File	Type	Description
`sd_xl_base_1.0_0_bf16.gguf`	BF16	Best quality, largest size
`sd_xl_base_1.0_0_Q8_0.gguf`	Q8_0	Great balance of quality and size ✅ recommended
`sd_xl_base_1.0_0_Q4_K.gguf`	Q4_K	Smaller size, good quality
`sd_xl_base_1.0_0_Q4_0.gguf`	Q4_0	Smallest size

Quick Start

1. Download the model

bash

# Recommended — Q8_0
wget https://huggingface.co/kostakoff/stable-diffusion-xl-base-1.0-GGUF/resolve/main/sd_xl_base_1.0_0_Q8_0.gguf

# Other quantizations:
# wget https://huggingface.co/kostakoff/stable-diffusion-xl-base-1.0-GGUF/resolve/main/sd_xl_base_1.0_0_bf16.gguf
# wget https://huggingface.co/kostakoff/stable-diffusion-xl-base-1.0-GGUF/resolve/main/sd_xl_base_1.0_0_Q4_K.gguf
# wget https://huggingface.co/kostakoff/stable-diffusion-xl-base-1.0-GGUF/resolve/main/sd_xl_base_1.0_0_Q4_0.gguf

2. Build stable-diffusion.cpp

Requirements: CUDA-capable GPU, CMake ≥ 3.18, CUDA Toolkit

bash

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

Version used for conversion and testing:

stable-diffusion.cpp version master-520-d950627, commit d950627

3. Start the server

bash

export CUDA_VISIBLE_DEVICES=0

./stable-diffusion.cpp/build/bin/sd-server \
  -m ./sd_xl_base_1.0_0_Q8_0.gguf \
  --vae-on-cpu \
  --listen-ip 0.0.0.0 \
  --listen-port 8081

⚠️ The --vae-on-cpu flag is required! The VAE decoder consumes up to 10 GB of VRAM when converting the latent representation to PNG. Offloading VAE to CPU makes it possible to run the model on most consumer GPUs.

4. Generate an image

bash

curl -s http://127.0.0.1:8081/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sdxl",
    "prompt": "A lovely cat<sd_cpp_extra_args>{\"seed\": 357925}</sd_cpp_extra_args>",
    "n": 1,
    "size": "1024x1024",
    "response_format": "b64_json"
  }' | jq -r '.data[0].b64_json' | base64 --decode > out.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 sd_xl_base_1.0_0.9vae.safetensors weights using the built-in sd-cli conversion tool:

bash

# Q8_0
./stable-diffusion.cpp/build/bin/sd-cli -M convert \
  -m ~/llm/models/sdxl/sd_xl_base_1.0_0.9vae.safetensors \
  -o ./sd_xl_base_1.0_0_Q8_0.gguf -v --type q8_0

# BF16
./stable-diffusion.cpp/build/bin/sd-cli -M convert \
  -m ~/llm/models/sdxl/sd_xl_base_1.0_0.9vae.safetensors \
  -o ./sd_xl_base_1.0_0_bf16.gguf -v --type bf16

# Q4_0
./stable-diffusion.cpp/build/bin/sd-cli -M convert \
  -m ~/llm/models/sdxl/sd_xl_base_1.0_0.9vae.safetensors \
  -o ./sd_xl_base_1.0_0_Q4_0.gguf -v --type q4_0

# Q4_K
./stable-diffusion.cpp/build/bin/sd-cli -M convert \
  -m ~/llm/models/sdxl/sd_xl_base_1.0_0.9vae.safetensors \
  -o ./sd_xl_base_1.0_0_Q4_K.gguf -v --type q4_K

License

This model inherits the license of the original — CreativeML Open RAIL++-M