Implementation:Alibaba MNN CMake Build Diffusion

Field	Value
implementation_name	CMake_Build_Diffusion
schema_version	0.3.0
impl_type	API Doc
domain	Stable Diffusion Deployment
stage	Engine Compilation
source_file	CMakeLists.txt (L75, L119-123, L846-852), transformers/diffusion/engine/CMakeLists.txt (L1-30)
external_deps	CMake (>= 3.10), C++14 compiler, make or ninja
last_updated	2026-02-10 14:00 GMT

Summary

This implementation documents the CMake build invocation required to compile MNN with Stable Diffusion inference support. The build produces the diffusion_demo CLI binary and a libMNN shared library containing the diffusion engine, OpenCV image I/O, and optional GPU backend support.

API

cd build
cmake -DMNN_BUILD_DIFFUSION=ON \
      -DMNN_BUILD_OPENCV=ON \
      -DMNN_IMGCODECS=ON \
      [optional flags] \
      .. && make -j32

Key Parameters

CMake Flag	Default	Description
MNN_BUILD_DIFFUSION	OFF	Master switch: enables diffusion engine and diffusion_demo build
MNN_BUILD_OPENCV	OFF	Enables MNN's built-in OpenCV-compatible image API (auto-forced ON by MNN_BUILD_DIFFUSION)
MNN_IMGCODECS	OFF	Enables image codec support (JPEG/PNG). Required alongside MNN_BUILD_OPENCV for diffusion
MNN_LOW_MEMORY	OFF	Enables low-memory weight loading for quantized models (auto-forced ON by MNN_BUILD_DIFFUSION)
MNN_SUPPORT_TRANSFORMER_FUSE	OFF	Enables fused transformer attention kernels (auto-forced ON by MNN_BUILD_DIFFUSION)
MNN_OPENCL	OFF	Enables OpenCL GPU backend (Linux, Android)
MNN_METAL	OFF	Enables Metal GPU backend (macOS, iOS)
MNN_SEP_BUILD	ON	When OFF, links all components into a single libMNN.so
MNN_CPU_WEIGHT_DEQUANT_GEMM	OFF	Enables CPU weight dequantization GEMM kernels for quantized models

Inputs

The MNN source tree (root directory containing CMakeLists.txt)

Outputs

diffusion_demo -- Standalone CLI executable for text-to-image generation
sana_diffusion_demo -- Standalone CLI executable for Sana diffusion models
libMNN.so (or libMNN.dylib on macOS) -- Shared library with diffusion support compiled in
libdiffusion.so (only when MNN_SEP_BUILD=ON and MNN_BUILD_SHARED_LIBS=ON) -- Separate diffusion shared library

CMake Auto-Forced Dependencies

When MNN_BUILD_DIFFUSION=ON, the following flags are automatically forced (CMakeLists.txt L119-123):

IF (MNN_BUILD_DIFFUSION)
    set(MNN_LOW_MEMORY ON CACHE BOOL "<docstring>" FORCE)
    set(MNN_SUPPORT_TRANSFORMER_FUSE ON CACHE BOOL "<docstring>" FORCE)
    set(MNN_BUILD_OPENCV ON CACHE BOOL "<docstring>" FORCE)
ENDIF()

The diffusion library is conditionally included only when all three prerequisites are met (CMakeLists.txt L846):

IF(MNN_BUILD_DIFFUSION AND MNN_BUILD_OPENCV AND MNN_IMGCODECS)
    include(${CMAKE_CURRENT_LIST_DIR}/transformers/diffusion/engine/CMakeLists.txt)
    IF(NOT MNN_SEP_BUILD)
      list(APPEND MNN_TARGETS diffusion)
      list(APPEND MNN_OBJECTS_TO_LINK $<TARGET_OBJECTS:diffusion>)
    ENDIF()
ENDIF()

Diffusion Engine Build Targets

From transformers/diffusion/engine/CMakeLists.txt:

# Diffusion library (shared, static, or object depending on build config)
add_library(diffusion ...)

# CLI demo executables
add_executable(diffusion_demo
    ${CMAKE_CURRENT_LIST_DIR}/diffusion_demo.cpp
)
target_link_libraries(diffusion_demo ${MNN_DEPS})

add_executable(sana_diffusion_demo ${CMAKE_CURRENT_LIST_DIR}/sana_diffusion_demo.cpp)
target_link_libraries(sana_diffusion_demo ${MNN_DEPS})

Usage Examples

Linux build with CPU only:

mkdir build && cd build
cmake -DMNN_BUILD_DIFFUSION=ON \
      -DMNN_BUILD_OPENCV=ON \
      -DMNN_IMGCODECS=ON \
      -DMNN_SEP_BUILD=OFF \
      ..
make -j32

Linux build with OpenCL GPU acceleration and quantization support:

mkdir build && cd build
cmake -DMNN_BUILD_DIFFUSION=ON \
      -DMNN_BUILD_OPENCV=ON \
      -DMNN_IMGCODECS=ON \
      -DMNN_OPENCL=ON \
      -DMNN_SEP_BUILD=OFF \
      -DMNN_CPU_WEIGHT_DEQUANT_GEMM=ON \
      ..
make -j32

macOS build with Metal GPU acceleration:

mkdir build && cd build
cmake -DMNN_BUILD_DIFFUSION=ON \
      -DMNN_BUILD_OPENCV=ON \
      -DMNN_IMGCODECS=ON \
      -DMNN_METAL=ON \
      -DMNN_SEP_BUILD=OFF \
      ..
make -j32

Notes

The MNN_IMGCODECS flag is not auto-forced by MNN_BUILD_DIFFUSION and must be explicitly set. Without it, the build guard at L846 will prevent the diffusion engine from being compiled even if MNN_BUILD_DIFFUSION=ON.
When MNN_SEP_BUILD=OFF, the diffusion object files are linked directly into the monolithic libMNN shared library, simplifying deployment.
The build_lib.sh script (L339) provides reference build configurations for Android, iOS, macOS, and HarmonyOS platforms.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment