Stable Diffusion v1-5 GGUF Model Card
Quantized versions of stable-diffusion-v1-5 in GGUF format for use with stable-diffusion.cpp.
At the time of publishing, no ready-made GGUF weights for SD1.5 were available for sd.cpp runtime — so here we are.
Sample generation: "A lovely cat" · seed 357925 · Q8_0 · 512×512
Available Quantizations
| File | Quantization | Description |
|---|---|---|
v1-5-pruned_bf16.gguf | BF16 | Near-lossless, largest file ⚠️ See note below |
v1-5-pruned_Q8_0.gguf | Q8_0 | High quality, ~half the size of bf16 |
v1-5-pruned_Q4_K.gguf | Q4_K | Balanced quality/size |
v1-5-pruned_Q4_0.gguf | Q4_0 | Smallest, fastest, slight quality loss |
⚠️ BF16 known issue: The
bf16variant produces a black image when used with SD 1.5 in stable-diffusion.cpp. This appears to be a bug specific to SD 1.5 — other model architectures are not affected. UseQ8_0if you need maximum quality without quantization loss.
Quick Start
1. Download the model
wget https://huggingface.co/kostakoff/stable-diffusion-v1-5-GGUF/resolve/main/v1-5-pruned_Q4_0.gguf
# Other quantizations:
# wget https://huggingface.co/kostakoff/stable-diffusion-v1-5-GGUF/resolve/main/v1-5-pruned_Q4_K.gguf
# wget https://huggingface.co/kostakoff/stable-diffusion-v1-5-GGUF/resolve/main/v1-5-pruned_Q8_0.gguf
# wget https://huggingface.co/kostakoff/stable-diffusion-v1-5-GGUF/resolve/main/v1-5-pruned_bf16.gguf
2. Build stable-diffusion.cpp
Requirements: CUDA-capable GPU, CMake ≥ 3.18, CUDA Toolkit
git clone https://github.com/leejet/stable-diffusion.cpp
cd stable-diffusion.cpp
git submodule init
git submodule update
mkdir build && cd build
cmake .. -DSD_CUDA=ON
cmake --build . --config Release
This was tested on commit d950627 (version master-520-d950627). Check your version with:
./build/bin/sd-cli --version
3. Start the server
export CUDA_VISIBLE_DEVICES=0
./stable-diffusion.cpp/build/bin/sd-server \
-m /path/to/v1-5-pruned_Q4_0.gguf \
--vae-on-cpu \
--listen-ip 0.0.0.0 \
--listen-port 8081
The server exposes an OpenAI-compatible /v1/images/generations endpoint.
4. Generate an image
curl -s http://127.0.0.1:8081/v1/images/generations \
-H "Content-Type: application/json" \
-d '{
"model": "sd",
"prompt": "A lovely cat<sd_cpp_extra_args>{\"seed\": 357925}</sd_cpp_extra_args>",
"n": 1,
"size": "512x512",
"response_format": "b64_json"
}' | jq -r '.data[0].b64_json' | base64 --decode > output.png
Extra parameters are passed via <sd_cpp_extra_args> as a JSON snippet embedded directly in the prompt field.
How the weights were created
Converted from the original stable-diffusion-v1-5/stable-diffusion-v1-5
# Q4_0
./stable-diffusion.cpp/build/bin/sd-cli -M convert \
-m ~/llm/models/sd1.5/v1-5-pruned.safetensors \
-o v1-5-pruned_Q4_0.gguf -v --type q4_0
# Q4_K
./stable-diffusion.cpp/build/bin/sd-cli -M convert \
-m ~/llm/models/sd1.5/v1-5-pruned.safetensors \
-o v1-5-pruned_Q4_K.gguf -v --type q4_K
# Q8_0
./stable-diffusion.cpp/build/bin/sd-cli -M convert \
-m ~/llm/models/sd1.5/v1-5-pruned.safetensors \
-o v1-5-pruned_Q8_0.gguf -v --type q8_0
# BF16
./stable-diffusion.cpp/build/bin/sd-cli -M convert \
-m ~/llm/models/sd1.5/v1-5-pruned.safetensors \
-o v1-5-pruned_bf16.gguf -v --type bf16
License
This model inherits the license of the original: CreativeML Open RAIL-M