Generate images.
On your Mac.
No cloud.

LumenForge supports over 18 architectures across image and video generation: Stable Diffusion 1.5, SDXL (base and refiner), FLUX dev/schnell/Kontext, FLUX 2, Chroma, Lumina 2, Mugen, Anima, Qwen Image and Qwen Image Plus, Z-Image, ERNIE Image and ERNIE Image Turbo, Wan 2.2, and LTXV. Models can be in safetensors format, GGUF (quantized), or MLX. The app detects the architecture and format from file metadata — you don't need to tell it what kind of file you're loading.

LumenForge does not distribute or bundle model weights. You download them yourself from Hugging Face, Civitai, or other sources and place them in your models folder (the path is shown in the app's settings). The app scans the folder on launch and detects everything automatically. This approach means you own your models outright — they're not tied to LumenForge, and they'll keep working even if you switch tools later.

16 GB of unified memory handles SD 1.5 and most SDXL models comfortably. FLUX models in full precision benefit from 24–32 GB. GGUF quantized variants of FLUX (4-bit or 8-bit) run on 16–24 GB with some quality tradeoff at lower bit widths. Wan 2.2 and LTXV video generation work best at 32 GB or more — video models are large. An M2 Pro or M3 with 36 GB is the sweet spot for running FLUX and video without constraints.

The Mac App Store build includes the core generation engine — all model architectures, all inference engines (diffusers, mflux, MLX, stable-diffusion.cpp), and the full UI. The three optional capability packs (ControlNet preprocessors, AI upscaler, Prompt Enhancer) are offered as in-app downloads from that build. The direct download from highroadsoftware.com includes all three packs pre-bundled so you don't need separate downloads. Both builds are notarized by Apple and have identical generation capabilities.

Nothing from your generation workflow leaves your machine. Prompts, model files, generated images, and workspace settings all stay in local storage. LumenForge's only network connection is between its Swift frontend and its Python backend — both processes running on your Mac, communicating over localhost. There is no LumenForge server. No telemetry, no analytics, no usage reporting. The privacy manifest submitted to Apple confirms: zero data collection.

Yes. LumenForge supports text-to-video and image-to-video using Wan 2.2 and LTXV models. There's a dedicated video generation view with controls for duration, motion strength, and video settings. Video generation is slow on any local machine — that's the nature of it. But there's no timeout, no billing per second, and no queue. You can start a generation before bed and check results in the morning. 32 GB RAM is recommended for Wan 2.2.

GGUF is a quantized model format from the stable-diffusion.cpp project. Quantization compresses model weights to 4-bit or 8-bit precision, so a model that would normally need 24 GB runs on 10 GB instead. Quality degrades slightly at lower bit widths — 8-bit is usually hard to distinguish from full precision; 4-bit is more noticeable. GGUF models run via LumenForge's integrated stable-diffusion.cpp engine with Metal GPU acceleration. MLX is Apple's own machine learning framework, optimised for the unified memory architecture of Apple Silicon. MLX models are fast and memory-efficient on M-series Macs. Use them when you have an MLX-format model and want the lowest possible latency.

Yes. Once you have models downloaded, LumenForge works entirely offline. Generation, history, workspaces, LoRA loading, and all UI features work without any network connection. An internet connection is needed only for the initial app download and for downloading optional capability packs or new model files — none of which are required to run the core generation.