void MeshCollector::append(const TileSpec &tile, const video::S3DVertex *vertices,
 		u32 numVertices, const u16 *indices, u32 numIndices)
@@ -24,6 +39,7 @@ void MeshCollector::append(const TileLayer &layer, const video::S3DVertex *verti
 	const u16 aux = layer.texture_layer_idx;
 
 	u32 vertex_count = p.vertices.size();
+	assert(vertex_count + numVertices <= U16_MAX);
 	for (u32 i = 0; i < numVertices; i++) {
 		p.vertices.emplace_back(vertices[i].Pos + offset, vertices[i].Normal,
 				vertices[i].Color, vertices[i].TCoords, aux);
diff --git a/src/client/meshgen/collector.h b/src/client/meshgen/collector.h
index 876338baa9..1a1b3fa453 100644
--- a/src/client/meshgen/collector.h
+++ b/src/client/meshgen/collector.h
@@ -19,6 +19,10 @@ struct PreMeshBuffer
 	PreMeshBuffer() = default;
 	explicit PreMeshBuffer(const TileLayer &layer) : layer(layer) {}
 
+	bool empty() const {
+		return indices.empty();
+	}
+
 	/// @brief Colorizes vertices as indicated by tile layer
 	void applyTileColor()
 	{
@@ -33,6 +37,10 @@ struct PreMeshBuffer
 				c->getBlue() * tc.getBlue() / 255U);
 		}
 	}
+
+	/// @brief Append another buffer to this one
+	/// @return false if index would overflow
+	bool append(const PreMeshBuffer &other);
 };
 
 struct MeshCollector
diff --git a/src/client/tile.h b/src/client/tile.h
index 5ec61c3c72..46fc57a7c1 100644
--- a/src/client/tile.h
+++ b/src/client/tile.h
@@ -167,6 +167,23 @@ struct TileLayer
 	bool has_color = false;
 };
 
+template<>
+struct std::hash<TileLayer>
+{
+	// All layers equal according to TileLayer::operator== will have the same
+	// hash value according to this function.
+	std::size_t operator()(const TileLayer &l) const noexcept
+	{
+		std::size_t ret = 0;
+		for (auto h : { l.texture_id, l.shader_id, (u32)l.material_flags }) {
+			ret += h;
+			ret ^= (ret << 6) + (ret >> 2); // distribute bits
+		}
+		return ret;
+	}
+};
+
+
 // Stores information for drawing an animated tile
 struct AnimationInfo {