Focused FLUX.2 Klein inference engine. Fork of stable-diffusion.cpp stripped to a single model family, with a persistent JSON-over-stdio engine that keeps models in VRAM between generations.

stable-diffusion.cpp supports dozens of model architectures and sampling strategies. condenser.cpp trades that breadth for depth: one model family (FLUX.2 Klein), two frontends (CLI and engine), and a C API designed for embedding into desktop applications. The engine (cn-engine) runs as a child process, accepts NDJSON commands on stdin, and streams progress and results back on stdout — no HTTP server, no dependencies beyond the GPU driver.

• FLUX.2 Klein text-to-image and image-to-image (4GB and 9GB GGUF variants)
• Persistent engine — load once, generate many, with faster repeat generations vs cold-start CLI
• Prompt conditioning cache — same prompt with different seeds skips the text encoder entirely
• Reference image latent cache — same reference image across img2img runs skips the VAE encoder
• Multi-backend — Vulkan, CUDA, Metal, CPU (and experimental ROCm, SYCL, OpenCL)
• VRAM offloading — run on 4-8GB GPUs by keeping idle model components on system RAM
• Flash attention — reduced memory footprint and faster inference where supported
• C API — clean C interface (condenser.h) for embedding into any language

git clone --recursive https://github.com/jcluts/condenser.cpp
cd condenser.cpp

cmake -B build -DSD_VULKAN=ON
cmake --build build --config Release

cmake -B build -DSD_CUDA=ON
cmake --build build --config Release

cmake -B build -DSD_METAL=ON
cmake --build build --config Release

cmake -B build
cmake --build build --config Release

Binaries are output to build/bin/. See docs/build.md for advanced options and platform-specific notes.

./build/bin/cn-cli \
  --diffusion-model model.gguf \
  --vae ae.safetensors \
  --llm qwen.gguf \
  --prompt "a cat on a windowsill" \
  -W 1024 -H 1024 \
  --steps 4 \
  --seed 42 \
  --offload-to-cpu --fa \
  -o output.png

cn-engine is designed to be spawned as a child process by a parent application. It reads JSON commands from stdin and writes JSON responses to stdout. All log output goes to stderr.

# Quick test
echo '{"cmd":"ping","id":"1"}' | ./build/bin/cn-engine
# → {"id":"1","type":"ok","data":{"status":"pong"}}

# Interactive session
./build/bin/cn-engine
{"cmd":"load","id":"1","params":{"diffusion_model":"model.gguf","vae":"ae.safetensors","llm":"qwen.gguf","offload_to_cpu":true,"flash_attn":true}}
{"cmd":"generate","id":"2","params":{"prompt":"a sunset over mountains","width":1024,"height":1024,"seed":42,"steps":4,"output":"output.png"}}
{"cmd":"generate","id":"3","params":{"prompt":"a sunset over mountains","width":1024,"height":1024,"seed":99,"steps":4,"output":"output2.png"}}
{"cmd":"quit","id":"4"}

The second generate is fast — the model stays loaded and the prompt conditioning is cached from the first run.

See tools/engine/README.md for the full protocol reference, caching behavior, and integration examples (Python, Node.js).

See docs/INFERENCE_PARAMETERS.md for the complete parameter reference.

• FLUX.2 Klein (4GB and 9GB GGUF variants)
• FLUX.2

Based on stable-diffusion.cpp by leejet.