sd-cli

TLDR

Generate an image from a text prompt

$ sd-cli -m [model.safetensors] -p "[a photo of a cat]" -o [output.png]

Generate with specific dimensions

$ sd-cli -m [model.safetensors] -p "[prompt]" --width [512] --height [512] -o [output.png]

Set generation parameters

$ sd-cli -m [model.safetensors] -p "[prompt]" --steps [20] --cfg-scale [7.0] --seed [42] -o [output.png]

Use a specific sampling method

$ sd-cli -m [model.safetensors] -p "[prompt]" --sampling-method [euler_a] -o [output.png]

Generate with negative prompt

$ sd-cli -m [model.safetensors] -p "[prompt]" -n "[blurry, low quality]" -o [output.png]

Image-to-image generation

$ sd-cli -m [model.safetensors] --img2img [input.png] -p "[oil painting style]" --strength [0.75] -o [output.png]

SYNOPSIS

sd-cli -m model -p prompt [options] -o output

sd-cli is the command-line interface for stable-diffusion.cpp, a lightweight C/C++ implementation of Stable Diffusion using the ggml tensor library. It runs image generation models on CPU and GPU without requiring Python or heavy ML frameworks.
The tool supports Stable Diffusion 1.x, 2.x, SDXL, and Flux model architectures. Models in safetensors or gguf (quantized) format can be loaded directly. Quantized models significantly reduce memory usage while maintaining reasonable quality.
Text-to-image generates from a prompt. Image-to-image transforms an existing image guided by a prompt. The strength parameter controls how much the original image is altered.
Build from source using CMake. GPU acceleration is available through CUDA, Metal, and Vulkan backends.

PARAMETERS

-m, --model FILE

Path to model weights (.safetensors or .gguf).

-p, --prompt TEXT

Text prompt for image generation.

-n, --negative-prompt TEXT

Negative prompt to guide away from.

-o, --output FILE

Output image path.

--steps N

Number of sampling steps (default: 20).

--cfg-scale FLOAT

Classifier-free guidance scale (default: 7.0).

--seed INT

RNG seed (-1 for random).

--width PX

Image width in pixels.

--height PX

Image height in pixels.

--sampling-method METHOD

Sampling method: euler, eulera, heun, dpm2, dpm++2sa, dpm++2m, dpm++2mv2, lcm.

--img2img FILE

Input image for image-to-image generation.

--strength FLOAT

Denoising strength for img2img (0.0-1.0).

--threads N

Number of CPU threads.

--rng TYPE

RNG type: std_default, cuda.

CAVEATS

Must be built from source. Large model files required (several GB). Quality depends on model choice and parameters. GPU acceleration requires backend-specific build flags. SDXL and Flux models need more memory than SD 1.x.

HISTORY

stable-diffusion.cpp was created by leejet in 2023, inspired by llama.cpp's approach of bringing ML models to C/C++ with minimal dependencies. It brought Stable Diffusion inference to resource-constrained environments and platforms without Python support.