High-performance diffusion model inference in pure Go.

Features

• Image and video generation
• Image editing
• High-resolution upscaling (Neural ESRGAN models + Latent-space methods)
• Callback support for progressive previews during inference
• Model conversion (to SafeTensors / GGUF, optional VAE merging, tensor type rules)
• Hardware-accelerated inference (CUDA, Metal, Vulkan, ROCm and CPU)
• Minimal performance overhead compared to C/C++

Quick start

gosd library is a set of pure Go bindings(no CGO) for stable-diffusion.cpp.
Equip your Go programs with image and video generation — hassle-free.

Installation

go get github.com/l8bloom/gosd

to add gosd to your Go module.

After that, the only thing left is to get the stable-diffusion shared libraries. There are multiple ways to do it, and is directed with the underlying OS and hardware.

The simplest approach is to download an official release from the stable-diffusion project that matches your system. The gosd library is designed to be agnostic regarding which specific build you choose though.

# fetch the driver for your distribution, eg. on Ubuntu:
sudo apt update
sudo apt install mesa-vulkan-drivers vulkan-tools

# fetch the Vulkan loader
VULKAN_VER="1.4.341.1"
mkdir -p /tmp/vulkan && cd /tmp/vulkan \
    && wget -O /tmp/vulkan/vulkan.tar.xz "https://sdk.lunarg.com/sdk/download/${VULKAN_VER}/linux/vulkansdk-linux-x86_64-${VULKAN_VER}.tar.xz" \
    && tar -xf vulkan.tar.xz \
    && rm -rf vulkan.tar.xz

# update the env and run commands to confirm successful installation
. /tmp/vulkan/${VULKAN_VER}/setup-env.sh
vulkaninfo
vkcube

# clone the stable diffusion
git clone --recursive https://github.com/leejet/stable-diffusion.cpp
cd stable-diffusion.cpp

# build the sd
cmake -B builds/vulkan -DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON -DSD_BUILD_SHARED_GGML_LIB=ON
cmake --build builds/vulkan --config Release

# gosd is ready now, export the lib root and try out examples
export GOSD_DYN_LIB="$(realpath builds/vulkan/bin/)"
git clone https://github.com/l8bloom/gosd && cd gosd
go run examples/image_gen/image_gen.go

Examples

Image generation

Generate an image in a few lines:

package main

import sd "github.com/l8bloom/gosd/pkg/gosd"

func main() {
	// Load the dynamic libraries
	if err := sd.Load(); err != nil {
		panic(err.Error())
	}

	// Create and configure the inference context
	ctxParams := sd.ContextParamsInit()

	// Declare models
	ctxParams.DiffusionModelPath = "/tmp/stable.diffusion/flux-2-klein-9b-Q8_0.gguf"
	ctxParams.VAEPath = "/tmp/stable.diffusion/diffusion_pytorch_model.safetensors"
	ctxParams.LLMPath = "/tmp/stable.diffusion/Qwen3-8B-Q8_0.gguf"

	ctx := sd.NewContext(ctxParams)
	defer sd.FreeCtx(ctx)

	// Initialize image generation parameters
	imgParams := sd.ImageGenParamsInit()

	// Prompts
	imgParams.Prompt = "An orange cat on palm beach playing with oranges."
	imgParams.NegativePrompt = "mascots, watermark, signature"

	genImage := sd.GenerateImage(ctx, imgParams)
	genImage.SavePNG("output.png")
}

Result:

Image generation with a preview

examples/callbacks/image_gen/image_gen_with_callbacks.go shows image generation with a preview callback set.

1st image

4th image

10th(last) image

ControlNet Preprocessing

Canny Edge Detection: Extract structural outlines from any image to guide ControlNet inference with pixel-perfect precision.

Generated Image

Canny Preprocessed

ControlNet Output

With A futuristic office with neon lights prompt, see examples/system/canny/preprocess_canny.go for more details.

examples folder has more snippets showcasing classic use-cases.

stable-diffusion.cpp documentation provides more insights into library's features, examples, models etc.

Environment variables

• GOSD_DYN_LIB indicates root of stable-diffusion shared lib(.so, .dll etc.)

You may need to extend OS search path to load libraries sd depends on.
E.g. for Linux deployments: export LD_LIBRARY_PATH=$GOSD_DYN_LIB:$LD_LIBRARY_PATH

Portability

CI/CD pipelines regularly test CPU-based inference on Linux, Windows, and macOS.
GPU acceleration via Vulkan and AMD ROCm/HIP stack are being tested on privately hosted hardware, with verified compatibility on Linux.

✅ = regularly tested

Thanks

• C/C++: stable-diffusion.cpp, llama.cpp
• Go: ffi, purego
• Hugging Face