gguf-diffusion

Image generation GUI for GGUF diffusion models, packaged for Python. The GUI runs in your browser against a local server; generation is done by the diffusion.cpp engine, compiled during pip install and bundled with the package as a single binary. Model and image files are referenced by filesystem path through a built-in file browser — nothing is uploaded or copied to temp storage.

Install

pip install gguf-diffusion

Building the bundled engine requires a C/C++ toolchain and CMake ≥ 3.15 (on Windows: MSVC Build Tools). The engine source is resolved from a sibling ../diffusion checkout, a vendored vendor/diffusion copy (see scripts/vendor_engine.py), or GGUF_DIFFUSION_ENGINE_DIR. GPU backends are opt-in at build time via the engine's own options:

CMAKE_ARGS="-DSD_CUDA=ON"    pip install gguf-diffusion   # NVIDIA (CUDA toolkit)
CMAKE_ARGS="-DSD_HIPBLAS=ON" pip install gguf-diffusion   # AMD (ROCm/HIP)
CMAKE_ARGS="-DSD_METAL=ON"   pip install gguf-diffusion   # Apple (macOS)
CMAKE_ARGS="-DSD_VULKAN=ON"  pip install gguf-diffusion   # Vulkan

Usage

python -m gguf_diffusion             # launch the GUI in the browser
gguf-diffusion                       # same
gguf-diffusion --port 8643 --no-browser

GUI features (as in the desktop app's diffusion panel):

• txt2img with the full model stack: --model / --diffusion-model, VAE, external text encoders (--clip_l, --t5xxl, --llm, …), additional models (ControlNet, TAESD, upscaler, PhotoMaker, …), tokenizer packs
• image inputs: init image (img2img), mask (inpainting), end frame, control image, reference images
• sampling controls: CFG scale, steps, size, seed, batch count, all engine sampling methods and schedules, flash attention, low-VRAM flags
• live progress and engine log, output gallery, saved workflows (localStorage + JSON export/import), copyable/editable CLI command

Engine CLI passthrough (runs the bundled diffusion binary):

python -m gguf_diffusion engine -- --diffusion-model model.gguf -p "a cat" -o cat.png
gguf-diffusion engine -- --help

How it works

• pip install compiles the diffusion.cpp engine (static libdiffusion + static ggml linked into one CLI executable) via scikit-build-core and installs it into the package's bin/ directory.
• gguf-diffusion starts a stdlib HTTP server (default port 8643) serving the static GUI and a small JSON API, and opens the browser.
• Each generation spawns one engine process; the server parses its progress bars, streams the log to the GUI, and lists the produced images.
• File selection uses a server-side directory listing (/api/browse) so the GUI gets real filesystem paths — no drag & drop uploads of multi-GB models.